Estimated reading time: 10 minutes

A single dose of DT120 lowered depression scores 14.2 points more than placebo after one week. Twelve weeks later, the separation was still 7.3 points, in a trial that did not require prior treatment failure. The efficacy signal pointed toward major depression at large.

The delivery model did not. Each patient occupied a treatment station for most of a clinical day.

In an earlier essay, I argued that Spravato’s success rests less on its receptor pharmacology than on its delivery: a two-hour observation window that permits staggered arrivals, a monitoring role that can be handed to clinical staff, and a physician who can be decoupled from the chair. DT120 poses the inverse. Its efficacy reaches toward a broad depression population; its delivery remains interventional. Emerge reported an extraordinary number, but it was aimed at the wrong room.

I. The Incumbent’s Room

Spravato, in the economics that actually run it, functions as a billing architecture built around a molecule. A certified center acquires esketamine, supervises its administration, monitors the patient for at least two hours, and bills through a pathway that combines the drug with professional and observation services. Depending on the payer and coding pathway, the modeled contribution margin before fixed overhead is roughly three to six hundred dollars per session. With a maintenance patient returning every few weeks, the arithmetic rewards throughput.

Throughput is possible because the observation staggers. One medical assistant can monitor several recliners while each station cycles through its two-hour window. A clinic may turn the same station three or four times in a day, with each arrival overlapping the last patient’s tail. The contribution margin that keeps an interventional program solvent is a per-station-day number, and that number depends on the stagger.

Spravato’s economics strengthen as maintenance extends. The patient returns, the station turns, and the billing recurs. That structure helps explain how the drug generated approximately $1.7 billion in 2025 sales while reaching only a small fraction of the treatment-resistant population. A treatment monetized through recurrence does not need broad penetration to become a blockbuster.

II. What Emerge Actually Showed

The first week’s 14.2-point separation is among the largest early antidepressant signals reported after a single administration. The effect attenuated toward the Week 6 primary endpoint, where the placebo-adjusted difference was 8.1 points, then settled at 7.3 points by Week 12. Rapid onset, partial fade, persistent separation.

Twenty-four percent of treated patients reached remission at Week 6, compared with three percent on placebo. That is a striking number in a field where remission is scarce. It also means three in four treated patients did not remit.

The population has been described too broadly and too narrowly. Emerge did not require prior treatment failure, leaving open the possibility of a broad MDD label. Yet it was not a front-line cohort. Eighty-three percent of participants entered in the severe MADRS range, and 53 percent had received two or more prior antidepressants. Definium characterizes that subgroup as having failed two prior treatments, and the Week 6 effect held within it.

Emerge therefore enrolled a mixed population: broader than formal treatment resistance by protocol, but with roughly half the cohort carrying the treatment history that often routes patients into interventional care. That overlap with Spravato’s population is where the threat to the incumbent begins.

III. The Eight-Hour Room

My reflex assumption about LSD is a ten-to-twelve-hour ordeal. The trial complicated it. Patients were assessed hourly from hour five against a structured end-of-session checklist. The average time to meeting discharge criteria was 5.8 hours, the median was 5.1, and every patient had cleared by hour eight. Clearance does not mean every subjective effect had ended, but the Phase 3 monitoring window had a defined outer boundary.

Measured in annual chair-hours per patient, DT120 may even look efficient. One full-day administration could consume less monitored time than a year of recurring Spravato maintenance. That comparison has circulated since the readout. It uses the wrong denominator.

The binding variable is turnover.

A treatment station is a patient-ready dosing space: a private room, bay, or recliner, plus the monitoring capacity assigned to it. A Spravato station may turn three or four times in a treating day. A DT120 station holds one patient from morning until discharge and turns once. A clinic could run several DT120 sessions simultaneously, but only by dedicating several stations and enough shared or individual monitoring to cover them. The commercial REMS will decide whether staffing or floor space binds first.

Commercially, DT120 fits neither familiar box. Emerge used no adjunctive psychotherapy during dosing, so psychedelic-assisted therapy does not describe it. Yet an eight-hour session with structured monitoring and occasional staff intervention is not ordinary pharmacotherapy either. It is a full-day drug-administration event.

DT120 needs to be understood not as one patient per clinic, but as one turn per treatment station per day.

That’s the figure that matters most to the clinics driving adoption.

IV. The Deliverable Market

Definium has already sized its market. Its materials trace a funnel from 50 million American adults with anxiety or depression, to 26 million diagnosed, to 13 million taking prescriptions, to 4.2 million who have failed two or more treatments, deduplicated for patients who carry both diagnoses. That 4.2 million is the company’s target.

One percent is 42,000 patients. Using Spravato as an annual pricing surrogate, Definium translates that cohort into a roughly two-billion-dollar opportunity. The slide carrying the calculation is titled “Modest Adoption Supporting a Blockbuster.” It asks the reader to picture the smallest plausible market share and see that even this supports a large business. As a way to make a drug sound conservative, it works.

So what happens when the same funnel is translated into a clinic schedule?

Each of those 42,000 patients needs one monitored station-day for the initial dose. The extension protocol then permits as many as four symptom-triggered, open-label retreatments. Depending on how often patients are actually redosed, a cadence no one yet knows, Definium’s illustration produces between 42,000 and 210,000 station-days a year.

Spread across 250 treating days, that requires roughly 170 to 840 occupied stations on an average weekday. The low end is absorbable. The high end requires a new delivery system.

Set that requirement against the infrastructure that exists. I previously argued that the commercial success of specialized treatments like Spravato is constrained by power-law or Pareto mathematics. Prescription-claims data cited by AtaiBeckley suggest that roughly 500 to 600 high-throughput sites account for about three quarters of national Spravato volume. Give each of 600 sites one dedicated DT120 station, operating every treating day, and the network produces approximately 150,000 station-days a year.

That is a scenario, not a national ceiling. Two stations per site would double it; a restrictive staffing ratio would pull it down. The argument does not require 150,000 to be a hard cap. Against that modeled capacity, one administration for each of Definium’s 42,000 patients consumes 28 percent. One retreatment consumes 56 percent. The five-session ceiling consumes 140 percent.

And that is before a single operator has agreed to convert a profitable, staggered Spravato station into one that turns once.

The addressable market is the illness. The deliverable market is the station-days.

Two recent developments sharpen the contrast. Compass has entered rolling review with a psilocybin program that met the primary endpoints in two Phase 3 trials and reported six-month data in July. Lilly has agreed to acquire AtaiBeckley for approximately $2.8 billion upfront, with another $1 billion contingent on development and regulatory milestones. AtaiBeckley’s lead asset, BPL-003, produced its Phase 2b results after an average clinic visit of approximately two hours.

Neither event proves that duration will decide the category. Lilly is buying a pipeline, not a stopwatch. The transaction is nevertheless consistent with the premium the field has placed on shorter treatment windows and delivery models that resemble an existing interventional workflow.

The station constraint also cuts against Spravato. Limited DT120 adoption could still weaken the incumbent because Spravato’s economics depend on repeated maintenance. A durable episodic alternative would give payers a comparator they do not currently have. DT120 may never win enough stations to displace Spravato directly and still weaken its pricing power.

That possibility depends on the value of a single session. A throughput-bounded treatment can still support a large business if one administration is worth what several staggered sessions were. The room-day problem therefore routes into the trial’s other unresolved question: how much of Emerge’s signal belongs to pharmacology, how much to expectancy, and how much to their interaction?

V. What a Session Is Worth

Emerge deserves credit for using central raters blinded to both treatment allocation and visit number. That design narrows a familiar source of bias in psychiatric trials: the assessor who infers, from the patient’s manner or the timing of a visit, which condition the patient received and scores accordingly.

It cannot blind the patient to the dosing day.

Emerge compared DT120 with an inert placebo, a control strategy the FDA’s final psychedelic guidance now describes as potentially problematic for efficacy assessment. A patient who feels nothing can infer that they received nothing. DT120’s active condition announces itself. In Definium’s Emerge presentation, 64 percent of treated patients had an event coded as illusion, roughly one in nine experienced visual hallucination, and others reported euphoric mood, anxiety, or transient blood-pressure elevation.

Functional unblinding is therefore a reasonable presumption, even though the public materials do not report formal blinding assessments. The placebo arm improved by 5.2 MADRS points at Week 6, below the eight- or nine-point movement commonly seen in modern antidepressant trials. That pattern is consistent with expectancy concentrating in the active arm and disappointment accumulating among patients who feel nothing. It does not prove that explanation.

Definium has treated the concern as a design problem rather than dismissing it. The company points to the dose-response relationship in its Phase 2b anxiety study as evidence that the effect tracks the drug, not merely the knowledge of having received a psychedelic. Ascend, the confirmatory MDD trial, randomizes patients to 100 micrograms, 50 micrograms, or placebo. The middle dose should make assignment harder to infer while providing a second test of dose response.

That design tracks closely with the FDA’s July 2026 guidance. The agency recommends central raters blinded to allocation and visit number, direct assessment of blinding and expectancy, characterization of dose response, and complementary designs that can reduce the bias created by inert placebo. Ascend can improve causal interpretation. Its 50-microgram arm still has pharmacology of its own, so it cannot cleanly partition the molecule from the occasion.

Until Ascend reports, the most defensible reading of Emerge is a large treatment effect whose pharmacologic and expectancy shares remain unresolved. The patient’s ability to perceive the intervention is not a contaminant that disappears after approval. It travels with the drug into the clinic, where regulators, payers, and clinicians will judge the treatment as a whole.

This is where expectancy becomes a pricing question. An all-day station has to be financed by a benefit large and durable enough to price like a procedure rather than a refill.

A full-day treatment only works if one session is worth several turns.

If Ascend confirms that the effect survives a serious attempt to confound assignment, the per-session value may carry the economics and the binding constraint becomes the build. If it does not, the throughput burden becomes much harder to finance.

A positive Ascend would still leave three variables unresolved. The first is retreatment cadence: every additional administration multiplies the station-day burden. The second is staffing: current FDA trial guidance expects two monitors for the full session, including a licensed lead with graduate-level psychotherapy training, though trial guidance is not a commercial staffing rule. Shared coverage across several stations would loosen the capacity limit; a trial-like ratio would make labor bind as tightly as floor space.

The third is reimbursement. The 0820T family of continuous-monitoring codes remains Category III, which means the codes track an emerging service without assuring broad payment. An eight-hour session has to be reimbursed as a procedure, not absorbed as an unusually long office visit. The deliverable market remains a function of station-days, staffing, and price.

Return to the room. One administration, most of a clinical day, one patient in a station until a checklist allows discharge. Emerge’s efficacy points toward a broad depression population; its delivery remains narrow.

The regulatory path is accelerating, but rooms and trained staff are not.

Ascend has not reported, and no one has yet built the capacity DT120 would need. We have a drug for the bell curve and a delivery model for the power law. The distance between them is the whole question.

Disclosure. I am a board-certified psychiatrist practicing interventional psychiatry at a Spravato-certified center, where I run esketamine and TMS programs. I have no financial relationship with Definium Therapeutics. The clinic-economics figures here come from public billing guides and modeled reimbursement rates consistent with operational experience, not from proprietary contract terms.

Dark Matter Pragmatism examines the forces shaping medicine before patients ever reach the treatment room. If you found this useful, consider subscribing.

Subscribe now

Works cited.

• Definium Therapeutics, “Positive Topline Results from Phase 3 Emerge Study of DT120 ODT in Major Depressive Disorder,” June 22, 2026.

• Definium Therapeutics, corporate and Investor & Analyst Day presentations, 2026.

• AtaiBeckley, Virtual Investor Day presentation, March 6, 2026, including Forian and Komodo Health prescription-claims data.

• Eli Lilly and Company, “Lilly to Acquire AtaiBeckley to Advance Therapies for Treatment-Resistant Depression and Other Mental Health Conditions,” July 16, 2026.

• Compass Pathways, FDA rolling-review and Commissioner’s National Priority Voucher announcement, April 24, 2026; and COMP006 six-month data, July 7, 2026.

• U.S. Food and Drug Administration, Psychedelic Drugs: Considerations for Clinical Investigations, final guidance, July 2026.

• Johnson & Johnson, Fourth-Quarter and Full-Year 2025 Results.

• Centers for Medicare & Medicaid Services, HCPCS coding discussion for G2082 and G2083, including the esketamine, professional service, and two-hour observation bundle.

• American Medical Association, Category III CPT codes, including the 0820T series for continuous in-person monitoring.

• Prior Dark Matter Pragmatism essays: A Flawed Blockbuster and Its Challengers, The Power Law of Psychiatric Prescribing, The Elegant Challenger, The Route Problem in Psychiatry, and The Placebo Problem in TMS.

That is not a complaint about a bad instrument. The PHQ-9 is one of the most extensively validated and widely used questionnaires in medicine. It is a confession about a category error that medicine makes thousands of times a day without noticing: treating a screening number as if it were a measurement of the thing it screens for.

I. The contamination is by design

Four of the nine PHQ-9 items ask about the body: sleep, energy, appetite, and psychomotor change. Five ask about the mind: interest, mood, self-worth, concentration, and thoughts of death. In a patient whose conditions each independently flatten energy and wreck sleep, the four somatic items are close to pinned before the interview begins.

The scale’s somatic tilt is deliberate. It follows from a diagnostic philosophy with a name. The DSM-5 and the post-stroke depression literature endorse the inclusive approach: count somatic symptoms toward depression even when a medical condition could plausibly explain them, because excluding them does not improve diagnostic accuracy and does cause clinicians to miss depression that would have responded to treatment [1]. The logic is sound, and the stakes behind it are real. Depression in the medically ill worsens adherence, lengthens recovery, and raises mortality. A scale tuned to over-include is a scale tuned to avoid the most dangerous error in screening, the false negative.

The trouble begins when the same number is asked to do a second job. Cancer-screening data show how the two jobs separate. In an analysis of 4,705 patients at a German cancer center, the non-somatic items carried the diagnostic weight for major depression, while the somatic items added accuracy mainly for milder, less specific depressive presentations [2]. The cognitive items did most of the diagnosing; the somatic items carried real signal and real illness burden at once. Same questionnaire, two readings, one combined score that cannot tell you how much of it is which.

II. What the patients are actually saying

When researchers stop scoring the PHQ-9 and start listening to how patients fill it out, the contamination becomes audible. A 2023 study in the Clinical Journal of Pain ran cognitive interviews with thirty-three high-impact chronic pain patients as they completed the scale. Three items held up: depressed mood, worthlessness, and thoughts of being better off dead were read as intended in better than nineteen of twenty cases. The rest drifted. Patients took the questions about fatigue, concentration, and sleep and answered them about their pain [3]. The instrument asked about depression. The patients reported their bodies and the score reads identically either way.

The long COVID literature shows the same gap at scale, and shows researchers reacting to it. In a registry of 1,022 post-COVID patients, structured PHQ-9 screening flagged signs of depression in 82.5 percent, median score 11; only 26.7 percent described themselves as depressed [4]. A separate post-COVID cohort went further: its investigators recorded the PHQ-9 but used it, in their own words, descriptively only, on the grounds that the scale’s fatigue, sleep, and cognitive items have such low specificity in this population that the total cannot be read as a depression measurement [5].

This is the part worth sitting with. In the research setting, where the question is studied with care, the same instrument that anchors depression care in the primary-care clinic gets quietly demoted to a description of how sick someone feels. The same number gets generated in both cases.

III. The number escaped the chart

A clinician can hold all of this in mind and still be overruled by the system the number lives in. Over the last decade the PHQ-9 stopped being only a screen and became an outcome.1 The National Committee for Quality Assurance now builds depression quality measures directly on it, for screening, for follow-up, and for monitoring the score over time [6]. Medicare goes one step further and attaches money to it: under MIPS Measure #370, the share of a clinician’s patients whose PHQ-9 falls below five at twelve months is a remission rate that feeds the Merit-based Incentive Payment System, where it adjusts how much that clinician is paid [7]. In the collaborative care model that Medicare reimburses, the same number is the treat-to-target; care managers escalate treatment until the score comes down, and the program’s case for itself rests on doing exactly that. One large collaborative-care cohort reported a mean time to remission of 86 days, against 614 days in usual care [8].

That last figure is the strongest argument for measurement-based care, and it is a real argument. Structured tracking catches deterioration and suicidality that unstructured judgment misses; clinical gestalt drifts, forgets, and flatters itself. But notice what has happened to the number along the way. A screening score built to over-include somatic symptoms has been promoted to a treatment target, then to a quality metric, then to a term in a payment formula. Goodhart’s law is the economist’s name for what follows: once a measure becomes a target, it stops being a good measure.

In the patient with five overlapping diseases, the target is worse than imperfect; it may be unreachable by construction. Four of the nine items are held up by medical illness that no antidepressant will move. “Treat to PHQ-9 below five” asks the clinician to drive a score beneath a floor the patient’s biology has already poured. The two ways to chase it are both bad: escalate antidepressants against a somatic baseline that will not yield, or accept that a well-managed patient will register as a permanent treatment failure on the metric that grades the clinic.

So is the number contaminated, or is it doing exactly the job it was built for?

IV. The strongest version of the other side

The case against everything above is better than it first appears, and it deserves articulation.

Start with the screen-versus-measurement distinction, which cuts both ways. The PHQ-9 was never meant to replace the diagnostic interview or to plan treatment on its own; it was built as a DSM-criteria-based module to flag depression and follow it over time. Faulting it for failing to diagnose is faulting a thermometer for not naming the infection. A competent clinician was always supposed to do the diagnostic work in the interview, and the measurement-based-care guidelines say so explicitly, instructing clinicians to rule out medical conditions that mimic or mask depression. On this reading, a PHQ-9 of 18 in a complex patient is doing its only job perfectly: it is saying look harder.

The somatic items are also not noise everywhere. In cancer survivors, an item-response analysis found that fatigue tracked the underlying depression signal closely and that somatic items discriminated mild depression better than the mood items did [9]. The inclusive approach earns its keep in exactly the populations it was built for. Whether the body items are signal or contamination is not a fixed property of the scale; it depends on how much somatic disease the patient is carrying, and chronic-pain and post-viral patients sit at the far end of that range.

The cleanest objection is biological. The whole maneuver above assumes that “depression downstream of pain” and “independent depression” are separable kinds. The newest evidence suggests they may not be. A 2026 paper in Science, pairing human neuroimaging with animal models, offered mechanistic evidence for one pathway by which chronic pain can become depression: microglial remodeling in the hippocampal dentate gyrus, a physical change in the circuitry mood depends on [10]. By the time a patient has carried pain for years, the secondary depression may have become as biologically real as any primary one. Demoralization and anhedonia even predict each other over time. The sorting the interview is supposed to perform may be sorting a distinction the brain has already dissolved.

V. What the interview is still for

Grant all of it, and the clinical problem does not dissolve; it sharpens. Separating the strands is a treatment decision, not a taxonomic one: the strands respond to different levers.

Here psychiatry has a distinction older and more useful than any cutoff score. Demoralization is not the same thing as depression, though the two often travel together. The demoralized patient has lost morale, not the capacity for pleasure; offered a good day, she can still enjoy it, and her core trouble is a sense of being unable to cope, what the literature calls subjective incompetence [11]. The depressed patient has lost the capacity itself; the good day arrives and lands on nothing [12]. The levers diverge from there. Demoralization lifts with restored meaning, restored competence, and better disease control. Anhedonia is the symptom that tracks the mesolimbic dopamine system, the part of depression that antidepressants and neurostimulation are actually built to move.

No PHQ-9 item draws this line. The most useful question in the encounter is one the scale never asks: if you woke tomorrow with no pain and a full tank of energy, would the things you used to love still land? A yes points toward demoralization riding on top of unrelenting disease, and toward disease management and meaning as the first move. A no points toward something no amount of pain control will reach. The honest question in the room is which part of her suffering is most movable, and by what. The answer reorders the treatment plan. The score of 18 does not contain it.

The comorbidity argues for taking this seriously rather than less. When chronic-pain conditions overlap, the psychological picture is not the sum of the parts but a heavier, distinct presentation; comorbid patients cluster into a more vulnerable profile than either condition alone predicts [13]. Fibromyalgia and depression share so much of their affective signature that the distinguishing features come down to where the weight falls: more rumination and self-blame in depression, more pain-specific pathology in fibromyalgia [14]. Those are differences a careful interview can find and a sum score cannot.

Coda

So I come back to the 18, and to the admission that I cannot read it. That admission is the right place to start, not the wrong one. The inclusive scale is built for the average patient, and most days it serves that patient well; it catches the depression a busy clinician would miss and tells you whether treatment is working. The patient in front of me is not average. She is five diseases deep, and the number that screens the population cannot plan her care, cannot define her remission, and should not be allowed to set the terms of either.

The scale measured her suffering honestly. It was never measuring her depression alone.

Dark Matter Pragmatism examines the systems shaping modern medicine, from the exam room outward. If this was useful, consider subscribing.

Subscribe now

References

1. Robinson RG, Jorge RE, Starkstein SE. Poststroke Depression: An Update. J Neuropsychiatry Clin Neurosci. 2024;36(1):22–35.

2. Grapp M, Terhoeven V, Nikendei C, Friederich HC, Maatouk I. Screening for depression in cancer patients using the PHQ-9: the accuracy of somatic compared to non-somatic items (n=4,705, National Center for Tumor Diseases, Heidelberg). J Affect Disord. 2019;254:74–81. doi:10.1016/j.jad.2019.05.026.

3. Aagaard A, Ravn SL, Andersen TE, Vaegter HB. Interpretation of the Patient Health Questionnaire 9 in High-Impact Chronic Pain: Do We Measure Depressive Symptoms the Way We Think? Clin J Pain. 2023;39(10):501–515.

4. Longterm course of neuropsychological symptoms and ME/CFS after SARS-CoV-2 infection: a prospective registry study (n=1,022). Eur Arch Psychiatry Clin Neurosci. 2023. doi:10.1007/s00406-023-01661-3.

5. Long-term symptom severity and clinical biomarkers in post-COVID-19/chronic fatigue syndrome (PHQ-9 used descriptively only owing to symptom overlap). eClinicalMedicine. 2023.

6. National Committee for Quality Assurance. HEDIS Measurement Year 2026 depression measures: Depression Screening and Follow-Up for Adolescents and Adults (DSF-E); Utilization of the PHQ-9 to Monitor Depression Symptoms (DMS-E).

7. CMS Quality Payment Program. MIPS Quality Measure #370 (CBE 0710): Depression Remission at Twelve Months. Remission demonstrated by a PHQ-9 or PHQ-9M score below 5 at twelve months (±60 days), for patients with an initial score above 9.

8. Time to Remission for Depression with Collaborative Care Management in Primary Care (mean 86 days vs. 614 days in usual care). J Am Board Fam Med; AIMS Center, University of Washington.

9. The value of distinct depressive symptoms (PHQ-9) to differentiate depression severity in cancer survivors: an item-response approach. Psychooncology. 2019.

10. Ding M, Xiang S, Zhang Y, et al. From chronic pain to depression: neurogenesis-driven microglial remodeling in the hippocampal dentate gyrus. Science. 2026;391(6791):eaee6177. doi:10.1126/science.aee6177.

11. de Figueiredo JM, Frank JD. Subjective incompetence, the clinical hallmark of demoralization. Compr Psychiatry. 1982;23(4):353–363.

12. Clarke DM, Kissane DW. Demoralization: its phenomenology and importance. Aust N Z J Psychiatry. 2002;36(6):733–742. (See also Thom R, Silbersweig DA, Boland RJ. Major Depressive Disorder in Medical Illness. Psychosom Med. 2019;81(3):246–255.)

13. Cangelosi M, Cavicchioli M, Mesce M, et al. Beyond Pain: Psychological Profiles in Chronic Migraine Compared With Fibromyalgia and Their Comorbidity. Ann NY Acad Sci. 2026;1560(1):e70304.

14. Renz MP, Schmidt H, Drusko A, et al. Neural, Psychological, and Daily Life Evidence for a Transdiagnostic Process of Affective Dysregulation in Depression and Chronic Widespread Pain. Pain. 2025;166(12):e788–e806. doi:10.1097/j.pain.0000000000003800.

I signed up for the whole week mostly to catch up on CME, get back to visit friends and family in California, but also because over the last few months, I feel like my role in clinic visits is increasingly that of “second opinion to the machine.”

I have mixed feelings about my patients using AI in their care. On one hand, it is only fair that they empower themselves with it; most clinicians already do, with our own scribes and support tools. I gladly embraced the role of “the human in the loop” back when that meant sounding smart after quietly asking OpenEvidence about a drug interaction or an obscure side effect in the middle of a telehealth call. Other times I get “looped in” only when the patient and their machine need me to sign a disability form or send a controlled substance.

I do not mind fielding the occasional out-of-office call about something mundane. The calls I have started to worry about are the ones I should be getting and might no longer get, because the patient’s machine took the question first and decided, on its own, that it could handle it.

I have more to write on the uncomfortably intersection of AI and Psychiatry. For today, the symposium walk-through, in running order:

Opening Remarks:

1. Lloyd Minor, Dean of the School of Medicine. Set the tone every medical conference opening now requires: enormous promise, paired with a warning about “devastating consequences” and a call for unimpeachable guardrails. The chaperone arrived before the guest did. Every enthusiasm at this conference came with a minder.

2. Laura Roberts, Chair of Psychiatry. Framed severe mental illness as common, early, disabling, and underserved, then asked that AI transform the field responsibly rather than merely produce tools. A reasonable ask. Whether “responsibly” and “at scale” aren’t mutually exclusive was the unspoken tension for the next four hours.

3. Kilian Pohl, AI4MH co-director. Provided the most useful stat of the morning: the FDA has authorized more than 1,400 AI-enabled medical devices, and not one of them targets psychiatry. His explanation is that machine learning needs clean, objective ground truth, and psychiatry does not have any. Radiology gets the cool tools; psychiatry gets the epistemology seminar. I think he is right, and he’s getting at the deepest problem in our field.

Keynote Panel: Responsible Deployment and Priorities

4. Ehsan Adeli, AI4MH co-director. Argued that clinical quality is more than sounding empathic; good therapy needs agenda-setting, guided discovery, risk recognition, escalation, and boundaries, and his group built a system called TherapyGym to score chatbots on those skills rather than on warmth.

The part the models have already mastered, in other words, is the part that was never the hard part. His closing framing: the real question is not whether AI can sound like a therapist, but whether it can support safer, better, and more equitable care.

5. Brandon Staglin, One Mind. The lived-experience anchor, and the most human stretch of the day. He described his own recovery from schizophrenia and contrasted two technologies he met along the way: multiplayer online games, which offered a thinner substitute for connection, and cognitive remediation training, which helped rebuild it, after which he could finish an Asimov novel again. Technology, he said, is a double-edged sword. He had held both edges.

6. Vaile Wright, American Psychological Association. The institutional case: no equitable market without a regulatory system that earns the trust of providers, patients, and, she stressed, payers, because you cannot scale for what payers will not cover.

Then the sharper line, and the most quietly devastating of the morning: what does it mean when our number one relationship is with a device? She also worried that AI is taking up all the air in the room while every other reform the system needs goes unfunded.

Yes, but I think the deeper problem is that even if we had well-funded, safe, AI digital therapeutics, I don’t think my patient’s would switch to them from Claude.

Academic Research: Foundations and Frontiers

7. Alexis Hiniker, University of Washington. “Friend or Frenemy? AI Chatbots and Teen Mental Health.” A human-computer-interaction researcher who co-designs and field-tests technology with teens and families rather than theorizing about them from a distance. The framing matters: for a large share of adolescents, the companion chatbot is already the confidant, not a hypothetical risk to be managed.

8. H. Andrew Schwartz, Vanderbilt. “Explaining GPT’s Schema of Depression.” His lab is doing really cool work in computational psychology. This is the talk I’ve re-watched several times and not stopped thinking about. Rather than ask whether GPT can detect depression, he asked what model of depression it had built. He then mapped how GPT correlates the symptoms to one another. It scores near the ceiling for accuracy, yet its internal map is off in important ways: it overweights psychotic symptoms and underweights suicidal ideation.

A system that grades depression about as well as our instruments do, while quietly treating the one symptom that can kill as less central, is not a reassuring combination. And it learned that map somewhere. Most likely from us.

9. Shannon Wiltsey Stirman, Stanford. Made the access-and-implementation case: LLMs for screening, waitlist support, between-session work, therapist training, and relapse prevention, with clinicians in the loop rather than replaced. Her recurring caveat was the honest one: the longitudinal data that would tell us whether any of this helps people across time mostly do not exist yet.

Industry and Translation: What It Takes to Deploy

10. Sara Johansen, OpenAI. Product policy lead for mental health. Placed ChatGPT use on a spectrum from wellness to distress to crisis, and distinguished model policy (how the model answers) from product policy (the systems around it: parental controls, crisis routing, account-level safety). What stood out, from inside the largest deployment on earth, was the stated position that the company does not want its product to replace human connection or care. The question the room repeatedly didn’t ask out loud: what happens when it becomes a replacement anyway?

11. Jina Suh, University of Washington. “Beyond the Conversation.” The reframe that did the most work all day. General-purpose AI has quietly become a utility that people in distress reach for at scale, and we keep evaluating it one conversation at a time while the user is living a journey. So: evaluate the journey, not the conversation; treat the whole sociotechnical arrangement as the unit of action, not the model alone; and measure whether a person can find a path to safety from wherever they actually are. This is the move that dissolves much of the apparent paradox, and deserves its own post.

Policy and Ethics: Governing AI in Mental Health

12. Assemblymember Mia Bonta, California. Walked through the state’s legislative push. AB 489 would stop an AI from implying it is a licensed clinician; AB 1979 would require a human to exercise professional judgment when AI informs a clinical decision. These are truth-in-advertising and accountability rules, not bans, and the first is hard to argue with, which is presumably why it is the easy one. The harder questions, privacy, liability, language bias, and engagement design aimed at minors, she named but did not resolve. No one at the conference did.

13. Nicole Martinez-Martin, Stanford. The explicit DEI angle: marginalized communities carry two histories at once, a history of being underprotected, neglected and undertreated, and a history of being overexposed, surveilled and experimented on. Mental health AI can repeat either one, which is why “give everyone access” and “protect the vulnerable” can both be right and still pull in opposite directions. (One data point cited at the symposium, from Common Sense Media: 72% of teens have used an AI companion, and 12% have turned to one for emotional or mental-health support. The deployment is already here.)

14. Jane Kim, Stanford. “Keeping Humans in the Loop.” The cleanest statement of the measurement problem: every claim about whether these tools are safe or useful is only as good as the evaluation behind it, and expert human review does not scale. The field’s shortcut is to let models grade other models. Kim’s point is that you cannot simply hand the judging to the defendants; she proposed actual sample-size math for how much human review a given precision target requires. We have decided to grade the machines using the machines. Someone should at least do the statistics on it.

The vibe in the room

If there was one through-line, it was a conflict almost every speaker seemed to carry at once: AI is dangerous and we must protect vulnerable patients from it, and AI is essential and we must get it to the vulnerable patients the system has already abandoned. The same patient sits at the center of both sentences. Nobody resolved it, and I do not think anyone could in a morning, because it’s more of a fault line running through the field itself.

That is the thread I want to pull in future posts, in no particular hurry:

• The ground truth problem. Why mental health AI is harder than radiology AI, and why Pohl’s missing benchmark is psychiatry’s oldest problem wearing a new coat.

• Sounding like a therapist is the easy part. Why empathic language and actual treatment come apart, and the specific clinical situations where the most satisfying answer is the wrong one.

• The two duties. Protect versus access, bioethics versus public health, and why the room oscillated between them without quite noticing.

• The mirror. The real reason, I think, that the psychiatrists were the most uneasy people in the building. It was not the machine. It was the reflection.

More soon.

-Adam

Notes: AI for Mental Health Symposium, Stanford Health AI Week, June 1, 2026. Talk characterizations are paraphrased from the session; quotations lightly so. FDA figure from the FDA’s AI/ML-Enabled Medical Devices tracker (more than 1,400 cumulative authorizations through end-2025, roughly three-quarters in radiology). California bills referenced: AB 489, AB 1979, SB 976.

Between 5 and 25% of antidepressant placebo recipients discontinue because they cannot tolerate the side effects from a substance that does nothing.

A new Lancet Psychiatry review argues that the prescriber is part of the mechanism.

Estimated reading time: 12 minutes

When I tell a patient that the medication I am about to prescribe has a roughly one-in-three chance of causing nausea, fatigue, or sexual dysfunction, I have always assumed I am being honest. The label says so. The trials reported it. The patient has a right to know.

What I had not understood, until I read a May 2026 Lancet Psychiatry review by Matthew Burke and colleagues, is how much of that conversation is itself a clinical intervention, and in which direction it is moving the patient.

Across drug trials covering 250,000 patients, 49.1% of patients in placebo arms reported side effects they unknowingly attributed to an inert substance. In antidepressant trials specifically, the rate is 44.7%. Between 5 and 25% of patients in placebo arms of antidepressant trials discontinue the inert treatment because they cannot tolerate the side effects they are experiencing from it. In placebo arms of antipsychotic trials, 68% of patients report side effects. In fibromyalgia trials, 67.2%, with roughly 10% discontinuing inert treatment.

These are not the numbers from one outlier study. They are the consistent finding across hundreds of trials and hundreds of thousands of patients. A substantial fraction of what clinicians have learned to call drug side effects is generated by the same expectancy and contextual machinery that produces placebo response, operating in the opposite valence. The Burke paper is the most recent and most prominent statement of this in psychiatry, and the clinical implications are bigger than the field has reckoned with.

I have written before about placebo as biology: the idea that the contextual response to a treatment is not noise to be subtracted from the real signal but is itself a psychobiological event operating through the same circuits that interventions target. Nocebo is the other half of that argument. The biology that allows expectation, ritual, and relationship to help patients can also work against them, and the prescriber is not a neutral observer of the process.

The prescriber is one of its primary inputs.

I. What the Numbers Look Like

The nocebo literature has been accumulating quietly for two decades, and Burke and colleagues have done the field a service by assembling it in one place. A 2018 systematic review by Howick and colleagues found that 70 to 80% of side effects reported in active drug arms of trials are also reported in placebo arms, raising the obvious question of how many of those side effects are pharmacologically caused at all. The 49.1% placebo-arm side-effect rate across 250,000 patients is not a number that varies much by disease area or treatment class. It is a property of the trial encounter itself.

The discontinuation data are the part that should make every prescriber pause. In antidepressant trials, 5 to 25% of patients receiving an inert pill stop taking it because of perceived side effects. In fibromyalgia trials, roughly 10%. These patients are not malingering. They are not exaggerating. They are reporting genuine subjective experience, generated by the brain’s ordinary interpretation of ambiguous bodily signals in the presence of strong negative expectations.

The clinical implication is straightforward. Some unknown but non-trivial fraction of the patients in any psychiatric practice who have failed an SSRI because they could not tolerate it, who have refused a trial of a second antidepressant because of what happened with the first, who have generated chart notes describing them as treatment-resistant or medication-intolerant, are carrying a clinical history that was partially generated by the counseling and the cultural priors surrounding the drug rather than by the drug itself. We do not know how to identify which patients these are. We do know the population-level rate is substantial.

The directional asymmetry between placebo and nocebo discussions in clinical training is worth naming. Most psychiatry residents have been told at least once that placebo response is large in depression trials and complicates the interpretation of efficacy data. Few have been told that nocebo response is comparably large, that it drives discontinuation, and that the informed-consent practices they were taught are part of the mechanism. The placebo side of the conversation has been politicized into a debate about whether antidepressants work. The nocebo side has been mostly ignored.

So how do words on a consent form become symptoms in a body?

II. How the Mechanism Works

The neurobiology of nocebo is the mirror image of the neurobiology of placebo. Burke and colleagues describe both phenomena as instances of the same predictive-processing architecture: the brain generates top-down predictions about future bodily and emotional states, those predictions function as priors that shape how ambiguous interoceptive signals get interpreted, and the resulting perception is a Bayesian compromise between the prior and the incoming signal.

In the placebo direction, a strong positive prior (this treatment will help me) recruits frontolimbic networks that include the dorsolateral prefrontal cortex, the rostral anterior cingulate, the default mode network, and downstream regulation through the amygdala, subgenual cingulate, and ventral striatum, producing measurable improvement in symptoms. In the nocebo direction, a strong negative prior (this treatment will make me feel worse) recruits overlapping networks in the opposite valence, amplifying ambiguous bodily signals into experienced symptoms that feel as real to the patient as any pharmacological side effect.

The sources of negative priors are well characterized. A 2023 meta-analysis of 73 studies by Rooney and colleagues found that exposure to negative suggestions about a treatment produced moderate worsening of side effects (Hedges’ g = 0.4), and direct observation of others experiencing side effects produced a larger effect (g = 0.6). Media coverage works similarly. New Zealand patients on antidepressants reported more side effects and lower drug effectiveness after nationwide media coverage of antidepressant adverse events; the drug had not changed, but the priors surrounding it had. A 2024 study of patients on metoprolol found that 32% reported erectile dysfunction when explicitly informed of that side effect, compared to 13% when the information was omitted. The drug, again, had not changed. The only modified variable was the counseling.

The misattribution mechanism deserves separate attention because it is so easy to underestimate. In any given week, 894 of 1,000 otherwise healthy people report a median of five benign bodily symptoms: a headache, some fatigue, gastrointestinal irregularity, a poor night’s sleep, transient concentration difficulty. These are the baseline noise of being human. In the absence of a treatment expectation, they get attributed to ordinary life. In the presence of a strong negative prior generated by side-effect counseling, the same symptoms get attributed to the drug. Once that attribution forms, it tends to be sticky. The patient now has direct personal experience of the drug causing the symptom, which functions as a much stronger prior than any future counseling can overcome. Telling someone that “most side effects are transient” might be code for “your bodily symptoms will soon regress to the mean.”

This is not a story about suggestible patients. It is a story about how brains work. The same patient, given the same drug, will report different rates of the same side effect depending on what they were told before they started taking it. The variation is not in the patient. It is in the priors the patient is using to interpret what they feel.

III. Informed Consent

The standard informed-consent script in psychiatric prescribing is something most clinicians do on autopilot. The medication is named, the indication explained, and a list of side effects is recited, usually in some rough order of frequency or seriousness. The patient is asked to repeat what they heard and if they have questions. They might get a handout or package insert. The prescription is sent.

What the Burke review makes uncomfortable to ignore is that this consent conversation is a contextual intervention, and the intervention is being delivered in the nocebo direction. The recitation itself shapes the prior. The prior shapes what the patient subsequently reports. The reported side effects then feed back into chart documentation, into the clinician’s sense of the drug’s tolerability, into the patient’s future medication decisions, and ultimately into the trial-derived side-effect rates that prescribers cite to the next patient. The loop is self-reinforcing and largely invisible.

The ethical tension here is real and worth stating plainly. Informed consent exists to protect patient autonomy, and patients have a genuine right to know what they are taking. Withholding side-effect information to reduce nocebo response is not ethically neutral; it shades into paternalism that the field has been working for fifty years to leave behind. A 2019 systematic review by Webster and Rubin found that simply omitting side-effect information produces large reductions in nocebo response across multiple drug classes, but the authors are appropriately cautious about endorsing the practice. Doing no harm is not the only ethical principle in play. So is respecting patient autonomy.

What the Burke paper proposes, building on earlier work by Wells and Kaptchuk and on expert-consensus statements by Evers and colleagues, is a more nuanced position. Not all side effects are equally informative, and not all patients are equally vulnerable to nocebo response. The principle of contextualized informed consent is that the clinician adapts the level of detail in the side-effect discussion to the specific drug, the specific side effect, and the specific patient. Severe, specific, or actionable side effects (a 1% risk of severe hyponatremia on an SSRI, the QTc prolongation risk on a stimulant, the rare but real risk of agranulocytosis on clozapine) get full disclosure, because the patient needs the information to act on early warning signs. Non-specific symptoms that are common in the general population and likely to be transient (mild fatigue, brief headache, low-grade nausea in the first week) get a more general framing: the patient is invited to report anything unusual without being primed to expect any specific sensation.

At first glance, this looks like a Trojan horse for paternalistic non-disclosure. But it may be more a recognition that the side-effect recitation has both a benefit (informed autonomy, ability to recognize warning signs) and a cost (nocebo generation), and that the cost is being almost universally ignored. The Wells and Kaptchuk argument is that informed consent has historically been treated as if its only ethical dimension is autonomy, and that the non-maleficence dimension, the harm done by the counseling itself, is real and underweighted.

The position is not without critics, and the literature acknowledges that more research in clinical settings is needed before contextualized consent becomes standard practice. But the directional evidence is unambiguous: how a prescriber talks about side effects determines, in part, how many of those side effects the patient will experience.

IV. Nocebo-Aware Prescribing

If the side-effect counseling is itself a contextual intervention, the question is what to do about it. The Burke review identifies three approaches that have moved past the conceptual stage into something resembling clinical guidance. None of these should be read as clinical advice or my endorsement, but as thought experiments of how taking nocebo effects seriously would change the informed consent conversation.

The first is positive framing. A 2019 review of six studies by Barnes and colleagues found that telling patients “90% of people do not experience this side effect” produced meaningfully different reporting than telling them “10% of people experience this side effect,” despite the two statements being mathematically identical. What changes is the direction of the prior. The numerator becomes the exception rather than the expectation. The patient enters treatment with a default assumption of tolerance rather than vulnerability. The denominator is the same. The reframing is honest. The effect on subsequent side-effect reporting is real.

The second is contextualized informed consent itself, as outlined above. Full disclosure of specific, identifiable, actionable risks. General framing for non-specific symptoms that are common in the population regardless of treatment. An invitation for the patient to report anything unusual rather than a primed list of expected sensations. This requires the prescriber to actually think about the specific drug and the specific patient rather than reciting the package insert, which is the point. The current standard of care, in which every patient receives the same recitation regardless of their somatic baseline or prior medication history, is not informed consent in any meaningful sense. It is a liability ritual.

The third, and the most underused, is brief nocebo education. The intervention is simple: before starting a medication, the prescriber tells the patient that negative expectations themselves can produce side effects, that this is brain-based and involuntary rather than a sign of weakness or imagination, and that the same biology that allows positive expectations to help can also work against them. A 2019 trial by Pan and colleagues found that a brief leaflet explaining the nocebo effect reduced nocebo-generated headaches in healthy volunteers. A 2021 trial in gastrointestinal cancer patients found that pre-chemotherapy nocebo education reduced fatigue, pain, and other non-specific side effects from the chemotherapy itself. A small pilot study found that adding the simple statement “these side effects are also reported in placebo arms of trials” to standard informed consent significantly reduced subsequent side-effect reporting on tricyclic antidepressants.

Nocebo education has a second benefit that is harder to quantify but worth naming. It treats the patient as a collaborator in the treatment rather than a passive recipient of it. The patient is told something true and useful about their own biology. The clinician is acknowledging that the visit itself is a meaningful contextual input rather than a delivery mechanism for pharmacology. The relational frame is different. The therapeutic alliance research suggests that the relational frame matters in its own right, independent of the specific informational content. Nocebo education may work partly because the patient feels less alone with the experience of side effects when they occur.

Of these, brief nocebo education feels to me the most actionable. It is fast, it is ethically defensible, it requires no new training infrastructure, and the supporting evidence is consistent across drug classes and patient populations.

V. What This Implies About Drug Labels

The deeper implication of the nocebo literature is one the field has been reluctant to name. If a meaningful fraction of placebo-arm side effects is generated by the trial encounter itself, then the side-effect rates reported in drug labels are not pure measurements of drug pharmacology. They are measurements of drug pharmacology embedded inside a particular informed-consent environment, with all the expectancy generation that environment entails.

This is not a critique of the FDA labeling system, which is a reasonable solution to a difficult regulatory problem. It is an observation about what those numbers mean. When a label says that 22% of patients on a drug report nausea, the prescriber typically reads this as “the drug causes nausea in 22% of patients.” The Burke data suggest the more accurate reading is something like “in the trial environment, with the trial counseling and the trial expectation architecture, 22% of patients on the drug reported nausea, and 11% of patients on inert treatment in the same trial also reported nausea.” The pharmacologically attributable rate, the rate that is not explained by the contextual machinery, is the difference between those two numbers rather than the higher one. For many psychiatric drugs, that difference is smaller than the headline figure suggests.

This is also why the typical clinical conversation about side effects, in which the prescriber recites the label number to the patient, ends up amplifying the prior in ways the label cannot anticipate. The trial-derived rate already included the nocebo contribution that the trial counseling generated. The clinical recitation then adds its own additional nocebo contribution on top. The patient ends up with a prior that is stronger than the pharmacology alone would justify, and the prescriber, looking at the subsequent side-effect report, has no way of separating what the drug caused from what the counseling caused.

A substantial fraction of psychiatric drug intolerance is generated by contextual machinery that the field does not measure, does not name, and does not teach prescribers to recognize. It distorts trial estimates, clinical decision-making, patient histories, and ultimately the cumulative picture of what counts as treatment-resistant illness. It is invisible because the field has not built the instruments to see it, and the instruments would be cheap to build. Brief nocebo education in informed consent. Tracking of side-effect reporting before and after framing changes. Differential analysis of label rates and placebo-arm rates as a routine part of clinical counseling rather than a research curiosity.

None of this requires the field to wait for new pharmacology, new biomarkers, or new neuromodulation devices. It requires the field to take seriously what its own placebo literature has been saying for two decades.

What the Burke paper makes clear is that the prescriber is not standing outside the mechanism, dispensing pharmacology to a passive recipient. The prescriber is inside the mechanism, generating priors that the patient’s brain will use to interpret every subsequent bodily sensation, every transient symptom, every ambiguous interoceptive signal that the medication may or may not have caused. The conversation is not preliminary to the treatment but is part of how the treatment is being delivered.

For most of medicine, the side effects that patients report are accepted as a transparent readout of what the drug is doing. In psychiatry, where the gap between trial conditions and clinical reality is widest, and where the pharmacological effect sizes are most modest, that assumption costs the field more than it usually admits. The drug does what it does. The label says what it says. What happens in between is closer to the actual mechanism than the field has typically allowed itself to admit.

Dark Matter Pragmatism explores the systems shaping modern medicine, from the exam room outward. If you found this useful, consider subscribing.

Adam D. Borecky, MD is a board-certified interventional psychiatrist in Southern California and Western Michigan. He writes Dark Matter Pragmatism, a long-form essay series on the forces that shape medicine before patients ever walk through the door. He has no financial conflicts of interest to disclose.

Works Cited

Barnes K, Faasse K, Geers AL, et al. Can positive framing reduce nocebo side effects? Current evidence and recommendation for future research. Front Pharmacol. 2019;10:167.

Burke MJ, Sandra DA, Peciña M, et al. Harnessing placebo effects and mitigating nocebo effects: implications for clinical practice in psychiatry and medicine. Lancet Psychiatry. 2026;13(5):413-425.

Cocco G. Erectile dysfunction after therapy with metoprolol: the Hawthorne effect. Cardiology. 2009;112(3):174-177.

Evers AWM, Colloca L, Blease C, et al. Implications of placebo and nocebo effects for clinical practice: expert consensus. Psychother Psychosom. 2018;87(4):204-210.

Howick J, Webster R, Kirby N, Hood K. Rapid overview of systematic reviews of nocebo effects reported by patients taking placebos in clinical trials. Trials. 2018;19:674.

MacKrill K, Gamble GD, Petrie KJ. The effect of television and print news stories on the nocebo responding following a generic medication switch. Clin Psychol Eur. 2020;2:e2623.

Michnevich TL, Hendi A, Clinic A, Oechsle BK, Stein A. Preventing adverse events of chemotherapy for gastrointestinal cancer by educating patients about the nocebo effect: a randomized-controlled trial. BMC Cancer. 2021;22:1008.

Mitsikostas DD, Mantonakis L, Chalarakis N. Nocebo in clinical trials for depression: a meta-analysis. Psychiatry Res. 2014;215(1):82-86.

Pan Y, Kinitz T, Stapic M, Nestoriuc Y. Minimizing drug adverse events by informing about the nocebo effect: an experimental study. Front Psychiatry. 2019;10:504.

Petrie KJ, Faasse K, Crichton F, Grey A. How common are symptoms? Evidence from a New Zealand national telephone survey. BMJ Open. 2014;4:e005374.

Rooney T, Sharpe L, Todd J, Tang B, Colagiuri B. The nocebo effect across health outcomes: a systematic review and meta-analysis. Health Psychol. 2023;43(1):41-57.

Webster RK, Rubin GJ. Influencing side-effects to medicinal treatments: a systematic review of brief psychological interventions. Front Psychiatry. 2019;9:775.

Wells RE, Kaptchuk TJ. To tell the truth, the whole truth, may do patients harm: the problem of the nocebo effect for informed consent. Am J Bioeth. 2012;12(3):22-29.

Four days into treatment, Carolyn looked like herself again.

That was the problem.

She was elderly, thirty years sober, with MS and a second chronic pain condition. Her life had been held together by routine, social support, and the hard-earned discipline of someone who understood relapse in more than one form. Then a disease flare arrived, followed by prednisone, and the equilibrium cracked. We started the Medicare-authorized standard thirty-six TMS sessions and I gave her my rehearsed line about “time-to-response” taking weeks. I prayed both that it would work and that we could hold out that long. She was rapidly losing hope.

But inexplicably, not even four days had passed and her brightness returned. By the end of week one, her PHQ-9 scores and my own HAM-D ratings were already in remission range. Over the next six weeks, the numbers did not bounce back upward. Two years later, she was still doing well.

This kind of situation makes me uncomfortable. Everything I understood about how TMS is supposed to work biologically made her response feel too early, too clean, and too durable to fit comfortably inside the story I usually tell patients.

I. The Clean Story

The clean story begins with the coil.

Depression is described as a disorder of brain circuits. The left dorsolateral prefrontal cortex is underactive, or at least insufficiently engaged in the regulation of mood networks. The coil delivers magnetic pulses. Those pulses alter cortical excitability. Repeated sessions change network behavior. Over time, depressive symptoms improve.

This explanation has several virtues. It is plausible, teachable, and gives patients a reason to try something when medications have failed them. It lets clinicians speak in the language of mechanism rather than desperation. It allows a psychiatric treatment to feel modern in a field that often still depends on drugs discovered before most of our patients were born.

It also leaves out much of what happens.

A patient does not receive a coil in isolation. She receives a mapping session, a chair, a sound, a schedule, a technician, a psychiatrist’s recommendation, a series of weekly psychometric testing meetings, and a story about why this treatment is different from the medications that came before it. She receives the visible seriousness of a device. She receives daily repetition. She gets out of her house every day. She receives the sense that something real is finally being done.

The clean story starts at the coil. The clinical story starts earlier.

The strongest skeptical case against TMS does not begin with the claim that TMS does nothing. That is too crude, and in my experience too easy to disprove at the level of individual patients. The stronger argument is that the field still does not know, with the confidence implied by its adoption, how much of the observed treatment effect comes from direct neuromodulation; how much comes from expectancy; how much comes from repeated clinical contact; and how much comes from the kind of patient who is able and willing to attend thirty or more sessions in the first place.

Is the coil being asked to explain too much?

II. The Sham Problem

Placebo control is easy to describe in pharmacology. It is harder to build in neuromodulation.

A pill placebo can be made to look and feel like a pill. The patient swallows it. The sensory experience is nearly identical. If the blinding works, the difference between the groups can be attributed, with appropriate caution, to the active medication.

TMS is different. Active stimulation is loud. It produces a tapping sensation on the scalp. It can contract facial muscles. It has a physicality that is difficult to fake. A sham condition has to mimic the sound, sensation, discomfort, and procedure without producing the same cortical effect.

Duecker and Sack make this point in their critique of sham TMS. The intended neural effects of TMS are accompanied by auditory and somatosensory effects, including the clicking sound, skin stimulation, and peripheral nerve activation. Sham methods are therefore asked to do two jobs at once: reproduce the experience of treatment while withholding the intended brain stimulation. The authors argue that sham TMS is limited as a full control condition and should be paired with other control strategies rather than treated as a complete solution.[1]

That is the central methodological problem. Tilt the coil and you may still stimulate cortex. Use a shielded sham coil and you may lose the scalp sensation. Add surface electrical stimulation and the control condition becomes physiologically messier. Reduce the sensory intensity and patients may suspect they are not receiving the real treatment. Preserve the sensory intensity and the control may no longer be neutral.

The control has to be convincing without becoming active.

This matters because depression trials are already vulnerable to expectation. A patient who believes a treatment may finally work can improve. A patient who suspects he has been assigned to the dead arm of a device trial can fail to improve for reasons that have little to do with cortical excitability.

A 2018 meta-analysis of sham response in rTMS depression trials found a large sham response across 61 randomized trials, with a placebo effect size of Hedges g = 0.8. The sham response was also associated with improvement in the active treatment group, suggesting that placebo response may be a component of therapeutic response rather than a nuisance sitting neatly outside it.[2]

This makes TMS’s clinical effect difficult to decompose.

Active-arm improvement can be real while still being shaped by the ritual around the active condition. Sham-arm improvement can be real while still failing to capture the full psychological force of receiving the treatment everyone in the room believes is active.

This is not scandalous in the way that psychiatry’s more smug critics want it to be.

Expectation, attention, relief, prediction, and meaning are not contaminants from outside the organ being treated. They are part of the organ being treated.

III. Expectation Enters the Scanner

A new JAMA Psychiatry study released this week makes the expectancy problem harder to ignore.

The study was not a definitive trial of TMS for depression. It did not ask whether a full clinical course of TMS treats major depressive disorder. Its question was narrower and, in some ways, more interesting: can theta-burst stimulation over the dorsomedial prefrontal cortex alter the brain’s response to antidepressant expectancy?

Participants received different stimulation conditions, including intermittent theta-burst stimulation, continuous theta-burst stimulation, and sham. After stimulation, they underwent an fMRI task designed around placebo antidepressant cues. The authors reported that intermittent theta-burst stimulation increased dorsomedial prefrontal and default mode network activity during expectancy cues, and that this activation was associated with greater expectancy-related mood improvement.[3]

The finding is more inconvenient than simply equating TMS’s effect with placebo. Expectancy is not merely subjective noise. It can be experimentally manipulated. It can be observed in brain networks. It can change mood.

The usual mistake is to treat placebo as the opposite of biology. But placebo effects are biological events. They are not imaginary because they involve belief. They are not irrelevant because they involve expectation. The brain is a prediction machine, and psychiatric symptoms are often altered by what the patient expects, fears, notices, and rehearses.

Placebo is not what remains after biology is removed. It is one of the ways biology enters the room.

For TMS, this creates a difficult interpretive problem. The device may stimulate a target. The procedure may also stimulate expectancy.

In my clinic, these do not arrive separately; they sit together in the same chair.

IV. Mechanism Does Not Make the Problem Disappear

The lazy version of this argument1 would be simple: TMS is a technocratic placebo.

I do not think that is true.

The same week that expectancy became harder to ignore, mechanism also became harder to dismiss. Recent mechanistic work on accelerated intermittent theta-burst stimulation has begun mapping circuit-level and cell-type-specific pathways that could plausibly explain antidepressant effects. A Cell in-press summary described an optogenetic model of accelerated intermittent theta-burst stimulation that points toward cell-type-specific plasticity and a fronto-insular circuit involved in antidepressant effects in humans.[4] A related preprint reports that accelerated iTBS altered activity in specific prefrontal neuron populations in a mouse model and that inhibition of dorsomedial prefrontal intratelencephalic neurons blocked antidepressant-like behavioral effects.[5]

This is not trivial. If TMS produces reproducible circuit changes, and if those changes map onto antidepressant response, the field has more than a theatrical device with good marketing. There seems to be a biological signal whispering through the noise.

While mechanism does not make the placebo problem disappear, it certainly makes it more interesting.

A treatment can have a real mechanism and still have an exaggerated clinical narrative. It can alter brain circuits and still depend on expectation. It can produce cellular changes and still be delivered through a ritual that shapes the patient’s experience of illness. Drugs do this too. The difference is that device procedures make the ritual visible.

The machine is in the room. Everyone can hear it.

That visibility is part of the treatment, whether or not we know how to measure it.

V. What the Clinic Sees

Erika had the kind of depression that makes every treatment discussion feel like a negotiation with injury.

She was in her early forties with severe treatment-resistant depression and anxious distress. She had side effects to everything. She told me at our first meeting (with a hint of self-satisfaction) that her last four psychiatrists had eventually thrown their hands in despair. Eventually, after much persuasion, she agreed to try TMS. Her insurer approved it within hours, mostly because her list of prior treatment failures was so long it might as well have been a separate medical record.

She did not enter treatment as an enthusiast. If anything, she seemed almost offended by the possibility that it might work. She noticed every small discomfort. She reported every minor side effect. Weeks passed. Her self-report scales did not improve. Some worsened.

Then, around treatment twenty, the staff began noticing something.

She was brighter. Not dramatically. Not in a way that would make a case report. But she began cracking jokes. She asked my medical assistants about their lives, their classes, their graduate school applications. She seemed, for moments at a time, able to peer outside the small locked room of her own suffering.

Her clinician-rated scores began to improve. She insisted she was not improving.

The staff noticed before the patient did.

That is one of the things a TMS clinic sees that a trial may flatten. The clinic sees the patient thirty or forty times. It sees posture, eye contact, irritation, punctuality, clothing, jokes, silence, and whether the patient asks someone else a question for the first time in months. Measurement-based care is important. I use it. But a PHQ-9 is not the only instrument in the room.

This is also why the placebo problem cannot be dismissed as an academic nuisance. TMS clinics are repeated observational systems. They are places where patients return again and again, often after years of failed treatments, and are watched by people who expect change but are also trained to detect its absence.

This does not mean the clinic ritual explains the effect. Erika may have been responding to stimulation. She may have been responding to the structure. She may have been responding to both. She may have been improving before she could feel it. The point is that the treatment she received was not only a magnetic pulse; it was a sequence of repeated encounters organized around the possibility of improvement.

VI. The Commercial Drift

Once a treatment becomes a billable protocol, the system starts behaving as if the active ingredient has been settled.

This is not unique to TMS. It is how healthcare works. A code exists. A machine is purchased. Staff are trained. Patients are scheduled. Prior authorizations are submitted. Clinics learn which insurers pay and which ones resist. A therapeutic claim becomes an operational pathway.

The clinic is not paid to separate direct cortical stimulation from expectancy, repeated contact, staff attention, symptom tracking, and spontaneous recovery. It is paid to deliver the session. The commercial system therefore tends to treat the package as unified and settled. The coil becomes the product because the coil is the part that can be studied, bought, scheduled, billed, marketed, and regulated.

The rest of the treatment is harder to invoice.

This creates a subtle drift. Scientific uncertainty remains, but operational confidence grows. A clinic that runs TMS every day begins to speak about it with the confidence of routine. A patient hears that confidence. A device company packages it. A payer reimburses it, or refuses to. A treatment whose mechanism is still debated becomes part of the ordinary machinery of psychiatric care.

That may still be good for some patients. Carolyn is not an abstraction. Neither is Erika. What is not good is when the field becomes less curious once the workflow works.

VII. The Patients Who Receive Nothing

There is another group of patients this essay has to include.

They do everything right. They show up. They tolerate the tapping, the noise, the scheduling burden, the hope. They are remapped. Their clinicians request extensions. The staff are kind. The ceremony is intact. The expectation is present. The machine turns on, and nothing changes.

These patients matter because they prevent the argument from becoming sentimental. If ritual explained everything, they should improve. Many do not. If expectation were enough, they would not remain ill. If staff attention were the active ingredient, every well-run clinic would produce better outcomes than it does.

The coil may not be enough. The ritual may not be enough. Sometimes none of the ingredients are enough. I really wish this were not true. All of my clinical readers know that it is.

That is why the answer cannot be simple skepticism or hype. TMS sits in the uncomfortable middle, where many psychiatric treatments sit: biologically plausible, clinically useful for some, overclaimed by some advocates, underexplained by the evidence, and experienced by patients as more than its mechanism.

VIII. The Better Future

The future of TMS would do well to avoid dependence on louder remission claims or label expansions2. What if we could better predict which patients are more likely to have a Carolyn, Erika, or non-responder experience?

This is why preliminary biomarker work matters, despite my constitutional skepticism. A 2026 Journal of Affective Disorders study examined mismatch negativity as a possible predictor of response to dorsomedial prefrontal iTBS in major depressive disorder. The authors found that larger baseline mismatch negativity amplitudes were associated with greater symptom improvement after active iTBS, but not after sham stimulation. They also cautioned that the finding needs replication in larger samples.[6]

Biomarker findings always need this and rarely get it.

But this is the right kind of question. A treatment with a modest average effect can be very valuable if we know whom to give it to. A treatment with a large marketed effect can disappoint badly if we give it to everyone who qualifies on paper. Psychiatry has lived this pattern for decades. The average effect is less important than the match between patient and treatment.

Dose-response work belongs in the same category. A 2024 JAMA Network Open dose-response meta-analysis found several significant dose-response associations across TMS and tDCS studies, often with bell-shaped curves rather than simple “more is better” patterns. The authors argued that these models need prospective validation in randomized, sham-controlled trials.[7]

This study is useful because it points in a more productive direction than another round of breathless device claims. The field needs clearer predictors; cleaner controls; practical dosing models; and durability data that matter after the acute series ends.

But predictors create their own dark matter. It is the same issue I wrote about in my last essay about inflammatory biomarkers:

Who orders the test? Who pays for it? Does it delay care? Does it change the decision? Can the result be explained to the patient? Does it improve outcomes enough to justify the added friction? A beautiful biomarker is not a clinical tool until it survives scheduling, insurance, workflow, and the ordinary impatience of a suffering person.

If TMS becomes more precise, the value may shift from the machine to the selection rule.

That would be a better field; a more honest one.

IX. The Honest Version

Carolyn got better. I am not interested in explaining that away.

Erika may have improved before she could feel it. I am not interested in ignoring that either.

Other patients have sat through the same chair, the same ritual, and received nothing.

The honest version is less satisfying than the sales pitch and less simple than the skeptic’s dismissal. TMS probably helps some patients. Expectation probably helps some patients. Structure probably helps some patients. Sometimes they arrive together. Sometimes they fail together.

The field does not need to choose between mechanism and placebo, as if one makes the other impossible. It needs to learn how to talk about total treatment effects without pretending they are all coil. It needs sham conditions that answer the questions clinicians actually ask. It needs durability data that matter to patients after the acute series ends. It needs predictors that work outside research centers.

It needs to remain curious after the billing pathway has already been built.

Conflicts of interest: I have no financial relationships or stock ownership with any TMS or neuromodulation companies or manufacturers. I work as an independent contractor at a TMS clinic in Southern California and intermittently consult on the operations side for a few others.

Works Cited

[1] Duecker F, Sack AT. Rethinking the role of sham TMS. Frontiers in Psychology. 2015;6:210. doi:10.3389/fpsyg.2015.00210.

[2] Razza LB, Moffa AH, Moreno ML, Carvalho AF, Padberg F, Fregni F, Brunoni AR. A systematic review and meta-analysis on placebo response to repetitive transcranial magnetic stimulation for depression trials. Progress in Neuro-Psychopharmacology and Biological Psychiatry. 2018;81:105-113. doi:10.1016/j.pnpbp.2017.10.016.

[3] Snyder I, Handoko K, Neppach A, et al. Intermittent theta burst stimulation of the dorsomedial PFC and expectancy-driven placebo mood effects: a randomized clinical trial. JAMA Psychiatry. Published online May 2026.

[4] Fronto-insular circuit mechanisms of accelerated intermittent theta burst stimulation. Cell. In press summary, Cell Press feed, May 2026.

[5] Gongwer MW, Qi A, Enos AS, et al. A cell type-specific mechanism driving the rapid antidepressant effects of transcranial magnetic stimulation. bioRxiv. 2025. doi:10.1101/2025.01.29.635537.

[6] Auditory mismatch-negativity predicts response to dorsomedial prefrontal intermittent theta-burst stimulation in major depressive disorder. Journal of Affective Disorders. 2026;409:121908. doi:10.1016/j.jad.2026.121908.

[7] Sabé M, Hyde J, Cramer C, et al. Transcranial magnetic stimulation and transcranial direct current stimulation across mental disorders: a systematic review and dose-response meta-analysis. JAMA Network Open. 2024;7(5):e2412616. doi:10.1001/jamanetworkopen.2024.12616.

A patient sent me a new paper yesterday and asked if she should get her IL-6 levels checked before her esketamine induction. Her instinct that biology might predict whether a costly and time-consuming treatment will work was something I wanted to take seriously.

The paper was a 2026 preliminary study in the Journal of Psychopharmacology reporting that serum interleukin-6, a cytokine whose relationship to depression has been documented across hundreds of studies over more than two decades, may predict response to esketamine in treatment-resistant depression. This is not a bad paper. But “preliminary” is doing heavy lifting in that sentence. IL-6 has been preliminary in antidepressant response prediction across SSRIs, SNRIs, CBT, ECT, and now esketamine for decades. The marker keeps appearing but has not crossed the threshold to clinical utility. Those are two different statements, and the distance between them is the subject of this essay.

I. The Association Is Real

Start here, because this is where the cynicism breaks down.

IL-6 is not one of those psychiatric biomarkers propped up by a single optimistic cohort and a press release. It is the most consistently replicated inflammatory finding in depression research; meta-analyses confirm elevated circulating IL-6 in MDD patients relative to healthy controls with enough consistency that the question of whether the association exists is settled. The productive argument is about what the association means.

The more important evidence comes from two directions that are methodologically harder to dismiss. The first is the ALSPAC birth cohort study, which measured serum IL-6 in nine-year-old children and tracked outcomes into early adulthood. Children in the highest tertile (my new favorite word) of IL-6 at age nine had an odds ratio of 1.49 for depressive episodes by age 24, with a linear dose-response relationship across tertiles. This matters because it establishes temporal precedence: the elevated IL-6 came first, before the depression, which forecloses the simplest dismissal; that depressed people have elevated IL-6 because being depressed is physiologically stressful. More recent analysis using growth curve modeling across ALSPAC (n=4,835) and the UK Biobank (n=39,613) found that higher baseline IL-6 was associated with worse depression symptom trajectories across the life course, with associations strongest in the younger cohort and in women.

The second piece of evidence is Mendelian randomization. A bidirectional two-sample MR analysis using 135,458 MDD cases and 344,901 controls found an odds ratio of 0.85 per natural-log increase in IL-6 activity (95% CI: 0.75–0.96; p=0.007), with the direction indicating that higher circulating IL-6 activity increases depression risk. CRP did not show the same causal signal in this analysis (OR 1.06; p=0.36). Mendelian randomization uses genetic variants fixed at conception as instruments; those variants cannot be caused by the disease itself, which closes off the confounding pathways that plague observational data. Finding a causal IL-6 signal while CRP remains non-significant suggests IL-6 pathway activity is more etiologically proximal to depression than the downstream acute-phase reactant most clinicians actually measure.

This is a coherent biological story, not a series of coincidences. IL-6 does seem to be consistently and plausibly connected to depression.

II. Why You Still Cannot Act on It

The replication problem in IL-6 research is not that the average differences between depressed and non-depressed populations fail to replicate (they mostly do). The problem is that group averages and individual clinical decisions are different objects; the literature has spent twenty years treating the former as if it constituted evidence for the latter.

The distributions overlap too much. IL-6 is elevated by obesity, metabolic syndrome, cardiovascular disease, rheumatologic conditions, acute infection, chronic stress, smoking, and disrupted sleep, all of which co-occur at higher rates in depressed populations and all of which will generate elevated readings in patients who have never had a depressive episode. Confounding by BMI in particular is chronically undercontrolled in MDD biomarker studies; a clinician who treats treatment-resistant depression knows that the average BMI in that room does not look like a clinical trial sample. The within-group variance is wide enough that knowing a patient’s IL-6 level does not reliably tell you whether they are depressed, how severely, or how they will respond to any particular treatment.

Then there is the assay problem, which the popular coverage of this literature almost never addresses. There is no standardized IL-6 assay in clinical use. Measurements vary substantially across labs, across assay platforms, and across sample handling conditions: serum versus plasma, timing relative to meals, diurnal variation, freezing protocols. The research literature aggregating these studies is aggregating things that may not be measuring the same quantity with the same precision. A clinician ordering an IL-6 level from a commercial lab and comparing it to a research paper’s cutoff is not replicating what that paper did. This is more of a structural problem rather than a data gap.

The treatment-response prediction literature adds a third problem: directional incoherence. In some SSRI cohorts, higher baseline IL-6 predicts better response; in others, it predicts worse remission at twelve weeks. A large 2025 cohort study (n=1,086) found elevated baseline IL-6 was significantly associated with lower remission rates after twelve weeks of stepwise antidepressant pharmacotherapy. An earlier Japanese cohort found the opposite: SSRI responders had higher baseline IL-6 than non-responders, and the IL-6 fell in responders but not in treatment failures. Both studies are real. The moderating variables such as population characteristics, drug-specific effects, the non-linear relationship between inflammatory load and treatment benefit, assay differences have not been resolved. The signal as currently measured is not telling you the same thing in every context.

This doesn’t seem to be a situation where more data will sharpen a fuzzy picture; rather, the picture requires a different kind of measurement to become useful at all.

III. What the Number Does Not Tell You

Serum IL-6 measures a quantity. It does not tell you what that quantity is doing.

IL-6 operates through two biologically distinct pathways. In classic signaling, IL-6 binds to the membrane-bound IL-6 receptor expressed primarily on hepatocytes and certain leukocytes; this pathway is generally anti-inflammatory and regenerative. In trans-signaling, IL-6 binds to a soluble receptor circulating in plasma; the resulting complex can bind gp130 on virtually any cel, including neurons, microglia, and vascular endothelium that lack the membrane receptor, and initiates the pro-inflammatory cascades believed to be pathogenic in inflammatory depression. Trans-signaling is the principal pathogenic mechanism. Classic signaling may be neuroprotective. Standard serum IL-6 does not distinguish between them.

The tocilizumab data makes this clinically concrete. Tocilizumab is an anti-IL-6 receptor antibody used in rheumatology; it blocks both pathways. In a longitudinal cohort of allogeneic stem cell transplant patients (a population with high inflammatory load, high depression risk, and a clear treatment rationale for IL-6 blockade) preemptive tocilizumab was associated with significantly higher depression scores at day 28 compared to controls (β=5.74; p=0.03), with the effect persisting after adjustment for baseline symptoms. Blocking IL-6 signaling broadly made depression worse.1 The most defensible explanation is that classic signaling was providing neuroprotective activity that tocilizumab removed. This is why a proof-of-concept trial (the INSIGHT study) was designed specifically to test selective trans-signaling blockade; the hypothesis is that targeting only the pathogenic pathway might avoid the paradox that broad blockade produced.

The implication for a practicing clinician is specific: an elevated IL-6 level in a depressed patient is not an instruction to reduce IL-6. The marker points toward a mechanism; it does not specify a target that can be addressed with the tools currently in a psychiatrist’s hands. That distinction matters when a patient asks whether her number predicts her response to esketamine because even a validated predictive cutoff would not tell her which biological process that number reflects.

IV. The Esketamine Inheritance

Every new esketamine biomarker paper lands in a field that has been accumulating “preliminary” IL-6 results for two decades. That is less a reason to dismiss the new data than it is the context required to read it honestly.

The ketamine literature does contain a real signal. Yang and colleagues found that ketamine responders had higher baseline IL-6 than non-responders and showed a sustained IL-6 decrease post-treatment that non-responders did not. A separate ketamine study found only a small transient post-infusion IL-6 increase with no correlation to response. A 2024 review in Nature covering personalized ketamine and esketamine use for TRD concluded that blood-based biomarkers broadly have “limited usefulness” as response predictors; BDNF is the most consistent single-marker finding, at modest effect sizes; the review’s conclusion was that multivariate predictive models are where the field needs to go, not individual markers.

The ECT data is worth pausing on, because it runs in an unexpected direction. A 2025 meta-analysis in Brain, Behavior and Immunity found that peripheral inflammation (elevated CRP and IL-6) was associated with better ECT outcomes. One systematic review found that lower baseline IL-6 correlated with greater symptom reduction following ECT in some analyses. A clinician who treats patients with both esketamine and ECT is therefore holding two contradictory IL-6 heuristics simultaneously; the same cytokine may index better response to one intervention and worse response to another, through mechanistic pathways that the current literature does not resolve. That is a structural problem with single-marker predictive frameworks, not a problem that will be solved by more esketamine-specific data.

If the same biomarker predicts better ECT outcomes and better esketamine outcomes through apparently opposite mechanisms, the biomarker is not telling you what you think it is telling you.

The operational stakes are what the papers do not address. An esketamine induction is six to seven sessions over three to four weeks, two hours per session including the mandatory monitoring period, at an out-of-pocket cost that regularly runs into the thousands for patients without Medicaid coverage. A failed induction is not just a data point; it is months of a patient’s time, a real financial hit, and the particular kind of hope-depletion that comes from a treatment that was supposed to be different from the ones that failed before. The appetite for a screening tool that identifies likely non-responders before commitment to a full induction is not academic. It is the appetite of a clinician sitting with a patient who has failed three antidepressants and is asking, plainly, whether this will be different. That appetite is real. The tool is not ready. Both things are true.

If you work in interventional psychiatry or psychiatric drug development, the question in the next section is the one that actually matters for practice: does the clinical picture already tells you what the lab would say?

V. What the Inflamed Depression Concept Gets Right (and Doesn’t Solve)

In 2025, Andrew Miller at Emory published a formal call in the American Journal of Psychiatry to codify an inflammatory subtype of major depression, drawing on three recent publications from his group. The subtype is estimated to affect roughly 25–30% of MDD patients. It is characterized by elevated inflammatory markers in blood and CSF, a symptom profile dominated by neurovegetative features (anhedonia, fatigue, psychomotor slowing) and a trajectory of poor response to SSRIs and SNRIs. The claim is that this subtype is real enough and distinctive enough to warrant its own diagnostic category, with treatment stratification implications for ketamine, ECT, and anti-inflammatory augmentation.

This framing is more defensible than “check IL-6 before esketamine.” It is also, for a clinician who treats this population regularly, often recognizable before any blood draw.

The pattern is not subtle in its fully expressed form. Anhedonia-predominant, fatigue-heavy, psychomotor-slowed patients with metabolic comorbidities, disrupted sleep architecture, and a history of partial or absent SSRI and SNRI response. These patients present a coherent phenotype. BMI above 30, chronic pain, autoimmune history, metabolic syndrome: already on the intake form. The inflammatory phenotype is partially visible in the clinical history; whether a cytokine level adds incremental information beyond a careful structured intake is a question the biomarker literature has not seriously attempted to answer, and it is the question a clinician actually needs answered before building a new ordering workflow.

Where IL-6 may eventually earn a place is as a confirmatory layer within a multi-marker protocol, not as a primary screen. A 2025 cohort study found that adding IL-1β and IL-6 to hsCRP measurably improved prediction of twelve-week remission over CRP alone. The stratification workflow emerging in the literature positions hs-CRP ≥3 mg/L as the initial screen, with IL-6 and TNF-α as confirmatory and phenotyping layers. IL-6 as one component of a multi-marker inflammatory signature is a different claim, and a more defensible one, than IL-6 as a stand-alone esketamine predictor.

That difference matters when a patient is asking for a lab order.

How should I respond to her?

VI. The Gap Between True and Actionable

The honest answer has two parts. The first part she will find reassuring: she has found a real signal, the biology it points to is coherent, and the question she is generating is the right one. The second part is harder: she is asking it a few years before the infrastructure exists to answer it.

A “true” finding and an “actionable” finding are not synonyms. A biomarker earns clinical use when it measures what it purports to measure with sufficient reliability across real-world labs, discriminates at the individual level rather than the group level, produces a decision node that changes management in a defined and beneficial way, and comes with a protocol to follow when it is elevated. IL-6 partially satisfies the first of these; it satisfies the second poorly; it does not satisfy the third or fourth at all. There is no validated cutoff. There is no standardized assay in clinical use. There is no payer coverage for IL-6 as a treatment-selection tool in this context. There is no established protocol for what a clinician should do differently after receiving an elevated result other than to note it, which the clinical history often already suggests.

Psychiatric biomarker history is largely a record of these two categories being confused; the dexamethasone suppression test (See Awais’ excellent 2023 essay on this), once proposed as a diagnostic test for melancholic depression, is the textbook case. The inflammation story is more mechanistically coherent than most chapters of that history. The Mendelian randomization data, the longitudinal birth cohort work, the inflammatory subtype characterization coming out of Miller’s group represent genuine scientific progress, and they deserve a more honest accounting than “promising but preliminary” in the same paragraph used to dismiss twelve other candidates. The association has been confirmed. The clinical translation has structural barriers that a new preliminary paper does not clear. Those barriers are the assay problem, the overlap problem, the directional incoherence across modalities, and the absence of a selective treatment target.

The patient who sent me that paper and wants to know if she is an inflammatory phenotype is asking the right question. The answer that serves her best right now is a careful clinical phenotyping based on the history she has already given me; the cytokine level would add noise, not resolution. She came in asking about a biomarker. What I can tell her is that the clinical picture she described may already be the answer.

If this kind of analysis is useful to you, consider subscribing - Dark Matter Pragmatism is free to read and supported entirely by readers.

Subscribe now

Works Cited

1. ALSPAC birth cohort / IL-6 and childhood depression risk Zammit S, et al. Association of Serum Interleukin 6 and C-Reactive Protein in Childhood With Depression and Psychosis in Young Adult Life. JAMA Psychiatry. 2014;71(10):1121–1128. https://jamanetwork.com/journals/jamapsychiatry/fullarticle/1895671

2. IL-6 trajectories across the life course (ALSPAC + UK Biobank) Ronaldson A, et al. Associations between IL-6 and trajectories of depressive symptoms across the life course. University of Edinburgh Research. https://www.research.ed.ac.uk/en/publications/associations-between-il-6-and-trajectories-of-depressive-symptoms/

3. Mendelian randomization: IL-6 → MDD causal evidence Hartwig FP, et al. Disentangling causal relationships between inflammatory markers and major depression. bioRxiv (preprint, peer-reviewed). https://www.biorxiv.org/content/10.1101/712133v2

4. IL-6 trans-signaling vs. classic signaling Rose-John S. IL-6 Trans-Signaling via the Soluble IL-6 Receptor. Semin Immunol. 2012. https://pmc.ncbi.nlm.nih.gov/articles/PMC3491447/

5. Tocilizumab associated with worse depression scores Felger JC, et al. The IL-6 antagonist tocilizumab is associated with worse depression in patients undergoing allogeneic hematopoietic stem cell transplant. Transl Psychiatry. 2021. https://www.nature.com/articles/s41398-020-01164-y

6. INSIGHT study protocol (selective IL-6 trans-signaling blockade) Kappelmann N, et al. Protocol for the INSIGHT study: a randomised controlled trial of single-dose tocilizumab in patients with depression and low-grade inflammation. BMJ Open. 2018. https://bmjopen.bmj.com/content/8/9/e025333

7. Preliminary evidence: IL-6 as esketamine response biomarker [2026 Journal of Psychopharmacology paper cited by patient]. https://journals.sagepub.com/doi/10.1177/02698811261443676

8. Yang et al. — ketamine responders and IL-6 Yang JJ, et al. Altered peripheral immune profiles in treatment-resistant depression. Transl Psychiatry. 2017. https://www.nature.com/articles/tp201731

9. Nature 2024 review: personalized ketamine/esketamine use, biomarker limitations Bahji A, et al. Personalized use of ketamine and esketamine for treatment-resistant depression. npj Mental Health Research. 2024. https://www.nature.com/articles/s41398-024-03180-8

10. ECT and peripheral inflammation: better outcomes (meta-analysis) [Brain, Behavior and Immunity 2025/2026 meta-analysis]. https://www.sciencedirect.com/science/article/pii/S0149763425000600

11. n=1,086 cohort: elevated IL-6 associated with lower 12-week remission [2025 cohort study, hsCRP + IL-1β + IL-6 prediction of antidepressant response]. https://www.sciencedirect.com/science/article/abs/pii/S0165032724020780

12. Miller 2025 — Advancing an Inflammatory Subtype of Major Depression Miller AH, et al. Advancing an Inflammatory Subtype of Major Depression. Am J Psychiatry. 2025. https://pubmed.ncbi.nlm.nih.gov/40329642/

13. Stratification workflow: hs-CRP screen + IL-6/TNF-α confirmation [Stratifying the Inflamed Endotype in Difficult-to-Treat Depression]. https://pmc.ncbi.nlm.nih.gov/articles/PMC12752927/

14. IL-1β and IL-6 added value alongside hsCRP for remission prediction [Additive value of IL-1β and IL-6 alongside hsCRP]. https://pubmed.ncbi.nlm.nih.gov/39732400/

15. IL-6 role in depression — mechanistic overview Sanacora G, et al. On inflammatory hypothesis of depression: what is the role of IL-6 in the middle of the chaos? J Affect Disord Rep. 2021. https://pmc.ncbi.nlm.nih.gov/articles/PMC7884972/

One found a 3.8-point improvement. The other found 15.5.

The differences in design are as striking as the differences in magnitude; the question is what each design is actually measuring.

I. The Two Numbers

Two psychedelic trials in treatment-resistant depression reported primary endpoint results within five weeks of each other. One found a MADRS improvement of 3.8 points over its comparator. The other found an improvement of 15.5 points. Both enrolled adults who had failed multiple oral antidepressants. Both used MADRS as the primary outcome.

The smaller number is COMP006, Compass Pathways’ Phase 3 confirmatory trial of psilocybin 25 mg reported in February 2026, meeting its primary endpoint and positioning the company for an NDA filing by the end of the year [1]. The larger number is GH001, GH Research’s Phase 2b trial of inhaled 5-MeO-DMT, published in JAMA Psychiatry in March 2026 [2]. Fifty-seven percent of active-arm patients were in remission at Day 8. The placebo arm moved 0.3 MADRS points in the wrong direction.

The differences in design are as striking as the differences in magnitude. A COMP360 session runs six hours with a trained therapist in the room; the psychoactive portion of a GH001 session is nine to fourteen minutes [2]. Compass built formal psychological support into its Phase 3 program; GH Research’s protocol excluded “planned psychotherapeutic intervention” by design [2]. One architecture treats the drug and the setting as inseparable. The other treats the setting as a confound to be engineered out.

Both sponsors have been consistent about what their design is for. The short-acting camp says its architecture makes the spreadsheet work. The long-acting camp says its architecture makes the treatment work. Both are right. The interesting question is what each design is actually measuring.

II. The Case the Short-Acting Camp Is Making

The infrastructure problem is no longer contested. A six-to-eight-hour psilocybin session with one-to-one therapist coverage occupies a clinic room, a clinician, and a support staff member for most of a workday; the economics only close at per-session prices that insurers have no history of paying. Short-acting developers took this constraint as their starting premise.

The argument has consolidated over the past eighteen months. Compass’s own Phase 3 data publication and the operational reality of Spravato’s in-office observation period have made it consensus among developers that session length is the binding commercial constraint on any psychedelic indication. The question each program now answers, implicitly or explicitly, is how far you can compress the session without compressing the drug effect.

GH001’s answer is the sharpest. The psychoactive peak of inhaled 5-MeO-DMT lasts nine to fourteen minutes per dose; the Phase 2b protocol allowed up to three sequential doses on a single day, gated on whether the prior dose produced a peak psychoactive experience [2]. Total in-clinic time, including post-dose monitoring, compresses to a fraction of a psilocybin session. AtaiBeckley’s BPL-003, an intranasal 5-MeO-DMT benzoate salt, operates on a longer arc; the company describes a roughly two-hour time to discharge [3]., Both sit on the short end of a spectrum where psilocybin occupies six to eight hours and LSD-based programs sit at eight to twelve.

The second design choice is the absence of formal psychotherapy. GH001’s protocol used trained medical personnel as session monitors, not psychotherapists [2]. BPL-003 follows a similar posture, using psychological support rather than manualized psychotherapy [3]. This is a direct structural response to language in the FDA’s June 2023 draft guidance on psychedelic drug development, which described the psychotherapy component as an uncharacterized variable that “complicates the assessment of effectiveness and presents a challenge for any future product labeling” [4]. The guidance explicitly recommended that the in-session monitor not also deliver post-session psychotherapy, on the grounds that a therapist who witnessed the acute drug experience cannot be blinded to treatment assignment in a subsequent session [4].

The third design choice concerns who rates the primary endpoint. The short-acting camp is not monolithic on this point, and the distinction matters. GH001 used remote independent MADRS raters against an inert placebo [2]. BPL-003 used centralized blinded raters against a 0.3 mg active comparator; 193 patients were randomized across 12 mg, 8 mg, and 0.3 mg arms [3]. The active comparator design is the FDA-recommended mitigation strategy named in the Lykos Complete Response Letter, intended to blunt the functional unblinding that distinguishing a psychoactive experience from an inert placebo makes nearly automatic [5]. The Phase 2b core data for BPL-003 remain in investor materials rather than peer-reviewed literature; the Phase 2a cohorts were published in Journal of Psychopharmacology in March 2026 and CNS Drugs in April 2026 [6].

The fourth piece is throughput. AtaiBeckley’s investor materials, including a March 2026 Investor Day deck, project three treatments per room per day for BPL-003 against one for long-session psychedelics [7]. The number is company projection rather than operational data, and should be read as such. The operational point it indexes, though, is not controversial. A shorter session, with no therapist required to remain in the room for the full duration of drug effect, permits a room utilization schedule that a full-day session does not.

The fifth is reimbursement architecture. The 0820T CPT code series, effective 2025, covers continuous in-person monitoring for an initial hour with add-on codes for each additional hour and for clinical staff monitoring under physician direction [8]. The code structure maps cleanly onto a short-session inhaled or intranasal psychedelic; it maps less cleanly onto a six-hour session in which the billing question is whether the physician is present at the bedside, adjacent, or merely onsite. The MDMA REMS language drafted for Lykos required that two healthcare providers be “onsite” during administration, not “at the bedside,” leaving room for a central monitoring interpretation that Cybin’s CEO Doug Drysdale has publicly described as a probable model for future psychedelic REMS [9]. Drysdale’s confidence on this point concerns CYB003, a deuterated psilocin analog whose session duration is four to six hours and which is a different pharmacological class than 5-MeO-DMT; the REMS architecture argument applies across the short-and-middle session space.

Christian Angermayer, the most vocal investor in the short-acting thesis, has made the room-economics version of this argument publicly, framing BPL-003 as the short-acting design that finally allows clinic operations to scale [10]. He is not wrong about the arithmetic. The AtaiBeckley slide deck assumes roughly 70% of Spravato patients on weekly 84 mg maintenance at up to $65,000 per year in WAC, and references the Forian/Komodo Health analysis that 500 to 600 clinics drive 75% of Spravato volume [7]. The short-acting commercial model is being pitched into a channel that already exists.

The regulatory point is the climax of the argument. Compass’s Phase 3 retained the one-to-one Certified Psilocybin Session Monitor structure that the FDA’s guidance treats as an uncharacterized variable; the company is now attempting to translate that trial architecture into a lighter commercial model. GH Research and AtaiBeckley ran trials that map directly onto their proposed commercial architecture. No therapist confound to separate. No in-session monitor moonlighting as an integration therapist. Blinded raters, or in BPL-003’s case centralized blinded raters against an active comparator, addressing head-on the problem that sank Lykos.

That problem is worth restating in its specifics. In Lykos’s pivotal Phase 3 trials, 94.2% of MDMA recipients correctly guessed their allocation; 75% of placebo recipients did the same; 40% of the cohort had prior illicit MDMA experience [5]. The FDA’s CRL identified, as a separate issue, that training manuals had instructed sites not to report positive or favorable effects as adverse events [5]. The short-acting trials are built to not reproduce this.

On paper, this is the cleanest drug development architecture psychiatry has seen in a decade. On paper.

III. What the Architecture Also Optimizes

Every design choice that makes the spreadsheet work also reduces expectancy variance in the trial.

Shorter session means less therapist contact time, which means less alliance effect. No psychotherapy means no integration scaffold shaping subjective response. Remote or centralized MADRS raters mean minimum assessor unblinding. A brief, unmistakable psychoactive peak means minimum ambiguity about who received active drug and who did not. Individualized dosing gated on whether the prior dose produced a peak experience means maximum biological activity in the active arm and, by construction, maximum certainty in both the patient and the rater about what was administered.

The same design choices that reduce noise in a trial also narrow the population in which the result is valid.

None of this is misconduct. Each decision has an independent methodological justification; several were directly recommended by the FDA. The point is that the effects are cumulative. A trial can simultaneously address the Lykos concerns and compress the range of expectancy effects that placebo-controlled depression trials historically produce.

The placebo responses track this. GH001’s placebo arm moved +0.3 MADRS points across a week [2]. COMP360’s Phase 2b placebo arms moved −1.2 and −3.7 [11]. A ketamine-for-TRD placebo arm typically moves around −7. A conventional antidepressant MDD placebo arm typically moves around −9. The gradient is orderly; it runs in the direction of expectancy and therapeutic alliance. Architectures that strip expectancy produce smaller placebo responses. The effect is a property of the design, not a failure of conduct.

The empirical baseline for this claim is now firmer than it was a year ago. Orsini and colleagues published the first systematic quantification of blinding integrity across psychedelic RCTs in JAMA Psychiatry in April 2026, covering 112 trials across psilocybin, LSD, ketamine, MDMA, ayahuasca, DMT, and noribogaine [12]. Only 29.5% of the trials formally assessed blinding integrity. Where it was assessed, participant and rater unblinding exceeded 90% in psilocybin, LSD, and ayahuasca trials, and exceeded 85% in inert-placebo MDMA trials [12]. The one trial that achieved genuine blinding, the 2023 Heifets-lab study administering ketamine under general anesthesia, found no MADRS difference between drug and placebo [13]. No control strategy in the literature has consistently achieved ideal blinding in an awake patient.

The Williams, Barnett, and Szigeti meta-analysis published in JAMA Psychiatry in March 2026 quantified what that baseline implies at the class level [14]. Across twenty-four trials, the authors compared psychedelic-assisted therapy against open-label antidepressants under equal unblinding conditions. The difference was 0.3 HAM-D points in favor of open-label antidepressants, statistically indistinguishable from zero. Both classes produced roughly twelve HAM-D points of improvement from baseline, which is clinically meaningful for both. The authors’ framing in the published conclusions was precise: their results argue against highly optimistic narratives surrounding psychedelic-assisted therapy and highlight the importance of blinding integrity; the drug class may still be valuable. Szigeti himself described the finding, on publication, as a “sobering reality check for psychedelic medicine” [14]. Whether the residual expectancy component in psychedelic trials is noise or part of the treatment mechanism is a deeper argument these trials are not designed to settle.

A precision worth keeping in view. GH001 and BPL-003 sit at different points on the expectancy-management gradient. GH001 used an inert placebo and remote raters; its trial is maximally vulnerable to unblinding critique. BPL-003’s 0.3 mg active comparator with centralized blinded raters is specifically structured to address that vulnerability. When the short-acting thesis is discussed as if it were one bet, this distinction collapses. It is two bets, made under the same commercial logic but with materially different methodological postures toward the blinding problem.

The other gradient worth naming is prior psychedelic experience in the enrolled cohort. Rates vary widely across modern trials: GH001 Phase 2b enrolled approximately ten to twelve percent prior-use participants; [15] Compass’s COMP360 Phase 2b enrolled six percent; [11] Carhart-Harris’s 2021 psilocybin-versus-escitalopram trial enrolled ninety-two percent; [16] Palhano-Fontes 2019 enrolled zero percent [17]. Prior experience predicts expectancy; expectancy predicts response. A cohort that has inhaled 5-MeO-DMT before is a cohort that knows what it is about to feel, and a cohort that knows what it is about to feel is a cohort whose blinding is structurally compromised regardless of who is doing the rating.

The architecture is coherent for the trial. It is less clear what it captures about the clinic. The short-acting design minimizes expectancy variance, minimizes assessor unblinding, and maximizes the probability that the drug effect, whatever it actually is, will show up as a large effect size.

So what happens when the architecture that produces the “better” trial result meets the patient the trial was not designed to enroll?

IV. The Chair the Prescription Would Be Written From

A trial tells you what a drug does in the population the trial enrolled. The question that matters clinically is how that population relates to the population the drug will have to treat.

The Hughes and Garcia-Romeu systematic review in eClinicalMedicine in July 2024 is the most complete available answer to that question for psychedelic trials [18]. Across thirty-nine psychedelic trials and 1,393 participants published between 1994 and 2024, 85% of enrolled participants were non-Hispanic White. In U.S.-only trials, the figure was 84.5%. Black participants were 2.9%, Hispanic participants 5.9%. The U.S. population those trials implicitly generalize to is 57.8% non-Hispanic White, 12.1% Black, and 18.7% Hispanic. There has been negligible movement in these figures between trials published before and after 2017.

For 5-MeO-DMT specifically, the numbers narrow further. The Reckweg Phase 1/2 trial of GH001 in 2023 enrolled sixteen participants, all of them White [19]. GH001’s Phase 2b enrolled 81, all of them White, split across European sites [2]. Whatever the drug does in those cohorts, the trial may not describe what the drug will do in the U.S. TRD population.

The demographic narrowing is downstream of an earlier narrowing. Read the eligibility criteria in order. GH001 required a current depressive episode of no more than two years, which excludes the chronic TRD patient for whom the previous episode has blurred into the current one across half a decade [20]. BPL-003 excluded prior non-response to ketamine, esketamine, ECT, VNS, or DBS, which excludes the patients for whom every prior interventional option has already failed [21]. BPL-003 excluded first-degree family history of psychotic or bipolar disorder, which excludes a substantial fraction of the patients most likely to present in a TRD clinic. BPL-003 excluded personality disorders, active substance use disorder, and any cardiovascular condition that would make an acute blood pressure excursion hazardous [21]. What is left after the list has done its work is a cohort whose depression is serious, whose comorbidity profile is clean, whose social stability is sufficient to complete a multi-visit protocol, and whose psychiatric history is compatible with consenting to an intense altered state of consciousness.

Self-selection narrows the cohort again. Participants in a psychedelic trial have to be willing to inhale or inhale-spray a substance that produces a peak psychoactive experience lasting anywhere from ten minutes to two hours. This is not the same population as patients willing to accept a sixth SSRI trial, or patients willing to begin twice-weekly ketamine infusions, or patients willing to accept a course of TMS. Szigeti and Heifets have argued in Biological Psychiatry: Cognitive Neuroscience and Neuroimaging that positive expectancy in psychedelic trial participants is itself a form of selection that inflates apparent efficacy; the patient who volunteers for the trial is, on average, a patient who already believes the drug might work [22].

The trial tells you what the drug does in a population the drug will struggle to reach.

The second movement is harder to quantify but arguably more consequential. Cultural framing shapes who is willing to consent to a given class of treatment before a clinician has a chance to discuss risks and benefits. Psilocybin’s public face over the last decade has been shaped by Michael Pollan’s How to Change Your Mind in 2018 and its Netflix adaptation in 2022 [23] the drug arrives in the patient’s imagination pre-framed as serious, investigated, and accessible to non-adventurous adults. Ketamine’s public face, after Matthew Perry’s death in October 2023 and the subsequent criminal charges against two physicians, arrives pre-framed as clinically useful and dangerous; [24] patients who consent to Spravato do so knowing the name of the drug and often with strong feelings about it, in both directions.

5-MeO-DMT arrives with a different cultural inheritance, and the inheritance is not a single voice. Pollan himself called it “the Everest of psychedelics,” a phrase intended to mark its intensity relative to psilocybin [23]. Mike Tyson’s description of the drug on the Joe Rogan podcast in November 2020, to an audience of roughly ten million listeners, framed it as ego-dissolution and rebirth in vivid, bodily language [25]. Bryan Johnson, the longevity entrepreneur whose public persona is organized around the measurement and optimization of biological aging, posted in early 2026 that 5-MeO-DMT had left him childlike and fresh three weeks out, with brain data he said matched his subjective experience; he described the limitation of ordinary consciousness as a rendering problem, comparing the mind to “a 480p screen trying to display an 8K image” [26]. Town & Country profiled the drug in January 2022 as the drug of choice for people with money seeking ultimate mind expansion [27]. The RAND survey of U.S. adults in September 2025 estimated past-year 5-MeO-DMT use at 0.30%, against psilocybin at 4.26% and ketamine at 1.26% [28]. The drug’s cultural footprint is small, vivid, and elite-coded in several different registers at once.

The cultural footprint is also, abruptly, joined by a policy footprint. On April 18, 2026, the administration issued an Executive Order directing the FDA to prioritize review of Breakthrough-designated psychedelic therapies, expanding Right-to-Try access to investigational psychedelic compounds, and allocating ARPA-H funds to match state-level psychedelic research programs [29]. The order accelerates attention and timelines. It does not change who consents to the treatment.

Consider the modal patient who sits in an interventional psychiatry clinic after two or three failed SSRI trials. She is sixty-something. She is on 25 mg of sertraline, a dose her primary care physician started her on four years ago and never titrated; her daughter is on the same medication at the same dose. She has heard of ketamine, has strong opinions about it, and is not sure whether those opinions would survive her reading the label carefully. She has not heard of Spravato. She has not heard of COMP360 by name. She has heard of 5-MeO-DMT exactly once, on a podcast her adult son sent her, and what she heard was enough to make her decline the conversation before it began.

She is not the exception in the clinic. She is the modal TRD patient. The likelihood that such a patient consents to inhale a drug whose cultural footprint consists of Tyson, Pollan, Johnson, Town & Country, and a set of recent federal headlines drops sharply relative to her likelihood of consenting to a ketamine infusion, and drops further still relative to her likelihood of accepting a fifth oral antidepressant. Multiply her across a practice and the constraint becomes visible. The infrastructure constraint is real. The recruitment constraint is upstream of it.

V. The Inelegant Bet

Flip the hierarchy. Compass’s 3.8-point MADRS delta looks modest until the question becomes who it applies to. COMP360’s Phase 2b enrolled 6% of its patients with prior psychedelic experience, the lowest proportion of any modern psychedelic TRD trial; [11] its cohort is more naive, more representative of the average TRD patient’s cultural distance from the drug class. The six-hour session with a therapist in the room is operationally worse on every AtaiBeckley slide. It is also the model that can absorb a patient who would otherwise never consent.

GH001’s 15.5-point delta describes the effect of inhaled 5-MeO-DMT in a self-selected cohort of Europeans who were willing to show up at sixteen trial sites across the continent to receive up to three doses of a drug with a ten-minute peak [2]. The cohort was uniformly White, entirely free of chronic TRD, and enriched for psychedelic-curious patients by the nature of the recruitment. COMP006’s 3.8-point delta describes the effect of psilocybin 25 mg in a larger, more diverse, partially treatment-naive cohort rated under centralized blinded conditions, with a placebo response that was not engineered to near-zero. Which magnitude better predicts what happens when the drug reaches an unselected TRD population is not a question the trial designs can answer. It is a question the difference between the trial designs produces.

Now the counterargument, stated in its strongest form. The short-acting camp’s case for solving the recruitment problem is not naive, and pretending otherwise understates what they are doing. Usona’s PSIL201 trial of psilocybin for major depressive disorder prescreened 1,529 patients, consented 347, excluded or withdrew 240 during screening, and randomized 104 [30]. The post-consent attrition rate was 69.2%. Very little of that attrition was patient refusal. The drivers were SSRI taper requirements, cardiovascular exclusions, comorbidity exclusions, and the time commitment required for preparatory psychotherapy sessions [30]. Short-acting architecture removes several of these. No full-day session means no cardiovascular exclusion for transient blood pressure changes lasting six hours; no preparatory psychotherapy means no week-long pre-treatment calendar to coordinate; no SSRI washout in some short-acting programs means a patient can remain on her baseline medication throughout induction. Remove those barriers and post-consent attrition should drop meaningfully.

The short-acting camp is solving a real and well-defined problem. They are solving the operational piece of the recruitment problem, not avoiding the recruitment problem.

GH Research has also articulated a defense of its Phase 2b design against the functional unblinding critique, and the arguments deserve direct engagement. The company’s case rests on three claims. First, the ultra-short duration of the psychoactive effect provides temporal separation from the primary endpoint; by Day 8, the acute experience is seven days in the past, and what patients and raters are assessing is a sustained mood state, not an acute drug effect [31]. Second, the Cohen’s d of approximately 2.0 on the primary endpoint exceeds, by a substantial margin, the expectancy-attributable magnitude that Szigeti’s meta-analysis estimates for the class; even granting a real expectancy contribution, a residual pharmacological effect remains [2, 4]. Third, efficacy at Day 8 did not correlate with the number of prior treatment failures, which the company advances as evidence against a purely expectancy-driven mechanism, on the logic that heavily refractory patients typically carry lower expectancy and still responded at similar magnitude [31]. Each of these arguments is internally coherent and grounded in real constraints. None of them is a Phase 3 protocol. The agency will decide whether they are, in combination, a sufficient structural response to what Orsini, Szigeti, and the Lykos CRL have put on the table.

A sharper frame emerges from that steelman. Short-acting architecture solves the logistical friction component of the consent problem. It does not solve the cultural framing component or the consent conversation component. These components have traveled together in critiques of long-acting designs, but they are structurally distinct. One is plausibly addressable by commercial engineering. The other is not.

A model that maximizes per-room throughput is not necessarily the model that maximizes treated patients. The first is a question about an operational constraint. The second is a question about how many of the patients with TRD will walk through the door, sit down, and say yes.

Background evidence on the shape of that gap comes from Spravato’s access and attrition data. The Teeple analysis in 2023 of 534 first pharmacy claims found 34.6% approved on first submission, 46.3% rejected, and 19.1% abandoned; by the second session, 85.2% had eventually been approved, and 47.6% of patients completed the full eight-session induction [32]. A Frontiers in Psychiatry psychiatrist survey in 2026 found that 61.43% of psychiatrists cited in-office observation as the primary driver of patient hesitancy toward Spravato, ahead of cost at 57.40% and safety concerns at 55.60% [33]. External validation that the consent and access structure, rather than the molecule, is the operational bottleneck for the indication.

The durability question deserves a line. GH001’s open-label extension followed 23 remitters over six months; twenty of them needed retreatment, most received three or four retreatments over the six-month window [2]. Short-acting architecture that looks like an escape from Spravato’s maintenance model may, once real patients start cycling through it, reproduce that model at a different tempo. Per-session economics and per-patient lifecycle economics are different questions; the first favors short-acting designs, the second is an open empirical matter that the Phase 3 programs have not yet been sized to answer.

Sponsor-level divergence is visible here that the short-acting label obscures. GH Research, following the January 2026 lift of the FDA clinical hold on GH001, confirmed it would replicate the Phase 2b design for Phase 3: no psychotherapy, no disclosed active comparator, initiation targeted for H2 2026 [34]. A Guggenheim analyst note characterized the company as having “a clear strategy to address functional unblinding” without specifying what that strategy is [35]. AtaiBeckley exited its End-of-Phase 2 meeting in early March 2026 with FDA alignment on a dual-trial Phase 3 program using the active comparator and centralized rater design already deployed in Phase 2b [36]. Both sponsors are making short-acting bets. They are not making the same bet. GH Research is betting that the Phase 2b magnitude carries through to Phase 3 even if the expectancy asymmetry from the original trial is unchanged. AtaiBeckley is betting that methodological conservatism produces a more approvable package even at the cost of a smaller nominal effect size. The agency’s response to those two bets will shape the class.

The reader who sees only the short-acting elegance is missing the denominator. The reader who sees only the denominator is missing what the short-acting architecture genuinely solves.

The model that produces the cleanest trial result is not necessarily the model that treats the most patients.

VI. Coda

The commercial winner will not be the architecture with the best per-room-day math.

Drug development in psychiatry produces two kinds of information. The first is what a molecule does, in a cohort assembled to reveal what the molecule does. The second is what happens when that molecule enters a clinical system shaped by reimbursement, referral patterns, cultural framing, and the prior of a patient who has been on the same inadequate dose of sertraline for four years. The first kind of information is what a Phase 2b trial generates. The second kind is what the drug produces after approval, if approval arrives, over the decade that follows. The short-acting bet is a bet that the first kind of information is most of what matters. The long-acting bet is a bet that it isn’t. It’s the clinic that decides whether it matters.

The patient who will decide which bet was right exists in every interventional psychiatry practice. She is on a submaximal dose of an SSRI her daughter also takes. She has heard things about 5-MeO-DMT from her son’s podcast, none of them reassuring.

The spreadsheet does not see her. The clinician who might prescribe the drug does.

Disclosure: The author has no financial conflicts of interest with GH Research, AtaiBeckley, Compass Pathways, or Cybin. Key opinion leader advisory work with biotech investment firms informs the author’s perspective on commercial viability.

Dark Matter Pragmatism examines the structural forces shaping psychiatric drug development and clinic economics. If you found this useful, consider subscribing.

Subscribe now

Works Cited

1. Compass Pathways plc. COMP006 Phase 3 Topline Results: COMP360 Psilocybin Meets Primary Endpoint in Treatment-Resistant Depression. Press release, February 2026. ClinicalTrials.gov Identifier: NCT05711940. Industry source; not peer-reviewed.

2. Cubała WJ, Bajbouj M, Bauer M, et al; GH001-TRD-201 Investigators. GH001 vs Placebo in Patients With Treatment-Resistant Depression: A Randomized Clinical Trial. JAMA Psychiatry. Published online March 25, 2026. doi:10.1001/jamapsychiatry.2026.0096. ClinicalTrials.gov Identifier: NCT05800860.

3. AtaiBeckley. BPL-003 Clinical Trial Design and Phase 2b Results. Company website and investor materials, 2025–2026. ClinicalTrials.gov Identifier: NCT05870540. Industry source; Phase 2b core data (n=193) not peer-reviewed as of April 2026.

4. U.S. Food and Drug Administration. Psychedelic Drugs: Considerations for Clinical Investigations — Draft Guidance for Industry. June 2023. Docket FDA-2023-D-1987. Operative regulatory document; a finalized version has not been independently confirmed as of April 2026.

5. U.S. Food and Drug Administration. Complete Response Letter to Lykos Therapeutics regarding MDMA-Assisted Therapy for PTSD. August 2024. Specifics drawn from Psychopharmacologic Drugs Advisory Committee materials, June 4, 2024, and subsequent FDA communications.

6. AtaiBeckley BPL-003 Phase 2a open-label cohort publications: Journal of Psychopharmacology (March 2026, Cohort 1 monotherapy, n=12) and CNS Drugs (April 2026, Cohort 2 adjunctive-with-SSRI, n=12).

7. AtaiBeckley. Investor Day Presentation, March 6, 2026. Referenced slides include Slide 47 (Spravato maintenance dosing and WAC assumptions), Slide 49 (three treatments per room per day projection for BPL-003), and Slide 50 (Forian/Komodo Health analysis of 500–600 clinics driving approximately 75% of Spravato volume). Industry source; projections not peer-reviewed.

8. American Medical Association. CPT Code Series 0820T, 0821T, 0822T — Continuous In-Person Monitoring for Pharmacotherapy. Effective January 1, 2025.

9. Cybin Inc. / Douglas Drysdale. Public statements regarding psychedelic REMS architecture and the CYB003 program, 2025–2026.

10. Angermayer C. Public commentary on short-acting psychedelic architecture and clinic economics, 2025–2026.

11. Goodwin GM, Aaronson ST, Alvarez O, et al. Single-Dose Psilocybin for a Treatment-Resistant Episode of Major Depression. New England Journal of Medicine. 2022;387(18):1637–1648. doi:10.1056/NEJMoa2206443. (COMP360 Phase 2b reference for placebo arm performance and prior-use proportion.)

12. Orsini DK, et al. Blinding Integrity in Psychedelic Randomized Clinical Trials. JAMA Psychiatry. Published online April 15, 2026. doi:10.1001/jamapsychiatry.2026.0255. PMID: 41984443.

13. Lii TR, Smith AE, Flohr JR, Okada RL, Nyongesa CA, Cianfichi LJ, Hack LM, Schatzberg AF, Heifets BD. Randomized trial of ketamine masked by surgical anesthesia in patients with depression. Nature Mental Health. 2023;1(11):876–886. doi:10.1038/s44220-023-00140-x. PMID: 38188539. ClinicalTrials.gov Identifier: NCT03861988.

14. Williams ZJ, Barnett H, Szigeti B. Psychedelic Therapy vs Antidepressants for the Treatment of Depression Under Equal Unblinding Conditions: A Systematic Review and Meta-Analysis. JAMA Psychiatry. Published online March 18, 2026. doi:10.1001/jamapsychiatry.2025.4809. PMID: 41848744.

15. GH Research. Phase 2b Demographics Data (GH001-TRD-201): prior psychedelic use approximately 10.0% in the GH001 arm and 12.2% in the placebo arm, presented at ASCP 2025 and included in JAMA Psychiatry 2026 publication (see reference 2). Demographics from conference presentation; figures also tabulated in the peer-reviewed paper.

16. Carhart-Harris R, Giribaldi B, Watts R, et al. Trial of Psilocybin versus Escitalopram for Depression. New England Journal of Medicine. 2021;384(15):1402–1411. doi:10.1056/NEJMoa2032994.

17. Palhano-Fontes F, Barreto D, Onias H, et al. Rapid antidepressant effects of the psychedelic ayahuasca in treatment-resistant depression: a randomized placebo-controlled trial. Psychological Medicine. 2019;49(4):655–663. doi:10.1017/S0033291718001356.

18. Hughes ME, Garcia-Romeu A. Ethnoracial inclusion in clinical trials of psychedelics: a systematic review. eClinicalMedicine. 2024;74:102711. doi:10.1016/j.eclinm.2024.102711. PMID: 39050106.

19. Reckweg JT, van Leeuwen CJ, Henquet C, van Amelsvoort T, Theunissen EL, Mason NL, Paci R, Terwey TH, Ramaekers JG. A phase 1/2 trial to assess safety and efficacy of a vaporized 5-methoxy-N,N-dimethyltryptamine formulation (GH001) in patients with treatment-resistant depression. Frontiers in Psychiatry. 2023;14:1133414. doi:10.3389/fpsyt.2023.1133414. PMID: 37409159. ClinicalTrials.gov Identifier: NCT04698603.

20. ClinicalTrials.gov. A Trial of GH001 in Patients with Treatment-Resistant Depression. NCT05800860. Eligibility criteria as posted.

21. ClinicalTrials.gov. BPL-003 Efficacy and Safety in Treatment-Resistant Depression. NCT05870540. Eligibility criteria as posted.

22. Szigeti B, Heifets BD. Expectancy Effects in Psychedelic Trials. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging. 2024;9(5):512–521. doi:10.1016/j.bpsc.2024.02.004. PMID: 38387698.

23. Pollan M. How to Change Your Mind: What the New Science of Psychedelics Teaches Us About Consciousness, Dying, Addiction, Depression, and Transcendence. New York: Penguin Press; 2018. Netflix documentary adaptation, 2022.

24. Matthew Perry death and subsequent criminal proceedings involving two physicians, October 2023 onward. Summarized from contemporaneous journalistic coverage; no single citation.

25. The Joe Rogan Experience, Episode 1532, featuring Mike Tyson. November 2020.

26. Johnson B. Public social-media posts on 5-MeO-DMT experience and subsequent neurophysiological measurements. Early 2026.

27. Town & Country. Feature on 5-MeO-DMT and elite psychedelic use. January 2022.

28. RAND Corporation. National Survey on Psychedelic Use Among U.S. Adults (n=10,122). September 2025.

29. The White House. Executive Order: Removing Barriers to Psychedelic Drugs as Potential Treatment for Serious Mental Illness. April 18, 2026. https://www.whitehouse.gov/fact-sheets/2026/04/fact-sheet-president-donald-j-trump-is-accelerating-medical-treatments-for-serious-mental-illness/

30. Usona Institute. PSIL201 (Psilocybin for Major Depressive Disorder) CONSORT flow and screening attrition data. JAMA, 2023. Specific citation pending.

31. GH Research. Announces Publication of Phase 2b Results for Mebufotenin (GH001) in JAMA Psychiatry and Reports New Finding of Severity-Independent Efficacy in TRD. Press release, March 25, 2026. Supporting post-hoc analysis (severity-independent efficacy) reported in Psychopharmacology Bulletin (forthcoming). Industry/peer-reviewed; see reference 2 for the primary Phase 2b publication.

32. Teeple A, Joshi K, Zhdanava M, Pilon D, Caron-Lapointe G, Lefebvre P. Access and real-world use patterns of esketamine nasal spray among patients with treatment-resistant depression covered by private or public insurance. Current Medical Research and Opinion. 2023;39(8):1167–1174. doi:10.1080/03007995.2023.2239045. PMID: 37492015.

33. Frontiers in Psychiatry. Psychiatrist perception survey on esketamine patient hesitancy drivers. 2026. Full citation pending.

34. GH Research. Announces FDA Lifts Clinical Hold on GH001, Clearing Path for Global Phase 3 Initiation in 2026. Press release and SEC Form 6-K, January 5, 2026.

35. Guggenheim Partners. Analyst note on GH Research following FDA clinical hold lift. January 2026.

36. AtaiBeckley. Announces Successful End-of-Phase 2 Meeting with FDA for BPL-003. Press release, March 2026.

I recently wrote about the economics of Spravato, and the response was stronger than I expected. The most common pushback was simple: “That’s not how most clinics operate.”

That’s true. Most psychiatric clinics are not running high-throughput interventional programs. Most psychiatrists are not navigating REMS workflows, staggered monitoring schedules, and buy-and-bill reimbursement. Most are doing what the field has always done: managing depression and anxiety with a familiar set of medications, one patient at a time.

But that objection points in the wrong direction.

The question is not what most psychiatrists do. The question is which psychiatrists determine what happens next.

When you look closely at how new psychiatric drugs actually enter practice, a pattern emerges that should be more widely understood. In the case of Spravato, roughly 500 to 600 clinics appear to drive approximately 75% of national prescription volume, out of thousands of certified treatment sites. I work inside one of these clinics. What I am about to describe is the view from the node.

That concentration is not an outlier. It is the structure of the system.

I. The Distribution

The pattern extends well beyond Spravato.

When Axsome Therapeutics launched Auvelity, a branded antidepressant with a novel NMDA mechanism, it did not build a sales force to reach the hundreds of thousands of clinicians legally permitted to prescribe antidepressants. It built one sized to reach approximately 44,000 physicians, because its internal analytics showed that cohort writes more than 80% of all branded antidepressant prescriptions in the United States. The rest of the prescriber universe was, from a commercial standpoint, irrelevant.

Open Payments data tells the same story from a different angle. A study in Psychiatric Services analyzing industry payments to 56,955 psychiatrists from 2015 to 2021 found that the top 10% of psychiatrists received 93.6% of all non-research industry payments. The top 1%, just 427 individuals, received 74.7% of the total, with a median compensation of $362,631. The bottom 90% received a statistically negligible share.

BCG puts it bluntly: 30 to 40% of physicians drive 80 to 95% of sales. After the first two to three years post-launch, more than 90% of growth comes from physicians who already adopted. The rest of the prescriber base is commercially irrelevant.

To put this in familiar terms, we often talk about American income inequality as severe, using the Gini coefficient as our benchmark. By that same measure, prescribing inequality is not just high, it is dramatically worse.

The Gini coefficient for drug prescribing across Medicare Part D providers is 0.759; American household income inequality sits at 0.49.

This is not a normal distribution. It is a power law. And in a power law system, the median clinician is largely irrelevant to the outcome.

If prescribing were evenly distributed, you would expect industry attention to be evenly distributed as well. It is not.

The payment data is not a coincidence. It is a map. Pharma directs its resources toward the same concentration curve that the prescribing data reveals, because it has learned, through decades of commercial experience, that broad outreach is mathematically futile.

But this raises an obvious question: if the distribution is this skewed, what does it actually describe? The answer requires a distinction that most discussions of prescribing behavior fail to make.

Psychiatry does not have one prescribing market. It has two.

The first is the broad market: primary care physicians, generics, SSRIs, legacy stimulants, trazodone at bedtime. General practitioners write roughly 59% of all psychotropic prescriptions in the United States. This market is stable, vast, and operates on something closer to a bell curve. Most prescribers cluster near the mean. Volume is dispersed.

The second is the innovation market: specialists, branded drugs, novel mechanisms, procedural treatments. This market is narrow, volatile, and governed by the power law. It is where Spravato, Auvelity, Caplyta, Cobenfy, and every psychedelic in the pipeline will succeed or fail. It determines which drugs survive their launch window, which companies remain solvent, and which patients gain access to options beyond the generic formulary.

The first market is a bell curve. The second is a power law. Nearly every important decision in psychiatric drug development is made as though the first market is what matters. It is not.

II. What the “600 Clinics” Actually Are

The 600 clinics are not simply practices with more patients. They are systems.

Walk into one and you see multiple treatment rooms running simultaneously, staggered patient arrivals timed to maximize room utilization, medical assistants managing two-hour observation periods while the physician bills separately in an adjacent room. There is a front-desk coordinator who knows which commercial insurers require which prior authorization forms and how to appeal a denial within 48 hours. There is a scheduling template built around the specific time blocks that REMS-monitored treatments demand. There is a referral pipeline, sometimes formal and sometimes organic, that routes treatment-resistant patients from surrounding practices toward the clinic that can actually deliver the branded intervention.

These clinics have payer fluency. Their staff know which CPT codes produce clean claims, which modifiers are required for buy-and-bill drugs, and which insurers will approve a branded medication on the first submission versus which ones require two rounds of documentation and a peer-to-peer call. They have absorbed the operational friction that would overwhelm a two-physician practice with one medical assistant.

And they have referral gravity. Once a clinic becomes known in a metropolitan area as the place that offers Spravato, or TMS, or clozapine management, or any other treatment that most practices cannot deliver, patients begin arriving who have already exhausted the standard formulary. Primary care physicians learn which specialists can actually follow through on a branded recommendation. Therapists learn where to send the patient who needs more than talk therapy and sertraline. The clinic does not advertise this capability. It accumulates through a thousand quiet routing decisions made by other clinicians who tried once, hit the operational wall, and decided it was easier to refer. The referral pipeline, once established, is self-reinforcing. The clinic that already treats 30 Spravato patients has more operational fluency than the clinic considering its first. The next referral flows to the same place.

This is not a difference of clinical philosophy. It is a difference of infrastructure.

Consider a specific example. A psychiatrist who runs a Spravato program is not philosophically committed to esketamine. She built the program because she had the rooms, the staff, the payer relationships, and the patient volume to make it viable. Her colleague across town may hold identical clinical views about esketamine’s role in treatment-resistant depression. He may have read the same TRANSFORM-2 data, attended the same conference presentations, and reached the same conclusions about which patients might benefit. But he practices in a three-room office with one MA and no dedicated prior authorization staff. She prescribes Spravato. He refers patients to her. The drug reaches the patient not because of a difference in clinical judgment but because of a structural fact.

What distinguishes these clinics is not opinion. It is position within the system.

III. The Filters

If the distribution is this concentrated, the next question is mechanical: why? The answer is not conspiracy. It is selection. Five structural filters operate simultaneously, and a clinician must pass through all of them to enter the power-law distribution. Fail one, and the drug might as well not exist.

Operational friction. Spravato requires REMS certification of both the prescriber and the facility, two-hour in-clinic monitoring per session with mandatory vital signs, and a documentation workflow that produces a paper trail for every administration. A single Spravato patient occupies a treatment room for roughly 2.5 hours. A 15-minute medication management appointment generates comparable relative revenue with a fraction of the complexity. The barrier is not knowledge. It is workflow tolerance. Clozapine is the cautionary precedent: 60% of prescribers reported that its REMS requirements delayed care, and the drug remained dramatically underused for decades despite unmatched efficacy for treatment-resistant schizophrenia. The REMS was not eliminated until 2025.

This is not just a regulatory detail. It is a geographic and structural constraint.

The map tells the story more efficiently than the data: roughly 40,000 general psychiatrists scattered across the country, and a far smaller number of certified treatment centers clustered in metropolitan corridors.

Economic fit. Many novel psychiatric treatments operate under buy-and-bill reimbursement, meaning the clinic purchases the drug, administers it, and submits claims retrospectively. A denied claim means the clinic absorbs the full cost. Reimbursement lags of 30 to 90 days create cash flow pressure that solo practitioners and small groups cannot sustain. Payer mix compounds the problem; a clinic whose patients are predominantly on Medicaid or narrow-network commercial plans may find that a branded drug is technically available but practically unprescribable, because the margin after acquisition cost and administrative overhead is negative. A drug can be clinically effective and financially non-viable in the same building.

Risk posture. Novelty tolerance varies enormously across clinicians, and the variance is not random. A Harvard study tracking adoption of three novel cardiovascular drugs found that 63.8% of physicians adopted zero new drugs in the first 15 months after introduction. Only 1.2% adopted broadly across multiple novel agents. Prescribing a new drug that produces an adverse event carries a different medicolegal texture than prescribing a generic that produces the same event. The clinician who chose the novel agent will be asked to justify the choice. The clinician who chose the generic will not. Most clinicians who avoid new drugs are not wrong. They are conservative, and the system rewards that conservatism.

Information exposure. Pharmaceutical representatives do not visit all psychiatrists equally. They visit the ones whose prescribing data, payer mix, and patient volume make them commercially viable targets. IQVIA and Komodo Health sell prescriber profiling products that rank clinicians by decile; a D1 prescriber receives frequent rep visits, sample deliveries, and invitations to speaker programs, while a D7 prescriber may never see a rep at all. Conference attendance, advisory board invitations, and continuing medical education sponsored by industry cluster awareness among the same subset that passes the other filters. The effect compounds through peer networks: 84 of 103 frequently prescribed drugs show diffusion patterns consistent with social contagion, meaning a clinician embedded in a network of early adopters encounters new drugs under fundamentally different conditions than one practicing in relative isolation. A psychiatrist whose three closest colleagues are prescribing Auvelity will hear about the drug in a hallway conversation. A psychiatrist whose colleagues are not will hear about it, if at all, from a journal article she reads six months later. Awareness does not diffuse evenly. It clusters.

Patient stream. A psychiatrist in community practice sees anxiety, adjustment disorders, mild-to-moderate depression, ADHD follow-ups. The patients who need Spravato, clozapine, or a novel antipsychotic rarely walk through this door. Treatment-resistant populations concentrate at referral centers, academic programs, and specialty clinics; the distribution of clinical severity mirrors the distribution of prescribing, because the patients with the most complex conditions are routed toward the clinicians with the most infrastructure. The clinician who sees these patients is already inside the power-law distribution before she prescribes anything novel. She does not need to be convinced that a new drug is worth trying. Her patients have already failed everything else. Some clinicians do not just prescribe differently. They see different patients. And the patients they see are the ones whose treatment decisions generate the data points that the entire commercial apparatus is built to capture.

Pass all five filters, and you enter the distribution. Fail one, and the drug might as well not exist.

IV. The Taxonomy

If the distribution is this concentrated, and the filters this structural, then how does the market actually classify the clinicians who pass through? And would the clinicians recognize themselves in that classification?

There are two taxonomies of the psychiatric field. One belongs to the prescribers. The other belongs to the commercial apparatus that determines which drugs reach which patients. They describe the same people. They do not describe the same thing.

How prescribers see themselves. Clinicians tend to describe their prescribing in clinical terms: conservative or early-adopting, generalist or subspecialized, guideline-driven or experience-informed. They notice who among their peers reaches for a branded atypical as a first-line augmentation and who exhausts the generic formulary before considering anything with a copay card. These distinctions are real. They shape whether a patient hears about Auvelity in month two of treatment or month twelve. And they are almost entirely invisible to the commercial infrastructure that determines which drugs reach which patients.

How the market sees them. From the perspective of anyone launching a drug, psychiatry does not look like a profession. It looks like a segmentation problem.

The non-participant prescribes generics from a stable, familiar formulary. SSRIs, bupropion, legacy stimulants. Will not adopt a new drug regardless of evidence, rep contact, or access. Invisible to the market.

The deferred adopter is guideline-driven and risk-averse. Will prescribe a branded drug in year three or four, once it is no longer novel, once the safety data have matured, once a colleague she trusts has used it. Important for maintenance volume. Not decisive for launch success.

The receptive generalist listens to reps, reads the trials, tries new drugs selectively. Moves early adoption curves but lacks the infrastructure to prescribe at volume. This is the mid-decile target, and the stakes are high: studies of antipsychotic prescribing show that the average psychiatrist directs roughly 66% of all prescriptions within a drug class to a single favored agent. Win this clinician’s loyalty and you own the patient panel.

The infrastructure node is the high-throughput clinic. Multiple rooms, dedicated staff, operational workflows built to deliver complex treatments. This is where drugs become real; where the gap between trial protocol and clinical delivery is actually bridged.

The concentrated specialist runs a TRD clinic, an OCD program, a clozapine service, an interventional psychiatry practice. Sees the patients the trials were designed for. Small volume relative to general practice, but the patients who walk through this door are the ones the drug was built to treat.

The academic signal generator has low personal prescribing volume and high influence on the prescribing behavior of others. Publishes, speaks at conferences, advises companies, sits on guideline committees. Shapes belief, not volume.

Drug adoption is driven disproportionately by the receptive generalists, infrastructure nodes, and concentrated specialists. The rest of the field follows, slowly or not at all.

But within this segmentation lies a distinction that the commercial apparatus has only recently begun to internalize; a distinction that is, I think, the most important structural observation in this essay.

The most important clinicians are not necessarily the highest-volume prescribers. They are the most connected.

A high-volume prescriber practicing in relative isolation is an island: high revenue per prescriber, but no multiplier effect. Changing her prescribing behavior yields linear returns. A well-connected prescriber embedded in an academic system or referral network is a hub: moderate personal volume, but exponential network spillovers. Converting the hub yields cascading adoption among the peers who observe and trust her clinical decisions. A high-volume island can be commercially irrelevant. A well-positioned hub can reshape a market. Research on drug diffusion through physician networks found that each time a targeted physician receives industry compensation, peer use of the promoted drug increases by 2% on average, with peer spillovers accounting for a full quarter of total prescription increases. Pharmaceutical companies disproportionately target specialist physicians with many peers, confirming that at least some commercial strategies have internalized the volume-influence distinction.

Janssen classified 886 national key opinion leaders by commercial utility: “Advocates,” “Neutrals,” “Opponents.” They invested $1.3 million building solicitation programs exclusively for the Advocates. A Cephalon internal strategy document, later released in litigation, stated the logic as plainly as anyone ever has: “A small number of prescribers (<500) generate large dollar volume.”

The prescriber’s self-taxonomy and the market’s taxonomy rarely overlap. A clinician who considers herself a careful, evidence-based generalist may be classified by the commercial team as a non-participant whose decision-making is commercially irrelevant. A clinician who considers himself an independent thinker making individualized prescribing decisions may be, from the field force’s perspective, a mid-decile target whose entire panel is the prize.

V. This Is Not About Spravato

The instinct, at this point, is to localize. Spravato has a REMS. Spravato requires in-clinic monitoring. Spravato is a special case.

It is not.

Auvelity has no REMS requirement. It is an oral tablet prescribed like any other antidepressant. And yet Axsome built its entire commercial strategy around the same concentrated prescriber minority described in Section I, because the power law holds regardless of delivery complexity. Caplyta entered the market as an adjunctive treatment for bipolar depression and MDD; its early prescribers were the same subset of psychiatrists who were already prescribing branded atypicals for mood disorders. Vraylar’s adoption for bipolar depression and MDD augmentation followed the same infrastructure nodes that had absorbed the prior generation of branded antipsychotics.

Who actually writes these prescriptions? Not psychiatrists as a group. A subset.

The concentration is not just structural. It is temporal.

The first 90 days post-launch are disproportionately predictive of a product’s long-term commercial trajectory. Early prescriber adoption patterns are extremely difficult to reverse. A separate analysis of eight drug launches in primary care found that adoption follows an exponential diffusion curve that permanently plateaus between month 6 and month 23. Once the curve flattens, it does not resume climbing.

This means the power law has a temporal corollary. The same 600 clinics that determine volume also determine trajectory. A drug that lands with the right 5 to 10% of the prescriber universe in its first quarter will track toward commercial success. A drug that misses them will plateau, and its commercial ceiling may be permanently set. By the time most clinicians encounter a drug, the curve has already bent.

Two countertrends deserve acknowledgment.

GLP-1 receptor agonists are the apparent exception: broad diffusion, more than 7% of all U.S. prescriptions by late 2025. But the conditions that permitted that diffusion (exceptional efficacy, enormous primary care applicability, no REMS, no monitoring burden) do not exist in psychiatry. GLP-1 drugs clarify the rule rather than contradicting it.

Prescribing by nurse practitioners and physician assistants is growing at 11.8% annually for antipsychotics, while psychiatrist-written prescriptions are declining. This is a real structural shift. But NPs and PAs are largely diffusing into community care and maintenance prescribing; they are not, in the main, building the high-resource specialty hubs that adopt novel mechanisms. The broad market is democratizing. The innovation market is not.

VI. Why This Matters

The power law is not merely a description of prescribing behavior. It is the operating system of psychiatric drug development. Its implications are uncomfortable precisely because they are structural.

Pharma does not behave as if the average psychiatrist matters. Its entire commercial architecture assumes they do not. The Axsome targeting model described earlier is not an outlier; it is the standard playbook. A survey by ZS Associates found that 94% of senior pharma executives consider Key Account Managers a top strategic priority. The KAM architecture exists for a specific reason: the individual prescriber has lost commercial relevance. The institutional account is the unit of sale. The field force does not try to reach the profession. It tries to reach the nodes.

Payers reinforce the concentration, whether they intend to or not. Step therapy and prior authorization do not flatten prescribing behavior. They select who participates. A solo practitioner hits the prior authorization wall and reverts to the generic formulary. An institutional hub punches through it with dedicated staff, peer-to-peer calls, and appeals infrastructure. Prior authorization does not prevent branded prescribing; it routes branded prescribing through the clinics with the administrative capacity to survive the process. The ostensible purpose is cost containment. The structural effect is concentration.

Infrastructure determines which drugs survive. A drug that requires a two-hour observation period will succeed only if enough high-throughput clinics can absorb that time. A drug that requires a licensed therapist in the room will succeed only if enough clinics can afford that staffing model. A drug that requires cold-chain logistics and buy-and-bill will succeed only if enough clinics can manage the financial risk. Infrastructure is not a downstream implementation detail. It is the primary constraint on whether a molecule that works in a trial becomes a treatment that works in the world. In a previous essay, I asked how many drugs psychiatry has classified as pharmacological failures that were actually engineering failures. The power law adds a corollary:

How many drugs failed not because clinicians rejected them, but because the right clinicians never encountered them?

Psychedelics will follow the same pattern. If psilocybin receives FDA approval, it will concentrate in the same structural minority that currently delivers Spravato, and for the same reasons: REMS constraints, monitoring requirements, staffing costs, and reimbursement friction. The 600-clinic model is not a description of Spravato’s market. It is a prediction about every novel psychiatric treatment that requires infrastructure to deliver.

VII. Return to the Clinic

A patient sits across from me and asks about Caplyta. She saw the commercial; the one with the woman peeling away layers of grey. She has been on sertraline for three years and feels flat, emotionally blunted in a way she can describe precisely but that her previous psychiatrist attributed to residual depression. She does not think it is residual depression. She thinks it is the sertraline. She wants to try something different.

I know three things simultaneously: that lumateperone might help, particularly if what she is experiencing is SSRI-induced emotional blunting rather than incomplete remission; that the prior authorization will take five to seven business days and has a meaningful chance of being denied on the first pass, requiring a letter of medical necessity and possibly a peer-to-peer call; and that in a practice without dedicated support for that process, the likelihood of the patient actually receiving the medication drops sharply.

Not because the drug is wrong for her, and not because the clinician is unwilling, but because the path between “this might work” and “the pharmacy will fill it” becomes unreliable without the infrastructure to carry it through.

A different patient, later that afternoon. Treatment-resistant insomnia. Years on zolpidem, escalating doses, tolerance that has turned a sleep aid into a dependency he is now afraid to stop. He has tried melatonin, hydroxyzine, and CBT-I with partial compliance. An orexin antagonist becomes the next step; lemborexant, probably, based on the next-day residual data and the fact that the mechanism is entirely distinct from anything he has tried. Not because a rep pitched it. Because the clinical question (can we get you off a drug that stopped working two years ago?) demands a different pharmacological approach, and the prior authorization pathway for this drug class is one that can be navigated reliably based on his insurance.

One clinician can make that recommendation because the payer relationships, prior authorization workflow, and clinical volume are in place to absorb the friction of getting a branded drug covered. Another clinician, trained in the same program and treating a similar population, reaches for trazodone. Not because trazodone is the better drug, but because the pathway is shorter.

Both of these prescribing decisions feel clinical. Both feel individualized. Both were shaped by forces that were in place long before the patient sat down.

The referral network that routes treatment-resistant patients to one clinic rather than to another ten miles away. The staffing model. The payer mix. These are not clinical skills. They are structural positions. They are the reason clinicians appear in one decile rather than another.

The patient experiences it as a recommendation. The clinician experiences it as an independent clinical judgment. The market experiences it as a data point in a distribution that was skewed before either of them walked into the room.

Not by consensus. By concentration.

Adam D. Borecky, MD is a board-certified psychiatrist practicing in Southern California and Western Michigan. He has no financial conflicts of interest with Axsome, Intra-Cellular Therapeutics, AbbVie, or any company whose products are discussed in this essay. Dark Matter Pragmatism is an independent publication.

Works Cited

1. Forian/Komodo Health Prescription Data, AtaiBeckley Investor Day (March 6, 2026). 7,000-8,000 interventional psychiatry clinics have prescribed Spravato; 500-600 clinics drive 75% of prescription volume.

2. Axsome Therapeutics Q2 2023 Earnings Call / Investor Disclosures. ~44,000 physicians write >80% of branded antidepressant prescriptions. SEC filing: EX-99.1.

3. Havlik P, et al. “Industry Payments to Psychiatrists, 2015-2021.” Psychiatric Services. 2024. Top 1% of psychiatrists received 74.7% of all industry payments; top 10% received 93.6%.

4. Boston Consulting Group. “Three Steps to a More Productive Pharma Sales Team.” BCG Publications, 2018. 30-40% of physicians drive 80-95% of sales; >90% of growth post-launch comes from existing adopters.

5. Gini coefficient data: Bress AP, et al. “Visualizing Nationwide Variation in Medicare Part D Prescribing Patterns.” BMC Medical Informatics and Decision Making. 2018;18:128. G = 0.759 for individual drug prescribing.

6. Levine Taub A, Kolotilin A, Gibbons RS, Berndt ER. “The Diversity of Concentrated Prescribing Behavior.” NBER Working Paper 16823. 66% of antipsychotic prescriptions directed to a single favored drug per physician.

7. Shin SS, et al. “Patterns and Predictors of Physician Adoption of New Cardiovascular Drugs.” Healthcare. 2018;6(2):48. 63.8% adopted zero new drugs in first 15 months; 1.2% were broad adopters.

8. Agha L, Zeltzer D. “Drug Diffusion Through Peer Networks: The Influence of Industry Payments.” American Economic Journal: Economic Policy. 2022. 2% peer increase per payment; 25% of adoption from spillovers.

9. Gac P, et al. “Pharmaceutical Industry Use of Key Opinion Leaders to Market Opioids.” PMC. 2024. Cephalon internal document: “A small number of prescribers (<500) generate large dollar volume.” Janssen KOL classification: 886 national KOLs; $1.3M in Advocate solicitation programs.

10. ZS Associates. Key Account Manager survey: 94% of senior pharma executives consider KAMs a top strategic priority.

11. FDA eliminates clozapine REMS program requirement. FDA announcement, March 2025. 60% of prescribers reported REMS-related delays (STAT News, February 2025).

12. IQVIA Modern Launch analysis (2024). Oncology: 70% of promotional engagement concentrated on 17% of providers (2,747 physicians).

13. Shifts in Antipsychotic Prescribing by Clinician Type for Medicare Part D Beneficiaries. AAPP Perspective. 2026. NP prescriptions grew 11.8% annually; psychiatrist prescriptions declined 3.2% annually.

14. GLP-1 prescription volume: Forbes, January 2026. Anti-obesity and anti-diabetic prescriptions account for approximately 7% of all U.S. prescriptions.

15. Pharmaceutical Launch Strategy analysis: “The first 90 days post-launch are disproportionately predictive of a product’s long-term commercial trajectory.”

16. Spravato REMS program: spravatorems.com. Certification requirements for prescribers and facilities; two-hour monitoring mandate.
    
    

The clinical model may work at the level of the patient. The question is whether it works at the level of the room.

I. The Practitioner’s Spreadsheet

There is an uncomfortable question beneath any Spravato program: what is a course of treatment worth to the clinic delivering it?

“A course of treatment” here means something concrete. Industry billing data tell a consistent story: buy-and-bill reimbursement’s net margins land somewhere between $300 and $600 per treatment, depending on dose and payer mix [25]. Over a maintenance year of 20 to 25 sessions, that translates to roughly $10,000 to $15,000 in gross profit per patient [26].

The more important number is not per session or per patient. It is per room. A well-run treatment room, operating at 60-80% capacity, can cycle multiple patients through staggered two-hour observation windows, generating several thousand dollars in profit per day. These numbers are consistent with my own experience prescribing Spravato.

Most clinicians never need to think about these numbers directly. The economics are handled elsewhere; billing, administration, someone whose job it is to make the margins work so the doctor does not have to. It is easy to preserve a certain narrative about the work: physician and patient aligned, everyone else somewhere outside the room.

I did not start out differently. I built a Spravato program because I wanted to offer something new to patients who had run out of options. Early on, we experimented with different delivery models before settling on buy-and-bill, which turned out to be both more scalable. At some point, the spreadsheet stops being abstract. It becomes part of how the treatment actually functions.

This essay’s central question is simple to state and difficult to resolve:

We are still operating in something like fog of war with psychedelics in psychiatry. Two JAMA Psychiatry papers published on March 18, 2026 illustrate why.12

Neither paper refutes psilocybin’s antidepressant activity. Together they suggest that the measured advantage over existing treatments is less decisive than the bullish narrative implies. This essay is not a pharmacometric exercise. It is an attempt to think past the efficacy question into the economic realities of getting new treatments to patients.

If psilocybin’s symptom-score separation is modest, then the value proposition must be something else: durable remission from fewer sessions. And the question confronting clinics is economic.

What does it cost to deliver a treatment whose clinical case rests on fewer visits per year rather than a larger effect size per visit?

II. How a Flawed Drug Built a Blockbuster

Spravato’s registration efficacy package was mixed to modest.3

None of this prevented $1.7 billion in global revenue in 2025 [6]. Over 200,000 patients treated worldwide. And yet the drug has reached only about 3% of the estimated three million American adults with treatment-resistant depression. A blockbuster by revenue; barely a footnote by penetration.

The explanation is structural advantage. Some comparative analyses suggest broadly similar antidepressant effects between intranasal esketamine and IV ketamine [7]. But IV ketamine has no FDA approval for depression, no standardized billing codes, and no insurance coverage pathway for most patients [8] [9]. By the end of 2023, office-based ketamine prescribing declined while Spravato kept climbing.

The clinic-level economics explain this. A typical Spravato patient receives about 22 sessions per year: eight during induction, four in the second month, roughly ten during maintenance. Medicare reimburses the bundled G2082 code at a national average of roughly $950; major commercial payers average somewhat less, though buy-and-bill arrangements with separate drug codes often push total reimbursement higher [25]. Drug acquisition runs $590 to $885 per kit. The structural insight is the decoupling described in the opening: the MA watches patients during monitoring while the physician bills E&M elsewhere, and staggered scheduling lets a single room cycle multiple patients per day. That separation is the invisible engine of Spravato’s economics, and it is the thing most at risk when a new treatment enters the picture.

But the revenue model has a ceiling, and honesty requires naming it. In the Clarivate claims analysis, only 43.2% of patients completed the full induction phase [10]. Six-month persistence ranged from 25% to 37%. SUSTAIN-3 provides the optimistic ceiling: among trial completers who stayed on maintenance, median exposure was 37.7 months and remission rates rose from 35.6% to 46.1% [11]. The gap between that figure and the real-world persistence rate is one of the most instructive tensions in interventional psychiatry. The patients who respond to and stay on Spravato do fairly well. But this is a fraction of the potential number of TRD patients.

That tension defines the baseline any new treatment must compete against. Not a theoretical number; a real one, tempered by real-world attrition. These are the economics a new treatment must match, or explain why it does not need to.

III. What Psilocybin Asks of a Clinic

What is fundamentally different about psilocybin vs Spravato at the operational level?

The psilocybin-assisted treatment model, as developed in Compass Pathways’ COMP360 program and others, is by any clinical standard well designed. A patient takes a single 25 mg oral capsule. Post-administration monitoring lasts six to eight hours. Rescue medication use in prior trial samples was rare, under 1.5% [12].

The retreatment cadence is where the model diverges from everything interventional psychiatry has built its revenue around. In COMP005, 70% of patients were re-dosed at weeks 10 to 14 [13]. For modeling purposes, I assume two induction doses and one to two later retreatments annually, or roughly three to four sessions per year. The exact commercial regimen is not yet established, but these numbers are consistent with the emerging trial data. Compared to Spravato’s 22 sessions, that is an 85% reduction in visit frequency. For patients, this is unambiguously better. Fewer clinic visits, fewer hours of post-dose monitoring, fewer days rearranged around a treatment schedule. If treatment burden is the metric, psilocybin is the superior product by a wide margin.

The staffing model is the second divergence. Compass’s Psychological Support Model specifies licensed mental health professionals, not MAs [12]. In recent commercial messaging [24], Compass has described a lighter version: the patient self-administers and is monitored by a trained licensed clinician, with no psychotherapy during dosing. That is a meaningful shift from the trial model. It does not solve the economics. It changes them.

Line up the parameters against the Spravato model described above. Psilocybin: six-to-eight-hour monitoring window instead of two. Licensed clinician present instead of an MA. Physician or therapist potentially re-coupled to the treatment room instead of free to bill elsewhere. Three to four sessions per year instead of 22. The exact REMS requirement is not yet known; Compass’s trial model and the FDA’s signals from the Lykos PDAC proceedings [14] both point toward qualified clinical oversight, but the final staffing mandate could be lighter or heavier than what I model here.

Taken together, this is a different operational model.

Given these constraints, what does the math actually require?

IV. The Breakeven Math

Two comparisons matter. Per-patient annual parity: if a Spravato patient generates $10,000 to $15,000 in annual profit, a psilocybin patient at three sessions per year must generate roughly $3,300 to $5,000 per session to match. At four sessions, the target drops somewhat. The 0820T-series psychedelic monitoring codes [15] might produce $1,300 to $1,600 for eight hours at modeled rates; these are projected figures, not established payment. The remaining $2,000 to $3,300 must come from drug margin via buy-and-bill.

Per-room-day parity is harsher. A busy Spravato room cycling multiple patients through staggered two-hour windows can produce several thousand dollars in daily profit. A single psilocybin session occupying that room for the full day must match that figure. This is almost certainly unachievable under current assumptions. A licensed therapist at several times the MA hourly rate for eight hours adds staffing cost; and the physician may be partially or fully tethered to the psilocybin session depending on what is required in the 0820T bundled codes.

The honest assessment:

And the distinction matters, because it determines how clinics will absorb the treatment. No rational operator will convert a high-utilization Spravato room into a psilocybin room. The opportunity cost is too large. What clinics will do instead is dedicate an underused room to psilocybin on designated days, absorb a revenue reduction to maintain clinical completeness, or decline to offer the treatment and refer out.

None of these produces the rapid infrastructure-wide adoption that bullish models assume. They produce patchwork adoption: a few hundred motivated clinics, unevenly distributed, serving a fraction of the TRD population that could benefit.

Some argue that short-acting psychedelics (5-MeO-DMT) will rescue psychedelics from the constraints of the psilocybin model [16]. Its in-clinic stay is approximately two hours, which matches Spravato’s logistical profile almost exactly. But session duration is only half the variable. If REMS requires a licensed therapist regardless of how long the session lasts, a two-hour psychedelic session with a licensed clinician is still more expensive than a two-hour Spravato session with an MA. The question is not just how long the session is. It is who must be in the room.

So why would any clinic add this if doing so produced a throughput collapse?

One answer is in the billing codes, the staffing rules, and the decisions that payers and regulators have not yet made.

At that point, the question is no longer what the model requires, but how clinics and payers respond to it.

V. The CPT Code Bet and Clinic Conversion

Psilocybin’s economic narrative rests on the new Category III psychedelic monitoring codes (0820T series), allowing clinics to bill hourly for continuous in-person monitoring [15]. Compass has stated that these codes mean “provider economics will not be impacted regardless of treatment time required” [24]. Category III codes carry inherent uncertainty; adoption is slower and less consistent than Category I.

I have some experience (scars) from this particular species of optimism.4 For a treatment whose viability depends on those codes closing the gap per session described above, “historically slower and less consistent” is not a reassuring qualifier.

The rollout story Compass is telling is broader than the caricature I sketched earlier [24]. The company is not simply assuming that high-volume Spravato rooms will flip wholesale into psilocybin rooms. It is building multiple entry points: large interventional networks like Greenbrook, decentralized therapist practices through Journey Clinical, and safety-net CCBHCs through partners like HealthPort.

Credit where it is due. The room-economics problem remains.

The high-volume Spravato center still has the most to lose. Even under the lighter commercial model, a psilocybin session still occupies a room for most of the day. The mixed-service independent practice, probably accounting for a quarter or less of total Spravato volume, is the more plausible early adopter; it may have underused space and tolerance for a lower-yield room if the clinical case is strong. Academic centers and CCBHCs adopt for mission and funding reasons; their economics are structurally different.

On the patient side, the psilocybin-aligned newcomer is the clean additive case: no cannibalization, new revenue. The Spravato maintenance patient seeking de-escalation remains the sharpest problem: a clinical win and a revenue loss. The KOL panels stepped around that case by arguing the market is not zero-sum, that patients may layer TMS or Spravato between psilocybin sessions. Fair enough. But those are answers to a broader market question. They are not answers to the narrower question of what a clinic does when one of its highest-yield maintenance patients asks to de-intensify.

The partial responder may be the friendliest archetype commercially: COMP360 as a first move, Spravato or TMS layered around it, perhaps a sequence in which the clinic makes money on more than one modality while the patient gets closer to remission than any single treatment delivered alone.

Compass is counting on several things to go right at once: a lighter monitoring model than skeptics assume, payer behavior friendlier than Category III billing typically gets, and enough incremental patient flow to soften cannibalization. A more plausible story than the caricature. Also a more conditional one.

Same drug class. Different room.

VI. The Payer Wedge

The preceding sections assume that psilocybin and Spravato compete for patients. But there is a second-order effect that does not require psilocybin to be adopted at scale. It only requires psilocybin to be approved.

Spravato’s revenue model depends on chronic maintenance: 22 sessions per year, reimbursed by insurers who have built prior-authorization infrastructure around the assumption that no durable alternative exists. The moment a durable episodic alternative receives FDA approval, that assumption changes. Not because payers will immediately prefer psilocybin; they may not. But because the existence of an approved episodic treatment gives payers a tool they currently lack: bargaining power over Spravato’s maintenance economics. Payers are already using step therapy, duration caps, and tight clinical criteria on Spravato [18]. An approved episodic competitor gives them additional justification to sharpen those tools. Step-therapy edits could require psilocybin before extended Spravato maintenance. Utilization reviewers could scrutinize maintenance duration more aggressively. Reimbursement rates could face compression. None of these require payers to believe psilocybin is clinically superior. They require only that an approved alternative exists, because the existence of a choice is the prerequisite for cost-containment pressure.

The first exclusivity step-down arrives January 2028 [17]. Core patents extend into the 2033-to-2040 range and could delay generic competition through litigation. But the expiration gives payers additional ammunition in contract negotiations. If generic esketamine enters the market within a few years of psilocybin’s approval, clinics face erosion from both directions: payer pressure on Spravato reimbursement rates from above, and generic competition on drug margins from below.

The institutional logic is consistent with how payers have historically handled high-cost chronic therapies when episodic alternatives emerge. The tightening will be incremental, not a cliff. But savings accrue to payers, not providers.

VII. What I’m Watching For

Five variables will determine whether psilocybin-assisted treatment is commercially viable in the clinics that currently deliver interventional psychiatry, or whether it becomes another treatment that works but that patients cannot access at scale.

The first is the REMS staffing requirement. MA versus licensed therapist; this single regulatory decision will do more to determine psilocybin’s commercial trajectory than any efficacy readout. The second is actual payer reimbursement for the 0820T-series codes. Category III codes with no payment history are promises, not revenue. The third is drug pricing and buy-and-bill margin: whether Compass prices COMP360 for clinic-level profitability or payer palatability. The fourth is DEA scheduling and state-level implementation; friction that does not appear in national adoption models.

There is a fifth variable I have not addressed in detail because it is mechanistic rather than economic, but its economic consequences are real. Psilocybin’s Phase 3 program required withdrawal from serotonergic antidepressants before dosing. Whether this is pharmacologically necessary remains an open question.5 Meanwhile, esketamine has demonstrated monotherapy efficacy without any oral antidepressant requirement [23]. If psilocybin’s labeling requires SSRI discontinuation and esketamine’s does not, the delivery friction compounds everything else in this essay: additional clinic visits for taper management, a window of untreated or undertreated depression before the dosing session, and patient attrition during the gap.

I started this essay with a spreadsheet. Small clinics like where I work might be willing to absorb a revenue reduction for the right patients. But willingness to take a loss is a personal clinical decision, not a scalable business model. Seven thousand treatment sites will not adopt a treatment that costs them money. The question is whether the reimbursement architecture can be built fast enough and priced generously enough to prevent psilocybin from becoming another entry in the catalog of treatments that psychiatry could not deliver; not because the pharmacology failed, but because the economics did.

The clinical model may work at the level of the patient. The question is whether it works at the level of the room.

Disclosure: The author is a board-certified psychiatrist who is an independent contractor with a Spravato-certified treatment center. He has no financial conflicts of interest with Compass Pathways or Johnson and Johnson. Economic figures cited in this essay are drawn from publicly available billing guides, CMS fee schedules, and published practice-economics analyses; they do not represent proprietary contract rates.

Works Cited

[1] Mertens LJ, Koslowski M, Betzler F, et al. Efficacy and safety of psilocybin in treatment-resistant major depression: the EPISODE randomized clinical trial. JAMA Psychiatry. Published online March 18, 2026. doi:10.1001/jamapsychiatry.2026.0132

[2] Williams ZJ, Barnett H, Szigeti B. Psychedelic therapy vs antidepressants for the treatment of depression under equal unblinding conditions: a systematic review and meta-analysis. JAMA Psychiatry. Published online March 18, 2026. doi:10.1001/jamapsychiatry.2025.4809

[3] Popova V, Daly EJ, Trivedi M, et al. Efficacy and safety of flexibly dosed esketamine nasal spray combined with a newly initiated oral antidepressant in treatment-resistant depression: a randomized double-blind active-controlled study. Am J Psychiatry. 2019;176(6):428-438. (TRANSFORM-2; NCT02418585)

[4] Fedgchin M, Trivedi M, Daly EJ, et al. Efficacy and safety of fixed-dose esketamine nasal spray combined with a new oral antidepressant in treatment-resistant depression: results of a randomized, double-blind, active-controlled study (TRANSFORM-1). Int J Neuropsychopharmacol. 2019;22(10):616-630. (NCT02417064)

[5] Ochs-Ross R, Daly EJ, Zhang Y, et al. Efficacy and safety of esketamine nasal spray plus an oral antidepressant in elderly patients with treatment-resistant depression (TRANSFORM-3). Am J Geriatr Psychiatry. 2020;28(2):121-141. (NCT02422186)

[6] Johnson & Johnson. Q4 and full-year 2025 earnings results. January 2026. https://investor.jnj.com

[7] Elmosalamy A, Tarikogullari I, Patarroyo-Rodriguez L, et al. Intravenous ketamine versus esketamine for depression: a systematic review and meta-analysis. Ther Adv Psychopharmacol. 2025;15. doi:10.1177/20451253251394127

[8] Compass Pathways January 2026 investor webinar, citing spravatohcp.com data pulled December 29, 2025 (~6,800 Spravato treatment clinics in U.S.). See also Jefferies analyst report (~2,800 certified treatment centers by narrower definition).

[9] IQVIA esketamine and ketamine dispensing data, projected through end-2023.

[10] Clarivate claims analysis: Spravato real-world induction completion and persistence data.

[11] Wajs E, Aluisio L, Grunfeld J, et al. Esketamine nasal spray plus oral antidepressant in patients with treatment-resistant depression: assessment of long-term safety in a phase 3, open-label study (SUSTAIN-2). J Clin Psychiatry. 2020;81(3):19m12891. Long-term maintenance data from SUSTAIN-3 extension analyses.

[12] Compass Pathways. Compass Psychological Support Model (CPSM): staffing requirements, training framework, and fidelity assessment for COMP360 psilocybin therapy sessions.

[13] Compass Pathways. COMP005 Phase 3 trial (NCT05624268): topline results including retreatment cadence data.

[14] US Food and Drug Administration. Psychopharmacologic Drugs Advisory Committee (PDAC) meeting, June 4, 2024. Discussion of therapist qualifications and REMS requirements for psychedelic-assisted therapy (Lykos/MDMA context).

[15] AMA CPT Category III codes: 0820T (continuous in-person monitoring, first hour, physician/QHP); +0821T (each additional hour); +0822T (clinical staff under physician direction, per hour). Effective 2025.

[16] ATAI Life Sciences / Beckley Psytech. BPL-003 (5-MeO-DMT): Breakthrough Therapy designation; Phase 3 initiation anticipated Q2 2026; approximately 2-hour in-clinic stay.

[17] Spravato patent estate: controlling regulatory exclusivity expires January 17, 2028. Core patents extend through approximately 2033-2040.

[18] Commercial payer Spravato coverage policies: step-therapy requirements, duration caps, and clinical criteria variation documented across major commercial plans (2024-2026).

[19] Goodwin GM, Croal M, Feifel D, et al. Psilocybin for treatment resistant depression in patients taking a concomitant SSRI medication. Neuropsychopharmacology. 2023;48:1492-1499. doi:10.1038/s41386-023-01648-7

[20] Becker AM, Holze F, Grandinetti T, et al. Acute effects of psilocybin after escitalopram or placebo pretreatment in a randomized, double-blind, placebo-controlled, crossover study in healthy subjects. Clin Pharmacol Ther. 2022;111:886-895.

[21] Gukasyan N, Griffiths RR, Yaden DB, Antoine DG, Nayak SM. Attenuation of psilocybin mushroom effects during and after SSRI/SNRI antidepressant use. J Psychopharmacol. 2023;37:707-716. doi:10.1177/02698811231179910

[22] Erritzoe D, Barba T, Greenway KT, et al. Effects of discontinuation of serotonergic antidepressants prior to psilocybin therapy versus escitalopram for major depression. J Psychopharmacol. 2024;38:458-470.

[23] Janik A, Qiu X, Lane R, et al. Esketamine monotherapy in adults with treatment-resistant depression: a randomized clinical trial. JAMA Psychiatry. 2025;82(9):877-887. doi:10.1001/jamapsychiatry.2025.1317

[24] Compass Pathways. Virtual KOL Event: Commercial Preparations for COMP360 Psilocybin for Treatment-Resistant Depression and Clinical Trial Plans for Post-Traumatic Stress Disorder. January 7, 2026. Webinar and slide deck. Includes fireside chat with Greenbrook Mental Wellness Centers, Journey Clinical, and HealthPort.

[25] Spravato reimbursement and margin data synthesized from: Finnastra, “Spravato Billing & Operations Guide” (2025), reporting reimbursement of $1,300-$1,600/session, drug cost ~$875/dose (84mg), net margin $300-$600/session; CMS G2082 national average payment rate ~$952.59; myfcbilling.com (2026), reporting major commercial payer G2082 average ~$857. Drug acquisition cost range of $590-$885 per dose consistent across multiple publicly available clinic billing resources.

[26] Ketamine Academy, “Why Your IV Ketamine Practice Needs Spravato” (June 2025), estimating ~$16,900 in gross profit per patient completing the full first-year Spravato protocol (34 sessions). Annual maintenance figures vary by session frequency and payer mix.

Estimated reading time: 11 minutes

I. The Mechanism That Keeps Failing to Arrive

I have never prescribed brexanolone. I have prescribed zuranolone twice, both times with the particular caution of a clinician working from a label rather than from experience. The allopregnanolone mechanism — GABA-A positive allosteric modulation by an endogenous neurosteroid — has been validated in controlled settings repeatedly over the past decade. It has been clinically usable for my patients approximately zero times.

The mechanism works. Allopregnanolone’s role in the neurobiology of postpartum depression is about as well-established as anything in psychiatric pharmacology gets. The problem has never been the molecule. The problem has been everything between the molecule and the patient.

Brexanolone was the first attempt. The FDA approved it in March 2019 for postpartum depression — the first drug ever approved specifically for that indication. The delivery requirements were extraordinary: a 60-hour continuous intravenous infusion, administered only in certified healthcare facilities, under a restricted REMS program requiring continuous pulse oximetry and sedation checks every two hours. Patients had to be accompanied during any interaction with their child. A total of 140 patients received the drug across the entire clinical trial program.

For most psychiatrists, brexanolone existed as a journal article. You could read about it. You could not use it. Sage Therapeutics eventually stopped marketing it and requested the FDA withdraw approval — not for safety or efficacy reasons, but because the drug was no longer being sold. The FDA completed the withdrawal in April 2025.

Zuranolone was supposed to fix this. Approved in August 2023 as the first oral medication for postpartum depression — a pill, fourteen days of treatment, no infusion center — the field exhaled. The relief was premature. Zuranolone is a synthetic neuroactive steroid, not allopregnanolone itself. The efficacy data for major depressive disorder were modest: a 1.8-point separation from placebo on the HAM-D-17 in the WATERFALL trial, and a failed primary endpoint in the MOUNTAIN trial. The FDA issued a Complete Response Letter for MDD on the same day it approved the PPD label. And clinically, many postpartum women decline it for reasons that don’t appear in the trial data — concerns about sedation interfering with nighttime infant care, or uncertainty about breastfeeding safety. The ambition of an oral neurosteroid for the twenty million Americans with major depression remained unfulfilled.

Now there is a third attempt. Seaport Therapeutics’ GlyphAllo is an oral prodrug of allopregnanolone itself, attached to a lipid scaffold designed to reroute absorption through the lymphatic system rather than the portal vein. A Phase 2b trial in major depressive disorder began dosing in July 2025.

Three companies. Three delivery strategies. The same mechanism each time. The pharmacology has been ready for years. What keeps failing is the route.

II. The Hidden Step

The most common reason my patients stop a medication is not that it doesn’t work. It’s that it makes them nauseous.

This mundane observation shapes more treatment outcomes than most receptor pharmacology does. A patient who discontinues sertraline at week three because of GI distress never reaches the week six timepoint where the SSRI might have separated from placebo. The side effect that drives the most treatment failure in outpatient psychiatry is the gut telling the patient that something is wrong before the brain has a chance to notice that something is right.

Much of this traces back to first-pass metabolism. When you swallow a pill, the drug crosses the intestinal wall and enters the portal vein — the vessel that carries everything absorbed from the gut directly to the liver before it reaches the body. The liver metabolizes a fraction of the drug on this initial pass, sometimes a large fraction. What survives is what the brain eventually sees. A drug that loses 90 percent of its dose to hepatic metabolism needs to be dosed ten times higher, which means more drug sitting in the gut, more metabolic activity in the liver, and more of the tolerability problems that drive patients out of treatment.

But there is an alternative route. Long-chain dietary fats are packaged by intestinal cells into large particles called chylomicrons, which are too big to enter the portal blood capillaries. Instead, they drain into the lymphatic system and reach systemic circulation without ever passing through the liver. If you could design a drug that mimics a dietary fat closely enough to enter this pathway, you could bypass first-pass metabolism entirely. That is the pharmacokinetic idea behind Seaport’s Glyph platform — and the context that makes the next two examples intelligible.

III. Two Drugs, Two Failure Modes

Allopregnanolone: Killed by Administration

Allopregnanolone has near-zero oral bioavailability — the Zulresso label puts it below five percent. The only way to achieve therapeutic exposure was a 60-hour IV infusion. From that single pharmacokinetic fact, every subsequent constraint follows: inpatient setting, continuous monitoring, REMS certification, supervised mother-child contact. Each requirement is individually reasonable. No single one is excessive. But the aggregate effect is a drug that almost no one could access. The number of facilities willing to certify for a 60-hour postpartum depression infusion was always going to be small. The number of women who could arrange 2.5 days away from a newborn — at a certified facility that might not be in their city — was always going to be smaller still. The molecule worked. The system around the molecule did not.

Agomelatine: Killed by the Liver

Every psychiatrist encounters agomelatine in the literature eventually. A melatonin MT1/MT2 agonist and 5-HT2C antagonist — a genuinely novel dual mechanism unlike anything available in the United States. In the Cipriani et al. 2018 Lancet meta-analysis — 522 trials, 116,477 participants, 21 antidepressants — only two drugs had fewer dropouts than placebo: agomelatine and fluoxetine. Patients randomized to agomelatine were less likely to quit for any reason than patients randomized to sugar pills.

This is a drug that American psychiatrists cannot prescribe.

No New Drug Application was ever submitted. Novartis ran U.S. Phase III trials and in October 2011 discontinued development, citing disappointing results and hepatic concerns. The hepatic problem is a first-pass metabolism problem. Agomelatine is metabolized predominantly by CYP1A2 — roughly 90 percent of its clearance. At the doses required to overcome first-pass extraction, some patients develop transaminase elevations. European regulators mandated liver function monitoring: tests before initiation, periodic checks during treatment, hard contraindication above three times the upper limit of normal.

The monitoring is clinically manageable — not onerous by the standards of clozapine or lithium. But it introduces the particular kind of friction that accumulates quietly until it reshapes prescribing behavior. A drug requiring blood draws before and during treatment requires lab orders, follow-up appointments to review results, documentation of monitoring compliance, and a plan for what to do when a transaminase comes back elevated. Each step is minor. In aggregate, they make it slightly easier to skip in favor of an SSRI that requires no monitoring at all.

The CYP1A2 dependence compounds this. Ciprofloxacin and estrogen-containing oral contraceptives can increase agomelatine exposure severalfold. Smoking induces CYP1A2 and reduces it. The drug’s first-pass vulnerability creates problems in two directions: hepatotoxicity risk from hepatic exposure and exposure variability from enzymatic dependence.

The cascade: transaminase elevations trigger monitoring. Monitoring introduces friction. The CYP profile adds complexity. Friction and complexity reduce European uptake. Reduced uptake eliminates the commercial incentive for a U.S. filing. American patients are locked out of a drug that works, that most would tolerate well, and that occupies a distinctive pharmacological niche.

The irony: agomelatine’s tolerability profile is excellent — except for the liver. The liver problem is a first-pass problem. Route the drug around the liver, and you would preserve the tolerability advantage, potentially reduce the monitoring burden, and attenuate the CYP1A2 dependence.

So what happens when you try to route around these failures?

IV. What Glyph Claims and What It Has Shown

Seaport’s Glyph platform attaches a drug to a lipid scaffold that mimics a dietary triglyceride, redirecting absorption through the lymphatic route described in Section II. The prodrug enters the chylomicron pathway, reaches systemic circulation via the thoracic duct, and releases the parent drug — having largely bypassed the liver.

The early data suggest this works pharmacokinetically. Phase 1 data for GlyphAllo showed roughly ninefold greater allopregnanolone exposure compared to oral allopregnanolone without the modification, and pharmacodynamic markers confirmed GABA-A target engagement. In a proof-of-concept stress test study, it significantly blunted cortisol response versus placebo. For GlyphAgo, preclinical data showed 67 percent lymphatic transport versus less than one percent for agomelatine alone; in monkeys, oral agomelatine was undetectable in plasma while the prodrug at equimolar doses produced measurable levels.

These are real pharmacokinetic gains. They are not clinical proof. The Phase 2b trial in depression (BUOY-1) is enrolling roughly 360 patients with a six-week primary endpoint. The Phase 1 for the agomelatine prodrug is underway. Whether lymphatic rerouting translates into antidepressant efficacy, reduced monitoring requirements, or changed prescribing behavior remains unanswered.

V. Why Route Determines Destiny

The delivery failures described above are not isolated misfortunes. They are expressions of systemic forces — regulatory, economic, behavioral — that determine which drugs reach patients.

When a drug produces transaminase elevations, a predictable institutional cascade begins: monitoring thresholds, periodic lab mandates, boxed warnings. Each decision is defensible. The cumulative weight reshapes prescribing behavior in ways regulators do not model. And the cascade runs in one direction — there is no mechanism for removing monitoring requirements once prescriber behavior has adapted.

Oral drugs scale; IV drugs do not. In my own practice, even Spravato — a nasal spray with established reimbursement and six years on the market — strains clinic operations because of its two-hour observation requirement. A patient arrives, self-administers under observation, and must be monitored before discharge. The treatment occupies a room and a medical assistant’s attention that could accommodate two or three medication management visits. If a nasal spray still creates this kind of logistical friction, a 60-hour IV infusion never had a chance. Pharmaceutical development preferentially pursues oral formulations because oral is commercially viable at scale. The drugs that reach the market are a filtered sample — filtered by what can be swallowed.

And clinicians avoid complexity. A psychiatrist managing 400 patients makes prescribing decisions under time constraints that favor simplicity. A drug requiring baseline labs needs only to add one task to a compressed encounter to tip behavior toward the easier option. Patients vote with their adherence; GI side effects are the behavioral expression of a pharmacokinetic problem.

These forces are invisible in treatment algorithms and as consequential as anything that happens at the synapse.

VI. The Other Reading

Everything to this point builds a sympathetic case for why Glyph’s approach is mechanistically serious. That case is incomplete.

Glyph prodrugs are new chemical entities with composition-of-matter patent claims — the strongest form of pharmaceutical IP. The agomelatine patent is titled “Lipid prodrugs of agomelatine.” These are not method-of-use patents on an old drug. They are new molecular entities built to capture known clinical signals inside fresh patent protection.

Readers of this blog will recall LB Pharma, which methylated a 40-year-old European antipsychotic and presented it as next-generation. The parallel is real but imprecise. LB Pharma’s methylation produced a pharmacokinetic change difficult to distinguish from noise. Glyph’s glyceride scaffold produces ninefold exposure increases and 67 percent lymphatic transport. LB Pharma’s methyl group was a patent strategy with a pharmacokinetic rationale. Glyph is a pharmacokinetic innovation with a patent consequence. The arrow of causation points in different directions, even if the commercial destination looks similar from a distance.

But genuine innovation does not exempt the platform from scrutiny.

The hepatotoxicity gap. Clinical liver function data in humans for GlyphAgo do not yet exist. Rodent and human lymphatic physiology are not identical. Idiosyncratic liver injury can occur at exposures that look manageable on a PK curve. The monitoring elimination claim is a hypothesis, not a result.

CYP interactions: attenuated, not eliminated. Lymphatic rerouting bypasses first-pass CYP1A2 extraction, which is a real gain. But once agomelatine reaches systemic circulation, it is still a CYP1A2 substrate. A patient whose PCP prescribes ciprofloxacin, or a patient on estrogen-containing oral contraceptives, could still experience clinically meaningful exposure changes. Seaport has not published a human drug-drug interaction dataset. And Glyph does nothing about pharmacogenomic variability in CYP1A2 itself — the same kind of population-level metabolizer variance that led the field to prefer desvenlafaxine over venlafaxine in CYP2D6-dependent patients. Rerouting around the liver is not the same as eliminating metabolic vulnerability.

Food effects. Chylomicron formation depends on dietary fat. Quantitative fed-versus-fasted data have not been disclosed. If GlyphAllo requires a fat-containing meal, that is a compliance variable. The relevant comparison is not GlyphAllo versus brexanolone. It is GlyphAllo versus sertraline, which can be taken on an empty stomach with a glass of water.

The convenience test. Even if GlyphAllo works pharmacokinetically, does it change prescribing behavior in a market with thirty generic antidepressants? Phase 1 somnolence rates of 29 percent will require clinical management — evening dosing, patient counseling, dose adjustment. For GlyphAgo, agomelatine’s mechanism is novel, but its clinical profile amounts to a slightly better-tolerated antidepressant competing against cheap generics. That is a meaningful differentiator for the right patient — but it is not an unmet-need monopoly.

Glyph is, structurally, a strategy for converting European-approved or academically validated drugs into U.S.-patented, U.S.-priced assets. The pharmacokinetic modification is genuine. The commercial architecture is what makes the enterprise financially viable. These are not contradictory observations.

The interesting thing about Glyph is that both readings are simultaneously true.

VII. The Space Between the Molecule and the Patient

The allopregnanolone mechanism has been ready for over a decade. Three companies have spent the better part of a decade trying to deliver it. The first built an infusion protocol so restrictive that the drug was inaccessible. The second built a synthetic workaround that traded mechanistic fidelity for oral convenience and ended up with a narrower indication. The third is attempting to deliver the original molecule by a route the body already uses for something else.

Whether Seaport succeeds depends on data that do not yet exist. I have tried, in this essay, to take their claims seriously without taking them on faith.

But the questions this essay raises precede Glyph.

How many mechanisms disappeared not because the biology was wrong, but because the absorption was inefficient, the formulation required monitoring, or the first-pass metabolism profile made the commercial case unattractive? These drugs do not appear in meta-analyses. They appear as abandoned development programs, as withdrawn approvals, as drugs that exist in European formularies but not American ones. They are absent from the evidence base, and their absence is itself invisible.

The space between a molecule and a patient — the pharmacokinetic, regulatory, economic, and behavioral terrain a drug must cross before it becomes a treatment — is more decisive than most of psychiatry’s conversations acknowledge. It determines which drugs we can prescribe, which patients can access them, and which mechanisms get a fair hearing in clinical practice. I started this essay with a clinical observation: I have never prescribed brexanolone, and I have prescribed zuranolone twice. Those facts reflect my patients’ reality more honestly than any mechanism-of-action diagram. The pharmacology was ready. The route was not. Whether the route is ready now is a question that deserves better than hype or dismissal. It deserves the kind of careful, patient attention that the space between the molecule and the patient has always demanded and rarely received.

Dark Matter Pragmatism explores the systems shaping modern medicine from the clinic outward. If you found this useful, consider subscribing.

Subscribe now

Disclosure

The author has no financial relationships with Seaport Therapeutics, Sage Therapeutics, Biogen, or any other company discussed in this essay. No advisory fees, consulting arrangements, or equity positions influenced this analysis.

Works Cited

Cipriani A, Furukawa TA, Salanti G, et al. Comparative efficacy and acceptability of 21 antidepressant drugs for the acute treatment of adults with major depressive disorder: a systematic review and network meta-analysis. Lancet. 2018;391(10128):1357-1366. doi:10.1016/S0140-6736(17)32802-7

FDA. Sage Therapeutics, Inc.; Withdrawal of approval of a new drug application for ZULRESSO (brexanolone) solution. Federal Register. 2025;90(49):12162-12163. Docket No. FDA-2024-N-5852

FDA. Zulresso (brexanolone) prescribing information. 2019.

Sage Therapeutics and Biogen. FDA approves ZURZUVAE (zuranolone) for postpartum depression and issues Complete Response Letter for major depressive disorder. Press release, August 4, 2023.

Seaport Therapeutics. First patient dosed in Phase 2b BUOY-1 study of GlyphAllo (SPT-300) in major depressive disorder. Press release, July 17, 2025.

Lipid prodrugs of agomelatine. International Patent Application WO2025137181A1. Published June 26, 2025.

European Medicines Agency. Valdoxan (agomelatine) product information. Summary of product characteristics.

Adam D. Borecky, MD, is a board-certified psychiatrist in outpatient private practice splitting time between Southern California and Western Michigan. Dark Matter Pragmatism examines the invisible forces that shape how medicine works.

They may also increase the risk of sudden death.

Both findings are technically true. The explanation has less to do with pharmacology than with the ecosystem those drugs move through.

I. The Gatekeeper’s Confession

I prescribe stimulant medications for ADHD almost every day.

I will admit something that would probably make a quality improvement committee uncomfortable.

The way I conduct informed consent discussions about stimulant medications changes depending on who is sitting across from me.

The facts stay the same. The emphasis does not.

When the clinical picture is straightforward and I am confident a patient has ADHD, the conversation stays proportional. We talk about blood pressure monitoring, sleep, appetite, rare cardiac events, and what the literature actually shows. The tone is calm because the risk profile is well understood and the benefits can be substantial.

When the request arrives pre-assembled from TikTok, or the patient has already decided which medication they need, the conversation subtly shifts. I do not change the facts. I highlight different ones. The cardiovascular risk gets a longer paragraph. The addiction potential gets spelled out more explicitly. The monitoring requirements suddenly sound a little less optional.

Every clinician who prescribes controlled substances does some version of this calibration. Some call it clinical judgment. Others quietly acknowledge it is partly instinct. Very few of us examine what is actually driving it.

That omission matters because stimulant medications sit directly on one of psychiatry’s most unstable fault lines. They represent both some of the field’s clearest therapeutic successes and some of its most visible excesses.

On one side of that line, the benefits are undeniable. The marriage that stabilized because a husband could finally follow through on what he said he would do. The third grader whose daily meltdowns stopped once his attention could stay tethered to a classroom long enough to learn. The seventy-four-year-old man with melancholic depression whose executive function improved just enough to keep him living independently.

On the other side lies a more ambiguous reality. Demand for stimulant prescriptions that exceeds any plausible estimate of ADHD prevalence. An expanding population of adults who discover their diagnosis through social media before they ever meet a clinician. A growing suspicion among many psychiatrists that some portion of adult ADHD diagnoses reflects diagnostic drift more than a classical neurodevelopmental disorder.

Most prescribers carry all of this simultaneously. The drugs clearly help many people. The ecosystem surrounding them is far less tidy.

What sent me down the rabbit hole for this essay was a pair of studies that seemed impossible to reconcile.

One suggested stimulant medications reduce mortality. The other suggested they increase the risk of sudden death.

Same drug class. Same decade. Both published in respectable journals.

The contradiction is not actually about pharmacology. It is about something more fundamental. It is about the difference between studying drugs in controlled systems and watching them move through the far messier ecology of real-world medicine.

II. The Paradox

In March 2024 a Swedish research group published a large registry study in JAMA examining 148,578 individuals with ADHD.

Their finding was striking. Initiating ADHD medication was associated with a 21 percent reduction in all-cause mortality over two years.

The reduction was driven almost entirely by fewer unnatural deaths. Accidents, poisonings, and other events that disproportionately affect people whose impulse control is chronically compromised.

A separate Swedish cohort examined patients with amphetamine use disorder. Among nearly fourteen thousand individuals, treatment with lisdexamfetamine reduced mortality by 57 percent.

In most fields of medicine those are dramatic effect sizes.

Then another paper appeared.

A 2025 analysis using the FDA Adverse Event Reporting System suggested stimulant medications might increase the risk of sudden death roughly twofold. An electronic health record analysis embedded in the same publication suggested a three- to four-fold increase in all-cause mortality after stimulant initiation.

Same molecules.

Opposite implications.

For clinicians attempting to metabolize the literature between patient visits and half-finished CME podcasts, the result is a mild form of intellectual vertigo. For parents in custody disputes, the situation is simpler. Each side cites the study that supports their argument.

The studies are not actually contradicting each other about what stimulants do to the human body.

They are describing different ecosystems.

So how can the same drug appear both dangerous and protective?

The answer turns out to have less to do with pharmacology than with the ecosystem the drug moves through.

III. The Investigation

When two datasets produce opposite answers to the same clinical question, the natural impulse is to choose the one that aligns with your existing beliefs. That approach is efficient. It is also how physicians quietly drift into practicing medicine by vibes.

The more productive instinct is to treat the contradiction itself as diagnostic.

Start with the alarming signal.

The FDA Adverse Event Reporting System, or FAERS, has a respectable origin story. It grew out of the drug safety reforms that followed the thalidomide disaster in the early 1960s. The logic behind the system is straightforward. Clinical trials are small, carefully controlled, and relatively short. They frequently miss rare events that only appear once millions of people begin using a drug in the real world.

FAERS was designed to answer one question.

Is something happening more often than we expected?

It was never designed to answer the follow-up question.

Is the drug causing it?

The difference matters more than most people realize.

FAERS collects reports of adverse events from manufacturers, clinicians, and patients. Roughly ninety percent come from drug companies, which are legally required to submit serious safety reports within fifteen days. The remainder arrive through voluntary reporting systems like MedWatch.

The system captures the numerator. It knows how many adverse events were reported.

It does not know the denominator.

Ten reported deaths could represent ten deaths among ten million users. That is statistical noise. The same ten deaths among one hundred users would represent a catastrophe.

Without knowing how many people actually took the drug, you cannot calculate risk.

Layer on top of that several well-known reporting biases. Dramatic events get reported more often than mundane ones. Drugs involved in litigation generate more reports than quiet medications. Media coverage produces bursts of what epidemiologists politely call stimulated reporting. A 2022 cross-sectional study in the BMJ found that only about 30 percent of FAERS-generated safety signals are subsequently corroborated by published research.

The result is a system exquisitely sensitive to attention.

But dismissing FAERS entirely would be a mistake. The database is capturing something real. Not a clean causal signal, but a messy snapshot of the prescribing ecosystem.

That ecosystem includes misdiagnosis, stimulant diversion, rushed medication follow-ups, and prescriptions written for patients who never receive meaningful monitoring. In other words, it samples the drug as it exists in the wild.

Now contrast that with the Swedish studies.

Those analyses draw from national registries inside a healthcare system with universal coverage, consistent follow-up, and unusually complete data capture. Patients in those datasets tend to be properly diagnosed, regularly monitored, and treated within a fairly coherent clinical framework.

Both contain real plants. The environment surrounding them is radically different.

The Swedish studies also employed a particularly powerful analytic design known as a within-individual comparison. Instead of comparing people on stimulants to different people not on stimulants, the researchers asked a simpler question.

Does the same person experience fewer deaths during periods when they are taking their medication compared to periods when they are not?

Each individual effectively serves as their own control. Genetics, socioeconomic status, family background, and baseline severity all remain constant. The only variable changing over time is medication exposure.

It is not perfect. But it is one of the stronger designs available in observational epidemiology.

At that point the contradiction becomes easier to interpret.

The Swedish registries tell you what happens when stimulants are prescribed appropriately to people who actually have ADHD.

FAERS tells you what happens when the same drug class circulates through a much looser ecosystem.

Both signals are real. Neither is the whole story.

IV. The Population Reveal

The populations behind these datasets are fundamentally different.

The Swedish registries track individuals with confirmed ADHD diagnoses receiving treatment inside a structured healthcare system.

FAERS captures adverse events from anyone taking a stimulant for any reason. That includes diagnosed patients, undiagnosed users, diverted prescriptions, telehealth misdiagnoses, and people with unrecognized cardiovascular disease.

These groups are not interchangeable.

The mechanism by which stimulants reduce mortality in ADHD is also not mysterious.

Untreated ADHD is, at its core, a disorder of executive function. The most visible symptoms involve distractibility and poor attention. The most dangerous symptoms involve impulsivity.

Impulsivity accumulates risk over time. Motor vehicle accidents. Falls. Substance misuse. Risky decisions that compound across decades.

The mortality reduction observed in the Swedish study was driven largely by fewer unnatural deaths. That is exactly what you would expect if treatment improves impulse control.

Once that logic becomes clear, the paradox resolves itself.

If the drug’s benefit depends on correcting a specific neurocognitive deficit, prescribing the same drug to someone without that deficit removes the protective mechanism while leaving the pharmacology intact.

The drug did not change.

The patient population did.

The lisdexamfetamine data in amphetamine use disorder illustrates the point almost too neatly. In that cohort, individuals who had been harming themselves with uncontrolled stimulant use experienced dramatically lower mortality once the same pharmacological class was prescribed in a structured and monitored form.

Same molecule. Different system of control.

V. The Real Safety Signal

None of this means stimulant risks are imaginary.

It means they are specific.

Cardiovascular effects are the most discussed. They are also the most frequently misunderstood.

In children and adolescents, stimulants reliably produce small hemodynamic changes. Heart rate increases by roughly one to two beats per minute. Systolic blood pressure rises by about one to four millimeters of mercury. Large epidemiological studies looking for serious cardiovascular events in pediatric populations have consistently failed to find meaningful increases.

In working-age adults the signal becomes more nuanced.

A Swedish cohort of roughly 278,000 individuals with ADHD found that each year of stimulant exposure during the first several years of treatment was associated with an eight percent increase in cardiovascular disease risk. The effect appeared to plateau after several years.

These are modest changes. They matter more for patients who already have hypertension, diabetes, or other cardiovascular risk factors. For otherwise healthy adults they often remain clinically manageable.

The picture changes again in older adults.

A Canadian study of more than six thousand patients over the age of sixty-six found a forty percent increase in cardiovascular events within the first thirty days of stimulant initiation. The signal was driven primarily by arrhythmias and strokes. By six months the excess risk was no longer statistically significant.

The pattern is consistent.

Cardiovascular risk is real. It is age-dependent. And it is concentrated during the initiation window.

That is a monitoring problem more than an absolute prescribing prohibition. Start low. Titrate slowly. Check vitals regularly during the first month.

The more uncomfortable question is whether most prescribing environments actually provide that level of monitoring.

VI. The Consulting Room

All of this data eventually condenses into a very simple question.

What do you actually say to the patient sitting across from you?

The conversation changes depending on who that patient is.

The first version happens with a mother whose nine-year-old son was just diagnosed with ADHD. His older sibling had no academic difficulties, so the teacher’s phone call arrived as a shock. By the time she reaches my office she has already Googled phrases like “Adderall cardiac arrest children.”

What she needs to hear is that large population studies looking specifically for sudden cardiac death in children taking stimulants have not found evidence of increased risk. We will monitor his blood pressure and ask about family cardiac history. But the far greater danger may be allowing a capable child to spend several critical developmental years failing academically because his attention cannot remain anchored long enough to learn.

Executive function matures during a specific developmental window. Missing that window carries consequences that will never appear on an electrocardiogram.

The second conversation happens with a forty-two-year-old software engineer who began suspecting he might have ADHD after his son was diagnosed and responded well to treatment.

His question is direct.

How could a medication that raises heart rate possibly reduce the risk of death?

The answer is that heart attacks are not the main threat in his case. The bigger risk is the accumulation of impulsive decisions over decades. Reckless driving, substance misuse, professional instability. The mortality reduction observed in the Swedish study makes sense once you frame ADHD as a disorder that quietly amplifies everyday risk.

The stimulant does not fix the heart.

It improves the judgment that keeps the heart out of danger.

The third conversation is more complicated.

A seventy-one-year-old man with severe treatment-resistant depression has failed multiple antidepressants. His ability to manage daily tasks is collapsing. You are considering low-dose methylphenidate as an augmentation strategy.

Now the cardiovascular data becomes more relevant. Older adults do face higher initial risk when stimulants are started. The honest answer becomes conditional.

I think this might help you. But the first month requires close monitoring.

The alternative, watching untreated psychiatric illness slowly dismantle the remainder of his independence, also carries mortality risk. It is simply a form of mortality that does not trigger adverse event reports.

VII. The Cost of Misreading Data

There is a pipeline beneath almost every medication controversy in modern medicine.

A methodologically limited dataset produces a signal. The signal gets published because journals are appropriately cautious about drug safety. The publication generates headlines. The headlines reach patients, parents, legislators, and attorneys.

Somewhere downstream, someone who might have benefited from a treatment decides not to take it.

Not because the evidence said they should not.

Because the evidence was never translated honestly enough for them to understand what it meant.

Psychiatry has already seen this phenomenon.

After the FDA issued a black box warning in 2004 about possible suicidality risks with antidepressants in adolescents, prescribing rates for teenagers fell sharply. Over the following two years youth suicide rates rose for the first time in more than a decade.

The warning was well-intentioned.

The downstream consequences were real.

Patients today have unprecedented access to medical information. Soon they will have AI systems capable of synthesizing that information faster than any physician can read it. Yet the central question patients bring to the consulting room has not changed.

They want to know what the numbers look like when someone with experience applies them to a real human life.

“Have you seen this help someone like me?” “What would you do if this were your kid?”

Those questions are not requests for more data. They are requests for judgment.

Clinical judgment emerges when someone has spent enough time sitting across from patients to understand how statistical risks manifest in actual lives.

Between those two levels lies the space where medicine actually happens.

The precautionary principle exists to protect people from harm. Applied without careful interpretation, it can redirect harm toward patients whose suffering unfolds slowly enough to escape formal surveillance.

A child who fails out of school. A marriage that erodes under untreated executive dysfunction. An elderly man who quietly withdraws from the world.

None of those events appear in pharmacovigilance databases.

They simply occur on ordinary Tuesdays, in ordinary clinics, where no one is keeping score.

Dark Matter Pragmatism is a Substack about psychiatry, biotech, and the strange systems surrounding modern medicine.

If this piece was interesting to you, consider subscribing.

Subscribe now

I. Two Audiences, One Dataset

On February 17th, the COMP360 Phase 3 topline arrived. Within minutes, two very different audiences started reacting to the same dataset. In one corner of the internet, biotech investors parsed the press release, modeling peak sales curves and debating whether a ~3.7-point MADRS delta justified the company’s valuation. In the other corner, a patient I’ve been treating with Spravato (esketamine) for over a year messaged me a link and asked a much simpler question: “Should I switch?”

The message arrived between clinic sessions, sandwiched between a TMS follow-up and a ketamine intake. It wasn’t a theoretical question. It was a patient trying to decide whether the next phase of his life should involve a treatment he already knows works, or a six-hour monitored psychedelic session that, for now, only exists inside clinical trials.

The COMP360 trials are not just a biotech milestone. They may represent a new tool in the treatment-resistant depression toolkit I use every day. The task here isn’t to celebrate the press release or dismiss it. The task is to read the data the way a working psychiatrist has to read it: asking whether this is something I should actually prepare to offer in my clinic, or something I should keep watching from the sidelines for another cycle of evidence.

So before we get into the harder questions about durability, infrastructure, and where this might fit in real-world treatment, it’s worth starting with the most basic one: did the trials actually clear the bar?

II. What Replication Means - And What It Doesn’t

The first thing worth acknowledging about the COMP360 program is that it accomplished what psychedelic medicine has struggled to reach for decades: replication.

Compass ran two Phase 3 trials, COMP005 and COMP006, using slightly different designs that produced remarkably similar results. COMP005 showed a −3.6 point difference in MADRS change from baseline at Week 6. COMP006 showed −3.8. Both met their pre-specified primary endpoint with p-values below 0.001. The COMP005 primary endpoint (MADRS separation at Week 6) figure makes the result easier to see.

The figure below shows how depressive symptoms changed over time after dosing:

Figure: COMP005 primary endpoint showing −3.6 MADRS separation between COMP360 25 mg and placebo at Week 6.

The blue line represents the 25 mg psilocybin arm, while the gray line represents placebo.

For context, the pivotal TRANSFORM-2 trial for Spravato reported roughly a −4.0 MADRS difference at Day 28. That doesn’t mean the treatments are equivalent, as the trials were designed very differently. But it does provide a useful anchor for interpreting the magnitude of the COMP360 result.

On the narrow statistical question of whether psilocybin produces a measurable antidepressant effect in treatment-resistant depression, the answer is clearly yes.

That may sound modest. In the current psychedelic research climate, it is a meaningful milestone.

For the better part of a decade, the field has lived in an uncomfortable tension between extraordinary enthusiasm and relatively thin registration-quality evidence. The collapse of the MDMA program at Lykos reminded everyone how unforgiving the regulatory bar can be when enthusiasm outruns methodological rigor.

Against that backdrop, two positive Phase 3 trials with nearly identical effect sizes matter.

There is also a slightly awkward but important comparison worth making here. On the narrow question of pivotal trial consistency, Compass’s dataset is arguably cleaner than what got Spravato approved.

Spravato’s approval rested on three pivotal short-term trials: TRANSFORM-1, TRANSFORM-2, and TRANSFORM-3. Of those, only TRANSFORM-2 achieved statistical significance on the primary endpoint. TRANSFORM-1, the fixed-dose study, missed. TRANSFORM-3, which focused on an elderly population, also missed. The FDA ultimately approved the drug based on the totality of evidence, including the relapse-prevention results from the SUSTAIN-1 trial.

Compass now has two Phase 3 trials hitting their primary endpoint with almost identical deltas - a stronger starting position.

But replication alone does not answer the question my patient texted me.

The real question is whether the magnitude, durability, and practicality of COMP360 meaningfully change what I can offer patients who already have access to Spravato, TMS, and conventional pharmacotherapy.

That’s where the rest of the dataset becomes more interesting.

III. The Dose-Response Signal and Blinding — Why It Matters More Than the P-Value

The most analytically interesting feature of the COMP360 dataset is not the p-value. It’s the dose-response curve.

This matters because psychedelic trials have a structural methodological problem that everyone in the field understands but rarely says out loud: true blinding is almost impossible.

If a patient receives a 25 mg dose of psilocybin, they will almost certainly know it within the first hour. The perceptual effects are unmistakable. In the COMP006 trial, participants in the 25 mg arm reported visual disturbances, emotional lability, nausea, and other classic psychedelic effects. Treatment-emergent adverse events occurred in roughly 88% of the 25 mg group compared to about 56% in the 1 mg control arm. Even without prior psychedelic experience, it doesn’t take much sophistication to notice that you’re hallucinating.

Compass did attempt to mitigate this problem. Instead of a completely inert placebo, COMP006 used a 1 mg psilocybin dose as an “active control.” That choice helps somewhat. A low dose produces mild physiological sensations that can obscure group assignment. But realistically, the difference between 1 mg and 25 mg is not subtle.

So if expectancy effects were doing most of the work in this dataset, you would expect a particular pattern in the results. The high-dose and medium-dose arms should perform roughly the same. Both groups know they received an active psychedelic dose. The real separation would occur between the “active” arms and the minimal-dose control.

That is not what the data show.

The COMP006 trial provides the clearest visual example of this pattern.

The figure below shows how depressive symptoms changed over time across three dose arms: 25 mg, 10 mg, and a 1 mg active control.

Figure: COMP006 dose-response curve showing graded separation between 25 mg, 10 mg, and 1 mg arms through Week 6.

The key feature of this graph is the orderly separation between the three dose arms.

If expectancy effects were doing most of the work, we would expect the two clearly psychoactive doses — 25 mg and 10 mg — to perform similarly. Both produce unmistakable subjective effects.

Instead, the results follow a graded dose-response pattern that looks much more like pharmacology than suggestion. In drug development, that kind of orderly dose separation is reassuring. It suggests that something biological is happening underneath the subjective experience.

None of this eliminates the blinding problem entirely. Psychedelic trials will always carry a higher risk of functional unblinding than conventional pharmacology studies. Compass collected formal blinding integrity assessments in both trials, but those data won’t be available until after the studies fully conclude. When they are released, they’ll be worth examining closely.

That doesn’t mean the dataset is beyond criticism. But it does mean the most common critique of psychedelic research — that patients simply know they received the drug and therefore report improvement — has to grapple with a more complicated pattern than a simple active-versus-placebo effect.

IV. The Durability Question — Honest Edition

If replication establishes that COMP360 has an antidepressant effect, the next question is the one doing most of the commercial and clinical heavy lifting: durability.

This is also where the dataset becomes more complicated.

Much of the enthusiasm around psilocybin therapy rests on a simple narrative. A patient undergoes a single psychedelic session and experiences sustained relief from depression that lasts months. Compared to treatments that require continuous dosing, the appeal is obvious.

But when you look closely at the COMP360 Phase 3 program, the durability story is more nuanced.

A. The Threshold Problem

One of the most widely cited figures from the COMP005 trial is that roughly 25% of participants maintained a clinically meaningful improvement through 26 weeks.

The definition of “clinically meaningful” here is important.

Compass used a ≥25% reduction in MADRS score as the threshold for sustained benefit. That is a relatively permissive definition compared to the conventions used in most antidepressant trials. In treatment-resistant depression studies, a ≥50% reduction in MADRS is typically considered a response, and remission thresholds are even stricter.

To put that in clinical terms: a patient whose MADRS score falls from 32 to 24 would meet the ≥25% threshold while still remaining moderately depressed.

That doesn’t make the finding meaningless. Partial improvement still matters. But the headline number needs context.

The reason for the lower threshold appears to be methodological rather than promotional. The original analysis plan separated full responders (≥50%) from partial responders (25–49%). However, once the data were collected, the number of full responders was apparently too small to generate stable long-term curves on its own. The two groups were therefore combined.

That choice is defensible statistically. But the fact that it was necessary is itself informative. It tells us that the proportion of patients achieving conventional antidepressant response was small enough that those curves could not stand on their own — which is itself a data point about the treatment’s potency in this population.

At the Week 6 endpoint, using Compass’s own stricter remission definition — MADRS ≤12 with no individual item ≥4 — the number of remitters in the 25 mg arm was small: 11 out of 171 patients, roughly 6%.

More conventional response and remission rates will eventually be reported once the blinded period of COMP006 fully concludes. When those numbers arrive, they will be the most useful metrics for comparing COMP360 directly with existing treatments like Spravato.

B. The Retreatment Confound

There is another complication embedded in the durability curves.

The durability curve presented by Compass looks compelling at first glance.

The chart below shows MADRS score changes through Week 26 in the COMP005 trial.

Figure: COMP005 durability curve showing separation between COMP360 25 mg and placebo through Week 26.

At first glance, the figure suggests sustained improvement in the psilocybin arm compared with placebo through six months of follow-up.

But one detail in the slide deserves careful attention.

In the COMP005 study, about 70% of participants in the 25 mg arm received a second blinded dose during the follow-up period. Most of these retreatments occurred between Weeks 10 and 14, triggered by pre-specified criteria indicating that depressive symptoms were returning.

This means that the widely discussed 26-week durability curve is not purely the trajectory after a single administration. For most patients, it reflects the outcome of two psychedelic sessions spaced roughly three months apart.

That is still an attractive treatment model. Two or three monitored sessions per year would represent a dramatically lower treatment burden than most existing interventions for treatment-resistant depression.

C. The Antidepressant Confound

The final complexity is pharmacologic.

During the later phases of the COMP005 study, participants were permitted to initiate or adjust conventional antidepressant medications if their symptoms warranted it. The specific rates and timing of those medication changes have not yet been publicly detailed.

This introduces another layer of uncertainty when interpreting long-term outcomes. If a participant began an SSRI or another antidepressant during the follow-up period, any subsequent improvement becomes difficult to attribute exclusively to psilocybin.

For that reason, Compass has not performed formal inferential statistics on outcomes beyond Week 6. Later timepoints are presented descriptively rather than as hypothesis-testing comparisons.

That approach is methodologically appropriate. But it also means that the durability story currently rests more on visual trend data than on statistically confirmed endpoints.

D. My Read

None of this invalidates the durability signal.

The sustained separation between treatment arms at later timepoints would be difficult to generate purely by chance. Something real is likely happening.

The headline figure of '25% durable responders' should be read as a descriptive estimate using a lower-than-standard threshold, within a follow-up period that includes retreatment and potential medication changes — not a pre-specified finding.

The next wave of data from COMP006 Part B, which uses a planned two-dose protocol and a larger sample size, will be the most informative dataset for understanding how durable this treatment really is.

Until then, the honest answer to the durability question is not “six months after one session” but “meaningful improvement for a subset of patients, likely maintained with occasional retreatment.”

V. The Missing Slide — Functional Outcomes

There is one slide missing from the topline presentation that I was looking for in the investor presentation: functional outcomes.

Both Phase 3 trials pre-registered the Sheehan Disability Scale (SDS) as a secondary endpoint at Week 6. The SDS measures something slightly different from the Montgomery–Åsberg Depression Rating Scale (MADRS) that served as the primary endpoint.

MADRS asks whether depressive symptoms have improved.

The SDS asks a more practical question: is the patient’s life actually working again?

That distinction matters more than it might initially appear.

The MADRS is an excellent instrument for detecting treatment-induced symptom change, which is exactly why it has become the standard primary endpoint in antidepressant trials. But it was never designed to measure whether patients regain functional capacity.

The scale contains ten symptom items: sadness, pessimism, sleep disturbance, appetite change, suicidal thoughts, and several others. Only two of those items — lassitude and concentration difficulties — come close to approximating functional impairment. The rest capture emotional and vegetative symptoms of depression.

That means a patient’s MADRS score can improve substantially without their day-to-day functioning changing very much. A drop from 32 to 24 might reflect improved sleep and less sadness while leaving the patient still unable to concentrate at work or sustain normal routines.

This gap between symptom improvement and functional recovery is one of the most consistent findings in the depression treatment literature.

Across multiple studies of antidepressant treatment, roughly 40–60% of patients who meet symptom remission criteria still fail to achieve functional remission on the SDS. In other words, the two constructs are correlated — but far from interchangeable.

That’s why functional measurement has become standard practice in modern treatment-resistant depression trials. Programs evaluating therapies like Spravato routinely include the SDS alongside symptom scales, along with additional quality-of-life and productivity measures.

The COMP360 topline presentation focused entirely on symptom reduction, which is appropriate for a primary endpoint readout. But the absence of SDS results is still notable. If functional outcomes had shown dramatic separation between treatment arms, they would have been a natural addition to the slide deck.

That doesn’t necessarily mean the results were negative. It may simply reflect the normal sequence of clinical trial reporting, where secondary endpoints are disclosed later in conference presentations or peer-reviewed publications.

From the payer’s perspective (I can confirm from hundreds of “peer-to-peer” calls with insurers) this distinction is critical.

A therapy that produces a four-point shift on a symptom scale but leaves patients unable to return to work is a much harder economic proposition than one that restores daily functioning. When insurers evaluate coverage for a treatment that requires six hours of monitored administration and dedicated clinical infrastructure, the functional outcome data may ultimately carry more weight than the symptom scale itself.

If COMP360 truly produces durable transformation in patients’ daily functioning, the SDS should show it.

VI. A Quick Reality Check: What TRD Actually Looks Like

Before comparing COMP360 with the treatments I already use, it’s worth acknowledging a small mismatch between clinical trials and clinical reality.

Most treatment-resistant depression trials define TRD as failure of two adequate antidepressant trials in the current episode. That definition works well for regulatory purposes. But in specialty practice it’s closer to a lower bound than a typical patient profile.

By the time someone reaches an interventional psychiatry clinic, their treatment history is usually much longer.

Many patients have already tried four to six antidepressant strategies. Polypharmacy is common. It’s not unusual to see combinations like an SSRI or SNRI paired with a different class of antidepressant, an atypical antipsychotic, or a supervised summoning of the “old gods” of psychopharmacology (Nardil, Emsam, Parnate, etc).

Only after that pharmacologic ladder starts to stall do most clinicians begin discussing interventional options such as Transcranial Magnetic Stimulation, Spravato, or IV ketamine.

So when we ask where COMP360 might fit, the real comparison isn’t “third antidepressant versus psilocybin.”

The real decision point looks more like this:

Do we keep adding medications… or do we move to a different treatment modality entirely?

That is the clinical fork in the road where COMP360 will actually live.

VII. Leveling the Playing Field — COMP360 vs. What I Already Have

Once a patient reaches that fork in the road — the point where medication strategies start to stall and interventional treatments enter the conversation — the next practical question is simple: how does a new therapy compare with the tools we already have?

Compass anticipated this comparison and addressed it directly in their presentation, including a slide positioning COMP360 alongside Spravato, the current market leader among rapid-acting antidepressants.

Their comparison is summarized in the following slide from the presentation.

Figure: Compass Pathways slide comparing COMP360 with Spravato and traditional antidepressants.

At a high level, this table captures the company’s core positioning: similar efficacy with potentially greater durability and dramatically fewer treatment sessions.

That framing is understandable. But several subtleties emerge when you read the comparison from the perspective of someone delivering these treatments in clinic.

Effect Size

Let’s start with the basic efficacy numbers.

In the COMP360 Phase 3 trials, the placebo-adjusted difference in Montgomery–Åsberg Depression Rating Scale (MADRS) scores at Week 6 was about −3.6 in COMP005 and −3.8 in COMP006.

Those numbers are strikingly similar to the pivotal trial that led to approval of Spravato. In the TRANSFORM-2 trial, the Spravato arm separated from placebo by roughly −4.0 MADRS points at Day 28.

There are meaningful design differences between the programs. The Spravato trials required patients in both arms to initiate a new oral antidepressant simultaneously, which tends to inflate improvement in the control group. COMP005 used a pure placebo control, while COMP006 used a low-dose psilocybin comparator.

Those differences make strict cross-trial comparisons difficult. But at a high level, the magnitude of drug–control separation is remarkably similar across the two programs.

From a symptom standpoint alone, COMP360 does not appear dramatically stronger than the treatments we already have.

Speed of Onset

The slide also highlights the speed-of-onset advantage Compass emphasizes.

Their presentation points out that the COMP360 trials demonstrated statistically significant improvement the day after dosing.

That comparison juxtaposes an exploratory early timepoint for COMP360 with a pre-specified primary endpoint for Spravato, which makes the contrast less informative than it initially appears.

The primary endpoint in the COMP360 trials occurs at Week 6, whereas the pivotal endpoint for Spravato occurred at Day 28 in TRANSFORM-2. Early timepoints in the esketamine trials also showed rapid separation from placebo.

In practice, both treatments behave similarly in the clinic. Patients often report meaningful changes within 24–48 hours.

The important differentiator I’m watching for is what happens after that first improvement.

Durability and Treatment Burden

This is where the COMP360 story becomes genuinely interesting.

Maintenance treatment with Spravato typically involves weekly or biweekly dosing after the induction phase, which translates to roughly 20–25 clinic visits per year. The relapse-prevention study SUSTAIN-1 demonstrated that discontinuing esketamine leads to relapse in a large proportion of patients who had previously stabilized.

Real-world experience mirrors that pattern. Many patients remain dependent on regular maintenance sessions to preserve their gains.

Psilocybin therapy proposes a radically different cadence.

If the retreatment patterns observed in the COMP360 trials hold up, patients may require two or three sessions per year rather than two sessions per week during induction followed by ongoing maintenance.

Even if the antidepressant effect size is identical, that difference in treatment burden is clinically meaningful. Patients care deeply about how much of their life has to be structured around treatment visits.

From a patient’s perspective, three sessions per year versus twenty-plus is not a small distinction.

Safety

One additional piece of context worth mentioning briefly is safety.

Across both Phase 3 trials, COMP360 appeared generally well tolerated, with most treatment-emergent adverse events resolving the same day as dosing. The suicidality signal — often a concern in psychedelic trials — was rare and temporally clustered around dosing rather than emerging during follow-up, a pattern familiar to clinicians who manage transient distress during Spravato sessions. The DSMB reported no clinically meaningful imbalance in suicidality between treatment arms.

None of this eliminates the need for careful monitoring, but it suggests the safety profile is broadly consistent with what clinicians would expect from a supervised psychedelic intervention.

The Row Missing From the Table: Infrastructure

There is one comparison point that rarely appears in investor decks but dominates conversations among clinicians: the logistics of delivering the treatment.

A typical Spravato session requires two hours of monitoring under the current safety program. In a multi-room interventional clinic, that usually means cycling several patients through the same space over the course of a day.

The COMP360 protocol involves roughly six hours of monitored time per patient.

That difference has major implications for throughput. A room that might accommodate two or three Spravato sessions in a day may only support one psilocybin administration.

From the perspective of someone running a clinic, that changes the economics immediately. Staffing requirements, reimbursement structures, and physical space constraints all become part of the treatment equation.

I learned this lesson the hard way during the early rollout of Spravato in 2019 when insurers were slow to compensate the “buy-and-bill” codes JNJ rolled out. A therapy can have excellent clinical data and still fail to scale if the delivery model doesn’t align with reimbursement and clinic workflow.

Compass often references a long-term goal of thousands of treatment sites delivering COMP360. That ambition is plausible, but it depends on solving a set of logistical questions that are largely invisible in the clinical trial slides.

VIII. Where Does This Fit Into My Clinic?

At some point every clinical dataset stops being abstract and becomes practical. When I’m listening to a “study update” investor call on Quartr while driving to work, I’m not thinking about valuation models or peak sales curves. I’m mentally replaying the conversations I’ve had in my office that week and asking a much simpler question: would this actually change what I tell the next patient who walks through the door?

Right now, when I look at the COMP360 data, I can already picture at least three types of patients who might plausibly end up in that conversation.

Archetype 1: The Interventional-Naive Patient With Psychedelic-Aligned Priors

This patient shows up fairly often. They’ve usually failed several antidepressants, sometimes with one augmentation attempt already layered in. At that point their primary psychiatrist sends them to my clinic to talk about next steps.

I walk them through the menu: Transcranial Magnetic Stimulation, Spravato, occasionally IV ketamine, occasional ECT, sometimes more complicated augmentation trials.

A non-trivial percentage of patients recoil at ketamine. Sometimes it’s the cultural association. Sometimes it’s the logistics. I’ve had patients literally say, “Isn’t that the drug that killed Matthew Perry?”

Others hear 36 TMS sessions and immediately imagine spending the next two months commuting to a clinic every weekday.

But those same patients will sometimes say something very different when psychedelics come up.

Some have friends who swear by psilocybin retreats. Some tried mushrooms once in college and remember it as a meaningful experience. Others simply perceive a capsule administered in a medical setting as less invasive than a nasal spray protocol or a device attached to their head.

Some see psychedelics as more “holistic” or see it as a catalyst for deeper work in psychotherapy.

Clinical psychiatry is full of patient preferences that shape treatment engagement. If COMP360 becomes available, it would give me a credible option for a group of patients who currently decline the interventions we already have.

Archetype 2: The Spravato Maintenance Patient Looking for an Exit

The second group is the patient who is already doing well on Spravato, but struggling with the maintenance burden.

Biweekly monitored visits are not trivial. Patients have to arrange transportation, take time off work, plan around the transient dissociative effects. Over time, some of them begin asking the same question:

“Do I have to keep doing this forever?”

The relapse-prevention data from SUSTAIN-1 suggests that discontinuing esketamine leads to relapse in a substantial proportion of stabilized patients. I describe this as the “Achilles’ heel” of Spravato.

If psilocybin therapy can deliver comparable symptom stability with two or three sessions per year, the quality-of-life difference for this group would be substantial.

Archetype 3: The Partial Responder Across Multiple Modalities

The third group is the patient who lives in the gray zone between treatments who have partially or not responded to everything.

They’re on a complicated medication regimen that partially helps, produces side effects, and at this point nobody is entirely sure which drug is doing what — but everyone is too afraid to remove one in case the whole thing collapses like a pharmacologic Jenga tower.

For this group, COMP360 would not replace existing treatments. It would represent another mechanism to layer into the rotation — a different biological pathway, delivered in a fundamentally different way, that might unlock additional improvement where other approaches plateau.

That is often how progress in treatment-resistant depression happens: not with a single breakthrough, but with a widening menu of partially effective options.

IX. What I Still Need to See

If you read the COMP360 dataset the way a clinician has to read it — as a potential addition to a treatment program rather than a press release — the natural response is curiosity more than celebration or skepticism.

But there are still a few pieces of information that would meaningfully change how confident I feel about where this treatment belongs.

1. Conventional Response and Remission Rates

The single most useful numbers will be the conventional metrics used across almost every antidepressant program:

• ≥50% reduction in MADRS (response)

• MADRS ≤10–12 (remission)

Those are the numbers clinicians use when comparing treatments across trials.

Right now we have partial glimpses of these outcomes, but the full breakdowns for both Phase 3 trials are still forthcoming. Once those data are released, it will become much easier to compare COMP360 directly with treatments like Spravato or TMS.

2. The COMP006 Durability Dataset

The durability narrative currently rests heavily on COMP005, which included retreatment during the follow-up period.

The upcoming COMP006 Part B data, which follows a planned two-dose protocol in a larger population, will give us the clearest picture of what psilocybin maintenance actually looks like in practice.

Even though those results will likely remain descriptive rather than inferential, the shape of the curves will matter enormously. If the separation between treatment arms remains stable over time, confidence in the durability story will increase substantially.

3. Functional Outcomes

As discussed earlier, the results on the Sheehan Disability Scale remain absent from the topline presentation.

Those data will tell us whether improvements on the MADRS translate into improvements in the parts of life patients actually care about: work, relationships, and day-to-day functioning.

From a payer perspective, this may ultimately be one of the most important endpoints in the entire program.

4. Antidepressant Co-Treatment Rates

During the later phases of the COMP005 trial, participants were allowed to initiate or adjust conventional antidepressant medications.

Eventually we will need to see:

• how often those medication changes occurred

• when they occurred

• whether they differed across treatment arms

Without that information, interpreting long-term outcomes remains somewhat speculative.

5. Retreatment Frequency

The Phase 3 program suggests that a meaningful portion of patients opted for retreatment around the three-month mark.

What we still need to understand is the distribution.

Do most patients stabilize after two sessions per year?
Or do a substantial number require more frequent retreatment?

That question matters less for efficacy than for logistics and infrastructure — the practical realities of delivering this therapy at scale.

6. Blinding Integrity Data

Compass collected formal assessments of whether participants correctly guessed their treatment assignment.

Those results will be released once the trials are fully completed.

They won’t invalidate the findings if patients guessed correctly — that outcome is almost inevitable in psychedelic trials. But seeing the magnitude of that effect will help calibrate how much expectancy may have contributed to the observed outcomes.

X. Close — The Molecule Cleared the Bar

My patients with treatment-resistant depression have been waiting a long time for something to change.

The COMP360 trials suggest psilocybin-assisted therapy could be the next expansion, showing real signals for replication, dose-response, and promising durability.

But none of those things, by themselves, determine whether this therapy will actually reach patients. Between a successful Phase 3 trial and a functioning treatment program lies an entire ecosystem of decisions: regulatory review, REMS design, DEA scheduling, payer coverage, clinic infrastructure, training requirements, and reimbursement models.

Those factors will ultimately determine whether psilocybin therapy becomes a routine part of psychiatric care or remains a specialized intervention available in only a handful of centers.

For now, the most honest summary of the COMP360 dataset is a modest one:

Subscribe now

I. Opening: The Experiences We Don’t Ask About

Any psychiatrist who practices long enough learns a quiet truth: non-ordinary experiences are common. Patients sense presences, describe moments of unity, report religious insight, and carry vivid experiences from grief or crisis. They also learn quickly not to talk about them.

Patients fear these disclosures will be interpreted as pathology. Clinicians are trained to ask about hallucinations and delusions, not awe or transcendence. When such experiences surface, both parties often rush to resolve the ambiguity. The result is a mutual silence.

When patients do speak, they often describe these experiences as deeply important. Some are destabilizing. Others are morally organizing or consoling. Many shape a person’s life narrative for decades. The clinical problem is not their existence, but how to respond without either dismissing meaning or losing rigor.

I encounter these questions frequently. As a psychiatrist with religious commitments, I often see patients who bring such experiences not as symptoms to eliminate, but as events that demand interpretation. They want help understanding what happened without having its meaning stripped away.

This tension could once be managed quietly. It can no longer be ignored. Psychedelic-assisted therapies are now moving into mainstream psychiatry and reliably occasion experiences patients describe as sacred or transformative.

Psychiatry is now administering treatments that generate experiences it has never fully known how to talk about. If we do not develop a coherent framework, we risk disenchanting patients or abandoning clinical responsibility. What follows is an attempt to do better.

II. Anchoring the Problem in Real People, Not Abstract Theory

It is tempting to discuss mystical or anomalous experiences in abstract terms. In clinical practice, they rarely arrive that way. They arrive as people who are tentative, embarrassed, and unsure whether what they are about to say belongs in a psychiatrist’s office at all.

One example comes not from my clinic but from popular culture. Bob Gymlan has described vivid, recurring “lanky gray alien” figures from childhood that occasionally intruded into waking awareness (Gymlan, 2024). Clinically, what matters is not the imagery but the stance. He retains insight and interpretive flexibility, and he does not insist on literal belief.

From a diagnostic standpoint, nothing in this account requires psychosis. From a human standpoint, dismissing it as “just imagination” misses the point. Research on anomalous perceptual experiences suggests that unusual, vivid experiences can occur in psychologically intact individuals without predicting progression to psychotic disorder (Goulding, 2005; Dean et al., 2022).

In the clinic, the language is often more explicitly religious. One man with a history of alcohol and ketamine misuse recounted a vivid experience during a dissociative episode in which he felt confronted by something he described, without irony, as God. What mattered clinically was not the imagery, but that the experience left him sober, hopeful, and oriented toward a life he had previously given up on. Anthropological work suggests that such experiences do not imply psychopathology (Luhrmann, 2012; Luhrmann et al., 2010).

Then there are patients like Tim. Tim is an older man with medical comorbidities, including sleep apnea and a prior stroke. Under acute stress, he experiences terrifying episodes upon awakening: paralysis, pounding heart, confusion about whether he is dreaming or awake, and a powerful sense of threat. When I explained these episodes as REM intrusion, his anxiety decreased markedly. In his case, a material explanation was not disenchanting. It was therapeutic.

The literature supports this pattern. Sleep paralysis and hypnopompic hallucinations are common and frequently accompanied by fear and sensed presence (Cheyne et al., 1999; Sharpless & Barber, 2011; Denis et al., 2017). Cross-cultural studies suggest that while the imagery varies, the structure of the experience is remarkably consistent (Jalal & Ramachandran, 2017). For patients like Tim, demystification restores predictability and reduces secondary anxiety.

These cases illustrate a central dilemma. The same class of phenomena can demand radically different responses depending on the person experiencing them. Gymlan’s experiences call for interpretation without ridicule. Religious patients’ experiences call for respect without uncritical endorsement. Tim’s experiences call for explanation, reassurance, and symptom-focused intervention. A single interpretive stance fails all three.

If that is true, psychiatry’s discomfort is not because these experiences are rare or inherently pathological. It is because we lack a shared framework for responding to them. Before such a framework can be articulated, we need clarity about what neuroscience can and cannot explain about these states.

III. Mechanism Without Mockery: What Science Actually Explains

Before clinicians can respond wisely to mystical or anomalous experiences, we need clarity about what modern neuroscience does and does not claim about these states.

Over the last several decades, a converging literature across sleep science, computational neuroscience, psychosis research, and psychedelic science has clarified a crucial point: the human brain reliably enters states in which internally generated models dominate perception, affect, and self-experience.

Under ordinary waking conditions, sensory input strongly constrains perception. Under certain conditions, however, this balance shifts. When sensory precision is reduced or when prior models are excessively weighted, internally generated content can dominate conscious experience. This mechanism does not require psychosis but requires a change in state.

Experimental work demonstrates this directly. In a well-known study, healthy participants developed hallucination-like perceptions when perceptual priors were experimentally overweighted through conditioning, despite intact reality testing and no psychiatric illness (Powers et al., 2017). These findings support a graded, continuum-based model of anomalous experience rather than a categorical disease model (Corlett et al., 2009; Sterzer et al., 2018).

From this perspective, the question is not “why do some people hallucinate,” but “under what conditions do priors dominate perception?”

Sleep, REM intrusion, and the permeability of consciousness

Sleep research provides one of the clearest demonstrations of this principle. During REM sleep, internally generated imagery is vivid, emotionally salient, and narratively structured. Skeletal muscle tone is actively inhibited, and reality monitoring is largely offline. Normally, awakening proceeds gradually, allowing these systems to return to baseline together.

When that transition fails, REM features intrude into wakefulness. The result can include paralysis, vivid imagery, confusion about what is real, intense autonomic arousal, and a powerful sense of presence or threat. These phenomena are well described in the literature as sleep paralysis and hypnopompic hallucinations (Cheyne et al., 1999; Sharpless & Barber, 2011).

Importantly, the phenomenology is highly consistent. Large prospective and cross-cultural studies show that sleep paralysis frequently includes entity-like presences that feel external, intentional, and emotionally charged. While the specific imagery varies by cultural expectation, the structural features such as paralysis, sensed presence, threat, and boundary violation are remarkably stable (Cheyne, 2007; Jalal & Ramachandran, 2017). This strongly suggests that the brain is sampling a limited set of agentic models under conditions of reduced sensory constraint.

For patients like Tim, whose experiences are frightening and destabilizing, this mechanistic explanation is clarifying. Psychoeducation restores predictability and reduces catastrophic misinterpretation (Sharpless & Barber, 2011).

Psychedelics and the relaxation of high-level constraints

Psychedelics offer a different but closely related window into these mechanisms. Psilocybin, MDMA, and related compounds reliably alter large-scale brain network dynamics, particularly by reducing the dominance of high-level priors and increasing the influence of bottom-up signals and associative processes. One influential account describes this as a relaxation of top-down constraints, allowing for increased cognitive and emotional flexibility (Carhart-Harris & Friston, 2019).

Under these conditions, participants frequently report experiences of unity, sacredness, encounter, moral clarity, and insight. These experiences are not idiosyncratic. They are patterned, replicable, and strongly correlated with long-term therapeutic outcomes (Griffiths et al., 2006; Griffiths et al., 2018).

Crucially, participants do not lose insight into the experimental context. Many fully understand that a psychoactive compound was administered in a clinical or laboratory setting. Yet they continue to regard the experience as among the most meaningful of their lives. Long-term follow-up of the original Good Friday psilocybin experiment demonstrated that participants maintained the personal and spiritual significance of the experience decades later, despite explicit awareness of its pharmacological induction (Doblin, 1991). Modern replications have produced nearly identical results (Griffiths et al., 2006; Griffiths et al., 2018).

From a neuroscientific standpoint, this establishes a limited but important claim: certain brain states reliably generate experiences that humans interpret as deeply meaningful. Clinicians are increasingly encountering these states deliberately in practice.

Felt presence, agency detection, and meaning

One recurring feature across sleep paralysis, bereavement, psychosis-spectrum phenomena, and psychedelic states is the experience of a felt presence—the sense that “someone” or “something” is there, even in the absence of clear sensory input. Recent clinical reviews emphasize that felt presence is a transdiagnostic phenomenon, cutting across neurological, psychiatric, and non-clinical contexts (Barnby et al., 2023).

From a predictive processing perspective, felt presence reflects heightened agency inference under uncertainty. When sensory data are ambiguous and arousal is high, the brain’s threat-detection and social-cognition systems infer an intentional agent as the most plausible hidden cause. This inference can feel external, real, and emotionally charged without implying delusion or loss of insight (Corlett et al., 2009; Sterzer et al., 2018).

Mechanism explains how the experience arises. What it does not explain is why the person sitting across from the clinician still feels changed by it, oriented by it, or unsettled by it. Those downstream effects are what ultimately require clinical attention.

What mechanism does—and does not—settle

At this point, a familiar mistake often appears. Once a neural or computational mechanism is identified, the experience is assumed to be explained away. This is precisely the error William James identified under the label of medical materialism. To say that an experience arises from a brain state is not to say that it is trivial, deceptive, or empty of value (James, 1902/2002).

Understanding mechanism is necessary for safety, diagnosis, and treatment planning. It is not sufficient for integration. In practice, clinicians still have to decide how to speak, when to reassure, when to slow a patient down, and when to treat an experience as organizing rather than destabilizing. That work begins where explanation ends. The question then becomes how to respond in ways that preserve safety without foreclosing meaning.

IV. Respectful Naturalism: A Clinical Framework for Meaning Without Metaphysics

Neuroscience can tell us a great deal about how non-ordinary experiences arise. It cannot tell clinicians how to respond to them. That task is not metaphysical. It is clinical.

William James offered a distinction that remains unusually practical for psychiatry. The origins of an experience do not settle its value. Experiences are better evaluated by their effects on a person’s life and character, judged by their fruits rather than their causes or metaphysical status. This does not reject scientific explanation. It places it in its proper role.

In practice, this means abandoning the idea that the clinician’s task is to decide what an experience ultimately is. Psychiatry is poorly equipped to arbitrate metaphysics. The task is narrower and more demanding: to decide how an experience is functioning in this patient’s life, at this time, given their vulnerabilities, strengths, and context.

Seen this way, non-ordinary experiences tend to cluster into several broad clinical patterns.

Some experiences are frightening precisely because they are poorly understood. Tim’s REM-intrusion episodes are a clear example. For him, the experiences were terrifying rather than meaningful. In such cases, a clear neurophysiological explanation is not disenchanting. It is relieving. Demystification restores predictability and reduces secondary anxiety.

Other experiences are vivid, unusual, and emotionally charged yet occur in the context of intact insight and stable functioning. Bob Gymlan’s childhood experiences fall into this category. Here, reflexive reductionism is neither necessary nor helpful. The clinician’s role is not to resolve the experience, but to assess whether it organizes values and coexists with stable sleep, work, and relationships.

Still other experiences are interpreted explicitly through spiritual language and catalyze moral or behavioral change. In these cases, respectful naturalism involves working within the patient’s interpretive frame without becoming its guarantor. The clinician does not need to decide whether God spoke. The clinician needs to assess whether the interpretation promotes psychological integration or increases rigidity, guilt, fear, or destabilization.

Across all three patterns, risk is tracked in function, not theology. Warning signs include loss of insight, escalating certainty, impaired sleep, relationship rupture, and grandiosity. These markers matter regardless of whether an experience is framed as spiritual, neurological, symbolic, or psychedelic-induced.

This framework also clarifies why both reflexive reductionism and uncritical affirmation fail clinically. Explaining an experience as “nothing but brain activity” often leaves patients feeling subtly invalidated and alone with its meaning. Treating experiences as authoritative truths can harden into rigidity or grandiosity. Respectful naturalism avoids both errors by insisting that meaning be tested against ordinary life over time.

Psychedelic medicine makes this stance unavoidable. These treatments reliably occasion experiences patients experience as sacred, revelatory, or morally decisive. The clinical task is not to affirm or dismiss these interpretations, but to slow integration, protect functioning, and allow meaning to mature rather than crystallize prematurely.

V. A Practical Framework for Clinicians: Holding Meaning Without Losing the Frame

If the preceding sections argue that non-ordinary experiences are common, meaningful, and mechanistically intelligible, this section addresses the question clinicians ultimately care about most: what does responsible practice look like when these experiences walk into the room.

The challenge is not theoretical. It is practical, ethical, and increasingly unavoidable as psychedelic-assisted therapies move into routine psychiatric care. A framework that cannot be applied during intake, preparation, dosing, or integration will either be ignored or replaced by reflex. The aim here is to outline a stance that avoids both pathologizing meaning and romanticizing instability while maintaining clinical rigor.

The two recurrent clinical errors

Most clinical missteps around mystical or spiritually interpreted experiences cluster around two opposite but equally tempting errors.

The first is pathologizing meaning. This occurs when clinicians reflexively translate a patient’s experience into symptoms the moment spiritual or mystical language appears. The experience is reduced to anxiety, dissociation, psychosis risk, or drug effect. While such translations may be mechanistically accurate, studies in religion and mental health suggest that premature reduction often damages alliance, increases shame, and discourages future disclosure (Koenig, 2012; Luhrmann, 2012).

The second error is romanticizing instability. This occurs when clinicians treat mystical language as inherently profound or exempt from ordinary clinical scrutiny. In psychedelic contexts, this risk is especially acute, since experiences can feel urgent, authoritative, and morally binding. Failure to track sleep, functioning, and belief rigidity can inadvertently reinforce grandiosity or destabilizing interpretations, a risk well documented in psychosis-spectrum and substance-related literatures (Corlett et al., 2009; Sterzer et al., 2018).

A three-level clinical model: mechanism, meaning, integration

A simple but effective organizing tool is to separate clinical work into three distinct but interacting levels.

Level 1: Mechanism

This is where neuroscience, sleep medicine, and psychopharmacology operate most confidently. It includes altered brain network dynamics, predictive processing and precision weighting, REM intrusion and sleep fragmentation, autonomic arousal, and pharmacologic effects. Clear explanation at this level protects patients from catastrophic misinterpretation and unnecessary fear, as seen in sleep paralysis and REM intrusion syndromes (Sharpless & Barber, 2011; Cheyne et al., 1999).

Level 2: Meaning

This is the lived, first-person experience: visions, encounters, unity, dread, conviction, or moral clarity. This level deserves careful listening and descriptive fidelity without premature interpretation. Acknowledging meaning does not require endorsing metaphysical claims.

Level 3: Integration

This is where psychiatry exerts its greatest clinical leverage. Integration focuses on changes in daily behavior, effects on relationships and work, shifts in values and commitments, and flexibility versus rigidity over time. Meaning that cannot be integrated tends to either fade or destabilize. Integration that ignores mechanism risks harm. Psychedelic integration research increasingly emphasizes this level as central to outcomes (Bathje et al., 2022; Thal et al., 2024).

Clinicians can validate Level 2 experiences while remaining confident at Level 1 and anchoring treatment decisions at Level 3.

Enacting respectful naturalism in the room

In practice, respectful naturalism is conveyed less through theory than through language. Empirical and qualitative studies suggest that patients respond best when clinicians explicitly acknowledge the reality of the experience while deferring metaphysical conclusions (Luhrmann, 2012; Koenig, 2012).

Helpful clinical phrases include:

• “That sounds very real and important to you.”

• “Many people have experiences like this in certain brain states.”

• “We do not have to decide exactly what it was to work with what it did.”

• “Let’s pay attention to how this affects your life over time.”

What is intentionally absent is metaphysical closure.

What clinicians should actively track

Risk assessment does not require theological judgment. It requires pattern recognition. Regardless of how an experience is interpreted, clinicians should monitor variables repeatedly associated with destabilization across diagnostic categories: sleep disruption or reduced need for sleep, escalating certainty paired with reduced tolerance for ambiguity, withdrawal from trusted relationships, and grandiosity or persecutory themes.

Psychosis-spectrum models emphasize that risk lies not in unusual experiences themselves, but in how beliefs are updated, fixed, and acted upon over time (Corlett et al., 2009; Sterzer et al., 2018). This framing allows clinicians to intervene without debating ontology.

Special implications for psychedelic medicine

Psychedelic-assisted therapies make this framework non-optional. Psilocybin- and MDMA-assisted protocols reliably occasion experiences described as sacred, revelatory, or morally authoritative, with intensity often correlated with therapeutic outcome (Griffiths et al., 2006; Griffiths et al., 2018). Integration research stresses that these experiences require careful containment rather than celebration or dismissal (Bathje et al., 2022; Thal et al., 2024).

In practice, this means preparing patients in advance for intense meaning without promising truth, discouraging major life decisions in the immediate aftermath, protecting sleep and routine, emphasizing humility and gradual integration, and pausing or spacing sessions if destabilization emerges.

Clinicians need not act as spiritual directors. They do need to act as stewards of meaning.

Leaving the door open without walking through it

A defining feature of respectful naturalism is its metaphysical humility. The clinician neither asserts nor denies spiritual reality. Instead, they acknowledge the limits of their role. Neuroscience can describe correlates. Psychiatry can assess risk and function. Ultimate meaning exceeds our jurisdiction.

This stance neither flattens experience nor inflates it. It creates space for meaning to mature slowly, tested against ordinary life rather than insulated from it.

VI. Conclusion: Re-Enchantment Without Naivety

Psychiatry finds itself at an inflection point.

Psychiatrists have always worked at the boundary between brain and meaning, whether we acknowledge it or not. Despite narratives of secular disenchantment, patients continue to bring experiences into our offices that feel sacred, revelatory, or morally decisive. Epidemiology tells us this is not aberrant. Neuroscience tells us it is lawful. Clinical experience tells us it can either integrate a life or destabilize it.

The question, then, is not whether such experiences belong in psychiatric care. They already do. The question is how we will meet them.

What the converging work of William James, Andrew Greeley, modern neuroscience, and contemporary psychedelic research makes increasingly clear is something both simpler and more demanding: non-ordinary states of consciousness are a normal feature of human psychological life. They arise under identifiable conditions. They recur across cultures and historical periods. They can heal or harm. And they matter deeply to the people who experience them (James, 1902/2002; Greeley & McCready, 1975; Griffiths et al., 2006).

A Jamesian stance of respectful naturalism offers a way forward. It allows clinicians to remain scientifically grounded while resisting disenchantment. It treats subjective experience as real without insisting on literal belief. It evaluates experiences by their fruits rather than their origins. And it insists that meaning must be tested and integrated in ordinary life—sleep, work, relationships, humility—rather than allowed to crystallize prematurely into certainty or chaos (James, 1902/2002; Bathje et al., 2022).

This stance is no longer optional. Psychedelic-assisted therapies will soon make encounters with awe, unity, presence, and moral insight routine within psychiatric practice. Participants already interpret these experiences spiritually, existentially, and personally, even when they fully understand their neuropharmacological basis (Griffiths et al., 2006; Griffiths et al., 2018; Doblin, 1991). Clinicians will be forced to respond to the consequences of these experiences rather than to their metaphysics.

That response requires a framework that avoids pathologizing meaning while avoiding romanticizing instability. It requires clinicians who can explain mechanism without mockery, who can listen without collusion, and who can protect patients from harm without stripping their experiences of significance. It requires humility about what science can answer and honesty about what it cannot.

The question, then, is not whether such experiences belong in psychiatric care. They already do. The question is whether we will meet them with fear, certainty, or care.

Respectful naturalism does not resolve the mystery. It refuses to close it prematurely. It leaves the door open without forcing anyone through it, and it grounds meaning in the slow, disciplined work of living well over time. As psychedelics and spiritual longing re-enter mainstream culture, psychiatry has an opportunity to model something rare: re-enchantment without naivety, and rigor without erasure.

That may be the most humane contribution the field can make.

Works & Further Reading

This essay draws on work across psychiatry, neuroscience, psychology, sociology of religion, and psychedelic science. I’ve grouped the references below so readers can follow the thread that most interests them. Some of these sources strongly support the framework I outline here; others push back against it in ways I find useful and clarifying.

Foundational texts: meaning, mysticism, and pragmatism

If there is a single intellectual ancestor behind this essay, it is William James.

• James, W. (1902/2002). The Varieties of Religious Experience.
  A classic for a reason. James treats mystical experience seriously without metaphysical dogmatism and insists on judging experiences by their effects on life rather than their origins. This book quietly underwrites much of modern thinking about spirituality and mental health.

• Luhrmann, T. M. (2012). When God Talks Back.
  An anthropologist’s careful account of how religious experience is cultivated, interpreted, and lived without assuming pathology or supernatural proof.

How common are mystical experiences, really?

These works establish that mystical or “ecstatic” experiences are not fringe phenomena.

• Greeley, A. M., & McCready, W. C. (1975). “Are we a nation of mystics?” New York Times Magazine.
  Based on National Opinion Research Center data, this piece startled readers by showing that a large minority of Americans report at least one mystical-type experience. Hard to find online, but widely cited and quoted online.

• Fox, J. W. (1992). “The structure, stability, and social antecedents of reported paranormal experiences.” Sociological Analysis.
  A sober sociological treatment of anomalous experience in ordinary populations.

Modern neuroscience: mechanism without metaphysics

These sources explain how non-ordinary experiences arise without claiming that mechanism exhausts meaning.

• Friston, K. (2010). “The free-energy principle.” Nature Reviews Neuroscience.
  The theoretical backbone of predictive processing models of perception and belief.

• Sterzer, P., et al. (2018). “The predictive coding account of psychosis.” Biological Psychiatry.
  A technical but important review showing how unusual experiences can arise along a continuum rather than as categorical illness.

• Corlett, P. R., et al. (2009). “From drugs to deprivation.” Psychopharmacology.
  Connects Bayesian models of perception to psychosis-spectrum phenomena.

Sleep, REM intrusion, and “felt presence”

These papers are especially relevant for experiences that feel paranormal but arise from sleep-wake boundary disturbances.

• Sharpless, B. A., & Barber, J. P. (2011). “Lifetime prevalence rates of sleep paralysis.” Sleep Medicine Reviews.
  Shows how common sleep paralysis is and how often it is misinterpreted.

• Jalal, B., & Ramachandran, V. S. (2017). “The ghostly bedroom intruder.” Frontiers in Human Neuroscience.
  An elegant explanation of why sleep paralysis so often produces the sense of a threatening presence.

• Barnby, J. M., et al. (2023). “The felt-presence experience.” The Lancet Psychiatry.
  A modern synthesis connecting felt presence across neurology, psychiatry, and non-clinical experience.

Psychedelics and mystical-type experience

These studies are central to the coming clinical shift discussed in the essay.

• Griffiths, R. R., et al. (2006). “Psilocybin can occasion mystical-type experiences.” Psychopharmacology.
  One of the most influential papers in modern psychedelic research.

• Griffiths, R. R., et al. (2018). “Enduring positive changes.” Journal of Psychopharmacology.
  Demonstrates that the meaning of psychedelic experiences persists long after the acute effects fade.

• Doblin, R. (1991). “The Good Friday Experiment: A long-term follow-up.” Journal of Transpersonal Psychology.
  A fascinating historical follow-up showing decades-long impact despite full awareness of pharmacological induction.

• Carhart-Harris, R. L., & Friston, K. J. (2019). “REBUS and the anarchic brain.” Pharmacological Reviews.
  A mechanistic account of how psychedelics relax high-level priors.

Integration, clinical practice, and guardrails

These are especially useful for clinicians working with spiritually meaningful experiences.

• Bathje, G. J., et al. (2022). “Psychedelic integration.” Frontiers in Psychology.
  A clear, clinician-friendly attempt to define what “integration” actually means.

• Thal, S. B., et al. (2024). “Therapeutic frameworks for integration sessions.” Clinical Psychology & Psychotherapy.
  Practical frameworks for working with post-psychedelic meaning without destabilization.

• Koenig, H. G. (2012). “Religion, spirituality, and health.” ISRN Psychiatry.
  A broad review of what psychiatry does—and does not—know about religion and mental health.

Public-facing and gray literature (including disagreement)

These pieces are not peer-reviewed, but they shape public and clinical discourse.

• Hayes, C. (2024). “Psychedelic experiences in clinical settings.” The Guardian.
  A skeptical, cautionary take that raises real concerns about framing, consent, and interpretation.

• Jalal, B. (2023). “Why sleep paralysis makes you see ghosts.” TIME.
  An accessible explanation aimed at the general public.

• Chacruna Institute essays on mysticism and integration.
  Spiritually open perspectives from within the psychedelic community that sometimes push further than clinicians should but are worth reading alongside more cautious perspectives.

A note on disagreement

Not all of these authors would agree with the framework I outline here. Some lean more strongly toward reductionism; others toward spiritual interpretation. I’ve included both intentionally. My argument is not that science proves spiritual reality, nor that mystical experience should be privileged over skepticism. It is that human meaning persists even when mechanism is understood, and that psychiatry must learn to work with that fact rather than pretending it away.

Complete Works Cited

Bathje, G. J., Majeski, E., & Kudowor, A. (2022). Psychedelic integration: An analysis of the concept and its practice. Frontiers in Psychology, 13, 824077.

Carhart-Harris, R. L., & Friston, K. J. (2019). REBUS and the anarchic brain: Toward a unified model of the brain action of psychedelics. Pharmacological Reviews, 71(3), 316–344. https://doi.org/10.1124/pr.118.017160

Castelnovo, A., Cavallotti, S., Gambini, O., & D’Agostino, A. (2015). Post-bereavement hallucinatory experiences: A critical overview of population and clinical studies. Journal of Affective Disorders, 186, 266–274. https://doi.org/10.1016/j.jad.2015.07.032

Cheyne, J. A., Rueffer, S. D., & Newby-Clark, I. R. (1999). Relations among hypnagogic and hypnopompic experiences associated with sleep paralysis. Journal of Sleep Research, 8(4), 313–317. https://doi.org/10.1046/j.1365-2869.1999.00165.x

Cheyne, J. A., & Girard, T. A. (2007). Paranoid delusions and threatening hallucinations: A prospective study of sleep paralysis experiences. Consciousness and Cognition, 16(4), 959–974. https://doi.org/10.1016/j.concog.2007.01.002

Cheyne, J. A., & Girard, T. A. (2009). The body unbound: Vestibular–motor hallucinations and out-of-body experiences. Cortex, 45(2), 201–215. https://doi.org/10.1016/j.cortex.2007.05.002

Corlett, P. R., Frith, C. D., & Fletcher, P. C. (2009). From drugs to deprivation: A Bayesian framework for understanding models of psychosis. Psychopharmacology, 206(4), 515–530. https://doi.org/10.1007/s00213-009-1561-0

Dean, C. E., Akhtar, S., Gale, T. M., Irvine, K., Grohmann, D., et al. (2022). Paranormal beliefs and cognitive function: A systematic review and assessment of study quality across four decades of research. PLOS ONE, 17(5), e0267360. https://doi.org/10.1371/journal.pone.0267360

Denis, D., & Poerio, G. L. (2017). Terror and bliss? Commonalities and distinctions between sleep paralysis, lucid dreaming, and their associations with waking life experiences. Journal of Sleep Research, 26(1), 38–47. https://doi.org/10.1111/jsr.12441

Doblin, R. (1991). Pahnke’s “Good Friday experiment”: A long-term follow-up and methodological critique. Journal of Transpersonal Psychology, 23(1), 1–28.

Fox, J. W. (1992). The structure, stability, and social antecedents of reported paranormal experiences. Sociological Analysis, 53(2), 151–167.

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Goulding, A. (2005). Healthy schizotypy in a population of paranormal believers and experients. Personality and Individual Differences, 38(5), 1069–1083.

Greeley, A. M., & McCready, W. C. (1975, January 26). Are we a nation of mystics? New York Times Magazine.

Griffiths, R. R., Richards, W. A., McCann, U., & Jesse, R. (2006). Psilocybin can occasion mystical-type experiences having substantial and sustained personal meaning and spiritual significance. Psychopharmacology, 187(3), 268–283. https://doi.org/10.1007/s00213-006-0457-5

Griffiths, R. R., Johnson, M. W., Richards, W. A., et al. (2018). Psilocybin-occasioned mystical-type experience in combination with meditation and other spiritual practices produces enduring positive changes. Journal of Psychopharmacology, 32(1), 49–69. https://doi.org/10.1177/0269881117731279

Gymlan, B. (2024, July 3). The chilling matter of extraterrestrials in bedrooms [Video]. YouTube.

James, W. (2002). The varieties of religious experience: A study in human nature. Modern Library. (Original work published 1902)

Jalal, B., & Ramachandran, V. S. (2017). Sleep paralysis, “the ghostly bedroom intruder,” and out-of-body experiences. Frontiers in Human Neuroscience, 11, 92. https://doi.org/10.3389/fnhum.2017.00092

Koenig, H. G. (2012). Religion, spirituality, and health: The research and clinical implications. ISRN Psychiatry, 2012, 278730. https://doi.org/10.5402/2012/278730

Kamp, K. S., & Due, H. (2019). How many bereaved people hallucinate about their loved one? A systematic review and meta-analysis of bereavement hallucinations. Journal of Affective Disorders.

Kamp, K. S., & Due, H. (2019). Retraction notice to “How many bereaved people hallucinate about their loved one? A systematic review and meta-analysis of bereavement hallucinations.” Journal of Affective Disorders, 250, 447. https://doi.org/10.1016/j.jad.2019.02.067

Luhrmann, T. M. (2012). When God talks back: Understanding the American evangelical relationship with God. Vintage.

Luhrmann, T. M., Nusbaum, H., & Thisted, R. (2010). The absorption hypothesis: Learning to hear God in evangelical Christianity. American Anthropologist, 112(1), 66–78. https://doi.org/10.1111/j.1548-1433.2009.01197.x

Powers, A. R., Mathys, C., & Corlett, P. R. (2017). Pavlovian conditioning-induced hallucinations result from overweighting of perceptual priors. Science, 357(6351), 596–600. https://doi.org/10.1126/science.aan3458

Sharpless, B. A., & Barber, J. P. (2011). Lifetime prevalence rates of sleep paralysis: A systematic review. Sleep Medicine Reviews, 15(5), 311–315. https://doi.org/10.1016/j.smrv.2011.01.007

Sterzer, P., Adams, R. A., Fletcher, P., et al. (2018). The predictive coding account of psychosis. Biological Psychiatry, 84(9), 634–643. https://doi.org/10.1016/j.biopsych.2018.05.015

LB-102, the centerpiece of LB Pharmaceuticals’ story, fits squarely into this pattern. It bills itself as the “next-generation antipsychotic,” potentially transformative for schizophrenia. But peel back the press releases, clinical trial decks, and S-1 filings, and what emerges is not a radical new treatment but something closer to a nostalgic reboot: a 1980s sedan with a fresh coat of paint, new hubcaps, and a higher sticker price.

This essay tries to make sense of LB-102 by translating corporate spin into everyday terms. We’ll explore what LB-102 is, what it does, why the company thinks it deserves attention, and why I think it’s more mirage than miracle. Along the way, we’ll touch on psychiatry’s history of me-too drugs, the market incentives that fuel them, and what it says about how innovation really works in mental health care.

Part I: Déjà Vu All Over Again

LB-102 is not a new molecule in any meaningful sense. It is N-methylamisulpride—literally amisulpride with an extra methyl group [3]. Amisulpride itself is hardly forgotten: developed by Sanofi in the late 1970s, it’s been prescribed across Europe as Solian since the 1980s. Meta-analyses consistently rank it among the more effective antipsychotics [20], [21], with a tolerability profile that is neither spectacularly better nor worse than peers. The reason it’s not in U.S. formularies is banal: Sanofi never bothered to push it through the FDA before the patent expired in 2008. Not scandalous, just economics.

LB Pharma’s bet is that methylating amisulpride improves its blood-brain barrier penetration [4], [5]. Better brain uptake, they argue, allows for lower doses, less drug floating around in plasma, and theoretically fewer systemic side effects. This is pharmacokinetic fine-tuning, not pharmacological reinvention. On paper, the logic is sound. In practice, the improvements so far appear incremental. LB-102 still binds to the same dopamine D2 and D3 receptors (with some 5-HT7 antagonism for flourish) [6], [7], [8]. It’s the same engine under a slightly more fuel-efficient hood.

Part II: The Trial Data—Significance Without Significance

LB Pharma hypes LB-102 as a “next-generation antipsychotic” [1], [2], [9]. Here’s what the data actually show:

• Phase 2 efficacy: In the pivotal NOVA study, a 4-week trial of 359 patients, LB-102 reduced PANSS scores by 5.0 points at 50 mg and 6.8 points at 100 mg compared to placebo [10]. For context, PANSS is a 210-point scale. These deltas are statistically significant but clinically modest. By contrast, Bristol Myers Squibb’s new muscarinic agonist KarXT (now branded Cobenfy) notched an 8.4-point advantage over placebo in 5 weeks [11]. In other words, LB-102 looks roughly equivalent to existing generics—not obviously superior.

• Negative symptoms and cognition: The company trumpets that LB-102 helps with negative symptoms (amotivation, blunted affect) and cognition [12], [13]. Reality: only the 50 mg dose showed a ~2.4-point benefit on the PANSS negative subscale, while the 100 mg dose missed statistical significance [14], [15]. Cognitive benefits required excluding outliers and pooling completers to generate significance [16], [17]. These are post hoc findings at best, wishful thinking at worst.

• Tolerability: LB-102’s supposed crown jewel is its side effect profile [18], [19]. But the data look like amisulpride’s greatest hits: prolactin elevation in most patients [22], [23], dystonia in one patient severe enough to be treatment-related [24], and four cases of acute dystonia in Phase 1 at higher doses [25], [26]. Patients on LB-102 gained about 2 kg more than placebo in 4 weeks [27], [28]—an annualized nightmare for anyone tasked with counseling patients on weight.

The result is a portrait familiar to psychiatrists: a drug that works, but not dramatically better than the dozen others already available.

Part III: The Clinical Trial as Performance Art

Why did LB Pharma run a 4-week trial when the standard in schizophrenia is 6 weeks [29]? Possibly because amisulpride tends to work quickly, and shortening the trial limits placebo creep. Possibly because fewer weeks means fewer dropouts and less cost. The risk is that you capture short-term efficacy without long-term durability. Unsurprisingly, the trial met its endpoint. Unsurprisingly, the deltas were modest. The company now hopes the FDA will treat this as one of two pivotal studies [30]. Maybe they will. But even if the FDA accepts it, the trial reads more like performance art than a robust test: engineered to generate significance, not to meaningfully shift clinical practice.

Then there’s the post hoc embroidery. Subgroup analyses of patients with higher baseline negative scores. Exploratory cognitive composites that only survive statistical significance after pruning. A small 100 mg cohort with outsized effects that the company highlights as if it weren’t underpowered. None of this is fraudulent; all of it is familiar. Every biotech digs through its data to find flattering angles. The problem is when those angles get marketed as gospel rather than tentative hypotheses.

Part IV: The Regulatory Bar (Low, but Not Nonexistent)

Could LB-102 get FDA approval? Absolutely. The bar for antipsychotics is lower than one might expect. Lumateperone (Caplyta) was approved in 2019 with underwhelming efficacy data, largely on the promise of tolerability. The FDA knows amisulpride is effective, and LB-102 looks similar enough. Two positive trials, one strong, one weake, are often enough for approval.

But approval is only half the story. The harder question is: will anyone prescribe it? And will insurers pay for it?

Part V: The Market—An Overcrowded Party

The U.S. schizophrenia market is quoted at $12 billion [31], but most of that pie is spoken for by cheap generics (risperidone, aripiprazole, quetiapine) and long-acting injectables like Invega. New entrants without clear differentiation are lucky to scrape 1% market share. History is littered with examples: ziprasidone (Geodon), iloperidone (Fanapt), asenapine (Saphris). All launched with some spin about tolerability. All languished [32], [33], [34], [35].

LB-102 arrives just as Cobenfy has changed the conversation. Unlike LB-102, Cobenfy actually represents a new mechanism, and psychiatrists are intrigued. Against that backdrop, LB-102 looks like a polite but forgettable guest arriving late to a party that has already crowned its star.

And insurers are unlikely to indulge. Step therapy rules mean patients must fail generics before trying expensive branded options. Why would an insurer pay for LB-102, a me-too benzamide with no clear superiority, at hundreds of dollars a month? The answer is: they probably won’t.

LB Pharma gestures toward developing a long-acting injectable (LAI) version. That could help with adherence, but LAIs are already a crowded space dominated by Janssen and Otsuka. A scrappy single-asset biotech is unlikely to unseat them.

Part VI: The Insider Deals

If the science doesn’t wow you, maybe the governance will. Except it doesn’t. From the beginning, LB Pharma’s insiders carved out perpetual royalties for themselves: up to 3.25% of LB-102 sales indefinitely [36], [37]. This isn’t a reward for risk; it’s a skim. The same insiders, CEO Heather Turner, co-founder Zachary Prensky, CSO Andrew Vaino, CFO Marc Panoff, also hold cheap equity bought at $1.50/share [38], [39]. If the IPO goes out at $10, that’s a 6–7x markup before patients or payers have weighed in.

Meanwhile, the company had a going-concern warning in 2025 with less than 12 months of runway [40]. They need a $100M IPO just to keep the lights on [41], and even that won’t fund commercialization [42], [43]. Which means more dilution ahead. In that context, LB-102 looks less like a medical mission and more like a liquidity event.

Part VII: The Broader Lesson - What Counts as Innovation?

To be fair, LB-102 is not worthless. If it’s approved, psychiatrists may find it a tolerable, once-daily, serviceable option. Incremental innovation has a place. Patients are heterogeneous, and one more choice can help. But that is a far cry from the company’s rhetoric of “next-generation,” “transformational,” or “first-in-class.”

What LB-102 really highlights is how low the bar for “innovation” often is in psychiatry. Add a methyl group, tweak the pharmacokinetics, rerun the same receptor blockade story, and you can market yourself as cutting-edge. It’s not fraud. It’s not salvation. It’s theater - financial, regulatory, and clinical.

Conclusion: The Mirage on the Horizon

LB-102 will probably get approved. It may carve out a small niche. But it will not revolutionize schizophrenia care. More likely, it will join the long list of well-publicized, modestly effective, commercially underwhelming antipsychotics that psychiatrists sometimes prescribe, insurers often restrict, and investors eventually forget.

The tragedy is not that LB-102 exists. The tragedy is that psychiatry’s bar for innovation is so low that simply methylating a 40-year-old drug can generate a nine-figure IPO. That may be good business, but it’s a mirage of progress that fades the closer you look.

Works Cited

[1] FierceBiotech coverage of LB Pharma’s “first-in-class/transformational” claims: https://www.fiercebiotech.com/biotech/lb-pharmas-twist-old-sanofi-drug-passes-schizophrenia-test-teeing-phase-3-push
[2] LB Pharma Investor Presentation (2025 Overview Deck): https://lbpharma.us/wp-content/uploads/2025/02/LB_Pharma_Corp_Overview-Deck-0205.pdf
[3] LB Pharma Corporate Presentation (2021): https://lbpharma.us/wp-content/uploads/2021/02/LB-Corporate-Presentation-February-2021-with-Appendix.pdf
[4] N-Methylamisulpride Wikipedia Entry: https://en.wikipedia.org/wiki/N-Methylamisulpride
[5] Pharmacokinetics of N-Methylamisulpride: https://en.wikipedia.org/wiki/N-Methylamisulpride
[6] Binding Profile of Amisulpride: LB Pharma PET Study (2024)
[7] Dopamine Receptor Occupancy PET Results: https://lbpharma.us/wp-content/uploads/2024/10/PET-NPP-paper.pdf
[8] In Vitro Binding Comparison (Amisulpride vs LB-102): LB Pharma Publications
[9] LB Pharma Claims of Strong Efficacy: https://lbpharma.us/lb-102/
[10] LB Pharmaceuticals Phase 2 Results Press Release: https://www.globenewswire.com/news-release/2025/01/08/3006104/0/en/LB-Pharmaceuticals-Announces-Positive-Topline-Results-from-Phase-2-Trial-of-LB-102-in-Schizophrenia.html
[11] Bristol Myers Squibb KarXT/Cobenfy Results: https://www.fiercebiotech.com/biotech/lb-pharmas-twist-old-sanofi-drug-passes-schizophrenia-test-teeing-phase-3-push
[12] Psychiatric Times Coverage of LB-102: https://www.psychiatrictimes.com/view/lb-102-a-new-novel-drug-for-schizophrenia
[13] LB Pharma S-1 Filings (2025): file://file-G4yHwPVUAy4FdnGZsG87vV
[14] PANSS Negative Subscore Data (LB Pharma S-1)
[15] Phase 2 Trial Details: file://file-G4yHwPVUAy4FdnGZsG87vV
[16] Cognitive Outcomes Analysis (LB Pharma S-1)
[17] Cogstate Battery Report (LB Pharma SEC Filings)
[18] LB Pharma Safety Claims: https://lbpharma.us/
[19] LB Pharma Globenewswire PR (2025): https://www.globenewswire.com/news-release/2025/01/08/3006104/0/en/LB-Pharmaceuticals-Announces-Positive-Topline-Results-from-Phase-2-Trial-of-LB-102-in-Schizophrenia.html
[20] Leucht et al. Meta-analysis of Antipsychotics (Lancet, 2019)
[21] Cipriani et al. Comparative Efficacy Studies of Schizophrenia Drugs (Nature, 2019)
[22] Prolactin Elevation Reported in Phase 2: LB Pharma S-1
[23] QTc Prolongation Disclosure: LB Pharma SEC Filings
[24] Phase 2 Adverse Events (One Dystonia Case): LB Pharma S-1
[25] Phase 1 Acute Dystonia: LB Pharma Clinical Trial Poster (2022)
[26] Phase 1 EPS Events Report: https://lbpharma.us/wp-content/uploads/2024/09/Phase-1-a-randomized-double-blind-placebo-controlled-study-of-the-safety-tolerability-pharmacokinetics-and-pharmacodynamics-of-LB-102-July-2022-Springer-Nature.pdf
[27] Weight Gain Reported: LB Pharma Globenewswire PR (2025)
[28] Comparative Weight Data: LB Pharma SEC Filings
[29] Clinical Trial Design Duration Standards (FDA)
[30] LB Pharma S-1: FDA Feedback on Pivotal Trial Status
[31] U.S. Schizophrenia Market Size Estimate: https://lbpharma.us/
[32] Historical Failures of Me-Too Antipsychotics: file://file-R8jKgcsMGWt2zEwTZLCJHF
[33] Historical Failures of Me-Too Antipsychotics: file://file-R8jKgcsMGWt2zEwTZLCJHF
[34] Commercial Underperformance of Antipsychotics: file://file-R8jKgcsMGWt2zEwTZLCJHF
[35] Commercial Underperformance of Antipsychotics: file://file-R8jKgcsMGWt2zEwTZLCJHF
[36] Insider Royalty Agreements: LB Pharma S-1
[37] Insider Royalty Amendments: LB Pharma SEC Filings
[38] Series C Funding Details: LB Pharma SEC Filings
[39] Share Price Data ($1.50/share): LB Pharma SEC Filings
[40] Going Concern Warning: LB Pharma SEC Filings
[41] IPO Filing (2025): https://www.renaissancecapital.com/IPO-Center/News/112955/Neuropsychiatry-biotech-LB-Pharmaceuticals-files-for-a-$100-million-IPO
[42] Insufficient IPO Proceeds Disclosure: LB Pharma SEC Filings
[43] Insufficient IPO Proceeds Disclosure: LB Pharma SEC Filings