In early August of 2026, just a few weeks ago, The Economist ran a thought-provoking piece titled: “The brain may be about to have its Ozempic moment.“ The piece highlighted the FDA approval of oveporexton, now marketed by Takeda as Orzeyful, for narcolepsy. The main idea is that similar to Ozempic, and its cousins, which have become household names marking a cultural moment, orexin agonists might follow a similar path. Meaning, while GLP-1 drugs started out as treatments for diabetes, they have now expanded into sleep apnea, cardiovascular and kidney disease, and might have utility even for substance use.

As you can immediately appreciate, I’m inspired by the Economist’s talent for attention grabbing headlines and have adopted it myself here. I even thought about bringing Limitless into the title, but thought it was an overkill and a potential distraction from what I really want to talk about– the promise for these medications in ADHD treatment.

Why ADHD?

My primary clinical interest (and expertise) is in psychotic disorders, broadly defined. I am not a child psychiatrist and don’t see the breadth of what people term ADHD. Discussions of construct validity aside, my primary exposure to ADHD is in the context of co-morbidity with psychotic illness.

There exists now a sizable population of humans on this planet (but maybe I’m mostly referring to planet USA, I don’t know) that present with a diagnosis of a psychosis spectrum illness as well as ADHD. Again, diagnostic validity aside, and the role of long term stimulant use in the evolution of the illness also aside, there is a practical conundrum facing psychiatrists prescribing stimulants in the context of a psychotic illness. Stimulants tend to worsen psychosis, both in the short term and perhaps in the long term as well.

On the other hand, stimulants are efficacious medications that help millions of people with ADHD live functional lives. Indeed, multiple studies indicate that they have the largest clinical effects in controlled studies and the real world. Nonstimulant alternatives are available, they just don’t tend to have the same efficacy. You often hear “Strattera does nothing”, but almost never the same about Adderall.

So it’s a real issue. My patients, at baseline, struggle with cognitive processing. In some cases they report issues that seem to overlap with those seen in ADHD. Given everything discussed, this population is in dire need for stimulant alternatives.

That I am certain of.

What Is Orexin and Why Does It Matter?

Orexin’s discovery is a remarkable story that I first heard about when I was in the middle of my PhD. It is one of those rare stories where two independent groups simultaneously identified the same system from completely different angles. In 1998, Emmanuel Mignot’s group at Stanford was studying narcoleptic dogs, trying to understand why these animals would suddenly collapse into sleep when excited. I actually learned some of the details about the story from attending Mignot’s lecture when I was a student. He ultimately revealed that through genetics, his team identified mutations in the orexin receptor as being responsible for the narcolepsy phenotype. Around the same time, a Japanese group led by Masashi Yanagisawa created knockout mice lacking the orexin peptide and discovered that these mice exhibited narcolepsy-like symptoms.

The convergence was pretty amazing, and led to the subsequent discovery that patients with narcolepsy type 1 lose the neurons that produce orexin. This is a relatively small population of neurons in the hypothalamus whose loss has a huge impact on the ability to maintain normal arousal states.

Orexin neurons project broadly to multiple arousal systems. They connect to the locus coeruleus (norepinephrine), ventral tegmental area (dopamine), dorsal raphe (serotonin), and basal forebrain (acetylcholine). In that sense, orexin may be thought of as a master coordinator for these systems, with its disruption leading to state instability.

I think the word arousal may be a bit confusing here because people tend to equate it with being awake. There is obviously much more to it than that. We switch between different states of (task) engagement throughout the day. Arousal systems participate in regulating these transitions, and orexin, being a master coordinator is well-positioned to play an important role.

Figure from: Sakurai, 2007 Nature Reviews Neuroscience

Orexin Agonists Enhance Human Cognitive Function

Indeed, a study published in JAMA Neurology last year showed the utility of oveporexton in enhancing cognitive function. The study treated 112 patients with narcolepsy type 1 with oveporexton or placebo for eight weeks and measured cognitive performance using several neuropsychological tests, including a psychomotor vigilance task, a continuous paired associate learning test and measures of executive function. Oveporexton improved attention, memory and executive function, with moderate to large effects across several of these measures.

Improving wakefulness is one thing, but improving cognition is potentially something broader. Of course, in somebody with narcolepsy type 1, it is difficult to completely separate the two. Still, this is partly the point, and the cognitive signal appears strong, inviting further research.

Also, we now know from the Phase 3 program that the cognitive findings were not limited to that study. Approximately 70 percent of patients receiving oveporexton reported no significant cognitive difficulties on the cognitive domain of the Functional Impacts of Narcolepsy Instrument, compared with approximately 15 percent receiving placebo. Objective measures of attention, executive function and memory showed improvements as well.

What’s remarkable is that these effects are not limited to narcolepsy type 1 which is characterized by lower orexin levels. Orexin agonists also appear to help patients with narcolepsy type 2, where the biology is less clear and orexin levels can be normal. Earlier work with danavorexton showed this and Alkermes has now reported similar findings with alixorexton, including improvements in wakefulness, cognition and fatigue.

Perhaps enhancing orexin signaling can help with cognition and task engagement across a broad set of conditions, similar to how GLP-1 agonists have broad efficacy across a variety of medical conditions. This could become relevant to ADHD-- after all, stimulants were developed for narcolepsy in the 1930s and later became standard ADHD treatments. So there is a historical parallel.

But there is a wrinkle…

One Wrinkle in the FDA Approval

There is one wrinkle that I did not anticipate when I wrote about this in January. Part of my enthusiasm for orexin agonists was the possibility that they could improve arousal and cognitive function without some of the liabilities that come with stimulants. The FDA label for oveporexton gives me pause.

In a human abuse-potential study in people with experience using CNS stimulants recreationally, oveporexton produced greater drug liking than placebo, with scores similar to or lower than phentermine, which is a Schedule IV stimulant. So there is clearly some abuse potential here. On the other hand, there was limited evidence of physical dependence after the drug was stopped, so exactly how much of a problem this becomes in the real world is something we will have to see.

There are also some side effects that make sense given what the drug is doing. Insomnia occurred in more than half of patients in the Phase 3 studies, although it tended to happen early and nearly two-thirds of cases resolved within a week. Frequent urination was also surprisingly common, also occurring in more than half of treated patients. Neither issue led to particularly high discontinuation rates, but clearly activating orexin is doing things beyond simply making people more cognitively engaged.

Should this temper our enthusiasm regarding orexin agonists being stimulant replacements in some populations? Perhaps, but oveporexton is just one molecule. Other molecules (e.g. Alkermes’ ALKS 7290) may have a different profile altogether.

Also, the question of psychosis remains unanswered. For patients with established psychotic disorders, this will be critical. If (some) orexin agonists can improve cognitive function and task engagement without making psychosis worse, it may address a large unmet public health need.

Looking Ahead

The obvious next thing to watch is Alkermes. They expect the initial proof-of-concept results for ALKS 7290 in ADHD by the end of the third quarter this year. It is a relatively small study, but it will be the first direct test of an orexin agonist in people with ADHD. This will also be a good test for the Ozempic analogy, at least in the sense of orexin agonists having utility beyond narcolepsy.

But maybe this is something that goes beyond psychiatry and is the start of an era of a different class of cognitive enhancers altogether. Modafinil (which is thought to partly work via the orexin system) is rumored to have inspired NZT-48 in Limitless. Maybe next generation orexin agonists will be closer to that.

It is certainly a space I plan to watch closely, and I suspect many of you might do so too.

If you found this analysis useful, please subscribe and share with others thinking about these questions. I’ll continue covering how neuroscience, clinical practice, and drug development intersect as we move toward more mechanistic approaches to psychiatry. Paid subscriptions support this work and help advance its broader aims.

If you’re interested in an expanded discussion regarding the neuroscience and precision psychiatry angles, please check out this post

A fairly common story I hear from families is the following: their child was doing reasonably well, started smoking a lot of cannabis, became paranoid or disorganized, and eventually ended up in the hospital. By the second or third hospitalization, which is when I meet with them, the family has usually reached the conclusion that cannabis triggered schizophrenia.

The legalization of marijuana in many countries, including the United States, has made this issue more timely than ever, with major debates regarding the public health repercussions. Indeed a substantial literature now links (heavy) cannabis use to psychotic disorders in one way or another. But as readers familiar with this newsletter may have come to appreciate, psychosis and schizophrenia are not interchangeable. Psychosis is a term we use to describe a state, where people may hear voices and have unusual beliefs. Schizophrenia is a syndrome that comes with other manifestations such as withdrawal and diminished cognitive resources, leading in some cases to progressive loss of function. Psychosis can occur in other illnesses too, such as bipolar disorder, major depression or can be limited to the periods following substance use.

During the first episode of psychosis, one cannot often tell what the long term illness trajectory will look like. A young person may present with paranoia, disorganization and heavy cannabis use, but only with time does it become clear whether the clinical picture would evolve into a bona fide psychotic illness. Age at onset can help with this distinction because different illness categories follow different patterns across the lifespan. A new study (August 2026) has attempted to investigate its link to cannabis.

The new study

The study in Biological Psychiatry took a broad look at when psychosis begins in people who use cannabis compared with those who do not. “Psychosis” here means whatever each contributing study called it, which was a mix of schizophrenia, affective and ‘unspecified’ psychosis, pooled together across studies. Within that definition, the basic finding is consistent with what has been reported before: cannabis use is associated with an earlier onset of psychosis. A meta-analysis on this exact question was done in 2011 and found cannabis users developed psychosis almost 3 years earlier than nonusers. The current study’s added value is that now it’s looking at this association across the lifespan, dividing it up into time windows. This is the type of more detailed analysis enabled by enough data to support the required statistical power.

The finding was a bit surprising: in studies whose participants generally developed psychosis at younger ages, the difference between cannabis users and nonusers was small. But in studies whose participants developed psychosis at older ages, the difference was larger.

What’s surprising, one could ask? Well, taking the findings at face value, many conclusions would be at odds with the usual way the cannabis-psychosis relationship is discussed. Much of the concern has centered on exposure during adolescence, when the brain is still developing and may be especially vulnerable to the effects of cannabis. The finding of this study doesn’t fit neatly, and one of the conclusions the authors suggested is that longer or more cumulative exposure could bring psychosis earlier in people whose illness would otherwise have emerged later in life.

For readers interested in the numbers: The meta-analysis included 149 studies and more than 70,000 participants. Across the full dataset, cannabis use was associated with psychosis 2.5 years earlier. The size of that difference changed considerably with the mean age of onset in each study. In samples with mean onset before age 20, there was no difference between cannabis users and nonusers. The difference increased to 1.6 years in samples with mean onset between 20 and 24, 4.4 years between 25 and 29, and 6.3 years in samples with mean onset at 30 or older. Mean age at onset remained associated with the size of the cannabis effect after accounting for the other moderators that survived the initial analyses.

Figure 3 of the paper, showing meta-regression of sample mean age at psychosis onset on the standardized mean difference (standard difference in means) of earlier mean age at onset (AAO) of psychosis associated with cannabis use.

What kind of psychosis?

Let’s look at this finding from a slightly different perspective.

As we’ve just discussed, psychosis refers to a broad mixture of conditions, and even when taking DSM categories at face value, the age of onset can be quite distinct across these categories. Schizophrenia, or at least what the bulk of what this term refers to, usually begins in late adolescence or early adulthood, and new cases become less common as people get older. Bipolar disorder (again, given all the diagnostic limitations in separating these illnesses), often begins early as well, but continues to appear across a much broader span of adult life.

What does that mean for this study? One possibility is that the type of patient being sampled across the age bins is not uniform. Even if we say that the proportion of schizophrenia diagnosis is equivalent across bins, we would still be faced with heterogeneity within that construct. To be concrete, readers of this newsletter may recall the piece,

Did John Nash Really Have Schizophrenia?, where a genetics paper suggested that schizophrenia can be thought of as a mixture of a neurodevelopmental variant that overlaps with autism and an affective variant that overlaps with bipolar disorder. So even without a formal bipolar diagnosis, this association could still mean that the bipolar-like variant is more susceptible than the autism-like one.

A large Danish registry study is consistent with this possibility. People with cannabis use disorder are more likely to subsequently receive a diagnosis of bipolar disorder, and the association is particularly strong for bipolar disorder with psychotic features. So perhaps the type of psychosis susceptible to cannabis is distinct from the neurodevelopmental variant regardless of the DSM label (think B-SNIP biotype 2 or 3).

Another Danish registry study followed people who had experienced substance-induced psychosis to see what diagnoses they received over the following years. While cannabis-induced psychosis carried the highest subsequent risk of later schizophrenia compared to any substance, the relationship changed with age. As people got older, an episode of substance-induced psychosis became progressively less predictive of schizophrenia. Bipolar disorder behaved differently, with the association persisting much later in life. This age pattern is reported across substance-induced psychosis broadly, so it may not be cannabis-specific, but it could be telling us something about the general susceptibility of a particular type of psychotic illness to external perturbations.

For readers interested in the numbers: Jefsen and colleagues studied more than 6.6 million people in the Danish national registers. Cannabis use disorder was associated with psychotic bipolar disorder at an adjusted hazard ratio of 4.05. Within bipolar disorder, the association was also significantly stronger for the psychotic than the nonpsychotic form, with a relative hazard ratio of 1.48. Starzer and colleagues followed 6,788 people with substance-induced psychosis. Among those with cannabis-induced psychosis, 41.2% subsequently received a diagnosis of schizophrenia, making cannabis the substance associated with the highest conversion risk. Age had a strong relationship with what happened next. Compared with people aged 16–25 at the time of substance-induced psychosis, the risk of later schizophrenia fell dramatically with increasing age and was lowest in those over 50. Bipolar conversion showed a different pattern and did not decline with age, with the oldest group showing the highest relative risk.

One Caveat

Now, the authors did look at whether the type of DSM diagnosis in each study influenced their main finding. Studies varied in how many participants had schizophrenia versus affective psychosis, and the authors asked whether that difference was related to the size of the cannabis effect. Studies with more affective psychosis did not, on average, show a larger difference in age of onset between cannabis users and nonusers.

At first glance, that looks like evidence against what I’ve laid out before. But “affective psychosis” as tested is a broad category that bundles psychosis associated with bipolar disorder and depression. Cannabis use disorder is associated with both, but in the Danish data psychosis specifically strengthened the association with bipolar disorder, while the association with psychotic depression was no stronger than with depression without psychosis. If the relevant phenotype is therefore something closer to bipolar psychosis rather than affective psychosis broadly, pooling these conditions into one category could dilute the signal before it ever reaches the moderator test. So this null result doesn’t rule out a bipolar explanation for the age result.

Along these lines, a second issue concerns questions that weren’t asked. Does the mixture of diagnostic labels change as the average age of the sample increases? And if it does, does some of the apparent relationship between age and cannabis go away once that changing mixture is taken into account? Those are not quite the same questions as asking whether studies with more affective psychosis have a larger cannabis effect, which is what was asked.

So there are really two versions of the analysis that would be nice to see at some point. First, using a category that isolates bipolar psychosis, does that proportion increase in studies where psychosis begins later? Second, if age and diagnostic composition are considered together, does age still carry the same relationship with the cannabis effect, or does some of it get absorbed by composition? If the age effect holds up under both, the cumulative-exposure interpretation becomes more convincing. If it doesn’t, at least part of what looks like an effect of age may reflect the changing mix of illness represented at different points in the lifespan.

For readers interested in the numbers: The authors tested the proportion of affective psychosis in each sample as a moderator of the cannabis-associated difference in age at onset, and the result was not significant (p = .25). They did the same for the proportion diagnosed with schizophrenia, which was also not significant. Their meta-regression strategy was to test potential moderators individually and then carry the significant ones into a multiple model. Because affective-psychosis proportion was not significant on its own, it was not entered alongside mean age at onset in the final model. As a result, the paper does not tell us whether the age coefficient changes after adjustment for diagnostic composition, or whether there is an interaction between age and the proportion of affective psychosis.

Cannabis, genes and causality

One might hope genetics could help sort some of this out. If people who are genetically more likely to use cannabis are also more likely to develop schizophrenia or bipolar disorder, that would provide a different kind of evidence than simply observing that cannabis use and psychosis occur together. The problem is that the genetic literature has not produced a particularly simple answer.

There is fairly consistent evidence that people with greater genetic liability to schizophrenia or bipolar disorder are also more likely to use cannabis. But that could reflect many things, including personality, cognition, early symptoms or other features of psychiatric vulnerability that make cannabis use more likely. Evidence in the opposite direction, from genetic liability to cannabis use toward psychiatric illness, has been less consistent.

This doesn’t mean cannabis has no causal effect. It points to a more complicated relationship than the usual question of whether cannabis “causes” psychosis. Someone already predisposed to psychiatric illness may be more likely to use cannabis, while cannabis exposure may then alter when the illness appears or what form it takes. Both directions can be true.

For readers interested in the numbers: Mendelian randomization studies (also discussed in the Nash piece) use genetic variants associated with an exposure, such as cannabis use, to ask whether genetically predicted differences in that exposure are associated with later disease. Jefsen and colleagues found evidence that genetic liability to bipolar disorder increased the likelihood of lifetime cannabis use, but did not find evidence for the reverse. Cheng and colleagues later reported evidence compatible with an effect of lifetime cannabis use on bipolar disorder, while also finding substantial shared genetic liability and evidence running from psychotic disorders toward cannabis use. Results for schizophrenia have varied similarly across studies depending on the cannabis phenotype, genetic instrument and analytic approach. Taken together, I don’t see how the genetics settle the argument in either direction. If anything, they reinforce the idea that cannabis use and vulnerability to psychotic illness are intertwined, which makes the answer to the specific question on everyone’s mind more interesting than a simple cannabis causes schizophrenia.

So what does this all mean?

So where does this leave the family sitting across from me? Their intuition that cannabis is relevant may be correct. We currently lack the measurements that would allow us to say what kind of psychotic illness their child has, whether cannabis preferentially perturbs one form of illness over another, or even whether the same relationship applies across patients. These distinctions become clearer with time, but often only after the illness has already declared itself.

For now, the practical conclusion may be much simpler. Once someone has developed psychosis, continued cannabis use is a bad idea. What we might want to resist is turning that sensible clinical advice into a prematurely specific causal story about schizophrenia. The emerging literature may ultimately be telling us as much about the different kinds of psychosis as it is about cannabis itself.

If you enjoyed this piece, please share it with your network. Consider subscribing to join the conversation.

I would like to highlight a piece I recently coauthored with one of my mentees, Nicole Sears NP, in Psychiatric Times. It is a qualitative description of a compelling clinical observation that has emerged from our experience treating more than 100 patients with xanomeline/trospium (XT; Cobenfy) in an acute inpatient setting.

Readers of this newsletter have seen this story develop over time. When I first wrote about our experience with Cobenfy, I described it as a “lightbulb” phenomenon. Meaning, in a subset of patients, often with prominent negative symptoms, the response seemed qualitatively different from what I was accustomed to seeing with conventional antipsychotics. Patients who had appeared profoundly withdrawn and disengaged became more socially connected, spontaneous and thoughtful. In some cases, the change occurred over just a few days.

I subsequently tried to quantify some of this in a paper that was later published in Nature Mental Health. The upshot is that negative symptoms were the strongest predictor to an XT response. It’s important to note that a response here isn’t necessarily limited to an improvement in negative symptoms, but it’s an aggregate clinical impression of the patient’s status, which would include things like positive symptoms burden and overall function.

That work was interesting to me in the context of precision psychiatry, because many patients showed no clear response within the window of evaluation. I have argued previously that one of the most useful ways to subdivide psychiatric illness may be through treatment response itself. If patients who look similar respond differently to medications, those differences may tell us something about the underlying structure of their illness (of note, this is not always true, as there could be simpler explanations related to how the particular drug is metabolized in different people, so this is a caveat). Nonetheless, the arrival of Cobenfy makes the potential for precision psychiatry especially interesting because it provides a unique pharmacological ‘knob’ and potentially help with stratification or subtyping.

But the quantitative paper could only capture part of what was observed. In a subset of responders, there is a striking change that several clinicians have independently described as a Cobenfy “clearing” or “awakening.” I think what everyone means by that is that, some patients become more socially engaged and spontaneous, but there can also be an apparent change in how clearly they think and reason. Although not exactly the same, the closest thing that comes to mind is Awakenings by Oliver Sacks, which was made into a feature film by Robert De Niro and Robin Williams.

I recommend watching this movie. It’s good. I also recommend reading the piece, which I pasted below for convenience.

Experience With Xanomeline/Trospium in an Acute Inpatient Setting Treating Individuals With Schizophrenia

Author(s)Nicole C. Sears, NP, Michael M. Halassa, MD, PhD

Key Takeaways

• FDA approval in late 2024 established xanomeline/trospium as the first nondopaminergic schizophrenia treatment, leveraging M1/M4-preferring muscarinic agonism to address symptom domains refractory to D2 blockade.

• Quantitative inpatient experience identified negative symptoms and stimulant-use history as response predictors, whereas intellectual disability signaled lower likelihood of benefit, supporting phenotype-stratified prescribing.

Real‑world inpatient use of xanomeline/trospium shows rapid awakening from negative symptoms in schizophrenia, with response predictors and striking case insights.

Anthony Mujica Viera/AdobeStock

Over the past year, we have treated over 100 patients with xanomeline/trospium (Cobenfy; formerly known as KarXT, referred to as XT in this article) in our acute inpatient setting, a large community hospital predominantly serving individuals with chronic and treatment-resistant conditions. Here is what we found.

History

The US Food and Drug Administration approval of XT in late 2024 introduced the first nondopaminergic treatment for schizophrenia. As an M₁/M₄-preferring muscarinic receptor agonist, this medication offers a mechanistically distinct approach to managing psychotic disorders. In theory, a different mechanism could be particularly useful for patients who show inadequate responses to traditional dopamine receptor 2 (D₂)–targeting antipsychotics.

The findings from 2 phase 3 EMERGENT trials of XT demonstrated significant reductions in overall symptoms of schizophrenia as well as major domains, including positive and negative symptoms,^1,2 but these results, although promising, represent group data that limit understanding of the degree to which XT has differential efficacy. As we await more definitive evidence-based data, we report here on our clinical experience prescribing XT in an acute psychiatric inpatient setting to share our observations.

Findings

Quantitative analysis of the clinical experience in our acute inpatient setting identified specific predictors of response.³ Patients with prominent negative symptoms showed the most robust improvements, as did those with histories of stimulant use. Conversely, patients with intellectual disability demonstrated markedly reduced responsiveness. Although the last observation should be interpreted with care given the low number of patients in this category, these findings generally align with emerging evidence that psychotic disorders represent biologically heterogeneous conditions requiring targeted, phenotype-specific interventions.⁴

The quantitative analysis involved identifying patients who responded to XT vs those who did not, based on routine clinical assessment. When we initially started administering XT, we were guided by 3 criteria: (1) residual psychotic symptoms despite adequate treatment with traditional antipsychotics, (2) prominent negative symptoms unlikely to respond to traditional antipsychotics, and (3) intolerance to traditional antipsychotics (in the form of extrapyramidal symptoms, excessive sedation, etc). Among such patients, approximately 40% were responsive.³ Importantly, enriching a subsequent cohort on the basis of identified features resulted in near doubling of response rate, highlighting the potential utility of such a precision stratification approach.³

Here, we build on this foundation and report on our qualitative experience using an illustrative case report. Secondary analysis of the three 5-week EMERGENT trials reported that, in a subset of patients, XT demonstrated a significant impact on cognitive symptoms, particularly when looking at patients with cognitive performance at pretreatment baseline more than 1 SD below the mean.^5,6 We have directly observed this in our patients and discussed similar anecdotal reports with several colleagues: a clinical phenomenon referred to as XT clearing or awakening. This pattern is characterized by unusually rapid emergence from profound negative symptoms in patients who were withdrawn, had flat affect, and had cognitively dulling who suddenly demonstrated social engagement, future-oriented thinking, and restored spontaneity shortly after the institution of XT. We estimate this to be more than 50% of patients who had positive responses to XT treatment in our selected cohort. The speed and quality of these responses suggest XT may be revealing preserved cognitive and social capacities that had been masked by illness (and/or by D₂ receptor blockade) rather than gradually restoring lost function.

FIGURE 1. XT Response Rate: Broad Criteria vs Precision Stratification

Periods of clinical improvement were observed in the 1990s with the introduction of second-generation agents.^7,8 Weiden and colleagues described improved functional outcomes and the emergence of psychological adjustment reactions as patients stabilized on newer medications. These reports emphasized the need for clinicians to anticipate emotional and behavioral changes accompanying recovery. However, they did not characterize a consistent pattern of rapid, within-days restoration of social engagement and goal-directed thinking in patients who were severely withdrawn. Our observations differ in the speed and qualitative nature of the response, occurring shortly after initiation of XT and often in individuals who had had failure with multiple dopamine-targeting treatments. Perhaps the closest qualitative pattern to our observation is what was depicted by Oliver Sacks’ work with levodopa.⁹

Case Example: Emergence and Resolution of Negative Symptoms

“Mark,” a 34-year-old man with a clinical diagnosis of chronic schizophrenia and prominent negative symptoms, was admitted to our inpatient unit following multiple failed trials of second-generation antipsychotics. During this hospitalization, he was initiated on clozapine due to persistent functional withdrawal and limited response to prior treatments. At baseline, he was socially isolated and minimally engaged, requiring prompting for basic activities of daily living and spending most of his time alone in his room or watching television.

Over the subsequent 2 weeks of hospitalization, the clinical picture deteriorated markedly. Despite trials of multiple antipsychotic combinations, Mark became increasingly withdrawn. He would not leave his room and refused to shower or engage with staff. When nursing staff attempted engagement, his responses became monosyllabic, at most 3 to 4 words. By this point, he had received olanzapine to address acute decompensation upon admission and clozapine for more effective symptom management, but without benefit.

Three weeks into his admission, we decided to simplify his regimen and switch over to a long-acting injectable formulation given concerns about nonadherence. We therefore cross-titrated olanzapine to risperidone 2 mg twice daily and decreased clozapine to 150 mg daily. This process took a week, after which we were convinced that Mark’s negative symptoms and overall functional status had not been touched by any of these medication changes.

Therefore, at that point and while he continued risperidone 2 mg twice daily and clozapine 150 mg daily, we added XT at the starting dose of 50 mg/20 mg twice daily, instructing nursing staff to make sure it was given outside of meals. We are fully aware that XT is recommended as monotherapy and that its combination with clozapine and risperidone is considered off-label. However, given the gravity of the situation, the trade-off appeared favorable and was thoroughly discussed with the patient to ensure informed consent.

Mark’s medication adherence remained inconsistent over the next day but stabilized by the second day. By the third morning, nursing staff immediately noted a dramatic transformation. He was spontaneously out of his room, appropriately dressed and groomed. Most strikingly, he agreed to meet with his treatment team, something he had refused for weeks. He appeared fundamentally different: He made direct eye contact, smiled, and engaged in conversation. The staff repeatedly asked us, “What did you do?” The contrast was stark; a patient who had been on the verge of requiring physical restraint, unable to pull himself from bed, was now pleasant, cooperative, and able to advocate for his needs.

FIGURE 2. Case Timeline: “Mark” and the XT “Awakening” Response

The quality of the conversation was remarkable. His thought process was clear and linear. He spontaneously explained that he had difficulty attending groups because he was claustrophobic and disliked feeling closed in with many people, an insight that demonstrated both self-awareness and capacity for abstract reasoning. When asked about his stated goal to “get a hobby, maybe a job,” he elaborated: He was considering working at a grocery store, bagging groceries or gathering carts, because he thought he would do well in that environment without feeling overwhelmed. This represented future-oriented thinking in the form of practical, grounded planning, which was completely absent just days earlier.

The response was sustained. He was continued on XT, increased to 100 mg/20 mg twice daily, with no reported adverse effects. He was discharged 1 week later. This case exemplifies the awakening phenotype: rapid emergence from profound negative symptoms, revealing preserved cognitive and social capacities that had been masked by illness. There is also the temporal relationship, deterioration despite adequate antipsychotic trials, then rapid response within 72 hours of consistent XT administration, which we had seen in many of the other cases. This pattern strongly suggested a specific effect on negative and possibly cognitive symptoms rather than simple resolution of positive symptoms.

Discussion

This case illustrates XT’s capacity to produce rapid, clinically meaningful improvements in patients with prominent negative symptoms, a domain historically resistant to pharmacologic intervention. The awakening phenotype we observed shares key features: rapid onset (typically 48-96 hours), transformation in social engagement and spontaneity, and emergence of future-oriented cognition.

These observations must be interpreted within limitations. Our cases represent select patients in an acute inpatient setting where negative symptoms were sufficiently prominent to warrant an XT trial. The concurrent use of other medications, although reflecting real-world practice, complicates attribution of specific effects. Nonetheless, the consistency of the awakening pattern across multiple patients, the rapid temporal relationship to XT initiation, and the alignment with our larger quantitative analysis strengthen confidence in these observations.

Concluding Thoughts

In our experience, XT represents a meaningful advance for patients with negative symptom–predominant psychotic disorders. The awakening phenotype—rapid restoration of social engagement, spontaneity, and goal-directed thinking—suggests this medication can reveal preserved capacities in patients who appeared treatment resistant to D₂ receptor antagonist approaches. Despite all the limitations of case reports and clinician observation, we feel it is very important to share this experience, which at this time offers a qualitative experience that was beyond the scope of the phase 3 clinical trials.

Ms Sears is a psychiatric-mental health nurse practitioner in the Psychosis Disorders Unit at Fuller Hospital in Attleboro, Massachusetts.

Dr Halassa is a professor at the Fralin Biomedical Research Institute at Virginia Tech, with appointments in psychiatry and behavioral medicine at the Virginia Tech Carilion (VTC) School of Medicine and in biomedical engineering at the College of Engineering. He is the inaugural faculty member of the Virginia Tech Patient Research Center at VTC, where he conducts patient-based research alongside his laboratory program. He is also a board-certified psychiatrist and maintains an active clinical practice treating patients with severe psychotic disorders.

References

1. Kaul I, Sawchak S, Correll CU, et al. Efficacy and safety of the muscarinic receptor agonist KarXT (xanomeline-trospium) in schizophrenia (EMERGENT-2) in the USA: results from a randomised, double-blind, placebo-controlled, flexible-dose phase 3 trial. Lancet. 2024;403(10422):160-170.

2. Kaul I, Sawchak S, Walling DP, et al. Efficacy and safety of xanomeline-trospium chloride in schizophrenia: a randomized clinical trial. JAMA Psychiatry. 2024;81(8):749-756.

3. Halassa MM. Preliminary real-world predictors of response to muscarinic targeting in psychosis. Nat Ment Health. 2025;3:1512-1518.

4. Halassa MM, Frank MJ, Garety P, et al. Developing algorithmic psychiatry via multi-level spanning computational models. Cell Rep Med. 2025;6(5):102094.

5. Sauder C, Allen LA, Baker E, et al. Effectiveness of KarXT (xanomeline-trospium) for cognitive impairment in schizophrenia: post hoc analyses from a randomised, double-blind, placebo-controlled phase 2 study. Transl Psychiatry. 2022;12(1):491.

6. Horan WP, Sauder C, Harvey PD, et al. The impact of xanomeline and trospium chloride on cognitive impairment in acute schizophrenia: replication in pooled data from two phase 3 trials. Am J Psychiatry. 2025;182(3):297-306.

7. Weiden PJ, Aquila R, Standard J. Atypical antipsychotic drugs and long-term outcome in schizophrenia. J Clin Psychiatry. 1996;57(suppl 11):53-60.

8. Weiden P, Scheifler P, Diamond R. My patient is better: now what? part I: managing psychological reactions. J Psychiatr Pract. 1998;4:175-181.

9. Sacks O. Awakenings. Gerald Duckworth & Co Ltd; 1973.

This observation has also become more interesting in light of what is happening with the broader muscarinic drug class. As I discussed recently in The Dawn of the Muscarinics, MapLight’s M1/M4 program has now produced an intriguing cognitive signal that resembles what had previously been observed with KarXT. In both programs, cognitive improvement appears at least partly separable from improvement in psychotic symptoms.

Whether those controlled cognitive findings have anything to do with the clinical “clearing” we are seeing remains to be determined. But it’s something the field should be pursuing. If a recognizable clinical and cognitive phenotype preferentially responds to muscarinic treatment, then treatment response itself can begin to reveal clinically meaningful subtypes within schizophrenia.

This is where the opportunity may be particularly interesting. We now have a new pharmacological knob to turn, and potentially several more muscarinic compounds coming behind it. Rather than asking only whether these drugs work on average, we can ask which patients respond and what exactly changes when they do, and whether those patterns generalize across compounds targeting the same system.

If they do, the “awakening” may turn out to be more than an unusual clinical observation. It may be an early glimpse of how new drug mechanisms can help us carve schizophrenia at its biological joints. That would move precision psychiatry for psychotic disorders from describing heterogeneity to engineering new treatment pathways.

If you liked this piece, please consider liking it and sharing it with your network. To become part of the conversation, you can hit the subscribe button.

Every few years, a psychiatric drug arrives promising to be the answer to whatever society has been missing. For the millions of patients with ADHD, a viable alternative to stimulant medications, which include amphetamine and methylphenidate, is an unmet need. Stimulants are first-line treatments, and have the greatest clinical efficacy. Yet, for many patients they are poorly tolerated , are inadequate for symptom control or too burdensome given their controlled-substance status.

On July 24th 2026, the FDA approved centanafadine (Brand name Simtriyo). It is thought to be a triple reuptake inhibitor (norepinephrine-dopamine-serotonin), and is marketed as a first in class medication based on that profile.

As I have written repeatedly throughout this newsletter, we absolutely need nonstimulant alternatives for ADHD treatment. Beyond the issues mentioned above, stimulants have a considerable abuse potential and may precipitate psychotic and/or manic symptoms in susceptible individuals.

Yet, despite 15 years of work setting up centanafadine as a nonstimulant, the FDA just labeled it as a stimulant! Let’s dig into what this means for patients, families, providers and people watching drug development in this space...

Where This Drug Came From

Centanafadine is a triple reuptake inhibitor. The idea of a reuptake inhibitor is not new to psychiatry, it underlies a large share of molecules thought to alter the availability of neurochemicals in the brain. For example, selective serotonin reuptake inhibitors or SSRIs are used for the treatment of depression and obsessive compulsive disorder. Double reuptake inhibitors, which additionally increase norepinephrine (SNRIs), are also used for depression among other indications. Even stimulants act as reuptake inhibitors, but for dopamine. Increasing the effective action of dopaminergic signaling is thought to underlie their therapeutic effects, as well as their abuse potential and their pro-psychotic effects.

Schematic showing SSRI action. From this article

The idea behind triple reuptake inhibitor started with the notion that SNRIs may be insufficient for treating symptoms that would otherwise benefit from boosting dopamine. Bupropion (Wellbutrin) is thought to boost dopamine, for example, and is thought to speed up the response to standard SSRI treatment, as does off-label use of stimulants in depression. Drug developers reasoned that combining all three targets into a single molecule might capture that same benefit without needing multiple prescriptions.

That reasoning produced an entire family of triple reuptake inhibitors, developed mostly for depression. Centanafadine is a member of that family, and its closest relative, amitifadine, was advanced towards FDA approval as a depression treatment, but never got there. Because centanafadine was engineered more toward norepinephrine and dopamine and less toward serotonin, it was redirected towards ADHD. Preclinical evidence from animal studies justified that switch. Centanafadine reduced hyperactivity in a rodent model of ADHD, and it did so without causing the excess movement in normal rodents that classic stimulants produce at comparable doses. Centanafadine also showed antidepressant-like activity in a standard rodent despair test, so it was thought to be potentially useful for the mood symptoms that commonly co-occur with ADHD. Thus, the preclinical evidence base moved centanafadine from a depression-focused side project into a dedicated ADHD development program. (I have my own qualms about these types of ‘preclinical’ rodent studies but let’s not go there right now).

For readers who want the corporate history: In 1998, DOV Pharmaceutical licensed the original triple reuptake inhibitor prototype from Wyeth and spent the next decade developing it for depression, partnering with Merck in 2004. After a string of failed depression trials, DOV was acquired in 2010 by Euthymics Bioscience. Euthymics inherited DOV’s whole pipeline, including both amitifadine, renamed EB-1010, for depression and centanafadine, originally EB-1020, for ADHD. The two compounds share essentially the same core chemical structure with only the attached side group different. From its first public unveiling in 2011, Euthymics described EB-1020 explicitly as a triple reuptake inhibitor optimized for ADHD “without drug abuse liability.” Centanafadine was later spun into a new company, Neurovance, which Otsuka Pharmaceutical acquired in 2017 for $100 million upfront and up to $150 million more in milestones, the money that funded the Phase 3 program that finally reached approval this year.

Is It a Non-Stimulant?

Stimulants are thought to work by increasing dopamine and norepinephrine, largely by blocking their reuptake and, in the case of amphetamines, by directly triggering extra release. This is also thought to explain their abuse potential: a fast, large rise in dopamine, particularly in the brain’s reward circuitry, is thought to make a drug reinforcing and prone to misuse.

Non-stimulant ADHD medications were developed specifically to sidestep this problem. Atomoxetine and guanfacine take different routes: atomoxetine is a norepinephrine reuptake inhibitor and guanfacine is a direct norepinephrine receptor agonist (alpha 2A receptors for people who are interested). Neither is thought to produce the rapid rise in reward-circuit dopamine associated with conventional stimulants. Centanafadine is thought to raise dopamine levels, but modestly and slowly through its extended-release formulation, with the idea that it would help ADHD symptoms without abuse potential.

Unfortunately, the regulatory abuse-potential studies were a bit more complicated. Regulators require abuse-potential testing using a fast-acting version of a drug because the concern is that some people might crush or chew the pill, defeating the extended-release mechanism. So the studies gave people with a history of misusing stimulants that fast-acting version of centanafadine and asked them to rate how much they liked it, a standard test for abuse potential. At the higher doses tested, the studies failed to show that participants liked it less than comparator doses of amphetamine and lisdexamfetamine. At lower doses, they liked it less than amphetamine, though still more than placebo. On the strength of this evidence, the FDA classified Simtriyo as a CNS stimulant and added a boxed warning for abuse, misuse, and addiction. Whether regulators ultimately treat it like a full-strength stimulant or place it in a less restrictive category is still undecided, and DEA scheduling will settle that question.

For readers who want the detail: The two human abuse-potential studies submitted to the FDA tested an immediate-release formulation, not the extended-release Simtriyo product that reaches patients; regulators generally require testing a drug’s fastest-acting form to establish worst-case abuse potential. In the first study, doses of 400 and 800 milligrams produced Drug Liking scores that were not statistically significantly less than 40 milligrams of amphetamine; the 800 milligram dose was also not statistically significantly less than 150 milligrams of lisdexamfetamine, a comparison that applied only at that higher dose. In the second study, doses of 200, 400, and 800 milligrams produced Drug Liking scores lower than 15 and 30 milligrams of amphetamine, though still higher than placebo. DEA scheduling has not been decided. Jefferies, the investment bank that covers Otsuka, has called scheduling the single biggest factor in whether the drug can compete commercially; a Schedule II classification, the same category as Adderall, might create a problem. Separately, in the two adult Phase 3 efficacy trials, patients who stopped taking centanafadine abruptly showed almost no withdrawal, which is good news.

What About Efficacy?

Effect size is the standard way trials measure how well a medication works against placebo. Centanafadine’s numbers follow a pattern familiar to anyone who has watched drug development closely: a strong result in an early, smaller trial, followed by a smaller effect once a larger, confirmatory trial was run. This is the same shrinkage we discussed in the last piece of this newsletter.

Centanafadine delivered a modest effect size in its confirmatory trials, on the low end of what is already available to adults with ADHD. Stimulants remain the most effective option by a clear margin, and atomoxetine shows a larger effect too. Guanfacine is not even approved for adult ADHD, so any comparison there is indirect. Nonetheless, centanafadine appears to have a relatively rapid onset. In one of the two confirmatory adult trials, it separated from placebo by week one. That is an interesting, if modest, potential advantage over atomoxetine, whose effects are slower.

Least-squares mean change from Baseline to day 42 in AISRS total score (primary endpoint). From this study

For readers who want the numbers: Neurovance’s original Phase 2b trial found an effect size of 0.66 on the ADHD Rating Scale at week three. The confirmatory adult Phase 3 program, run under Otsuka, produced effect sizes of 0.24 to 0.40 across the two studies and two doses tested. The best available network meta-analysis of ADHD medications puts adult, clinician-rated effect sizes at 0.79 for amphetamines, 0.49 for methylphenidate, and 0.45 for atomoxetine (Cortese et al., 2018). Atomoxetine can show an early signal but usually takes six to twelve weeks to reach its full effect, compared with centanafadine’s week-one separation from placebo.

Beyond Core Symptoms

Otsuka has also promoted a secondary finding from the Phase 3 program. The company pooled data from the two adult trials, 744 patients in total, split across centanafadine 200 milligrams, 400 milligrams, and placebo, and examined two patient-reported measures at week six: the Executive Functioning subscale and the Emotional Dyscontrol subscale of the Adult ADHD Self-Report Scale. On executive function, patients on centanafadine reported fewer difficulties with time management, planning and prioritization, task initiation and completion, and working memory, compared with placebo. On emotional dysregulation, they reported improvement on items covering emotional overactivity, affective lability, and anger outbursts.

I am not sure how much this differentiates centanafadine from what already exists (but someone can certainly correct me if I’m missing something). Guanfacine was developed and is already marketed on the premise that it acts on prefrontal, executive-function circuitry; a placebo-controlled crossover trial in children found that adding guanfacine to an existing stimulant significantly improved parent-rated executive function on the Behavior Rating Inventory of Executive Function, and adult trials have reported improvement on some BRIEF-A subscales as well. Atomoxetine has multiple placebo-controlled trials showing significant benefit on standard executive function measures. And a network meta-analysis in The Lancet Psychiatry found that both stimulants and atomoxetine improved emotional dysregulation compared with placebo at around twelve weeks.

Improvement in executive function and emotional dysregulation, in other words, already shows up across multiple existing ADHD drug classes. The available meta-analytic evidence suggests these effects are typically small to moderate and vary depending on which drug and which domain gets measured. But will there be a precision psychiatry angle here? Time will tell and I will be looking forward to whether that is something Otsuka (or others) will pursue.

What’s Left

There maybe an opportunity underneath all of this. Non-stimulant prescriptions made up only about ten percent of everything dispensed for ADHD in 2023, even after more than doubling since 2012. That is real room to grow.

The problem for new medications is that this space is already somewhat crowded, and there has got to be a clear added value for anything new. Generic atomoxetine, guanfacine, and viloxazine, sold as Qelbree, are all non-controlled substances with real-world track records. Qelbree alone has filled roughly three million prescriptions since its 2021 approval and keeps growing, and none of these drugs carry an abuse-potential boxed warning. Centanafadine’s effect sizes do not clearly beat any of them, and depending on how DEA scheduling lands, it may end up in a worse position than all three on the exact dimension, avoiding a controlled substance, that pushes most patients toward non-stimulants in the first place.

What is left may be more limited in scope: once-daily dosing, fast onset of action, and a different drug for patients who have already tried the existing non-stimulants and failed them for reasons that have nothing to do with abuse risk. That may be a niche, and while not “the next chapter beyond stimulants,” a particular subset of patients may find it beneficial.

If you enjoyed this piece, please consider hitting the like button. If you’re a non-subscriber, please consider subscribing (and commenting) to keep this conversation alive.

On July 27th, MapLight Therapeutics released topline* results from ZEPHYR, a Phase 2 trial testing a new schizophrenia medication called ML-007C-MA. The trial met its primary endpoint, and the twice-daily dose improved psychotic symptoms compared with placebo. The more interesting result, at least for me, was the prespecified cognitive endpoint. In a subset of patients with cognitive impairment, performance improved with treatment, and the size of that improvement was in the same range as what had previously been reported with KarXT.

I have argued in previous pieces that schizophrenia is a much broader illness than psychosis. Hallucinations and delusions often bring someone into a hospital and are the symptoms traditional antipsychotic medications treat most reliably. Cognitive and negative symptoms tend to determine whether that person returns to school, holds a job, maintains relationships and lives independently.

Traditional antipsychotic medications are thought to act via blocking of dopamine D2 receptors. KarXT, now marketed as Cobenfy, was historically important because it was the first muscarinic agent to be approved for treatment of adults with schizophrenia. More specifically, KarXT is thought to activate the M1 and M4 subtypes of muscarinic receptors in the brain, which are thought to impact all symptom domains in schizophrenia. The ZEPHYR result now provides another opportunity to ask whether this class may indeed have such broad efficacy.

There is also a commercial question, which is whether MapLight has produced a better version of KarXT. MapLight’s investor presentation encourages that comparison by placing ZEPHYR beside KarXT’s Phase 3 trials and emphasizing differences in tolerability, fasting requirements and titration. A phase-matched comparison gives a more accurate picture. When ZEPHYR is placed beside KarXT’s original Phase 2 trial, KarXT produces a larger antipsychotic effect and a more favorable adverse-event profile. The current evidence therefore gives KarXT the stronger overall clinical position, while the cognitive result from ZEPHYR strengthens the scientific case for the broader M1/M4 approach.

What Happened in ZEPHYR

For families following the development of a new schizophrenia medication, the practical questions usually concern the size of the clinical improvement, the kinds of symptoms that changed and the likelihood that patients will remain on the drug over time. ZEPHYR offers encouraging information in each area, along with several reasons for restraint.

The trial enrolled 307 adults with an acute exacerbation of schizophrenia and assigned them to placebo, ML-007C-MA 210/3 mg twice daily, or ML-007C-MA 330/6 mg once daily. After five weeks, the twice-daily dose produced a statistically significant improvement over placebo on the Positive and Negative Syndrome Scale, or PANSS, which is the standard symptom rating scale used in acute schizophrenia trials. The once-daily arm produced a smaller numerical improvement that remained below the prespecified statistical threshold.

The dosing result has practical implications. Once-daily medications are generally easier to incorporate into ordinary life, particularly for people whose illness already affects organization, memory and routine. Whichever regimen MapLight ultimately advances, the regimen actually supported by ZEPHYR’s data is the twice-daily one.

Now, let’s turn to the cognitive finding. A large proportion of people with schizophrenia have difficulties with attention, processing speed, working memory and the organization of thought. These difficulties are strongly associated with long-term functional outcomes. In ZEPHYR, participants who began the study with measurable cognitive impairment improved on a cognitive battery when treated with the twice-daily dose. The degree of cognitive improvement showed little relationship to the degree of improvement in their psychotic symptoms, which supports the possibility that M1/M4 agonists improve cognition in psychotic disorders independent of psychosis itself.

A reproducible cognitive benefit from two different muscarinic agents is potentially a big deal.

For readers who want the underlying numbers: ZEPHYR enrolled 307 participants across 25 U.S. sites, randomized 1:1:1 to placebo, ML-007C-MA 210/3 mg twice daily, or ML-007C-MA 330/6 mg once daily over a five-week blinded treatment period. On the primary endpoint, change in PANSS Total Score at Week 5, the twice-daily arm produced an effect size of 0.37 in the modified intent-to-treat analysis, corresponding to a 4.5-point advantage over placebo with p=0.015. The prespecified completer analysis yielded an effect size of 0.50 and a 6.0-point difference with p=0.002. The once-daily arm produced an effect size of 0.23 with p=0.110. Secondary measures in the twice-daily arm included an effect size of 0.48 on the Clinical Global Impression of Severity and 0.39 on the PANSS positive-symptom Marder factor. Among participants with baseline cognitive impairment, the Cogstate composite yielded an effect size of 0.51 with p=0.041. Cognitive change and PANSS change were uncorrelated in the active arm, r=-0.187 with p=0.314. Serious or drug-related severe adverse events were absent, gastrointestinal adverse events led to discontinuation in 2.0 percent of participants receiving the twice-daily dose, and the trial showed no clear metabolic, hepatic, urinary-retention or movement-related signal. The formulation also carries no fasting requirement.

The Muscarinics

KarXT was approved in September 2024 and became the first antipsychotic in more than seventy years whose primary mechanism does not involve direct dopamine D2 receptor blockade. It was the first muscarinic.

The development history of other muscarinic drugs has raised the question of whether both M1 and M4 activation is needed for clinical efficacy. Emraclidine, an M4-selective positive allosteric modulator, produced promising early results and then missed the primary endpoint in two Phase 2 trials. Direclidine, an M4-selective agonist, produced a statistically significant effect at the lowest of four doses in its Phase 2 study, with progressively less convincing results at the higher doses (strange, huh?). Like KarXT, ML-007C-MA is a dual M1/M4 agonist, and that appears to have worked!

Several interpretations remain possible. Dual M1/M4 activation may provide a more reliable antipsychotic profile than M4-selective approaches. M1 activity may contribute primarily to cognition while M4 accounts for more of the effect on psychosis. The dose-response properties of these receptors may also be unusually sensitive to exposure, timing and receptor desensitization, which could help explain the U-shaped result with direclidine and the separation between MapLight’s two regimens. A small number of trials leaves considerable uncertainty around each explanation.

The accumulating data still point toward a meaningful distinction between M1/M4 agonism and M4-selective strategies. KarXT has produced positive results across three acute trials. ML-007C-MA has now produced a positive Phase 2 result. Both compounds have also generated cognitive signals in participants who began treatment with cognitive impairment. Emraclidine failed in two Phase 2 trials, while direclidine’s result was concentrated at one dose. This pattern gives the dual-agonist hypothesis increasing plausibility, even though the separate contributions of the two receptors remain unresolved.

For readers who want more mechanistic detail: M4 receptors are expressed prominently in striatal circuits that regulate dopamine release, providing an indirect path toward control of psychotic symptoms. M1 receptors are heavily expressed in prefrontal networks, potentially impacting negative and cognitive symptoms. KarXT and ML-007C-MA stimulate both receptor types in the brain while pairing the central agonist with a peripherally acting anticholinergic intended to limit gastrointestinal and other peripheral cholinergic effects. Emraclidine modulates M4 selectively, and direclidine is an M4-selective agonist. AbbVie’s two EMPOWER trials of emraclidine missed their primary endpoints in November 2024. Direclidine’s Phase 2 trial tested four doses and produced statistical significance at the lowest dose, creating a U-shaped dose-response pattern. The field currently has evidence for multiple pharmacological routes into the muscarinic system, along with a limited ability to determine which receptor profile and exposure pattern will prove most reliable.

A Virtual Comparison With KarXT

Any comparison between ML-007C-MA and KarXT is necessarily indirect. The drugs have never been tested against one another in the same trial, and differences in patients, sites, placebo response, analytical decisions and study conduct can influence the apparent size of an effect. Cross-trial comparisons are still informative when studies are aligned as closely as possible.

MapLight’s presentation compares ZEPHYR with KarXT’s Phase 3 program, where the effect sizes were smaller than in KarXT’s original Phase 2 study and the rate of treatment-emergent adverse events was higher. This makes ML-007C-MA appear closer to KarXT on efficacy and more favorable on tolerability. The more comparable study is EMERGENT-1, KarXT’s five-week Phase 2 trial in 182 hospitalized patients with acute schizophrenia.

The efficacy difference is substantial. The twice-daily dose of ML-007C-MA produced an effect size of 0.37 in the primary modified intent-to-treat analysis. EMERGENT-1 produced an effect size reported as 0.75 in the trial’s original description and as 0.81 in a later pooled reanalysis, roughly double to more than double ZEPHYR’s number depending on which figure is used. KarXT’s effect subsequently settled at 0.61 and 0.60 in its two Phase 3 trials, with a pooled effect size of 0.65 across all three studies (Kaul et al., 2024).

Early clinical trials often generate larger estimates than later confirmatory studies, particularly when the sample is modest and the placebo response happens to be favorable, a pattern sometimes called the winner’s curse. KarXT followed that familiar trajectory, shrinking from roughly 0.75 to 0.81 in Phase 2 down to roughly 0.60 to 0.65 by the time its Phase 3 program was complete. ZEPHYR begins from a more modest estimate than KarXT’s own Phase 2 result, and if MapLight’s drug follows a similar trajectory, a future confirmatory trial is arguably more likely to land below 0.37. A successful Phase 3 program could still confirm a clinically useful antipsychotic effect, but the current data provide little basis for describing ML-007C-MA as more efficacious than KarXT.

Tolerability also looks different after phase matching. In EMERGENT-1, 53.9 percent of participants receiving KarXT and 43.3 percent receiving placebo reported a treatment-emergent adverse event, giving an active-placebo difference of approximately eleven percentage points. In ZEPHYR, those rates were 74.7 percent for the twice-daily ML-007C-MA arm and 48.1 percent for placebo, a difference of approximately twenty-seven percentage points, roughly two and a half times as large. Adverse events led to discontinuation in 3.3 percent of participants receiving KarXT in EMERGENT-1 and 7.1 percent receiving the twice-daily dose of ML-007C-MA in ZEPHYR (Correll et al., 2022).

These comparisons deserve the same caution as the efficacy comparison. Adverse-event collection and classification can vary among trials, and the severity and clinical consequences of the events may carry more practical meaning than the raw percentage of patients reporting any event. Most adverse events in both studies were mild or moderate and reflected the expected cholinergic effects of the drugs. ML-007C-MA also offers potential practical advantages through its lack of a fasting requirement and simpler titration.

The overall picture still favors KarXT on the evidence available today. It produced a larger effect at the same stage of development, maintained a clinically meaningful effect through two Phase 3 trials, and had a more favorable phase-matched adverse-event profile. ML-007C-MA has demonstrated enough efficacy to justify further development, and its eventual position will depend on whether the ZEPHYR result replicates in a confirmatory trial and whether its practical formulation advantages translate into better persistence in ordinary clinical use.

The Cognitive Signal Has Now Been Replicated

In 2022, researchers published a post hoc analysis of cognitive testing from EMERGENT-1 (Sauder et al., 2022). KarXT had little detectable cognitive effect across the full trial population, which included patients spanning a wide range of baseline performance. Among the participants who entered the study with clinically meaningful cognitive impairment, KarXT produced a moderate improvement compared with placebo. That improvement showed little relationship to the change in psychotic symptoms.

The result initially carried the familiar limitations of a post hoc subgroup analysis. The subgroup was small, the threshold for impairment was applied after the main trial had been completed, and participants selected for low baseline scores tend to improve partly through regression toward the mean. A later pooled analysis across the EMERGENT trials used a broader battery and a larger sample and found a comparable effect size of 0.54 (Horan et al., 2025).

ZEPHYR now extends the finding across molecules. MapLight prespecified cognitive performance as a secondary endpoint in participants with baseline cognitive impairment. The twice-daily dose produced an effect size of 0.51, close to the original KarXT estimate of 0.50 and the later pooled estimate of 0.54. Cognitive improvement again showed little association with improvement in psychotic symptoms.

The convergence is striking. KarXT produced the signal in an exploratory Phase 2 analysis and then reproduced a similar effect in pooled trial data using a broader battery. ML-007C-MA produced almost the same effect size in a prespecified analysis of a different Phase 2 trial. That two independent teams, years apart, converged on such a similar number makes the muscarinic-cognition link highly plausible.

From Drug Response to Precision Psychiatry

There is another way to interpret these results that connects to the problem of precision psychiatry.

Cognitive impairment by itself remains a broad feature. Two patients can receive similarly low composite scores for very different reasons. Only tasks amenable to computational dissection would be capable of revealing the underlying impairments.

This is where the drug-development and measurement problems begin to converge. Future trials could use tasks designed to separate these cognitive processes and determine which component changes with treatment. A reliable association between a particular computational parameter and response to an M1/M4 agonist would move the field closer to a treatment-relevant axis, with the medication serving as both an intervention and a probe of the illness structure.

The replicated cognitive signal is still several steps away from that standard. Nonetheless, its replication across KarXT and ML-007C-MA gives the field a stronger starting point for prospective stratification.

Reading the Market Reaction

MapLight’s stock closed roughly seventy percent lower on the day of the ZEPHYR announcement. The reaction makes sense from the perspective of investors who had expected a clearly differentiated successor to KarXT. The once-daily regimen missed the primary endpoint, the effective twice-daily regimen produced a smaller PANSS effect than KarXT’s own Phase 2 trial, and the gastrointestinal adverse-event burden remained substantial. Each of these weakens the case for an obvious best-in-class commercial profile.

The scientific interpretation can be more favorable without ignoring those concerns. Positive trials remain relatively rare in schizophrenia drug development, and ML-007C-MA produced a coherent effect across PANSS, CGI-S and positive symptoms, alongside a cognitive result that had already appeared with another dual M1/M4 agonist. Investors and scientists were, in effect, reading two different documents. One was pricing the probability that ML-007C-MA competes commercially with an approved product. The other was asking what the trial reveals about muscarinic biology and the possibility of reaching symptom domains that dopamine-blocking medications have served poorly.

Several analysts maintained favorable ratings after reducing their price targets, which is consistent with a reset in commercial expectations rather than abandonment of the program. MapLight still has a path into a confirmatory pivotal trial, along with the possibility of testing alternate once-daily regimens later. The cognitive signal may also have relevance to the company’s ongoing program in Alzheimer’s disease psychosis, where preservation of cognition carries obvious importance.

The Dawn of the Muscarinics

When Cobenfy first worked in a patient of mine,

I described it as a light bulb turning on. The empty, thousand-yard stare that psychosis leaves on a face lifted within weeks, replaced by someone visibly present again. At the time, that was a single clinician’s observation, vivid but impossible to generalize from. It could have been xanomeline-trospium specifically, some quirk of that molecule, that patient, that moment. ZEPHYR may have changed that. A second company, working from a different molecule built around the same receptors, has now produced a coherent antipsychotic effect and a cognitive signal that lands almost exactly where KarXT’s did. Does it also produce a lightbulb effect or what clinicians are now calling “the cobenfy clearing”? I don’t know, and time will tell.

But even if it did, we should not think that this means the muscarinic drugs available today solve schizophrenia. They do not work for everyone. Yet, something foundational is happening. For seventy years, psychiatry had exactly one lever for psychosis: block dopamine, and vary the side effects. There are now, demonstrably, at least two levers that work, built independently by different people. That is what a watershed looks like in this field: the moment a mechanism stops depending on whether a single company’s molecule happens to be good, and starts existing as a category clinicians can build on. The muscarinics are here to stay. What they will ultimately be able to do for patients is being written.

*Topline results are the initial, high-level summary of a trial’s main findings that a company puts out within days of a trial ending, well before the full dataset is published and reviewed by outside scientists.

If you found this piece useful, please consider subscribing and sharing it with someone who might find it clarifying. Paid subscriptions help support the time required to produce this work.

When a psychiatrist encounters a patient with serious mental illness, they attempt to narrow down the nature of the underlying process in order to treat it. Take psychosis for example, the term we use to describe conditions in which people lose touch with reality and experience functional decline. The Diagnostic and Statistical Manual (DSM) provides some guidance, pointing to psychosis spectrum disorders which include schizophrenia, schizoaffective disorder and bipolar disorder with psychotic features. In practice, the treatments of these conditions are overlapping and heavily reliant on individual prescriber experiences. Psychiatry doesn’t have the equivalent of a chest X-ray or an echocardiogram to objectively guide personalized treatments.

This can be a practical problem in many ways. In addition to the fact that ostensibly similar medications tend to have idiosyncratic responses in individual patients for reasons that are hard to study, up to one third of patients with psychotic illnesses do not benefit from common antipsychotics and require clozapine after a long period of trial and error. Clozapine is a highly effective antipsychotic but one that has a list of serious side effects and monitoring requirements that relegates it to a second or third line treatment. Some patients do not respond to clozapine either and may still respond to other types of treatment, but we have no way of knowing this beforehand.

The DSM was never designed to identify the biological processes responsible for mental illness, and there have been multiple recent efforts to do so. The B-SNIP (Bipolar-Schizophrenia Network for Intermediate Phenotypes) consortium has spent over a decade building a system to classify psychotic disorders for better treatment guidance. Taking a large group of patients spanning the psychosis spectrum, the consortium administered behavioral tests, measured brain signals via electroencephalography (EEG), and tracked eye movements. It asked if better categories emerged on the basis of these more objective measures. Three ‘biotypes’ were identified, and they did not respect DSM boundaries. Meaning, a patient with a diagnosis of schizophrenia and another with bipolar disorder could both be biotype 1, 2 or 3. This alone is a provocative finding, and perhaps congruent with the clinical intuition that many patients don’t really fit into the DSM labels that well anyway. But do these biotypes tell us which treatments to use for which patients?

LEFT: Psychosis cases with Biotype1 (red), Biotype2 (blue), Biotype3 (green) cluster distinctly when graphed by biomarker constructs. The gray cross/ellipse represents the biomarker profile of healthy controls, who scored similarly to Biotype 3 cases on measures of Cognitive Control. RIGHT: Psychosis cases similarly graphed by biomarker constructs show much less distinct clusters when grouped by the traditional diagnostic categories schizophrenia (sz), schizoaffective disorder (SZaff), and bipolar disorder with psychosis (BDP). Source: Brett Clementz, Ph.D., University of Georgia

This question is at the heart of precision psychiatry, a field increasingly pressed for delivering treatment-guiding tools. B-SNIP is one approach to precision psychiatry, emphasizing deep characterization to prospectively stratify patients and guide better treatment decisions. But there are other approaches too. In this newsletter I covered the use of guanfacine in a subtype of depression and bumetanide for a subtype of autism. These projects have taken clues from retrospective analysis of existing data, where particular patient features appeared to have predicted a specific treatment response. Ultimately, to guide care, these features have to prove their value in clinical trials that prospectively enrich them. Retrospective analyses can also complement approaches such as B-SNIP by providing candidate features for clustering or validation.

Now with all this, let’s take a step back and ask: how do these approaches compare to precision care that exists in other parts of medicine? We will take two examples and see what psychiatry may be able to learn from them.

The diabetes case: doing the sequence right and still coming up short

Clustering, which is relevant to B-SNIP and to the diabetes case we are about to discuss, is easiest to picture with just two measurements. Imagine taking a large group of people, and plotting each person as a dot in the two dimensional space of height and weight. Step back, and the dots will likely clump (cluster) into two visually distinct groups, small and light in one corner, tall and heavy in another. Clustering is a mathematical method of identifying those data-driven patterns, and can do the same thing with many features at once, grouping together nearby samples in that multidimensional space.

Two clusters in two dimensional weight/height space, unlabeled (left), labeled (middle). Four clusters in a multi-dimensional space (right). An individual sample in that space is visualized as a set of connected features.

The catch is that clustering requires judgment calls, since most methods will divide a population into a different number of clusters depending on what is specified ahead of time. Specifications include the choice of variables to include, how to scale them, and how to define “close,” since there’s more than one valid way to measure distance between sets of variables. Importantly, a clustering algorithm cannot tell whether the clusters are valid for the problem one is ultimately trying to solve. This is one reason why validation with measures different from the ones used for clustering is important.

In 2018, a clustering analysis of adult-onset diabetes generated excitement in precision medicine circles. The study utilized a large cohort of newly diagnosed patients and clustered them using six features: age at diagnosis, body mass index, a marker of long-term blood sugar control (Hemoglobin A1c), a blood marker of autoimmunity, and two indices related to insulin release and action.

Five clusters emerged and differed from one another in a manner that appeared promising. For example, one consisted of severely insulin-deficient patients who were young, lean, and positive for the autoimmune marker, essentially a category close to type 1 diabetes under a type 2 label. The clusters replicated in independent cohorts and later lined up with distinct genetic risk architectures.

But did they alter treatment decisions? A follow-up analysis found that simpler models predicted outcomes just as well or better. Separately, a 25-year follow-up study concluded that the clusters’ ability to predict complications like kidney and cardiovascular disease was insufficient to support using them in clinical practice. As of today, the five clusters identified in 2018 have not altered treatment guidelines.

The lesson here may be that even with reproducible, seemingly meaningful and externally validated clusters, altering treatment decisions is a much higher bar to clear.

The Asthma case: further progress

Severe asthma has a problem similar to one we discussed in the opening of this piece. Patients who look alike in terms of symptoms may respond differently to different interventions. Cluster analyses run on severe asthma patients by different groups using different specifications, kept converging on a common set of features: age of onset, obesity, allergic history, and a pattern of airway inflammation dominated by eosinophils, a white blood cell involved in allergic and parasitic immune responses. That last dimension matched what an earlier study using direct airway biopsies had already found: some patients with severe asthma had airways packed with eosinophils, while others did not.

That’s the distinction pharmaceutical development ultimately followed. Once drugs were developed to block the specific signaling molecule driving eosinophils, such as mepolizumab, the trials for those drugs enrolled patients using their blood eosinophil count. Patients above a given threshold responded reliably while those below it did not. This stratification is now part of the treatment guidelines: a patient generally needs a blood eosinophil count above a set threshold before one of these medications gets prescribed.

Why has this approach worked in asthma but not diabetes? I suspect it is because the key feature, eosinophil count, sits proximal to a disease-generating process in asthma. Being able to extract such variables prior to clustering is likely to be a driver of success.

B-SNIP Made Remarkable Progress

With these two examples in mind, we can start understanding what the B-SNIP consortium has built. Across the full program, the consortium assembled one of the most serious transdiagnostic psychosis datasets in the field. It included patients across the psychosis spectrum, along with healthy controls and first-degree relatives, acquiring several behavioral and neural measurements. Their initial discovery sample of 711 psychosis cases identified the three biotypes, and the finding was independently replicated in a sample of 717 cases. The group-level pattern across the two samples had an intraclass correlation of 0.95. Roughly 89% of individual patients landed in the same biotype when the classification rules built on one sample were applied to the other.

The three biotypes differ from each other in ways that look promising. Biotype 1 appears to be the most globally impaired group, with patients showing severe cognitive impairment, diminished background and sound-evoked EEG activity, and slower eye-movement responses. In essence, the brain appears under-responsive across several kinds of measurement. Biotype 2 patients are also cognitively impaired, but their physiological measures are different, showing elevated ongoing EEG activity and the worst performance on tasks requiring inhibitory control, where a person has to suppress an automatic response. Thus, Biotype 2 may reflect a different systems-level dysregulation than Biotype 1. Biotype 3 is the least impaired according to these measurements, with behavioral and EEG measures closer to healthy controls.

An intriguing finding is that some patients diagnosed with schizophrenia are categorized as Biotype 3, while some patients diagnosed with bipolar disorder are categorized as Biotypes 1 and 2. This is surprising because it runs against the intuition that schizophrenia is uniformly a condition of impaired cognition compared to bipolar disorder. The measures are stable across a year of follow-up, suggesting these are features of trait, not state.

B-SNIP Biotypes have external validators too. Structural MRI measures, which were not used for clustering, show different patterns across the groups, with Biotype 1 having the clearest and most diffuse structural signature. First-degree relatives show milder versions of some Biotype-linked patterns, suggesting that at least part of the signal may reflect familial or developmental liability. Also, in a practical move towards real-world adoption, the consortium distilled their measurements into a shorter battery termed ADEPT (Algorithmic Diagnostics for Efficient Prescription of Treatments) which can still identify Biotypes with reasonable accuracy.

The Limits of Measurements

But do the three biotypes reflect three illness-generating processes that may respond to different treatments?

The honest answer is that we simply don’t know, and a bit of methodology may help explain why. The B-SNIP consortium started as a discovery effort, collecting 44 separate measurements across cognition, EEG, and eye movements. Many of these measurements are correlated; performance on a verbal memory task predicts performance on a working memory task, since both rely on holding information in mind over a delay. Feeding all 44 raw measurements into a clustering algorithm as though they were independent would let a handful of correlated variables dominate the result by being counted several times. One fix is dimensionality reduction. Techniques such as Principal Component Analysis, or PCA, compress correlated variables into a smaller set of factors (or principal components). For example, height and weight are correlated, since taller people tend to weigh more. Plotting height against weight, a sample of people may fall roughly along a diagonal line running from short and light to tall and heavy. PCA finds that diagonal and reduces these two axes into a number line reflecting overall size. In B-SNIP, PCA turned the 44 measurements into 9 principal components, called bio-factors, before clustering.

Height and weight reduced to one Principal Component (left, middle). The same operation applied to many measurements (right).

One issue is that PCA has no way of knowing which trend in the data best addresses the question being asked. It finds whichever direction accounts for the most shared variance across the measurements it’s given, and in B-SNIP these measurements were the consortium’s best guess as to what would capture psychosis-relevant features. But if the dominant shared trend across cognition and EEG scores turns out to reflect something like general effort or arousal rather than a specific illness process, PCA will output that trend. And if these factors are fed into a clustering algorithm, the resulting clusters would reflect whatever the structure of those factors is, illness-generating or not.

Are there ways to get at disease-generating factor selection more directly? Computational psychiatry, the field whose goal is to describe psychiatric conditions in terms of specific alterations in information processing, aspires to capture illness-generating processes. It assumes that these processes will alter specific components of information processing sequences and aims to design behavioral tasks capable of exposing these components. Via computational models that can capture the behavior of individual subjects, it generates parameters that point to where the specific perturbations are. This can be diagnostic: for example, two patients with difficulty adjusting to new contingencies may have different underlying causes. One may have difficulty with uncertainty management and another with feedback processing, yet both may show similar scores on surface measures. Only a task designed with computational dissection in mind, and tools capable of that dissection will be able to tell the difference. B-SNIP has not taken that approach and it is unclear if its clusters group together distinct illness-generating processes as a result.

Despite this potential caveat, the B-SNIP team is positioned to make progress. The team is aware that treatment validation is ultimately most critical, and has formulated a clozapine hypothesis for biotype 1 based on the type of retrospective analysis we discussed earlier in this piece. Specifically, in existing data, biotype 1 patients on clozapine look closer to healthy controls on EEG measures than biotype 1 patients on other antipsychotics. To their credit, the B-SNIP team is now enrolling patients in a prospective randomized trial comparing clozapine with a standard antipsychotic in biotype 1 and biotype 2 patients.

A Path Toward Precision

But what if B-SNIP efforts eventually prove insufficient to alter treatment guidelines? Should we abandon stratification altogether? I think not, and I think the lessons from precision medicine point to attempting to get closer to the disease-generating process.

In asthma, this was key to successful stratification. In psychosis, the problem is much harder because the answers lie in the working of an organ that we do not understand nearly as well. Psychosis impacts how a person builds models of the world, updates them as evidence changes, handles uncertainty, and routes internally versus externally generated signals. I have repeatedly argued that precision psychiatry for psychotic disorders needs reliable ways to measure these computational elements. Along with their circuit implementation, these elements may provide ideal behavioral and neural features for clustering efforts. It is conceivable that the B-SNIP clusters, while reproducible, are not capturing the variance along a treatment-informing axis.

As mentioned in the previous section, computational psychiatry has already grappled with what tools and measurements would be helpful for this effort. Various groups have built tasks whose structure lets a fitted model dissociate distinct failure modes from one another. The rise of modern AI tools may enable us to go even further, and beyond traditional laboratory tasks. By developing more complex tasks closer to real-world conditions, we might get closer to computational components that matter most for psychotic illnesses, with computational models finally capable of interpreting that complexity. Along those lines, the finding that B-SNIP Biotype 3 overlaps with healthy controls on their batteries must be considered a real limitation. It suggests that their measurements may not be aligned with mechanisms that drive psychosis specifically.

Another important issue to address is validation. Treatment response is a form of diagnostic information, arguably the most direct kind. If patients who share a computational parameter or related neural readout respond similarly to a drug, behavioral therapy, or neurostimulation, the shared response is evidence that the measurement is treatment-relevant. With the rise of novel antipsychotic medications such as muscarinic agents, which may have effects distinct from traditional dopaminergic agents, differential treatment response may help us discover which features should be used for future stratification. In a sense, efforts to route patients to clozapine faster based on behavioral and neural measurements have adopted a similar logic. B-SNIP is now attempting to do exactly that, prospectively enrolling patients with the prediction that clozapine will be particularly effective for Biotype 1 based on retrospective analysis of treatment measures.

B-SNIP built something remarkable with the tools and questions available to them over the last decade, and my analysis is not at all meant to diminish what they have already accomplished. It is only meant to extend it and identify ways we can make further progress. Realizing precision medicine for psychotic disorders is a formidable challenge and will take enormous talent, resources and dedication. The upside is huge, with the possibility of restoring 1% of humanity to a level of functioning they currently lack.

I should note that wrote a piece about B-SNIP recently, which I encourage people to read. Our focus diverges in places, but the overlap reflects that stratification is an area of great interest and live discussion.

If you found this useful, I’d appreciate you liking it and sharing it with your network. While I keep the vast majority of my posts free, paid subscriptions are always appreciated.

Imagine trying to follow a conversation that is hard to follow. A friend is telling you a story, but you are not sure whether it is a joke, a complaint, or a confession. Each sentence moves you forward through a space of possible interpretations. This is one of many mental navigation tasks that we all go through when we reason, remember or daydream.

Navigation may be more than an analogy here. When we learn new things, our brains appear to build maps of their associations. Brain regions such as the medial temporal lobe, long known for their role in spatial navigation, are also engaged when we mentally navigate through memories, concepts, and imagined possibilities. But why store memories this way? It seems very different from how a laptop stores information, and it may even introduce new forms of confusion.

Maybe this is just how evolution set us up: maps are more powerful than lists. You can navigate them, find shortcuts, and generalize from one region to another. But a map is only useful if you can locate yourself within it. In physical space, doing so requires the brain to maintain two frames of reference simultaneously: where things are relative to the body, and where the body sits in the world independent of its current orientation. These are called egocentric and allocentric frames of reference. This can be a challenging transformation. Walking in a strong crosswind, one may angle their body into the wind while their actual path across the ground moves in another direction. The brain can track both signals separately, but knowing the true direction of travel requires converting them into a common frame of reference.

This is a general problem for animals. A fly in flight may be pointing one way while moving partly sideways, so it needs to perform the same transformation. Remarkably, because the fly brain has now been mapped at synaptic resolution, we can see exactly how its brain does it. The answer appears to involve phasor arithmetic, the mathematics associated with Euler’s formula, implemented in the connections between fewer than fifty neurons.

This example, which I will return to in detail, may help illuminate why the question of whether the brain is a computer is worthwhile. It has generated sophisticated debates among scientists and philosophers. Yet, in public discourse, people often get stuck on the image of a laptop, with a central processor, stored instructions, and symbols being moved around in memory. Let’s clear that up right away: that architecture is certainly not what brains are doing. But rejecting the brain-as-digital-computer does not answer the question of whether the brain is some type of computer.

Computation, in the broad sense, is a process that transforms inputs into outputs according to some rule. One useful question is whether neurons and circuits carry out identifiable transformations on their inputs while representing something relevant to the world they inhabit. At the level of single neurons, local circuits, neural populations, and large-scale networks, the answer is almost certainly yes.

This post will move through the brain at different scales, using examples that are common in computational neuroscience. We will start with single neurons and then move to small circuits, where we return to the fly example, discussing how the wiring of a small circuit appears to add and rotate directions. Finally, we will move to mammals, where internal maps suggest that the same basic logic may support memory and planning. The goal, in the end, is for you to have enough material to help make up your own mind as to whether you think the brain is a computer.

At the level of single neurons

TL;DR: Neurons are far more than simple switches. Each one weighs thousands of incoming signals before deciding whether to fire. The dendrites, the tree-like branches that receive those signals, turn out to be surprisingly powerful computational devices in their own right, capable of solving problems that computer scientists once thought required an entire network.

Neurons are complex machines in their own right. They are thought to communicate through spikes, brief electrical pulses that they generate. In sensory organs, spike rates may be directly related to physical attributes of the world: a neuron may fire more as the physical intensity increases. This is what is often meant by a neuron encoding or representing the stimulus. Turning the physical attribute of the stimulus into a voltage change is known as transduction (see Figure below).

Deeper into the brain, synaptic inputs arrive as small currents, and the receiving neuron integrates them into charge that raises its voltage. When the voltage rises sufficiently, it can cause a spike, but whether that happens or not depends on the strength, timing, and location of inputs across its dendritic tree. The mathematics of spike generation itself, how ion channels open and close to drive the membrane through the characteristic rise and fall of a spike, earned Hodgkin and Huxley a Nobel Prize in 1963 and remains one of the most beautiful models in neuroscience.

From synaptic integration and spiking emerges a surprisingly rich computational repertoire. A neuron can implement basic logic operations. Two inputs arriving simultaneously can summate more effectively than inputs arriving apart, making the neuron a coincidence detector, an AND gate. In other instances, either of two independent inputs can drive firing on its own, making the neuron an OR gate.

Having dendrites that can generate their own spikes (active dendrites) expands the computational repertoire of single neurons and, in humans, is thought to enable non-linear operations that have been classically thought to be beyond the capacity of single neurons. Specifically, having active dendrites allows a neuron to implement exclusive OR, or XOR, a linearly non-separable classification problem.

From Larkum 2022 Partly based on Gidon et al.

At the level of microcircuits

TL;DR: When neurons wire together into circuits, the computations they can perform expand dramatically. A handful of connected cells can extract edges from a visual scene, track where your eyes are pointing, or adjust sensitivity on the fly depending on context. In these instances, the circuit architecture can specify the computation.

When neurons are connected into circuits, the computational repertoire expands further. The circuit receives one kind of signal and transforms it into another, and the nature of the transformation starts to rely heavily on the pattern of connectivity rather than the properties of individual neurons.

The retina provides a clean entry point. Retinal ganglion cells typically represent the world in luminance differences: excitatory drive in a central region and inhibitory drive from a surrounding annulus. A uniform field produces little response because center and surround largely cancel, but edges may drive strong responses. Formally, this is spatial differentiation and the result is a representation of contrast. Local retinal wiring turns one variable into its spatial derivative.

Figure 1 of Zapp et al.

Circuits can also perform integration. A single neuron is a leaky integrator, which means it accumulates input but drifts back toward rest as fast as charge dissipates across its membrane, so it cannot perform true integration. True integration requires continued excitatory drive, or recurrent connectivity that counteracts the leak of single neurons. The oculomotor system is one of the clearest examples in neuroscience: moving the eyes requires a velocity command specifying how fast they should move, but holding them at the new position requires converting that velocity signal into a position signal specifying where they should remain. Brainstem integrator circuits perform this transformation. The calculus language here is quite real: position is the time integral of velocity, and the circuit computes exactly that.

Figure 1 of Seung et al. Signals in the Oculomotor System and the Architecture of a Recurrent Network Model. (A) Schematic of eye position and neural activity versus time in the oculomotor system.(B) Network model of the integrator based on recurrent excitation. All integrator neurons receive feedforward input from vestibular, excitatory burst, and inhibitory burst neurons. To make the diagram compact, only 5 integrator neurons are shown, but 15 were actually used in the simulations. Signals from the burst and integrator neurons are relayed by the motor neurons to the oculomotor plant, to produce eye position.

Circuits perform algebraic operations too. Gain modulation is a form of multiplication where one input scales the magnitude of another. A contextual signal can scale the response to a sensory input up or down depending on the animal’s behavioral state, attention, or expectation. In computational terms this is multiplicative gain control, one variable setting the gain applied to another, and it is found in many different domains. Our lab found it, for example, when studying how contextual signals from the thalamus can gate working memory representations in the prefrontal cortex. You can read the relevant piece here.

Divisive normalization is among the most thoroughly characterized computations in systems neuroscience. A neuron’s response to its preferred input is divided by the pooled activity of a surrounding population, mediated by inhibitory interneurons. The effect is to represent input across a wider dynamic range. It has been identified in visual cortex, auditory cortex, and olfactory systems, and appears to be a canonical computation the brain uses broadly.

The fly circuit: vector arithmetic

TL;DR: A fly’s brain solves a navigation problem that requires combining different directions into a common reference frame, and it does so using the same mathematics electrical engineers use to add wave signals. The key insight is that direction can be represented as the phase of a wave pattern distributed across neurons, and combining two such patterns produces a new pattern whose phase gives the correct answer. The wiring of the circuit was predicted mathematically and then confirmed directly by the connectome.

To understand what the fly circuit does, it helps to first review a piece of mathematics.

When you add two sinusoidal signals of the same frequency, the result is another sinusoid at that frequency, but with a new amplitude and phase. Tracking those two quantities directly with sines and cosines quickly becomes algebraically cumbersome. The reason is that a sinusoid is not fully described by a single number. At a given frequency, it has two key parameters: how large it is, and where it sits in its cycle.

Complex numbers are built for exactly this kind of problem: they provide a single algebraic object with two independent components. There is also a geometric picture that makes this concrete. A sinusoidal signal can be viewed as the shadow of a point rotating around a circle: the cosine gives the horizontal position of the point, and the sine gives its vertical position. The rotating point can be represented as an arrow whose angle encodes phase and whose length encodes amplitude. That arrow is a phasor.

Euler’s formula, e^(iθ) = cos(θ) + i·sin(θ), is the bridge between these descriptions. It says that a complex exponential contains both the cosine and sine components of a rotating signal. The angle θ encodes phase, and the magnitude of the complex number encodes amplitude. Multiplying by e^(iφ) rotates the arrow by φ, which corresponds to shifting the phase of the signal by φ. What would otherwise require trigonometric identities becomes simple multiplication.

This is why phasors are useful. Adding sinusoidal signals of the same frequency becomes equivalent to adding arrows in the complex plane. The resulting arrow gives the amplitude and phase of the combined signal directly. Electrical engineers love phasors for exactly this reason.

Figure from this nice tutorial

How is this relevant to the brain? Consider a ring of neurons, each with a preferred direction around the ring. If their activity forms a sinusoidal bump, with one neuron at the peak and activity falling off smoothly toward a trough on the opposite side, the pattern of activity across the population is itself a spatial sinusoid. The phase of this pattern, meaning the position of its peak around the ring, encodes a direction. The amplitude encodes the strength or confidence of the signal. This population-level sinusoidal pattern may be interpreted as a phasor: a direction represented as the phase of a sinusoid across a ring of neurons.

This idea was developed over years by the drosophila community. For example, in 2015, the Jayaraman lab found that neurons in the ellipsoid body are physically arranged in a ring, and as the fly changes its heading, a localized bump of activity moves around that ring. The bump is maintained by operations described in earlier sections: divisive normalization keeps the bump sharp and bounded, and winner-take-all dynamics, implemented through mutual inhibition, prevent multiple competing bumps from forming simultaneously. The bump persists in darkness without external input, which is the defining signature of a ring attractor, and snaps to the correct position when landmarks reappear. A ring attractor describes a robustly stable system that settles into one of a family of stable states and holds it.

From Figure 2 of Seelig and Jayaraman 2015. Activity bump moving around the ring that relies on the fly’s heading relative to landmarks rather than the landmarks themselves.

In 2022, Lyu, Abbott, and Maimon showed that the fan-shaped body downstream performs phasor addition on these types of representations. To compute allocentric traveling direction, the circuit must perform an egocentric-to-allocentric coordinate transformation by rotating, scaling, and adding four vectors related to the fly’s body-axis-referenced traveling direction. Direction is an inherently two-dimensional quantity: it has both a magnitude and an angle, which is efficiently represented by 2-dimensional vectors. The sinusoidal format of the neural signals does so via amplitude encoding vector length and its phase encoding direction. The fan-shaped body sums four such profiles, and because each can be understood as a phasor, their sum is a new sinusoidal profile whose phase encodes exactly the allocentric traveling direction that vector addition predicts. Neural circuits appear to perform vector arithmetic in a format that complex numbers describe naturally.

Illustration based on Lyu, Abbot and Maimon. Panel A shows the fly’s heading (H, pointing up) and its egocentric travel direction (T_ego, the green arrow at +45° relative to heading). The four dashed axes at +45°, +135°, -135°, -45° are the four PFN population axes onto which the travel vector gets projected. Panel B shows the decomposition. The egocentric travel vector is broken into four components (L1 through L4) along those four axes. Each component is then encoded as a sinusoidal activity pattern across the fan-shaped body columns, with the amplitude of each sinusoid proportional to the projection of the travel vector onto that axis. This is the phasor representation: each sinusoid’s amplitude is the vector length, its phase is fixed by the axis direction. Panel C shows that the h∆B neurons sum the four sinusoids to produce a single resultant sinusoid, and the peak of that resultant encodes the allocentric traveling direction T_allo on the world-referenced compass.

How was this figured out? The first step was functional: neural imaging via two-photon microscopy showed both sides of the computation simultaneously. Four input populations, each corresponding to one of the directional components, displayed sinusoidal activity profiles with amplitudes that varied according to the fly’s egocentric traveling direction. A separate population of output neurons displayed a new sinusoidal profile whose phase tracked the allocentric traveling direction. The inputs and output looked like phasors. The question was whether the wiring between them implements the arithmetic correctly.

A computational model made a specific prediction. The four input populations project from the protocerebral bridge, a bar-shaped upstream structure organized into columns where the heading direction signal is represented, into the fan-shaped body. Each population’s projection pattern determines where its sinusoidal peak lands across the fan-shaped body columns. When the output neurons sum the four profiles, the result depends critically on the relative positions of those peaks. Only one arrangement of peaks produces the correct allocentric direction. A different wiring pattern, placing any peak even slightly off, would point the output in the wrong direction. The model therefore predicted a specific set of anatomical offsets.

The hemibrain connectome, published in 2020 and analyzed in detail for the central complex in 2021, made it possible to check the actual wiring directly. The phase offsets in the anatomy matched what the computation required. The synaptic wiring is the computation, and the connectome was the proof.

Video Credit

From fly circuits to mammalian maps

TL;DR: In mammals, the brain may be using similar mathematics to the fly to build maps beyond physical space, extending to memories, concepts, and relationships. The same neural machinery that helps you find your way through a building may also help you find your way through a conversation. Exactly how our brains do it is one of the most interesting open questions in neuroscience right now.

The fly circuit is a powerful example because the computation is unusually clear. In the mammalian navigation system, the same mathematical logic may be at play. However, without the connectome, it must be inferred.

Mammals appear to have similar ingredients for navigation. Head direction cells, found in the anterodorsal thalamus and postsubiculum, fire when an animal faces a particular direction, and the population as a whole represents heading as a circular variable. These circuits have long been modeled as ring attractors, and the fly heading system gives a physically visible version of exactly that idea. Whether mammals implement the same attractor dynamics through analogous wiring is less completely resolved, but the computational motif appears conserved.

Grid cells in the medial entorhinal cortex add a second layer of organization. Each grid cell fires at multiple locations arranged in a repeating triangular lattice across space. Different cells within a module share the same spacing and orientation but have patterns shifted relative to one another. That shift is an important detail, because it means that the animal’s position can be represented as a phase within a repeating spatial pattern. Movement through the world becomes movement through phase, which is directly analogous to how the fly heading signal works: direction encoded as the position of a bump around a ring, updated continuously by self-motion.

The brain uses multiple grid scales, with different grid modules having different spatial periodicities. Combining phases across modules with different spacings can identify a unique location, much as overlapping rulers with different tick intervals narrow down a measurement. The hippocampus then may use this periodic infrastructure. Place cells, which fire in more localized regions, can be generated by combining grid cell inputs with different phases and spacings, each individually ambiguous but their combination selective. This gives the hippocampus a way to turn periodic coordinates into specific places, and helps explain how the same circuit that supports navigation also supports memory: a place is a set of associations.

In predictive map accounts, grid-like representations are treated as low-dimensional basis functions for predicting future states. The same mathematical logic, periodic structure, phase, and basis functions, extends from spatial navigation toward memory and planning. In a sense, many cognitive operations can be formulated as navigating through some sort of mental space that may rely on similar machinery to spatial navigation.

Indeed, human participants learning new conceptual relationships show hexagonal grid-like signals in entorhinal and prefrontal regions, and hippocampal theta oscillations track the semantic and temporal distances between memories during recall. Spatially periodic signals appear in both hippocampus and entorhinal cortex during navigation of abstract object spaces, with the hippocampal periodicity synchronized to the entorhinal signal in a manner consistent with a hierarchical transformation. These findings and related ones, suggest that the mathematical logic of representing relationships as phases in a periodic code and updating those phases to navigate between them, may extend from physical space into the space of meanings, associations, and predictions that the opening of this piece described. The conversation you were trying to follow may be navigated, at least in part, by some of the same machinery that helps you navigate the room you were sitting in.

Figure 1 of Park et al. Cognitive Maps in Everyday Decision Making. (A) Examples of zero-shot inferences in physical space, transitions between objects, and family trees. Representing abstract relationships as a cognitive map allows novel direct inferences that do not only rely on previously experienced associations. Black, experienced relationships; red, inferred relationships.(B) Bilateral ROIs generated independently from probabilistic maps. (C and D) Two hypotheses concerning how the brain could represent and flexibly switch between dimensions that characterize the same entities to guide inferences. (C) The brain could construct two separate maps for representing each 1D hierarchy learned on a separate day, and distinct regions could encode the 1D rank difference in the task-relevant dimension (D) and the 1D rank difference in the task-irrelevant dimension (I). (D) Alternatively, the brain could construct a unified map consisting of two dimensions and encode the inferred Euclidean distance (E) over the 2D representation

The fly gave us a beautiful transformation via the very wiring of its brain. Does our mental navigation rely as heavily on our connectome or on other features that may be more flexible? This is an open question and as connectomics make further progress in mammalian brains, we will be able to learn more.

So is the brain a computer?

Not if by computer we mean a laptop: a device with a central processor, stored instructions, and a clean separation between hardware and software.

But if computation means transforming inputs into outputs according to rules, while operating over representations that matter for perception, action, memory, and planning, then the answer looks different. Describing the brain computationally also lets us see deep links between things that otherwise appear unrelated. Who would have guessed that the machinery we use to navigate physical space might also help organize knowledge more generally? And that part of this machinery may be described using Euler’s formula seems both beautiful and strange.

Yet the hard philosophical question is not whether neural activity can be described with mathematics. Rivers and pendulums can be described with mathematics too. The harder question is whether the variables inside the system represent something for the organism, or whether that is a description we impose from outside on a physical process that simply runs. There is a reasonable answer to this. The fly’s fan-shaped body does not merely correlate with allocentric direction: disrupting it specifically impairs navigation while leaving other functions intact, and the signal is consumed by downstream circuits in ways that are sensitive to its content. A pendulum’s motion does not causally organize anything downstream based on what it means. That kind of functional organization, where a signal is used by the rest of the system in a content-sensitive way, is the strongest available criterion for genuine representation, and it is one the examples here appear to meet. Whether that criterion is sufficient, or whether genuine representation requires something more, is where serious people disagree. In essence, that debate is why the question is worthwhile.

So what do you think? Is the brain a computer?

Acknowledgement: I would like to thank Ana Lund for helpful comments.

If you found this piece useful, please consider subscribing and sharing it with someone who might appreciate it.

For those of us who grew up watching American television in the 1990s, Saved by the Bell occupies a strange and durable place in memory. The show was absurd, bright, and almost perfectly engineered for adolescence. So when a recent documentary revisited that world, I watched partly out of nostalgia and partly out of curiosity about what had happened to the young actors who had lived inside it.

What I learned about Lark Voorhies, the beloved actress who played Lisa Turtle, was surprising and heartbreaking. The documentary includes commentary from Dr. Drew Pinsky, who describes Lark as having been diagnosed with schizoaffective disorder, associated with what appeared to be years of subsequent disability.

Dr. Drew later said something to the effect of: schizoaffective disorder is very different from schizophrenia, with a much better outlook and a less downward spiraling trajectory.

Really?! Is that true Dr. Drew?

I should say upfront that I am not in a position to comment on Lark Voorhies’s diagnosis or clinical course, and that is not the point of this piece. The point is what the public discussion around her diagnosis reveals about how loosely we tend to talk about schizoaffective disorder, and how poorly that loose talk holds up against the actual evidence.

To me, Dr. Drew’s comment illustrated a confusion that runs surprisingly deep, even among people who speak authoritatively about mental health in public. Schizoaffective disorder is understood by some as a middle ground between schizophrenia, which is traditionally thought to have a deteriorating course, and bipolar disorder, which, at least in the textbook framing, carries a more episodic and recoverable course. But what is the evidence for that?

This piece is an attempt to explain what is known about how schizophrenia and bipolar disorder differ, why the distinction is often impossible to make early in the illness, what schizoaffective disorder actually names, and whether it represents a real biological category or something closer to a marker of genuine uncertainty.

Two illnesses, two different stories

Schizophrenia and bipolar disorder can both produce psychosis, and on the day someone is brought to the hospital in the middle of their first episode, the two illnesses can be genuinely impossible to tell apart from the surface presentation alone. Only time and careful longitudinal observation can.

In schizophrenia, the break from reality is rarely the true beginning of the illness, which is part of what makes it so hard for families to piece together in retrospect. There is usually a period, sometimes years long, during which something had already started to change: a previously engaged and social person becomes quieter, harder to reach, less able to sustain the academic and social demands that had come naturally before. Emil Kraepelin noticed more than a century ago that his dementia praecox patients deteriorated differently from his manic-depressive ones, and that the deterioration appeared to be independent of what happened to their psychosis. That observation has held up remarkably well, because cognitive decline in schizophrenia may begin before psychosis emerges, continues after it is treated, and is mostly resistant to antipsychotic treatment. After an acute episode of psychosis resolves, many are left with flatness, poverty of motivation and speech, difficulty initiating action and sustaining engagement that define the negative symptom picture that drives functional disability.

Bipolar disorder appears to be driven by a different kind of disruption. Lack of need for sleep is often the first signal of an impending manic episode, appearing days before mood changes are visible and representing one of the most reliable prodromes identified in the research literature. What follows can be frightening in its speed: energy surges, speech accelerates, confidence expands beyond the reach of judgment, and at the peak, psychosis can appear that is clinically indistinguishable from schizophrenia in the acute moment.

Longitudinally, however, many people with bipolar disorder return to something substantially close to their baseline function between episodes, and the negative symptom picture that defines so much of the schizophrenia course is largely absent (but note that some patients with bipolar disorder still describe mental fog following acute episodes that may be related to negative symptoms). Plus, cognitive impairment in bipolar disorder can be substantial, particularly in patients who have accumulated many psychotic episodes, indicating possible kindling. This is why I emphasize symptom control with my bipolar patients. Nonetheless, the mean tendency of these disorders to result in progressive vs. remitting/relapsing courses are distinct.

Which brings us to schizoaffective disorder, a diagnosis that exists precisely because the longitudinal separation described above often cannot be made at first presentation. The criteria require both persistent psychosis and major mood episodes accounting for the majority of the illness course, with psychosis occurring outside mood episodes as well, but applying those criteria in practice demands information that a first hospitalization rarely provides in full. Mood symptoms during acute psychosis are difficult to separate from the psychosis itself and antipsychotics can suppress mood components before the pattern has declared itself. The result is a diagnosis that is applied when the clinician has enough of both pictures to feel uncomfortable calling it either, which is a clinically honest position but one that gets mistaken for a stable biological description.

The diagnostic instability literature makes this problem concrete. A systematic review and meta-analysis found that 36% of patients initially diagnosed with schizoaffective disorder later received a different diagnosis, with the majority of those shifting toward schizophrenia rather than toward bipolar disorder. This is relevant for how we interpret confident public statements about schizoaffective disorder having a better trajectory than schizophrenia, because the prognosis of schizoaffective disorder as a category is only as meaningful as the category itself is stable. If a substantial portion of patients initially called schizoaffective eventually turn out to have schizophrenia, then reassuring a family at the time of diagnosis on the basis of the schizoaffective label is reassuring them on the basis of information that may not hold.

What genetics tells us about the boundary

The clinical picture described above, two disorders with distinct trajectories that converge at the acute psychotic episode and diverge over years, raises an obvious question about what is driving that divergence at a biological level. Genetics offers a partial answer, and the answer it offers has direct implications for how seriously we should take schizoaffective disorder as an entity.

A few posts earlier in this newsletter, I reviewed work by Watson and colleagues in a piece discussing the famous mathematician, John Nash (A Beautiful Mind). To briefly recap what is relevant here: Watson and colleagues used a statistical approach called GWAS-by-subtraction to decompose the genetic signal underlying schizophrenia into two formally independent components. One component, which they called PSYshared, captured what schizophrenia has in common with bipolar disorder, correlated positively with educational attainment. The other, the schizophrenia-specific residual, correlated negatively with IQ and educational attainment, pointed toward early neurodevelopmental mechanisms.

A more recent paper by Merola and colleagues took a step toward bipolar disorder using the same general logic. Their approach was to treat major depressive disorder (MDD) as a genetic filter, statistically removing the MDD-associated variance from the bipolar signal and asking what remains. What remains, they found, accounted for 81.5% of bipolar disorder’s genetic variance and correlated positively with educational attainment and subjective wellbeing, a profile that looks similar to Watson’s PSYshared component. The mania signal, once stripped of depression, does not look like a mood disorder in the ordinary sense. It looks like something that bridges mood and psychosis.

The bipolar I versus bipolar II distinction sharpens this picture considerably. The Merola analysis used the overall bipolar GWAS, which mixes both subtypes, but the broader genetic literature has been accumulating evidence that bipolar I and bipolar II are only partially the same entity. Bipolar I is substantially more genetically correlated with schizophrenia than bipolar II is, and when psychiatric disorders are clustered by their genetic architecture, bipolar I groups with schizoaffective disorder and schizophrenia, while bipolar II groups with major depressive disorder. Thus, when you subtract a neurodevelopmental component from schizophrenia you end up with a picture that appears highly overlapping with that of subtracting depression from bipolar disorder.

What this may say about schizoaffective disorder is that the category is probably sitting exactly on the part of the genetic landscape where our clinical language is least adequate. If schizophrenia and bipolar disorder each contain distinct genetic entities with overlapping components, then schizoaffective disorder is less likely to be a distinct entity located between them. It may instead contain patients with variable loading on these disorders’ latent components. Therefore, in one patient, the illness may eventually declare itself as bipolar with psychotic features. In another, the same early presentation may sit on top of stronger neurodevelopmental liability and evolve toward schizophrenia. In a smaller group, both liabilities may remain sufficiently prominent that the mixed picture is stable over time. Genetics does not make schizoaffective disorder disappear as a clinical reality, but it makes the DSM version of the diagnosis look biologically thin, and it shifts the meaningful question away from whether a patient occupies a midpoint between two diseases and toward which genetic and developmental liabilities are combining to produce the clinical picture observed.

When biology was asked to settle the boundary

The genetics described above raises an empirical question that a large research consortium spent years trying to answer through direct biological measurement. The Bipolar-Schizophrenia Network on Intermediate Phenotypes, known as BSNIP, set out in the early 2000s to enroll patients with schizophrenia, schizoaffective disorder, and bipolar disorder with psychosis, collect dense biological data across cognition, electrophysiology, eye movements, and brain imaging, and establish neurobiological markers that distinguished the diagnoses from one another. The founding assumption was that DSM categories, however imperfect, were tracking something real at the biological level and that the right measurements would reveal it.

That assumption did not hold. Across most measures the three diagnostic groups overlapped substantially, and biomarkers did not regularly discriminate between them. Schizophrenia and schizoaffective disorder were often difficult to distinguish biologically across the BSNIP measures, and bipolar disorder with psychosis, while trending somewhat less impaired on average, overlapped too heavily with both to permit clean separation. Tamminga and colleagues, summarizing the consortium’s experience, described the founding goal of establishing disease biomarkers for current psychotic diagnoses as poorly met.

What BSNIP found instead, when it pivoted from validating DSM categories to asking what the biological data itself revealed, were three biotypes defined by cognitive and neurophysiological profiles. Biotype 1 showed the most severe cognitive and sensorimotor impairment and the most widespread brain mass reductions. Biotype 2 showed an intermediate pattern with less extensive but still meaningful abnormalities. Biotype 3 was closest to healthy comparison subjects, with more localized reductions in brain mass. The critical finding was that all three DSM diagnoses distributed across all three biotypes. Schizophrenia patients still appeared in the least impaired biotype. Bipolar psychosis patients also showed up in the most impaired biotype. Schizoaffective disorder patients appeared throughout. When the same sample was organized by DSM diagnosis rather than biotype, the diagnostic groups looked more like a severity continuum than distinct biological entities, which is consistent with what the genetics suggests: the biology of psychosis does not respect the clean boundaries implied by the diagnostic manual.

What BSNIP has not yet delivered is the clinical tool families ultimately need: a way to assign individual patients to biologically meaningful groupings at first presentation and use that information to guide treatment. Recent work has begun developing lower-burden classifiers from clinical and cognitive measures that may prove useful for clinical trials, but prospective links to treatment response remain underdeveloped. The consortium changed the scientific question in a fundamental way. The answer that patients and families actually need is still being worked toward.

Where the science is headed

The studies described in this piece approach the problem at different levels, but they all circle the same limitation. Clinical trajectory studies tell us that schizophrenia and bipolar disorder often separate over time. BSNIP showed that when patients are grouped by cognition, physiology, and brain measures, the resulting biological categories do not map neatly onto DSM diagnoses. Together, these studies tell us that the current labels are useful but incomplete. They describe patterns in populations. They do not yet tell us what is happening in a particular patient sitting in a hospital after a first episode of psychosis.

A newer and more demanding line of work tries to move closer to that individual level. In one example, Raabe and colleagues took skin cells from patients with a schizophrenia diagnosis, reprogrammed them into neurons, and showed that gene-expression patterns in those neurons predicted individual differences in cognition, and prefrontal anatomy. The paper’s most striking result came from connecting synaptic density, measured directly in the neurons in culture, to EEG recordings taken from the patient’s scalp. The chain the paper is attempting to close runs from a patient’s own genetics through their neurons in a dish through their brain activity through their cognition.

These studies are early. The samples are small, and the field still needs replication, standardization, and proof that these cellular signatures can guide real clinical decisions. But the direction is important. The goal is to move beyond diagnosis as a group label and toward biology that can say something about the individual patient: what kind of illness process is present, which neural systems are most affected, and which treatments are more likely to work.

That brings us back to Lark Voorhies. What she needed, and what medicine could not yet give her, was a framework precise enough to say something meaningful near the beginning, rather than waiting for years of illness course to clarify the picture. A bipolar diagnosis later revised to schizoaffective disorder does not necessarily reflect bad clinical judgment. It reflects a system that still relies too heavily on time as its most powerful diagnostic tool.

The science described in this piece is the long work of building something better. Clinical course, genetics, BSNIP biotypes, and patient-derived cellular models are all attempts to replace retrospective certainty with prospective understanding. That future has not arrived. But the question driving it is the right one: can psychiatry learn to identify the illness architecture early enough to change the course, rather than merely naming it after the fact?

If you found this piece useful, please consider liking it, subscribing and sharing it with your wider network.

For more Saved By the Bell nostalgia, see Mark Wahlberg giving tribute to Mario Lopez here.

There is a robust ongoing public debate centered around the question of whether animal research is needed for understanding and treating human diseases. A subset of these discussions is motivated by the idea that animal research, in addition to simply being cruel, is unnecessary because it is readily replaceable by work on human organoids, computer simulations of a variety of disease processes and AI-based mining of public health data. As a consequence, policy appears to be changing across many parts of the western hemisphere. In Europe, the European Parliament has called for an action plan to accelerate the transition away from animal use in research, testing and education, while EU policy already states the long-term goal of phasing out animal use as soon as scientifically possible. In the US, similar pressure is building, with federal agencies and lawmakers urging greater public investment in non-animal methods such as organoids, organ-on-chip systems, computational modeling and AI-based approaches.

I could write multiple posts, and others have written books, to explain why animal research has been foundational to most scientific breakthroughs behind modern medicine. Yes, when cherrypicked, one can find examples of failures such as the lack of translatability of mouse models of sepsis, but on the flip side, the revolution in cancer, autoimmune and sleep disorders would not be possible without both the foundational discoveries conducted in animals and the significant translational pipeline for drug development that relies on their use.

But this post is not about these issues. Instead, it’s about the specific role of animal experiments in psychiatry.

It is no secret that the history of psychiatric drug development contains a long list of compounds that showed remarkable promise in animals but failed to address human clinical problems. Potential treatments for depression, anxiety and even cognitive dysfunction in psychotic disorders had pre-clinical signals based on animal research but never made it to clinical care. This graveyard is large enough that many stakeholders have drawn a seemingly-reasonable conclusion: the brain is not the kidney and modeling complex human behavioral abnormalities is not feasible in animals. In some instances, animal experiments are misleading. They generate hypotheses that consume years of development resources before showing utter failure in various stages of clinical development.

There is no doubt that one should be skeptical of animal models of psychiatric disorders. Afterall, many are the consequence of both the human mind and the human experience. How are we expecting a mouse (or a monkey) to model the trauma or grief associated with one’s loss of a career in the face of a global market crash? Animals cannot do that, at least not directly. But we should also be clear that neither can organoids, computer simulations or AI-mining of large datasets. So how can we make progress?

I like to approach the answer from a first principle: animal research for psychiatric disorders is a circuit reverse-engineering process. Meaning, we will not be able to model the human condition in animals (well, because it’s the human condition), but if we are able to localize a psychiatrically-relevant operation to a specific circuit, then we can study that circuit in animals under different perturbations. I have written about this in recent posts. The upshot is that when you start with a circuit suspected to be perturbed in a disorder and couple it with tasks capable of engaging that circuit in an interpretable manner, then you have a cross-species platform more likely to yield insights relevant to psychiatry.

In line with this thinking, I’d like to spend the rest of this post on a recently published paper that captures the spirit of this argument. It starts from a human finding in schizophrenia and reverse-translates it into mice. It then uses those mice to dissect a schizophrenia-relevant prefrontal circuit, tests whether a clinical drug can restore its function, and builds a predictive framework for understanding the drug’s mechanism of action.

That is the kind of experiment psychiatry needs more of. It does not ask mice to model the full human condition, but instead, to help us understand how to restore disrupted circuits we share with them. The paper also points in clinically-actionable directions: prospective patient stratification and biomarker development.

Let’s dig in…

Read more

Jaden was twenty-four when I first met him. It was his third admission to an inpatient psychiatric facility due to worsening psychosis. A typical story, in some sense, of medication non-compliance, escalating behavior at home and the family fearing for safety. A typically devastating story too, where just a few years earlier, this mild mannered and pleasant young man was going through high school, with all the promise and hope parents could have for their child. Smart, athletic and with big dreams, ready to take on college and beyond.

Shortly after college had started, Jaden began to struggle. Initially describing that things around him “just didn’t feel right.” He withdrew from his friends, and started missing classes. He then began interpreting ordinary events as carrying hidden meaning directed at him, and stopped leaving his apartment altogether. His roommate found him with a razorblade, cutting into his neck to find the hidden “thought transmitter” that had ruined his life. He did so just in time to call 911 and get him help.

Jaden did not like psychiatric medications. “Dr. Halassa, they just don’t work… I don’t really see the point.” His parents disagreed, and felt that there were some benefits. “You don’t seem to hear the voices anymore, that’s progress.”

“Mom, there is more to this than voices! I need to get my life back!”

And there it was. There is more to life than voices indeed. We do have medications that reduce psychotic symptoms, so from that perspective antipsychotic medications are helpful. However, in many people they do not restore functional capacity. The ability to reason, plan and interact with the world the way they did prior to their first psychotic break. Jaden’s parents agreed: “we would love to have our son back… will we?”

I cannot describe how gut-wrenching these conversations can be. It’s not that there are no examples of people overcoming psychotic illnesses and regaining function. There are, and I did write about that previously and discuss relevant literature. So there is certainly hope for people, but the issue is that we do not yet have a clear way of identifying the pathology in the broad entity we call schizophrenia, nor do we have effective medications that can target the broader syndrome beyond psychosis. To understand just how formidable this task is, I want to walk through what we currently think is going on, where the dominant theory falls short, and what a more complete picture might look like.

Schizophrenia beyond excess dopamine

For most of the past half century, dopamine sat at the center of the pathophysiological models for what we think schizophrenia is. The fact that giving people dopamine augmenting medications (like stimulants) precipitates psychosis and that traditional antipsychotic medications appear to work by blocking dopamine D2 receptors led naturally to the inference that excess dopamine was the problem. Plus, evidence indicated that people early in psychotic illness did have excess dopamine as measured through positron emission tomography (PET).

Shitij Kapur elaborated this connection via the aberrant salience hypothesis, which proposed that dysregulated dopamine signaling causes the brain to assign inappropriate significance to irrelevant stimuli. The world starts to feel charged with meaning that isn’t there. Random events feel connected. The mind builds explanatory narratives to account for these experiences, and because from the patient’s perspective the perceptions are real, those narratives have the internal logic of genuine beliefs.

As many readers of this newsletter know, I don’t particularly love the idea of molecular explanations of mental phenomena, so what are the circuit and computational correlates of this framework? Honestly, I’m not particularly clear on how exactly changes in striatal dopamine are supposed to relate to the experience of psychosis. There are many proposals that relate to the role of dopamine in assigning values to states, but no testable roadmap for how this translates to what people describe as hearing voices. Does this make the connection between dopamine and psychotic symptoms weaker? Possibly, but again, things that tend to increase striatal dopamine tend to increase psychosis (I’ve definitely seen people become psychotic on L-dopa who have never had a history of psychotic experiences, for example) and D2 blockers tend to make these symptoms better.

But the main issue is that schizophrenia is way more than voices, as Jaden pointed out. There are abnormal beliefs which we call delusions, many of which are not particularly ‘fixed’ by D2 blockade. Also, people experience loss of function due to changes in their ability to reason and think clearly. These are the cognitive deficits associated with the disorder which are also associated with ‘negative’ symptoms. Negative symptoms are hard to describe with natural language (you really have to see them to know what they are), but my best attempt at relaying them to you is the following: it’s the lack of initiation of action (and possibly thought), with what appears to be an intact ability to respond when input is provided.

Both cognitive and negative symptoms are linked to functional loss (loss of relationships, employment, education attainment) and both are thought to be a consequence of neural changes that are more causal to the illness than dopamine. Meaning, dopamine dysregulation and the positive symptoms appear to be secondary manifestations of something more fundamental that sits at the core of what we think the schizophrenia(s) are.

The Kraepelinian Core

Emil Kraepelin noticed the central problem more than a century ago, but lacked the tools we currently have to explain it. When he separated dementia praecox from manic-depressive illness, the distinction was primarily prognostic. Manic-depressive patients relapsed and remitted, while the dementia praecox patients deteriorated, often regardless of what happened to their psychosis. Kraepelin’s insight was that the trajectory was the illness, not the psychotic episodes that waxed and waned. He thought that an illness core may be eroding function steadily, in a way that was independent of psychosis altogether. That observation has held up across more than a century of follow-up data, and it is part of what makes schizophrenia so different from many other conditions we treat in psychiatry.

The evidence accumulated since Kraepelin points, imperfectly but consistently, toward the prefrontal cortex and the circuits that support how it does its work. The prefrontal cortex is responsible for the kinds of thinking that is most prized by us humans: maintaining goals across time, adapting to changing contexts, filtering out irrelevant information, and making decisions under uncertainty. Like all other cortical areas, the prefrontal cortex is built from both excitatory and inhibitory neurons; excitatory neurons are the main conduits of signals as they can drive one another (to sustain working memory or integrate signals across time for example) and project instructions to other areas, whereas inhibitory neurons control the activity patterns locally. There are many more excitatory neurons than inhibitory ones, but inhibitory function is critical for maintaining appropriate operating regime in cortical circuits, including operations like signal amplification and noise suppression. People have coined the term Excitatory/Inhibitory Balance (which was really developed for synaptic inputs on individual neurons, but gets used a lot more liberally to describe macroprocesses too) as a set point parameter for circuit function. In schizophrenia, converging evidence from postmortem studies, neuroimaging, and animal models points toward disruptions in inhibitory control that may give rise to loss of signal fidelity and accumulation of neural noise.

This formulation makes sense, broadly, when thinking about reasoning in schizophrenia and why it may be difficult for people to think clearly. Our lab has done a bit of work in this area (but we’ve also followed the literature very closely so it’s not just one experiment I’m highlighting). What we have observed is that patients with schizophrenia perform comparably to healthy individuals when task demands are straightforward, but some fail disproportionately as ambiguity increases. One view is that this is the consequence of a perturbed prefrontal network with particular failure modes that makes the brain susceptible to a variety of changes in the ability to reason, plan and initiate action. It also changes the ability to control downstream circuits including the dopaminergic system. Again, classical antipsychotic medications are able to address the dopaminergic abnormalities but do very little in addressing the prefrontal changes that may be at the core of these illnesses.

Schizophrenia may be hard to treat because we are not getting to its core.

Image illustrating prefrontal ‘top down control’ (prefrontal cortex: red), including that of dopaminergic function (ventral tegmental area: green). Modified from Ott and Nieder, 2019 TICS

Psychosis is not schizophrenia

If dopamine dysregulation is downstream of a more foundational core of schizophrenia, shouldn’t we be re-thinking the relationship between the two? Rather than being the defining feature of the illness, it becomes a clue: evidence that something has failed to regulate the dopaminergic system, resulting in surges that are associated with psychotic experiences. And one of the more interesting things about those surges is that they can apparently be triggered by multiple pathways.

Psychotic mania, which can happen in bipolar disorder, is a useful example. Molecular imaging studies find elevated dopamine receptor availability specifically in psychotic mania, with no significant difference between non-psychotic mania patients and healthy controls. A PET imaging study comparing bipolar disorder with psychosis directly to schizophrenia found that the dopaminergic signature in bipolar psychosis was comparable to what is seen in schizophrenia.

ADHD offers yet another route. In a subset of patients, particularly with high-dose amphetamines, psychosis can emerge. What’s intriguing about this is that there is a high co-occurrence of ADHD and bipolar disorder, which appears to be innate rather than treatment-driven. This suggests that some patients may carry a genetic liability spanning both conditions, placing them closer to the threshold to begin with. The diagnostic complexity these patients present clinically may reflect something important about their underlying biology that we have yet to understand.

What all of this suggests is that psychosis is better understood as a state than as the defining feature of any particular illness. In schizophrenia, the push toward that state may come from disrupted prefrontal function. In bipolar disorder and in stimulant-induced psychosis, it may come from other underlying mechanisms. This is a very important set of research questions that we need both public investment in and new talent to continue the quest of answering them.

But can excess dopamine drive schizophrenia?

This brings me to a question I find quite unsettling, and one the field has not fully worked through. If schizophrenia involves primary circuit disruption that produces dopamine dysregulation as a downstream consequence, what happens when you invert the sequence? When you start with dopamine and drive it hard enough, for long enough, we may be arriving at something that begins to resemble the underlying pathology we see in schizophrenia.

The animal literature on chronic amphetamine exposure offers some evidence that you can, at least partially. When dopamine is repeatedly and forcefully elevated, a secondary deficit in prefrontal function appears to emerge as a consequence. The cognitive deficits that follow, particularly in the kinds of flexible contextual thinking that depend on intact prefrontal function, begin to resemble what we observe in schizophrenia. In humans, chronic heavy methamphetamine use can produce persistent syndromes that extend well beyond the period of active drug use and include not just psychosis but negative symptoms and cognitive impairment. Strikingly, these cases are indistinguishable from schizophrenia that emerges the way it did in Jaden.

What it suggests is that dopamine dysregulation is not only a consequence of the primary pathology. When it is severe and sustained enough, it may begin to produce something like that pathology from the other direction. Whether that circularity has implications for how we think about illness progression in patients like Jaden, who experience repeated psychotic episodes with incomplete recovery between them, is very important. Because it helps inform whether psychotic episodes have a kindling effect, or not. Meaning, if people are slipping into psychosis (because of medication non-compliance for example), and when they are out of it, their frontal function is worse off than before, they need to know that so they are clear on the cost-benefits of taking medications, no matter how imperfect they currently are.

Map showing global prevalence of methamphetamine use. From this Lancet article.

For schizophrenia prevalence, check out this interactive map here.

Reasons for optimism

Jaden’s frustration was spot on. The voices were the part of his illness that our treatments could reach. They were not the part that determined whether he would get his life back.

Schizophrenia may be hard to tackle because psychosis, the feature that defines it in most people’s minds and that drives most of its treatment, is not the most devastating part of the illness. The illness may live in circuits that support how a person thinks, adapts, and engages with the world, circuits that were already showing signs of disruption before the first psychotic break and that tend to remain disrupted after psychotic surges are brought under control. The gap between what antipsychotics can do and what patients and families are actually asking for may be a consequence of having organized the entire treatment framework around a downstream signal while the upstream pathology has largely gone unaddressed.

But there is lots to be optimistic about.

The most significant development in schizophrenia pharmacology in decades is KarXT (Cobenfy), approved by the FDA in September 2024. Cobenfy combines xanomeline, a muscarinic cholinergic receptor agonist, with trospium chloride, which prevents xanomeline from acting outside the central nervous system and causing intolerable side effects. What makes it historically notable is simple: it is the first drug approved for schizophrenia in over seventy years that does not work by blocking dopamine D2 receptors. Cobenfy targets muscarinic M1 and M4 receptors, thought to act upstream of dopamine. M4 which is distributed subcortically, impacts dopamine release (and psychotic symptoms), while M1 which is enriched in the prefrontal cortex may get closer to the Kraepelinian Core of the illness.

This is consistent with both placebo controlled clinical trial data showing improvements on both negative and cognitive symptoms as well as preliminary real world reports. While we still need larger trials to confirm, the idea is that Cobenfy may be doing something different from traditional antipsychotics and that the patients most likely to benefit from that difference can be identified clinically. It is certainly going at least one step upstream of D2 blockers.

Other pharmaceutical companies are trying even further upstream, to target mechanisms of prefrontal excitatory/inhibitory balance more directly. One example I am familiar with is Ovid Therapeutics, which is developing a class of compounds that directly activate KCC2, the potassium-chloride cotransporter that controls chloride concentrations inside neurons. Chloride homeostasis is what determines whether GABA, the brain’s primary inhibitory neurotransmitter, can work effectively. Postmortem studies in schizophrenia have found reduced KCC2 in the prefrontal cortex, and animal models of schizophrenia-like pathology (in our hands) replicate the finding. Restoring KCC2 function would, in principle, address the inhibitory deficit directly rather than compensating for its downstream consequences. The first intravenous KCC2 activator, OV350, completed Phase 1 safety testing in late 2025 with a clean safety profile and electrophysiological signals consistent with central activity. An oral candidate, OV4071, is entering clinical trials in 2026, and will be applied to schizophrenia in 2027 if all goes well. I recently participated in an Ovid event and talked about our work here (you just need to enter some info to gain access to the full recording). Oliver (Ollie) Howes also spoke right before me and introduced the prefrontal E/I balance framework in relationship to his work.

None of this means we are close to having something to offer Jaden that addresses what he was asking for. The gap between what these drugs might eventually do and what he needed in that room remains large. But for the first time in a long while, the therapeutic programs being developed are starting to engage the biology that the evidence points toward, rather than cycling through variations on the same dopamine-blocking approach that has dominated since the 1950s. That is the right direction and for now, it’s the best we are all able to do.

If you found this piece useful, please consider subscribing and sharing it with someone who might find it interesting. Paid subscriptions help support the time it takes to produce this work.

Disclaimer: Jaden is a fictional character who draws upon a composite of patients I’ve interacted with over the years. I cannot overstate the sense of privilege I feel to have been a part of these folks’ and their families’ lives.

Illustration by Harol Bustos

“Dr. Halassa, you do not understand, but it is a new system that Elon Musk invented. Starlink… This is what they implanted in me, and I am getting zapped. You need to treat the zapping, not give me these ridiculous medications.”

As a practicing psychiatrist, I must have heard a version of this narrative hundreds of times. Delusions are one of the most common symptoms of schizophrenia, along with hallucinations, disorganized thinking, social withdrawal and flattened affect. Psychiatry has medications that sometimes control hallucinations, but delusions and other symptoms often persist despite treatment.

The idea of a delusion is quite profound; it touches on the very notion of what it is to believe something and how to adjust the belief when the evidence suddenly reverses. As a systems neuroscientist, when I hear these stories, I cannot help but wonder how they emerge. What type of brain process can possibly make someone believe something with absolute certainty and make it impossible for them to even consider alternative interpretations?

A framework I find useful here, and one I develop more extensively in a separate project I’m working on, is the coalition framework: the brain as a collection of distinct systems running in parallel, each representing the world differently, and each with its own strategy and timescale (for an early post about this, please see here). For many problems we face, multiple systems are active simultaneously, and the coherence of behavioral output (as well as appraisals of it) depends on mechanisms that arbitrate among them. Cognition, therefore, depends on coordination across these systems. The exact form of these systems is a work in progress, but it is clear for example that learning algorithms in cortex vs. basal ganglia vs. cerebellum are distinct, so the idea is that combinations of these microcircuit elements give rise to distinct macrocircuits with different algorithmic features adapted to distinct problems. We have recently discovered that associative thalamic circuits may arbitrate between different frontal areas adapted to learning associations from scratch versus adapting based on previously learned environmental structure (Wang et al., 2025).

This coalition framing has implications beyond psychiatry: if you take seriously the idea that we are all combinations of subsystems, then it may become less puzzling why we behave very differently in different social settings. One can think of social forces as external inputs that may synchronize different subsystems across people, giving rise to groupthink phenomena from components that wouldn’t be expected to do so under other models of function. However, I digress and the post here is not about this topic. Instead, it is about the loss of coordination in individuals and its relationship to mental health conditions.

Indeed, seeing someone in a psychotic state is among the most compelling evidence for this framework. The patient I described at the outset was not confused about what he believes is true. He was quite certain. Could this be related to changes in arbitration mechanisms that normally allow competing interpretations to be weighed appropriately? Did one system output become dominant in a manner that has become impervious to change? Understanding how that happens requires understanding what the coordination mechanisms are and where they live in the brain. That is where we turn to computational psychiatry and systems neuroscience to get some insight.

Computational psychiatry is a field that takes seriously the idea that mental symptoms reflect specific failures in the computations the brain performs. The field relies on tasks that isolate specific processes, such as how uncertainty is estimated, how evidence is weighted over time, or how competing memories are resolved. When someone with schizophrenia (or other psychiatric condition) performs these tasks, the pattern of errors can reveal something that clinical observations alone cannot. For example, in tasks requiring reasoning, we can ask if people use different strategies or take distinct steps in evaluating options that allow us to infer why they are more likely to exhibit delusional thinking. Readers of this newsletter will recognize that this level of description is algorithmic. It is the idea that a mental process can be captured by a set of operations that transform task relevant inputs to latent variables that drive behavioral outputs. Two patients carrying the same diagnosis may have distinct algorithmic failures that natural language, the substrate of clinical communication, would describe identically.

But to target algorithmic-level changes, one needs knowledge of the circuits that implement them. That is where systems neuroscience comes into play: the study of neural circuits and systems that give rise to specific operations. It has given us the notion of receptive fields, population dynamics and task-relevant manifolds, grounding abstract computational ideas in measurable neural activity. In essence, computational psychiatry needs systems neuroscience to make a difference in the real world.

Let’s make this very concrete: consider a decision-making task in which a cue guides a person’s attention to one of two targets, and then he or she makes a choice based on specific features of the attended target. When the cue is unambiguous, what to pay attention to is straightforward, and therefore decision-making is easy. But making the cue ambiguous renders attention and subsequent choice harder, a procedure that can be precisely titrated to quantitatively study how the brain handles uncertainty in decision-making. Instead of observing unstructured behavior and trying to figure out what went wrong, we can precisely isolate one specific operation, how beliefs get updated when evidence is uncertain, and measure how that process differs in someone with schizophrenia.

Real power emerges when we use the same task structure across species, testing humans, other primates and mice on versions of the same cognitive challenge. Although details may differ, the core computational problem remains constant, enabling us to observe similar neural operations across species, and how they implement different steps of the same algorithm. Crucially, we can do experiments in animals that would be impossible to conduct in humans, such as recording from specific cell types, manipulating their activity and decoding what information they carry. Because the animal is doing something constrained and measurable, we can link what is seen in the brain directly with what the animal computes.

Psychiatric neuroscience was developed outside of this tradition; researchers have tried to use animal models to capture the full mess of real-world symptoms: forced swim tests supposedly model despair, and elevated plus mazes supposedly model anxiety. But these tasks are too unconstrained, making it impossible to decode what specific computation has gone wrong, because the behavior reflects too many things at once. Some human neuroimaging has the opposite problem. For example, resting-state connectivity can show which brain regions are active together, but it cannot really tell what they are computing or communicating. Both approaches skip over the algorithmic level where the dysfunction driving symptoms lives.

Serendipitously, and using systems neuroscience approaches, my lab stumbled upon a domain in which well-controlled behavioral tasks in animals shed light on a brain-behavior relationship in schizophrenia. Over the past ten years, we studied how cognitive control is implemented in neural circuits, initially in mice and then in tree shrews. In these tasks, cues guide attention to targets and vary from trial to trial. Research in mice identified how neurons in the prefrontal cortex turn cues into attentional bias signals, which in turn influence processing in sensory circuits.

We discovered that the thalamus is critical for this prefrontal operation. The thalamus is traditionally thought to relay information from the periphery, but our experiments showed that the mediodorsal thalamus is turned inwards, extracting various signals from the vast data within the prefrontal cortex itself. This allows it to identify the right context and help the prefrontal cortex adapt when the context switches. It also identifies the degree of uncertainty in the cues it receives, and we found that it slows prefrontal decision-processing in a manner commensurate with these estimates. In essence, the thalamus provides a natural mechanism for filtering noise in decision-making.

When we developed this approach, we were not studying schizophrenia, but others had shown that people with schizophrenia exhibit altered resting-state functional connectivity between the thalamus and prefrontal cortex. These changes had no behavioral correlates, however. Once we understood what computation the mediodorsal thalamus performs, we could ask whether that same computation might be impaired in people with schizophrenia. Remarkably, we found that those with the disorder are worse at handling ambiguous cues: they perform normally when cues are clear but fail disproportionately as uncertainty increases. This behavioral pattern mirrors what we see when we optogenetically inhibit the mediodorsal thalamus in mice performing the same task. The deficit also correlates with thalamocortical resting-state functional connectivity in people with schizophrenia.

Graphical abstract of Huang, Wimmer et al., 2024

So we have convergence: the same behavioral deficit, analogous circuits and similar computational architecture across species. That gives us evidence that we have identified a specific computational alteration in schizophrenia, which opens the door to interventions that target the underlying circuits. Instead of broad pharmacological targeting, maybe we could target the circuits that implement noise filtering and contextual inference.

But that is just one piece. Delusions likely involve multiple computational alterations. How does evidence update beliefs? How is uncertainty represented and used? How does the brain assign credit in complex environments? These might be different changes in different people.

What systems neuroscience can help us do now is build tasks that dissect these operations separately. We have developed tasks in which we ask participants to arbitrate between different forms of uncertainty, including those involving sensory inputs and those involving the structure of the environment or its rules. This allows us to study how people and animals manage uncertainty, identify context and assign credit in naturalistic-like conditions.

Early data from this effort have confirmed our original findings from the simpler task: people with schizophrenia have difficulty managing uncertainty. However, it has also provided additional insight into the different types of algorithmic failures that people may experience. The deeper message from this line of work is that the question “why do patients with schizophrenia exhibit delusional thinking” may not have a single answer, and that this may be highly significant. Different people appear to carry different algorithmic failures that converge on similar clinical presentations. Identifying these fingerprints and mapping them onto neural circuitry is our challenge. However, we are optimistic that we (and the field more generally) are on the right track to identify these mechanisms, which will hopefully usher in a new era of precision psychiatry.

Superstar MD/PhD student Sahil Suresh presenting at the Schizophrenia International Research Society Meeting, 2026 in Florence. Early evidence for patient stratification based on an animal-inspired hierarchical decision making task.

If you enjoyed reading this post please consider sharing it with your network. Subscribe if you’d like to be part of the ongoing conversation.

A teacher notices that her student zones out several times a day. The episodes are brief and uneventful, but the child is struggling academically. The teacher calls the parents, who call the pediatrician, and the child is referred to psychiatry for an ADHD evaluation. The psychiatrist takes a careful history and notices that the child experiences abrupt episodes of zoning out with no memory of them afterward. An electroencephalogram (EEG), which measures electrical brain activity, reveals a striking pattern during these events: a generalized 3 Hz spike-and-wave discharge, the hallmark of childhood absence epilepsy.

Childhood absence epilepsy is one of the most common seizure disorders of early childhood, affecting roughly one in a thousand children. It is often misidentified as inattentive ADHD before a formal neurological evaluation is obtained. Absence seizures are global brain events that transiently suppress consciousness and happen to be visible on EEG. ADHD, like all psychiatric illnesses, has no equivalent test. We cannot simply order an EEG and rule in ADHD or determine what the treatment should be.

Or can we?

A paper just published in the Journal of Neuroscience, using high-density EEG found something intriguing: (at least some) ADHD brains show elevated localized sleep-like slow waves during wakefulness, and those events track the sluggish, variable side of task performance in patients. Could ADHD be, at least in part, a disorder of local sleep that EEG can detect?

What Is Local Sleep?

To understand what the new paper found, it helps to go back to a foundational observation published in Nature in 2011 by Vlad Vyazovskiy, Giulio Tononi, and colleagues at the University of Wisconsin. They were asking a basic neuroscience question: what happens to cortical neurons after prolonged wakefulness? In freely behaving rats, local populations of cortical neurons began going briefly offline, cycling through the same alternation between firing and silence that defines deep sleep, while the rest of the brain remained awake and the animal continued performing a task.

From Figure 2 of Vyazovskiy et al., 2011 Nature (LFP is local field potential, which is local EEG to some degree; MUA is multiunit activity, a measure of local spiking of many neurons (unsorted, so not single cell resolution)). The left panel shows global synchronized slow waves across two areas. The right panel shows one area going offline while the other maintaining a wakefulness signature.

This phenomenon is usually interpreted in the context of Tononi and Cirelli’s synaptic homeostasis hypothesis. The idea is that waking experience is associated with net synaptic strengthening, and sleep evolved in part to renormalize synaptic weights while preserving their relative strengths. From this perspective, slow waves are the physiological signature of that process: the alternation between neuronal firing and silence reflects cortical circuits doing maintenance that cannot happen efficiently while they are in continuous use. Local sleep during wakefulness is what happens when specific circuits accumulate enough load that they begin that process before the rest of the brain is ready to sleep. The phenomenon is use-dependent and locally regulated, tracking the recent history of individual cortical areas.

Figure panel from Jan Born’s review in 2023

What does this have to do with ADHD? Earlier in this newsletter, I covered a large-scale neuroimaging study suggesting that the core problem in at least some ADHD may be a failure of arousal regulation, the brain’s capacity to maintain a stable waking state under cognitive demand. Stimulant medications may work by stabilizing the arousal state that attention depends on. If that is true, then local sleep intrusions may be one consequence of a vulnerability that we see in ADHD. At least a subset of ADHD brains may slip more readily into local brief sleep-like states.

Pinggal and colleagues set out to test that idea directly.

What the New Study Found

Pinggal and colleagues recruited 32 adults with ADHD who had been taken off their medications and 31 neurotypical controls. Both groups performed a sustained attention task while wearing high-density EEG caps that recorded electrical activity across the scalp. At unpredictable intervals, participants were asked to report their mental state: were they focused on the task, had their mind wandered elsewhere, or had they blanked out entirely, aware of nothing in particular? The researchers measured the rate of local slow-wave intrusions during wakefulness and examined how it related to performance errors and subjective experience.

The results showed that the adults with ADHD had significantly more local slow-wave activity, concentrated over parieto-temporal regions. They also showed more errors on the sustained attention task, more mind wandering, more mind blanking, greater subjective sleepiness, and greater reaction time variability than controls, though mean reaction time overall did not differ between groups. The attention task was a standard sustained attention paradigm in which participants press a button in response to every digit on a screen except one designated target. This generates two error types: failing to respond when you should have, reflecting a lapse of engagement, and responding when you should have withheld, reflecting impulsivity. The slow-wave signal tracked the first type specifically, the disengaged lapses, along with reaction time variability and subjective sleepiness. It showed no relationship to impulsive responses, and slow waves did not associate with mind wandering in the combined sample. The critical finding was that mediation analyses, examining whether slow waves statistically bridge the link between ADHD diagnosis and attentional performance, were strongest for the lapse and variability measures. The slow waves appear to be carrying the disengaged, fatiguing side of ADHD’s attentional signature.

Key panels from Figure 5 of the new paper showing increased density and amplitude of local slow waves in ADHD compared to controls.

The distinction between mind-wandering and mind-blanking may be important to note. Mind-wandering is an active process where the brain is engaged with something other than the task. Mind-blanking is the absence of any reportable mental content. Both were elevated in the ADHD group, but it was the blanking state that tracked most closely with the slow-wave signal.

The study has real limitations though. Thirty-two participants per group is a modest sample. The ADHD group was 94% female and carried substantial comorbidity, with half meeting criteria for depression and nearly two-thirds for an anxiety disorder. The authors note this improves ecological validity, since real clinical ADHD rarely presents without comorbidities, but it limits how specifically these findings can be attributed to ADHD rather than to the broader profile. The 72-hour medication washout, while standard, leaves some residual stimulant effects possible. And the paper explicitly leaves open whether elevated slow-wave density in ADHD reflects a stable trait, a chronic hypo-arousal, or a faster-than-normal fatigue accumulation during sustained effort. The two accounts are not mutually exclusive, and the cross-sectional design cannot resolve them. That said, these types of exploratory studies are important to generate hypotheses that can be tested rigorously in follow ups.

A Possible Mechanism for “Consistently Inconsistent”

One of the most replicated findings in ADHD research has little to do with attention in the conventional sense. Across hundreds of studies, people with ADHD are not usually found to be uniformly slower or less accurate than controls. Instead, their performance is found to be more variable. Reaction times that are fast on some trials are followed by responses that are abnormally slow on others, with no obvious reason for the difference. This appears to be one of the largest and most consistent effect sizes in the ADHD literature, larger in many analyses than mean reaction-time differences. What mechanism can account for this feature?

Local sleep intrusions are a compelling candidate. An episodic process that silences cortical circuits for fractions of a second, distributed unevenly across a task session, may produce this pattern of performance that defines, at least some ADHD patients. The evidence is still early, but it is a reasonable mechanistic account that maps naturally onto one of ADHD’s most reproducible behavioral signatures: unstable rather than weak performance.

Now if we take this at face value, we can start asking: what sits upstream of that local arousal instability? In an earlier piece, I covered the case for orexin agonists as a candidate treatment class for ADHD. Orexin neurons in the lateral hypothalamus project broadly to every major arousal system in the brain: norepinephrine, dopamine, serotonin, and acetylcholine. Orexin does not drive arousal so much as coordinate its stability across these systems. When orexin signaling is disrupted, as in narcolepsy, the result is unstable arousal, a system that cannot maintain a consistent waking state. The parallel to what local sleep findings suggest about ADHD is hard to miss. If orexin helps suppress sleep-like intrusions during wakefulness, then it may be a viable intervention in a brain that experiences state regulation instability. Whether that hypothesis survives clinical testing will likely be known in the next year or two, and I will keep you posted as the data emerge.

So Is Some ADHD a disorder of local sleep?

ADHD is already associated with a disproportionate burden of sleep problems: delayed sleep phase, fragmented sleep architecture, and excessive daytime sleepiness. These have generally been treated as comorbidities, conditions that co-occur with ADHD without being constitutive of it. The local sleep framework suggests a different reading. Sleep difficulties and attentional difficulties in ADHD may not be two parallel problems that happen to share a patient. They may be different expressions of the same underlying neural circuitry. In fact, work from my group published in 2016 showed that common circuits regulate sensory processing across both sleep and attention, and that their disruption may explain why the two co-occur rather than attentional lapses being a consequence of poor sleep.

To be clear about what this does and does not mean: ADHD is not narcolepsy. The local sleep events in the Pinggal study are quantitative differences in a process that occurs in all brains under sufficient sleep pressure, not the dramatic state collapses seen in primary disorders of sleep regulation. The argument is not that ADHD belongs in sleep medicine. It is that, in a subset of patients, arousal regulation may provide a more useful framing than the standard attentional taxonomy from DSM.

If the arousal regulation is part of where the problem lies, that also opens a different set of intervention questions. In neurotypical adults, targeted auditory stimulation during sleep can enhance overnight slow-wave activity, and prior work suggests this is associated with reduced slow-wave intrusions during subsequent wakefulness. Whether this generalizes to ADHD populations is an open question, but the logic is coherent: if daytime local sleep intrusions reflect arousal-state instability that is partly downstream of fragmented slow-wave architecture overnight, then improving sleep quality from the bottom up may have effects on waking cognition that go beyond simply reducing fatigue.

For now, what the Pinggal study offers is something more modest but genuinely useful: a window into the neural substrate of a specific subset of attentional lapses in ADHD. Not a diagnosis, not yet a biomarker, but a candidate physiological mechanism that connects the sluggish, variable, fatiguing side of ADHD performance to a measurable EEG signal with a known biological interpretation. The child in the opening vignette had absence epilepsy, and an EEG told us that. For some children who have ADHD instead, the EEG may one day tell us something too. Understanding that, with precision, is where the next chapter of this story begins.

If you enjoyed this piece please consider sharing it with your community. If you’d like to stay informed, please consider subscribing to this newsletter.

Mental healthcare faces a workforce shortage. Globally, more than one billion people live with mental health conditions, yet the clinical workforce available to treat them remains limited. According to the World Health Organization, there is about 1 mental health worker per 10,000 people, and in many low-income countries there are fewer than one per 100,000. Even the United States, with all its resources, faces a shortage: 160 million Americans live in designated mental health professional shortage areas, and large parts of the country, particularly rural ones, lack a single practicing psychiatrist. As demand for mental health services has increased over the past several decades, the growth of the clinical workforce has not kept pace, contributing to long wait times and limited access to care.

Into this gap, large language models (LLMs) have arrived. Millions of people are already using general-purpose AI tools like ChatGPT for mental health support, without clinical validation, safety guardrails, or clarity on whether these tools help or harm. It is therefore important to highlight that the consequences for how health systems, regulators, and patients think about a technology already in widespread use are immediate and may not be compatible with a drawn out academic debate.

Many people have already covered the promises and pitfalls of LLMs in mental healthcare on Substack, so I will not rehash them here. What I will do instead is highlight a recently published paper in Nature Medicine whose press release made a striking claim: that AI has outperformed licensed therapists at delivering cognitive behavioral therapy (CBT). The claim was picked up as evidence that the therapist shortage may be solvable by current AI.

I am unconvinced by this claim and will address it in this piece. But there is an important NeuroAI angle to this paper that I found quite intriguing and would like to focus on it as well. The architecture introduced by this paper is a compelling exposition of the computational and neural requirements for reasoning (including how we ourselves reason). That is what makes it relevant to this newsletter and to the broader framework of Algorithmic Circuit Psychiatry.

What the study tested and how

Let’s face it: if you have used a general-purpose language model for anything requiring sustained, structured reasoning, you have probably noticed the problems. The model drifts, it picks up on the wrong cue, elaborates where it should probe, reassures where it should challenge, and loses track of where the conversation was supposed to go. For casual use this is a minor annoyance. As your therapist, this is a genuine failure mode, and can lead to really bad outcomes.

To be clear, these models were not designed to maintain goals, track interactions, or apply decision policies derived from clinical evidence. They were trained to generate plausible text given a context, which produces impressive fluency but does not, in and of itself, produce structured reasoning. So in a sense, the paper set out to ask: can we add components to the standard LLM transformer architecture to make it reason more like a human therapist? That is what they did: they added an input component that classifies incoming messages, inferring where the user’s presentation belongs in the broad clinical landscape (does it reflect anxiety, low mood, something more serious?), and an output layer that iterates over the LLM’s responses before they reach the user. This input-output wrap-around the LLM is referred to as the cognitive layer.

What is architecturally interesting (and the paper is only partially transparent about this because the full implementation is proprietary) is that the cognitive layer is a hybrid system composed of neural networks and more traditional AI components. The neural part consists of language models that have been fine-tuned on CBT-related material so they can produce responses consistent with therapeutic language and style. Surrounding this are several non-neural components that impose structure on the interaction. Some of these are classical embedding-based classifiers. These systems first convert text into numerical representations that capture semantic relationships between words and sentences, and then apply standard machine-learning methods, such as gradient boosting, to categorize the user’s message. In practice this allows the system to infer whether a message reflects anxiety, depressed mood, avoidance behavior, or another clinically relevant signal.

Additionally, there are the deterministic expert systems: rule-based decision tree systems written by clinicians directly from the CBT manual. Several parts of the cognitive layer use this approach, but the most central is the intervention selection module. This module maps the inferred patient state to a specific therapeutic action. It contains explicit if–then logic encoding clinical expertise. If a patient expresses a particular cognitive distortion, the system selects the corresponding CBT intervention. If the conversation has not yet established an agenda, it prioritizes agenda setting before moving further.

Figure 1 of the paper shows the components of the architecture. Note the small LLM in the middle (that’s where ChatGPT and its relatives would be). The authors highlight their cognitive layer features in this figure (input and output layers, in pink).

Of note, there was nothing special about the LLMs they used. They tried their architecture around four standard models that many readers will be familiar with: GPT-4, Claude, Gemini, and Llama 3, and their findings did not change based on which model they used. This strongly indicates that the cognitive layer did the actual heavy lifting. The standard LLMs used are all transformer-based architectures has become the standard method for developing generative AI systems (I will resist the urge to go off on a tangent related to world models).

So what is a transformer? At the most basic level, a transformer is a system that predicts what word or word-fragment should come next given everything that came before it. The technical term for these word-fragments is tokens, which are roughly syllable-sized chunks of text that the model processes as its basic unit of computation. The defining feature of the transformer architecture is the self-attention mechanism, which allows the model to dynamically weight relationships across long spans of text, so that a word near the end of a paragraph can be understood in light of something introduced at the beginning. This is what enables transformers to produce fluent, contextually sensitive language that can give a strong impression of understanding.

What transformers do not do, at least not natively, is maintain an explicit representation of task structure, track where a goal-directed interaction currently stands, or apply a formal decision policy. In domains like psychotherapy, where reasoning unfolds across structured stages and interventions must follow clinically defined principles, these absences become consequential. And this, in a nutshell is why the cognitive layer was so central.

The cognitive layer decomposes a therapeutic interaction into a set of structured computational stages that approximate the reasoning steps involved in CBT. To reiterate what was just explained above but slightly differently for this particular context: incoming user messages are first processed by safety classifiers that screen for risk and sensitive content. Clinical inference models then map the conversational content onto a probabilistic representation of the user’s psychological state, estimating where the person is across dimensions of cognition, emotion, and behavior. A decision module uses this inferred state to determine which stage of the therapeutic process the interaction currently occupies, and selects an appropriate CBT intervention accordingly. Only at this point does the language model generate text, and it does so under tightly constrained instructions that specify what the intervention should accomplish and how it should proceed. A final output evaluation stage screens the generated response against safety and clinical standards before it reaches the user, requesting regeneration if the output fails any check.

To evaluate the system, the authors ran two separate experiments. The first was a controlled experiment in which 227 participants conducted text-based sessions with one of three agent types: a licensed CBT therapist, a standalone language model, or the same language model operating within the cognitive layer. Participants did not know which type of agent they were talking to, and a separate group of 22 clinicians subsequently read the session transcripts without knowing their source. These clinicians scored each transcript on how faithfully the session followed standard CBT principles across dimensions like agenda setting, guided discovery, and intervention delivery. The cognitive layer outperformed both the standalone models and the human therapists on this measure, consistently across all four base models.

The second analysis was a large-scale observational study of nearly 20,000 sessions from close to 9,000 users of the Limbic Care app, a commercially available well-being application. Most of these users were members of the general public who had downloaded the app on their own, not patients in any formal clinical sense. A smaller group were NHS patients receiving therapy through UK mental health services, where the app supplemented their treatment. Within this broader sample, roughly 800 users had clinically significant anxiety or depression at the start and were followed over about ten weeks. Users who had the most engagement with the cognitive layer recovered at roughly twice the rate of those with the least engagement. A separate set of several thousand conversations where users rated sessions as helpful or not provided an additional measure of how the system was perceived in everyday use.

Results

In the first experiment, participants were recruited online, knew they were evaluating agents rather than seeking care, and the primary outcome measured protocol fidelity rather than whether anyone got better. This is because in the first experiment people did not come seeking mental health care. Also, the human comparator group was too small to support strong conclusions.

In the second experiment, the real-world component did track symptom change and recovery rates over time, but it was observational. The sample included a genuine clinical subgroup of NHS patients receiving therapy who used the app as a supplement to their treatment, but the majority were members of the general public who had downloaded a well-being app on their own. These facts constrain what the results as a whole can bear.

Within those constraints, a central finding from the first experiment is that performance was the same regardless of which base model the cognitive layer wrapped around. Claude, GPT-4, Gemini, Llama 3, it did not matter. This is the result that piqued my interest, not so much the comparison with human therapists (more on that in a bit).

The real-world data reinforces the relative lack of the LLM specific contribution to the outcomes. Across nearly 20,000 sessions, greater engagement with the cognitive layer was associated with better session quality, higher user ratings, and among the subset of users with clinical-level symptoms, greater symptom improvement and higher recovery rates over roughly ten weeks. Users in the highest engagement quartile recovered at nearly twice the rate of those in the lowest. The consistency of the relationship across different outcome measures and different user populations suggests that the cognitive layer does real work for real people.

What does this tell you about using ChatGPT for therapy? Well, it confirms what many people have already said: LLMs are good at converting structured instructions into fluent text (which is a formidable task, don’t get me wrong), but something else is needed to make it viable for supporting structured interactions in high-stake situations. The current paper shows that adding the cognitive layer was super important and this makes sense given what we just described about how it is built: classifiers inferring psychological state, deterministic decision trees selecting interventions, output evaluators iterating until criteria are met. With that in mind, one can start appreciating what reasoning requires: state inference, goal maintenance, policy selection, and output monitoring. These had to be engineered, and so far, artificial neural systems cannot seem to learn them the way humans do.

What does the architecture tell us about reasoning?

Looking at the components of the architecture introduced, one can see a potential NeuroAI story. Briefly, NeuroAI is a field that takes neural network models built for AI, in some conditions as a model of what the brain may be doing. This was a topic of recent discussion between and , and we’ve had some fun interactions back and forth since about whether, for example, deep networks reveal anything about primate fast visual object recognition works or whether LLMs are related to how the brain generates language as a vehicle for communication (a plug in for Ev Fedorenko, an absolute superstar in this domain).

So where is the NeuroAI angle in our current story? Well, the LLM itself is clearly inadequate and needs to be wrapped within an input and output of the cognitive layer. Very loosely speaking: one can view the input layer as early perceptual processing: classifying incoming signals, detecting clinical salience, filtering for relevance before anything reaches the generative system. The LLM maps loosely onto hippocampal function: pattern completion, context-dependent integration of disparate inputs, flexible retrieval from a vast prior distribution. The output layer, as mentioned earlier is an executive of sorts, perhaps mapping onto prefrontal systems in their monitoring role: evaluating outputs against goal-state criteria, detecting discrepancy, and sending corrective signals back until the response meets threshold. The architecture therefore captures something brain-like.

One limitation of the architecture is how it handles memory across a conversation. Each time the system produces a response, the different components analyze the transcript, select an intervention, and generate an answer. Once that answer is delivered, the internal process effectively starts over. The system reads the updated transcript and runs the same sequence of computations again. The system does not carry forward an internal state that has been shaped by the interaction as it unfolds. In technical terms, the architecture does not implement recurrent processing, where a system’s internal activity evolves continuously as new information arrives.

In contrast, in our brains, the prefrontal cortex continuously shapes what the hippocampus retrieves, how it encodes incoming information, and which representational configurations become accessible at subsequent moments. The feedback flows in both directions simultaneously, not episodically within a bounded correction cycle. A skilled clinician builds a patient model containing information the transcript cannot capture, because the clinician’s internal state is continuously shaped by the experience of the interaction as it unfolds. The hesitation before an answer, the change in register when a topic is approached, the way attention is drawn and released across the session: these leave traces in the clinician’s evolving internal state without necessarily leaving traces in the record of what was said. No system that reconstructs its patient model from the transcript at each turn can access this.

The real-world data in the paper offers indirect but consistent evidence of this constraint. The relationship between cognitive layer activation and clinical quality follows a square-root function, with diminishing returns setting in rapidly and most of the benefits present after relatively few module activations. A system whose critical contribution is establishing an initial state representation early would produce exactly this pattern. One that continuously deepens its model of the patient through the dynamics of the session would not. The architecture performs well when the task is to identify a presenting pattern and select an appropriate response protocol, and has less to offer when the task requires tracking how the patient’s state is evolving turn by turn, how early formulations need revision, or how the relationship between therapist and patient is itself becoming a clinical object.

The deeper distinction here is between computing a function and implementing a dynamical process. A feedforward system (which is what current AI architectures are for the most part, definitely the commercial ones) maps inputs to outputs. A recurrent system has an internal state that evolves continuously, such that the state at any moment depends on both the current input and the history of the system’s own prior states. The computation is a trajectory through a state space, and the trajectory carries information that no single function evaluation over a static input can generate.

Recurrence generates attractor dynamics. A recurrent system with appropriate connectivity has stable patterns of activity that the system naturally settles into when the input is consistent with a particular interpretation. Reasoning, on this account, is a settling process: the system moves through its state space until it reaches a stable configuration representing a coherent interpretation of the situation. The dynamics of convergence, how quickly the system settles, which transient states it passes through, whether it converges at all, are themselves informative about the structure of the problem. A feedforward architecture, however sophisticated, produces an output. A recurrent system arrives at a state through dynamics, and that process of arrival carries information the output alone does not.

Figure 3 of Dunker and Sahani. Different approaches towards studying neural population dynamics and computation. (a): Data analysis approaches require making assumptions to specify a parametric model of the underlying population dynamics. Different modalities of recorded neural activity can be used for model fitting to obtain a description of low-dimensional structure in the neural population and estimates of the model parameters. The recovered dynamical system can then be analysed in light of the task and computation of interest. (b): Task-optimised artificial network models learn a prespecified input–output transformation. Everything about the system is observable and can be analysed in detail to extract insight about the resulting dynamical system solving the task. However, the solution will depend on assumptions about the task and network design, as well as the cost function used for optimisation. (c): Mechanistic models are designed to implement a hypothesised solution to a given task, often taking biological constraints like cell-type structure into account. The design is often set up to capture selected qualitative features of neural data.

Recurrence also enables mutual constraint across levels of processing simultaneously that is not strictly sequential. What gets attended to shapes what gets inferred, and what gets inferred shapes what gets attended to, with these processes running in parallel and continuously influencing each other. The cognitive layer approximates this by running its modules sequentially over a shared external record, and the approximation is good enough to produce impressive performance on a structured clinical task. Sequential processing over a transcript and simultaneous mutual constraint between evolving representations are, however, different kinds of computation, and the difference becomes consequential precisely when the reasoning required is most difficult: ambiguous presentations, contradictory signals, interactions that shift register unexpectedly, cases where the clinician’s early formulation needs revision in light of what has emerged since.

What this means for the future of therapy and for mental health

What does this paper actually mean for the future of therapy? The cognitive layer is a serious engineering achievement, and the real-world data suggests it provides genuine value as a supplement to care, particularly for populations who might not otherwise access structured support. The NHS integration is notable. A system like this, used between sessions to reinforce therapeutic work, could extend the reach of a clinician’s time in ways that matter at scale. What the paper establishes is that AI-assisted mental health support, when constrained by explicit clinical scaffolding, is no longer a speculative proposition.

At the same time, it is important to be clear about what kind of intervention this probably represents. Systems like this deliver structured CBT techniques: identifying cognitive distortions, prompting reflection, suggesting behavioral exercises. That is useful, but it is only one component of psychotherapy. Therapy is also a relational process. Interaction between therapist and patient shapes how emotions are processed and stabilized in real time. Even in structured approaches like CBT, that relational dynamic is implicitly present. A text interface cannot replicate it. In that sense, systems like this may ultimately resemble self-help tools more than therapy in the traditional sense, closer to guided journaling, meditation, or structured reflection apps than to a clinical encounter.

There are also safety considerations. A language model without constraints behaves less like a therapist than like a generative associative system, capable of producing plausible responses but not necessarily grounded ones. When these systems are deployed without scaffolding, they can reinforce maladaptive interpretations or escalate emotional states rather than stabilize them. Reports of people developing increasingly delusional narratives through extended interaction with unconstrained chatbots illustrate this risk. has written about this and he is an authority on this, so check out his work. The cognitive layer is meant to prevent such outcomes by imposing clinical guardrails on what the model can say and when it can say it. Will it succeed? We shall see.

From a NeuroAI perspective, the architecture is interesting for a different reason. It was built to solve a practical engineering problem rather than to model the brain, yet it ends up partially recapitulating functional brain organization. The input classifiers resemble perceptual processing, the language model resembles a large associative memory system, and the output monitoring layer plays a role loosely analogous to prefrontal control. The fact that an engineered system designed to improve reasoning converges on something resembling this division of labor is worth noting. It suggests that the computational requirements for flexible, goal-directed reasoning may constrain the architectures that successfully implement it.

At the same time, the cognitive layer still lacks something central to biological cognition: continuous recurrent interaction between these components. In the brain, hippocampal retrieval and prefrontal control influence each other dynamically as an interaction unfolds. In this AI architecture the modules operate sequentially over a transcript, reconstructing the system’s interpretation of the conversation at each step rather than evolving a shared internal state over time. The system works well when the task is to identify a presenting pattern and apply a structured protocol. It is less clear how well it would handle situations where the interpretation itself must change as new information emerges.

That distinction has implications beyond AI. The cognitive layer embodies a familiar theory of failure: something is missing, and if the missing function can be supplied externally, the system works again. This modular model has implicitly guided much of biological psychiatry. Dysregulated dopamine signaling, impaired prefrontal gating, or miscalibrated predictive error signals are treated as missing or broken components that can be restored pharmacologically.

The recurrence argument suggests that some failures may instead arise from disruptions to the dynamics of the system itself. When feedback loops are altered, the system does not simply produce the wrong output from otherwise intact machinery. The entire landscape of possible mental states changes. Some states become unstable, others become attractors that the system repeatedly falls into. In such cases the problem is not the absence of a module but a shift in the dynamics of the system.

The cognitive layer is an impressive engineering solution precisely because it sidesteps that problem. It imposes structure externally on a system whose internal dynamics it cannot modify. That strategy works when the base system is intact but unspecialized. Whether it helps when the dynamics of the system itself are the problem is a different question entirely. For psychiatry, that distinction sits close to the center of what we still do not understand.

I would like to acknowledge for her peer review of this piece and the very thoughtful feedback throughout.

If you found this piece useful, please consider liking and subscribing. Feel free to share it with others who might find it interesting. Paid subscriptions help support the time required to produce this work.

Fast forward to my psychiatry residency, and I am now seeing many people who carry a diagnosis of schizophrenia. I can’t see a John Nash anywhere, and the whole story began to trouble me in a different way. For the patients I see, many had been cognitively compromised well before their first psychotic break, and after it, some appeared to decline further. Despite adequate treatment of psychotic symptoms, a substantial portion of patients never return to their baseline, never join the workforce, and never have families of their own. This is the devastating picture that dominates clinical training, and it sits uneasily alongside the story of a man who spent years still doing mathematics at Princeton after his psychosis had largely remitted. At some point I began to wonder: did John Nash really have schizophrenia?

The Nash question is in some ways unanswerable, and I have made my peace with that. In a previous piece, I argued that schizophrenia is almost certainly a collection of many entities, and that our diagnostic labels, schizophrenia, schizoaffective disorder, bipolar disorder with psychotic features, carve nature at joints that may not exist or be the best guides for treatment. These categories describe collections of behaviors that may not map on distinct mechanisms or generative processes. From that vantage point, asking whether Nash had “real” schizophrenia or bipolar disorder may be ill posed. The labels are placeholders for something we do not yet understand.

And yet the question refuses to go away, because the clinical differences are real even if the categories are artificial. It is true that I never saw something as dramatic as a beautiful mind, but some people with a schizophrenia diagnosis can have relatively intact cognitive capacity. These people are quite different from those whose cognition appears to deteriorate, akin to Kraepelin’s dementia praecox description. Something is producing that difference, even if our nosology cannot name it. A paper just published in Molecular Psychiatry by Watson and colleagues takes a step toward naming it, using genetics to do something that clinical observation alone has never managed: begin to show us what is actually varying underneath the label.

Two Trajectories Inside One Label

The previous piece I wrote on schizophrenia argued that the heterogeneity we observe clinically may reflect distinct generative processes rather than variation of a singular category. In other words, among the vast diversity within the psychosis spectrum, we can appreciate at least two distinct trajectories: patients who arrive at their first psychotic break already cognitively compromised, and who face a ceiling on recovery that most cannot exceed, and those whose cognition remains relatively intact, but with the reasoning changes that define psychotic illnesses more generally. If these were simply different severities of one process, you would expect them to respond similarly to treatment and follow predictable paths. They do not.

The question is what to do with that observation. Current clinical tools cannot resolve it. The acute presentation looks the same across both trajectories, and the longitudinal picture only becomes clear years later, often after multiple diagnostic revisions and treatment trials that may not have been appropriate for either. What we need is a way to probe the biology that is upstream of symptoms and independent of the categorical labels that symptoms generate. That is a harder problem than it sounds, and genetics is only one partial answer to it.

Heritability estimates for schizophrenia hover around ~64% (but removing shared environment, including that in the womb, reduces that further), so genetics alone cannot be the answer and the gap between genetic markers and any clinically actionable prediction remains vast. What genetics appears to be able to do, more modestly, is act as a probe for mechanistic coherence. If a diagnostic category is compressing patients with genuinely distinct underlying biology, the genetic signal inside that category should reflect the compression. It should look composite, pulling in different directions depending on which patients you happen to be sampling. And if you can decompose that signal, you can start to ask whether the clinical heterogeneity you observe maps onto something with a biological basis, even if you cannot yet name the mechanisms.

The cancer parallel is instructive here, even if imperfect. In chronic myelogenous leukemia, a cancer of the blood and bone marrow, decades of clinical heterogeneity (why some patients deteriorated faster, why responses varied) turned out to reflect, in a substantial proportion of cases, the presence of the Philadelphia chromosome. This is a translocation, or swap of genetic material, between chromosomes 9 and 22 that generates an abnormal fusion gene and drives uncontrolled cell growth. Once that mechanism was identified, a drug could be designed against it specifically. Imatinib, approved in 2001, transformed what had been a death sentence into a manageable condition for many patients. What is critical to understand here is that decomposing a clinically heterogeneous category by its underlying biology changed what was therapeutically possible. Whether that kind of decomposition is achievable in psychosis remains an open question, but Watson and colleagues take a step in that direction, demonstrating that the genetic signal underneath a single diagnostic label is not unitary.

What the Genetic Signal Inside “Schizophrenia” Actually Looks Like

Beginning in the late 2000s, genome-wide association studies (GWAS) made it possible to scan hundreds of thousands of genetic variants across large populations simultaneously, searching for common variants that each contribute a small increment of disease risk. For schizophrenia, the effort eventually identified hundreds of genomic loci. But rather than converging on a coherent biological picture, the signal appeared to pull in multiple directions at once, and understanding why took the better part of a decade.

The first hint of a composite signal came from an apparent paradox within the data itself. At the phenotypic level, schizophrenia is associated with lower educational attainment and cognitive decline. You would therefore expect the genetic risk architecture to show a negative correlation with both. It does not. The genetic correlation between schizophrenia risk and IQ is negative, as expected, but the correlation with educational attainment is weakly positive, meaning the same variants that increase schizophrenia risk also associate, at the population level, with staying in school longer. This counterintuitive finding was reported across multiple large studies and resisted easy explanation.

The resolution, proposed by Lencz and colleagues and subsequently replicated and extended, was that the aggregate GWAS signal was masking two biologically distinct subsets of variants. One subset showed the expected pattern: greater schizophrenia risk alongside lower cognition and lower educational attainment. These variants, when subjected to pathway analysis, pointed toward early neurodevelopmental mechanisms (genes related to how the brain was assembled during early development). A second subset showed the opposite: greater schizophrenia risk alongside higher educational attainment, and pointing towards mechanisms of synaptic function later on in life (mechanisms governing learning after birth). These two components were canceling each other out, resulting in the overall signal being near zero for educational attainment. This was an important conceptual advance, but the methods used could not formally separate and quantify the two components as independent.

A second line of evidence arrived from a different methodology entirely and pointed in a consistent direction. Rare variant studies, which examine low-frequency mutations with large individual effects, found that copy number variants and rare disruptive coding mutations associated with schizophrenia, at loci like 22q11.2, 1q21.1, and 15q13.3, are also enriched in autism and intellectual disability. These are high-penetrance risk factors for neurodevelopmental disruption that happen to increase the probability of psychosis among many other outcomes. Patients carrying heavier rare-variant burden tend to present with more cognitive compromise.

What the field lacked was a method capable of formally separating the two components of the common-variant signal into distinct genetic constructs, quantifying their independent relationships to cognition, and doing so in a framework clean enough to support downstream analyses like polygenic scoring and Mendelian randomization. That is the gap Watson and colleagues set out to close.

What Watson and Colleagues Add

Watson and colleagues took this unresolved picture and asked whether a newer statistical framework could do what prior approaches could not: treat the two components of the schizophrenia signal not as overlapping subsets but as formally independent genetic constructs. The approach they used, GWAS-by-subtraction implemented within a Genomic Structural Equation Modeling (SEM) framework, is worth understanding.

Genomic SEM works similarly in spirit to factor analysis: it models latent genetic dimensions across multiple GWAS datasets simultaneously. The specific application here is intuitive once explained. You start with the full genome-wide association signal for schizophrenia, which captures all common variants associated with the disorder. You then use the bipolar disorder GWAS signal as a genetic filter, statistically removing the variance that schizophrenia shares with affective psychosis. What remains after subtraction is the schizophrenia-specific component: genetic signal that predicts schizophrenia but is independent of affective psychosis risk. The complementary component, the shared portion, is what they call PSYshared.

The paper then asks three questions of these components: how do they correlate genetically with IQ and educational attainment (EA), do polygenic scores built from them predict educational outcomes in a large population sample, and does Mendelian randomization support a causal interpretation?

The genetic correlations are where the story sharpens. When you look at the overall schizophrenia signal, it appears to show almost no relationship with educational attainment. That turns out to be because two components are pulling in opposite directions and canceling each other out. The schizophrenia-specific component correlates negatively with both IQ and educational attainment, while the shared component correlates positively with educational attainment.

Polygenic scores built from each component and tested in nearly 382,000 people in the UK Biobank told the same story: people carrying more of the schizophrenia-specific component tended to have lower educational attainment, while those carrying more of the shared component tended to have higher educational attainment. Crucially, this was an entirely independent sample, meaning the opposing signals generalize beyond the original analysis.

Genetic correlations tell you that two things travel together in the genome, but not which way the arrow of causation points, or whether a third factor is driving both. Mendelian randomization addresses this by using genetic variants as a natural experiment. Because you inherit your genes at conception, before any life experiences accumulate, your genetic makeup cannot be caused by your educational outcomes. Watson and colleagues used this approach and found that genetic liability for the schizophrenia-specific component appears to drive lower cognition, while liability for the shared component appears to drive higher educational attainment. The caveat is that some variants may affect both psychosis risk and cognition through completely separate biological routes, so it is difficult to be certain the causal effect is running specifically through psychosis risk rather than through some other property of the same variants. That said, the Mendelian randomization results are broadly consistent with a causal relationship whose veracity will be tested in future experiments.

The final piece of the paper maps where in the brain these two genetic components are most active, using publicly available brain expression databases to ask which regions show the highest expression of the relevant genes. The shared component is predominantly expressed in cortical regions, particularly frontal cortex, consistent with the postnatally expressed synaptic biology that has been associated with affective psychosis in independent work. The schizophrenia-specific component shows a broader pattern, with expression extending into subcortical regions. This pattern is broadly consistent with the picture that has been building across the paper: the schizophrenia-specific component is not simply a more severe version of the shared one, but something with potentially distinct neural involvement. These analyses are exploratory, and the authors present them as such, but they add anatomical texture to the genetic analyses.

The Lesson and Nash

So where does this all leave us? Watson and colleagues provide important information about biological heterogeneity within the categorical label of schizophrenia. Maybe more importantly, they clarify the relationship between this category and two others, the psychosis spectrum and neurodevelopmental disorders.

One thing worth noting about the study is the notion of a schizophrenia-specific component. Methodologically, this component is the residual genetic signal that remains after removing shared variance with the broader psychosis spectrum. Personally, I would have probably just called it that, a residual, and not schizophrenia-specific. Because what really is schizophrenia? The neurodevelopmental framing is plausible and supported by converging evidence, but the label risks conferring more categorical definition than the method warrants.

Over many previous pieces in this newsletter, I have made a prediction that what we call schizophrenia will eventually give way to a set of subtypes defined by computational processes altered in individual people. These are algorithmic-level changes that ultimately impact how people build models of the world, update them, and approach reasoning challenges more generally. Two patients carrying the same diagnosis may have distinct algorithmic-level changes that would be described similarly by natural language, the substrate of clinical communication. Do genetics help with this computational dissection? Not directly, but they would add some confidence to conclusions drawn at this downstream process-level description. For grand challenges like schizophrenia, we will need multiple different methods to triangulate the generative process. Genetics underlie vulnerabilities, akin to the hyperparameters that set up model architectures in AI, but all the training data, inputs through early development and later fine-tuning via life, will determine the algorithmic-level changes that shape clinical presentation, and perhaps even treatment response.

So did John Nash really have schizophrenia? The film that brought his story to millions took considerable liberties with the clinical record: his hallucinations were auditory, not the elaborate visual conspiracies Hollywood rendered; and his greatest mathematical work was complete before psychosis ever emerged. There had been some discussion regarding whether his condition was truly schizophrenia or bipolar, but that debate may miss the point entirely. What's clear from the record is that Nash was not a man who did mathematics through schizophrenia, he was someone whose psychosis arrived after his most important work was done, perhaps ran an episodic illness and eventually receded, leaving behind someone who was diminished in some ways but still capable of serious thought. That trajectory does not sit outside the distribution of what we call schizophrenia. If individual people carry the neurodevelopmental and affective psychosis components in continuously varying proportions, then Nash's case stops being a paradox and becomes exactly what we should expect to find somewhere in that space. His case has been sitting at the edge of our nosology for decades, neither fully explained nor fully dismissed. Now we have tools to reframe the question entirely.

If you found this piece useful, please consider subscribing and sharing it with someone who might appreciate it too.

Acknowledgement: I am grateful for Dr. Cameron Watson’s input on this post. Please check him out and follow his work. I know I will!

Cameron is a Specialist Registrar in Psychiatry at the Maudsley Hospital in London and a UKRI Medical Research Council Clinical Research Training Fellow at the Institute of Psychiatry, Psychology and Neuroscience, King’s College London. His research integrates cognitive, genetic, and neurophysiological data to disentangle heterogeneity within psychotic disorders, with the ultimate aim of improving understanding of the diverse mechanisms driving these conditions.

Just a few weeks ago, a blockbuster study published in Nature demonstrated something I have rarely seen in my professional career: a science-backed brain intervention resulting in a massive treatment boost. By targeting a newly identified brain network, an international team of researchers doubled the effectiveness of brain stimulation for Parkinson’s disease.

Figure modified from Ren et al., 2026

I cannot overstate how massive of an effect that is. These aren’t subtle differences in rating scales but patients moving better, walking better, and living more fully (See Video).

My excitement goes beyond the specifics of Parkinson’s treatment to what the study demonstrates as an approach we could apply to serious mental illnesses. But to understand just how significant this breakthrough is, let’s first see where we are in psychiatry, what type of science was required in Parkinson’s disease and what we might want to do to achieve comparable success.

Why Brain Stimulation Outcomes Vary So Much

When transcranial magnetic stimulation (TMS) was approved for depression in 2008, it seemed like we finally had a non-invasive alternative for patients who can’t find the right medications. Early trials showed that less than half the patients responded, and two decades later those numbers haven’t dramatically changed. Focused ultrasound is another non-invasive method whose results also vary substantially. But even with invasive methods like deep brain stimulation (DBS), which works by inserting stimulating devices into the brain, some patients improve while others, with similar clinical features, do not.

The problem is unlikely to be in the devices themselves (when brain stimulation works, the effects can be substantial). But our inability to predict who will respond leaves us unable to explain treatment failures or adjust our approach systematically.

Two points may explain why outcomes vary so much. First, small errors in placement make a big difference in whether treatment works. Second, the brain regions we’re trying to target sit in different locations in different people.

The precision issue is striking. TMS effects fall off steeply with distance from brain surface, and being just millimeters off target substantially reduces efficacy.

The anatomical variability problem is equally important. Standard TMS protocols use simple rules: find the motor cortex by producing a thumb twitch, then measure a fixed distance to locate a psychiatrically-relevant area (e.g. dorsolateral prefrontal cortex in depression or schizophrenia). But functional brain organization varies substantially between individuals. Clinical trials demonstrate the consequences: MRI-guided targeting produced more than double the response and remission rates compared to the standard anatomical rule. Individualized functional MRI-maps result in even larger effects (highly relevant to this new blockbuster study).

Standard approaches use anatomical landmarks and population averages. The Parkinson’s study shows what becomes possible when targeting is based on newly discovered circuit architecture and science-backed precision mapping.

A Different Kind of Motor Map

Understanding why the Parkinson’s study worked requires understanding what Gordon, Dosenbach and colleagues discovered about motor cortex organization in 2023 (Gordon et al., 2023). The traditional view, taught in every neuroscience course, is that motor cortex contains a continuous map of the body from foot to hand to mouth, the famous Penfield homunculus that appears in every introductory textbook. That’s largely correct but seems to have been incomplete!

Dosenbach’s group used precision functional mapping to examine motor cortex organization in unprecedented detail, acquiring extended resting-state fMRI scans in individual subjects. This is the only way to get detailed individual maps and not rely on group averages. What they found was that motor cortex alternates between regions that control specific body parts and regions with completely different properties. These inter-effector regions sit between the foot, hand, and mouth representations and don’t activate during specific movements the way traditional motor areas do. Instead, they show strong functional connectivity to a set of brain regions involved in cognitive control, action planning, and arousal regulation.

From Fig. 1 of Gordon et al., 2023. 1, 3, 5 are classic hand, foot, mouth motor regions. 2, 4, 6 are the newly discovered inter-effector network regions.

The structural properties differ too: the inter-effector regions have thinner cortex compared to effector-specific areas, with higher white matter (neural cable) density consistent with broad connectivity. Functionally, they activate during whole-body movements like abdominal contractions or eyebrow raising that don’t map cleanly to a single effector.

This connectivity pattern makes the inter-effector regions particularly relevant to Parkinson’s disease. Human fMRI work showed that these regions connect to the ventral intermediate nucleus and centromedian nucleus of the thalamus, which happen to be major targets for deep brain stimulation and focused ultrasound treatments in movement disorders. They also connect to brain areas involved in autonomic control like anterior cingulate and pain processing like middle insula. This is consistent with earlier animal work which had found projections from motor cortex to internal organs including the adrenal medulla, a puzzling finding at the time but one that’s now easier to understand with the inter-effector discovery as part of a somato-cognitive action network (SCAN).

Figure 4 of Gordon et al., 2022

This new organizational framework provides a compelling link to the core features of Parkinson’s disease and ones that had been particularly difficult to unify: bradykinesia (reduced movement vigor), akinesia (failure to initiate movement), and the interplay between motor symptoms and cognitive or motivational state.

Circuit Dysfunction in Parkinson’s Disease

Parkinson’s disease involves degeneration of dopaminergic neurons in the midbrain and the standard model emphasizes effects on basal ganglia. This leads to a network dysfunction involving motor cortex, thalamus and various basal ganglia regions (e.g. putamen, globus pallidus, subthalamic nucleus). The blockbuster Parkinson’s neuroimaging study (Ren et al., 2026) showed that the basal ganglia network pathology involved a hyperconnectivity to the inter-effector regions, not traditional motor cortex.

They analyzed 166 patients with Parkinson’s and 60 age-matched controls with precision functional mapping data, plus 342 patients who had received deep brain stimulation with documented outcomes. The resting-state functional connectivity between cortical inter-effector regions and basal ganglia was significantly elevated in Parkinson’s patients compared to controls, and this hyperconnectivity was not seen for classic effector regions (which is surprising).

Every effective Parkinson’s treatment normalized this network-level signature. Deep brain stimulation and levodopa both reduced the hyperconnectivity, with the magnitude of reduction correlating with clinical improvement. The deep brain stimulation analysis revealed something particularly important: electrode locations that produced the best clinical results showed selective functional connectivity to cortical inter-effector regions rather than to standard motor areas, and this was true across different targets including subthalamic nucleus, globus pallidus interna, and ventral intermediate thalamus.

Electrocorticography recordings during DBS surgery in 17 patients verified this pattern. Single-pulse stimulation of the subthalamic nucleus evoked stronger cortical responses over inter-effector regions than over hand or foot motor areas. This convergence across modalities all pointing to the same circuit suggests this is the pathophysiological substrate worth targeting (See right below, Fig. 3 of Ren et al.).

The Trial

The TMS trial randomized 36 patients with Parkinson’s disease to receive stimulation targeted either to individually mapped inter-effector regions or to standard motor areas. For the inter-effector group, patients received 20-30 minutes of resting-state fMRI before treatment to identify their individual inter-effector regions based on connectivity patterns. The control group received standard targeting based on which limb showed the most symptoms: hand area stimulation for upper extremity symptoms, foot area for lower extremity symptoms.

Both groups received identical stimulation protocols over 14 consecutive days. The only difference was where the coil was placed.

After two weeks, the inter-effector group improved roughly twice as much as the control group.

Again, this is massive gain, with real people moving and living better.

The benefits weren’t limited to one type of symptom. Tremor improved, rigidity improved, movement speed improved, and walking and balance improved, with every component showing larger gains in the inter-effector group. This makes sense if the treatment is modulating a circuit that controls overall motor function rather than specific body parts, but the breadth of improvement remains striking.

In the patients who received follow-up brain scans, the abnormal hyperconnectivity between inter-effector regions and basal ganglia had decreased, and the degree of normalization correlated with clinical improvement. The control group showed no change in this connectivity despite some clinical benefit, suggesting the inter-effector targeting was normalizing the specific circuit dysfunction that defines the disease.

Focused Ultrasound Results

The same paper looked back at 10 patients who had received focused ultrasound treatment for Parkinson’s tremor. This procedure creates a small permanent lesion in the ventral intermediate nucleus of the thalamus. Standard targeting uses anatomical landmarks visible on MRI scans.

The researchers identified where the functional target should have been in each patient by mapping thalamic connectivity to cortical inter-effector regions, then measured how far the actual lesion was from this functional target. The distance ranged from 2mm to 8mm. Patients whose lesions landed closer to the functional target had better tremor control, with a clear linear relationship.

This retrospective analysis is small, but combined with the TMS trial and DBS electrode findings, it suggests a consistent pattern: functional targeting based on individual brain connectivity outperforms anatomical targeting based on population averages.

What This May Mean for Psychiatric Applications

The Parkinson’s study worked because of a specific sequence: first, discover the circuit architecture (Somato-Cognitive Action Network (SCAN) and inter-effector regions). Second, identify how that architecture changes in the disease (hyperconnectivity with basal ganglia). Third, target the dysfunctional circuit in individual patients. The results are dramatic perhaps because all three pieces are in place.

We don't have this level of circuit-level clarity for depression or schizophrenia yet. We don't know if there's a depression equivalent of SCAN, some set of regions with specific connectivity properties that become dysfunctional in these serious mental illnesses. The advantage, if we find such networks, is that functional connectivity patterns like SCAN can potentially be linked to computational descriptions of what these circuits actually do, creating bridges between network maps and algorithmic mechanisms.

But the challenge for Psychiatry may be even deeper. Parkinson’s has a fairly well-understood neurodegenerative cause which may make it behave like a single entity with a well-mapped network-level correlate that we (now) see in SCAN. Depression and schizophrenia may each encompass several entities, each potentially involving different network-level dysfunctions. We might need to discover multiple SCAN-equivalent architectures for these diagnostic labels.

Nonetheless, the Parkinson’s results show what becomes possible when you have all three pieces in place for a single pathophysiological entity. The challenge for psychiatry will be discovering the circuit architectures that define these disorders, determining which patients have which network-level dysfunctions, then developing targeting strategies based on that understanding. The precision mapping tools are finally in hand, but we need to figure out ways to put them to work. Only then are we able to know where to point our TMS coils and ultrasound beams.

If you found this analysis valuable, please share it with those who might benefit.

Not subscribed yet? Subscribe to get these analyses delivered to your inbox. Want to support this work? Upgrade to a paid subscription. It helps me dedicate time to translating research into actionable insights.

There has been a lot of discussion lately about what evidence-based medicine means, how it informs public policy and how it interacts with individual practice. Topics ranging from risk-benefits of vaccines to paracetamol in pregnancy to statins for cardiovascular protection. Some of these have fueled debates, discussions and even fallouts.

This post is not about any of these topics specifically. It’s about how the general problem of “evidence-based practice” applies to mental health. How it shapes conversations between providers, patients, family members and concerned members of the public. And how, in my view, it points to precision psychiatry as the way forward.

The Conversation

A few weeks ago, I had a conversation with one of my colleagues (and a dear friend) about this topic. I’ll call him Dr. Friend. I owe a lot of my clinical skills to his mentorship. He’s old school in the sense that he’s read almost every randomized clinical trial in psychiatry, and every time we discuss clinical cases (and we really get into it), he can cite the evidence base with impeccable clarity.

That day, we ran into each other in the hallway discussing a case where a colleague started a patient on methylphenidate for treatment-resistant depression. Dr. Friend was frustrated. “We have data,” he said. “We know what works for treatment-resistant depression. Why are clinicians constantly going off-label?”

I understood his frustration. Methylphenidate isn’t FDA-approved for depression. The RCT evidence is thin. But I also understood why the clinician made that call. The patient had failed three antidepressants and was profoundly impaired by ‘mental fog’: couldn’t focus at work, described her thinking as “moving through mud,” couldn’t make simple decisions. Standard treatment algorithms would say try lithium augmentation, atypical antipsychotics, switch to an older generation antidepressant, or consider neurostimulation. Our colleague went with a stimulant instead, and had told me privately that, “the mental fog felt more like ADHD than depression”.

Dr. Friend wasn’t convinced. So I asked him to imagine something and invite you to do the same.

The Thought Experiment

Imagine running a cancer trial with a novel medication where the primary outcome was survival and the main measurement for inclusion and tracking is weight. This is a world in which we somehow learned that cancers exist based on the common behavior of people afflicted, and where interpreting anatomical and histological data is lagging.

So we enroll people who are rapidly losing weight and perhaps had other people in their family experience such symptoms in the past to give us some more confidence. We also do a reasonable job keeping non-cancer illnesses out (infectious, autoimmune, etc…), so we end up with a mix of various types of cancers such as lung, gastrointestinal, and others.

The clinical trial results in a negative outcome. On average, people show no survival benefit from this drug compared to placebo. On average, they continued to lose weight just the same. The drug gets shelved.

Okay, back to the real world for a moment: what if I told you that this drug is imatinib (sold as Gleevec)? This is a precision treatment for gastrointestinal stromal tumors, or GISTs. These rare cancers arise in the digestive tract and used to have terrible outcomes. Before 2001, patients with metastatic GIST had a median survival under 18 months, and chemotherapy barely worked. Then came Gleevec, a targeted therapy. Response rates jumped to 50-70%. Median survival more than doubled. For the right patients, imatinib fundamentally changed the prognosis.

Sequential PET Scans Obtained in the Same Patient at Base Line (before Treatment, Panel A), 1 Month after Imatinib Treatment Began (Panel B), and after 16 Months of Continuous Treatment (Panel C). Figure 2 of this seminal NEJM paper

Continuing our thought experiment in the real world, let me be clear: no oncologist would design a trial that way. Even before molecular profiling existed, oncology had moved well past symptom-based classification. When someone presents with gastrointestinal symptoms and weight loss, oncologists perform imaging, biopsies, and distinguish between gastric, pancreatic, and colorectal cancer. They stage the disease and look at tumor characteristics. Only then do they select treatment.

By comparison, Psychiatry is a younger field because, in part, its neuroscience foundation is younger. This can result in practical issues akin to grouping people by weight in cancer. In some respects, it can even be worse: two patients can carry identical diagnoses without sharing a single symptom. Research examining DSM-5 diagnostic criteria found that 58.3% of diagnoses allow two people to meet criteria for the same disorder while having completely non-overlapping symptom profiles. This would result in grouping people by convergent symptoms (depressed mood, anhedonia, sleep disturbance) that can arise from different underlying dysfunctions. It’s the equivalent of grouping all coughs together without ever doing a chest X-ray to see if it’s pneumonia, lung cancer, heart failure, or asthma.

But this is not the end of our thought experiment. Would I really end on this type of pessimistic note? Absolutely not, and there is reason to be optimistic based on recent efforts.

Finding the Signal

My optimism stems from recent Precision Psychiatry studies tackling the issue of heterogeneity driving clinical trial ‘failures’ head on. I will highlight two examples.

The first is in the depression domain. In 2023, the Stanford Center for Precision Mental Health led a prespecified secondary analysis of data from a large antidepressant trial in over 1,000 patients with major depression. They found something intriguing. About 27% of patients showed a distinct profile: impaired performance on tests of cognitive control, reduced activation in frontal brain regions when subjects needed to perform action inhibition and planning. This group had significantly worse outcomes on standard antidepressants compared to patients without this profile.

At this point, one can hypothesize that the lack of response to standard antidepressants may be a signal, because it was selective to this group of individuals with the “cognitive biotype”. If this is the case, then perhaps these individuals would respond to another treatment approach altogether?

The Stanford researchers did exactly that and hypothesized that the measurements they ran to subtype patients were themselves mechanistic clues to base the treatment on. They asked if guanfacine, an alpha-2A receptor agonist thought to enhance prefrontal cortical function, would have utility in this subpopulation. They prospectively enrolled only patients who met criteria for the cognitive biotype based on imaging and cognitive testing.

The results: 76.5% achieved clinical response, 64.7% achieved remission. Among those who responded, 84.6% reached full remission. Compare this to the typical 33% remission rate with standard antidepressants in unselected populations. The drug restored prefrontal circuit function on imaging, and as brain activity normalized, both mood and cognitive performance improved.

This is potentially remarkable, but there are caveats. The entire trial is based on 17 patients and there is no placebo control, which constrains how much we can say about this specific problem. Nonetheless, the bigger issue that drives my excitement about it is the general approach. This is important and will continue regardless of whether the cognitive biotype in depression ends up being a treatment category that we use guanfacine (or stimulants, etc…) for.

Indeed, I am optimistic because this approach seems to generalize and that’s our second example.

Bumetanide was heralded as a treatment for autism, but had been shelved due to Phase 3 failures. This chemical is thought to modulate an ion exchanger that controls the distribution of sodium, potassium and chloride inside versus outside the cell. Distribution of ions in that way is important because it functions as a battery for powering metabolite transport across all organs. In the brain, it additionally sets the voltage of neurons so they can optimally spike. Bumetanide has a relatively selective impact on GABAergic signaling and neural inhibition. Proper GABAergic inhibition allows things like benzodiazepines to be sedative, which is a problem in some cases of autism because of ‘paradoxic activation’, suggesting that the ionic distributions may not be appropriately set up to power GABAergic inhibition the same way.

Early-phase (pilot and Phase 2) trials of bumetanide reported improvements on autism symptom scales in several cohorts, motivating larger Phase 3 testing. Yet results across studies have been heterogeneous, and later larger trials did not meet primary endpoints. Seems like the end of yet another promising mechanism of action when it meets the real world?

Not so fast. A group went back and reanalyzed the Phase 3 data using machine learning. Specifically, they used an algorithm called Q-Finder to search for patient subgroups defined by baseline clinical features that might have responded differently to treatment.

They found responder subgroups representing 20-40% of the trial population. These subgroups showed statistically significant treatment effects that were cross-validated between the two age cohorts. The children who responded shared specific baseline characteristics: mildly abnormal scores on the CARS domain measuring adaptation to environmental changes, often combined with other features like specific patterns on social communication scales.

If we believe that the drug worked for this subpopulation, then we may also consider that the overall negative result came from averaging their improvement with the majority who didn’t respond. Presumably, that majority didn’t require GABAergic adjustment that bumetanide targets.

The bumetanide story differs from Stanford’s in that Stanford moved from retrospective subgroup identification to prospective testing. No one has yet run a trial prospectively enrolling only children with the responder profile. But the principle is the same: psychiatric diagnostic categories contain multiple distinct biological entities, and when you average treatment effects across all of them, you risk losing meaningful signals.

Precision Psychiatry for Schizophrenia

The discussion above has been shaping my own thinking and research for the last few years. I will give two relevant examples.

Cobenfy (xanomeline/trospium) was approved by the FDA in 2024 as the first non-dopaminergic antipsychotic. It works through muscarinic acetylcholine receptors rather than dopamine blockade. The pivotal trials showed efficacy, but like all schizophrenia trials, they enrolled a heterogeneous population meeting DSM criteria.

I recently analyzed real-world data from an inpatient psychiatric unit where Cobenfy was added to existing treatment. About 40% of patients showed improvement, measured by standard clinical evaluation within the context of routine clinical care. That number alone tells you something: a drug that “works for schizophrenia” helps less than half of patients who have schizophrenia.

But wait a minute, using machine learning to analyze clinical features at baseline resulted in predictive factors of treatment response. Patients with prominent negative symptoms (flat affect, social withdrawal) were more likely to respond and so were those with stimulant use histories.

These findings were validated in an independent cohort, and used loosely for enrichment of subsequent patients. Doing so, resulted in near doubling of clinical efficacy. This pattern suggests that muscarinic targeting helps a specific subgroup characterized by certain clinical features. The drug may not work for “schizophrenia” as a unitary entity (because what is schizophrenia really?), but may work for patients whose psychosis involves specific circuit dysfunctions that muscarinic modulation can address.

This connects to another schizophrenia program being developed through OVID Therapeutics. They’re developing a series of compounds that address another ion exchanger (different from bumetanide, but similar general idea). Specifically these are direct KCC2 activators that restore GABAergic signaling when it is perturbed. Again, it is unlikely that everyone who meets DSM criteria for schizophrenia will benefit from this approach, but there is good evidence that some will. The good news is that even if these patients are not identified prospectively, the rise of machine learning tools promises to identify them retrospectively provided the appropriate measurements are collected.

I cannot overstate how exciting it is that this general thinking of subtyping extends to schizophrenia. Not only because of what reveals about the basic human condition (a topic I will write about in future posts) but also because of its profound impact on society.

Back to The Conversation

The flip side of this heterogeneity problem is that positive RCTs can overestimate real-world effectiveness. Psychiatric trials can exclude common comorbidities, recruit milder cases (or have patients and clinicians inflate severity scores to meet enrollment criteria), and use standardized outcome measures administered by trained raters. In clinical practice, you see the patients with complex comorbidities who were excluded from trials, and you’re tracking outcomes with whatever tools are practical in a busy clinic, not with the 17-item Hamilton Depression Rating Scale administered by a research coordinator.

So what’s the path forward? We need to stop pretending “major depression” or “schizophrenia” or “PTSD” are unitary disease entities. They’re not. They’re syndrome labels that group together people with diverse underlying pathophysiology. The RDoC framework from NIMH acknowledges this, advocating for research that cuts across diagnostic categories to study specific circuits and mechanisms. The biotype work I’ve described follows this logic.

The immediate practical question is whether we can identify these biotypes using accessible clinical or cognitive measures, or whether we need expensive neuroimaging that won’t scale to routine practice. The guanfacine trial suggests it might be both: they used fMRI to define the biotype, but 81% of patients who met circuit criteria also showed impairments on behavioral cognitive tests. If behavioral screening can identify most patients with circuit dysfunction, you could triage to neuroimaging only for borderline cases.

None of this means RCTs are useless in psychiatry. They remain the best way to establish whether a treatment has any signal at all, and do protect against hype and false positives. But we need to stop treating heterogeneous diagnostic categories as if they were biologically meaningful targets and start designing trials that either prospectively stratify by mechanism or collect enough data to identify responder subgroups retrospectively. The bumetanide reanalysis shows this is possible. The guanfacine trial shows prospective stratification may work (but take the specifics of that trial with a grain of salt for now).

I saw Dr. Friend a few weeks after our hallway conversation. “So you’re not saying RCTs don’t work in psychiatry,” he said. “You’re saying that sometimes they’re asking the wrong question.”

Exactly. They’re telling us about average effects in heterogeneous populations. What we need to know is which patients will respond. Those are different questions, and confusing them is why RCT-positive drugs often disappoint in practice while RCT-negative drugs might work beautifully for the right patient. The sooner we acknowledge that diagnostic heterogeneity breaks the standard evidence-based medicine playbook, the sooner we can build something better.

If you think this reframes what “evidence-based” should mean in mental health, please share it. That’s how these ideas spread. You can subscribe to receive future posts from this newsletter to stay in the conversation. As always, comments and thoughts are welcome.

Last year, I sat down with Ben Everett, host of the Journal of Clinical Psychiatry Podcast, to talk about my vision for where psychiatry could go next. The conversation lasted nearly two hours and ended up being a deep dive into Algorithmic Circuit Psychiatry, a framework I’ve been developing over the last several years to connect computational and systems neuroscience to clinical practice and treatment.

We covered a lot of ground. Why I wanted to be a physicist growing up. How I ended up in neuroscience. And how I ultimately found my way into psychiatry through an algorithmic and circuit-based lens.

I’ve done long-form conversations like this before but rarely felt compelled to share them widely. Listening to yourself talk for two hours is usually a humbling experience. This one felt different. There were moments of incoherence, including picking an algorithm that does have good neural mapping as an example of poor mapping. Still, on the whole, I felt I managed to articulate what Algorithmic Circuit Psychiatry is and why it offers a useful way to think about precision psychiatry and the future of mental health.

Links to the full episode:

Spotify

Apple

Below are highlights and broader context. As always, your feedback and thoughts are much appreciated.

From The Neural Code To The Severely Ill

COVID altered many lives, including mine. When labs shut down, I returned to inpatient clinical work after nearly a decade away. During that decade, I had spent most of my time thinking about thinking. My goal was to link neural machinery to cognition through a representational framework. Put more simply, I wanted to understand the neural code underlying our highest cognitive faculties.

Using tasks in animals and humans, we made progress on that front and continue to do so. Ben and I talk about this work on the podcast, initially in a way that feels circuitous. When I first listened back, my reaction was impatience: Why can’t I simply get to the point?! But by the end, the relevance was clear.

Coming back to inpatient psychiatry after years of systems and computational neuroscience changed how I listened. After working with generative models that attempt to explain how neural activity gives rise to reasoning and behavior, it becomes hard not to hear patients’ descriptions as outputs of a generative process. What kind of computation would have to change in order to produce this experience?

There were moments during that period that fundamentally changed me and set me on my current path. I’m deeply grateful to the patients who unknowingly did that work with me.

My research program evolved accordingly. I became convinced that conditions involving altered reasoning, psychotic illnesses in particular, are windows into human cognition itself. In fact, part of me thinks that it is the only way to truly understand the human condition. I’ll return to that idea in later posts.

Algorithmic Circuit Psychiatry

Back on the inpatient unit, I was seeing patients with profoundly altered reasoning. After years of building models to make sense of neural activity, I found myself asking the same question repeatedly: what computations are altered to produce this experience?

I turned to computational psychiatry expecting it to provide the tools I needed. Some of it was helpful, but it had no impact on my clinical decision making. I still used the DSM to categorize and treat. It is no wonder the vast majority of clinicians have not heard of computational psychiatry. This is where I felt the opportunity was in developing a framework that can ultimately change how people reason through their clinical cases and adjust their treatment selection.

Algorithmic Circuit Psychiatry starts from a different place. Start with algorithms, because they sit closest to symptoms and behavior. We now have artificial agents that exhibit algorithmic failure modes. With the right lens, some of those failures resemble what we see in human cognition. If you take that seriously, you can begin by diagnosing the algorithm.

By algorithm, I mean a structured computational process. In belief updating, for example: how does the system encode initial beliefs? How does it weight new evidence? What is the update rule? These questions can be formalized into process models that describe the computations the brain might be implementing.

The crucial constraint is that these models must have a form that facilitates neural mapping. This is where much of computational psychiatry falls short. Models can fit behavioral data beautifully while remaining disconnected from brain organization. Temporal difference learning is a rare exception where we have clear neural mapping through dopamine neurons and reward prediction errors for example. Many others, including drift diffusion, are far less clear.

When a model enables neural mapping, two paths open. In humans, you can look for biomarkers using EEG, fMRI, or other measurement tools to see whether algorithmic components are reflected in brain signals. In animals, you can pursue mechanisms directly using high-density recordings and causal manipulations to identify the circuits implementing those computations at a high mechanistic resolution (relevant to novel treatment development).

Put together, this is Algorithmic Circuit Psychiatry. An algorithmic problem identified in patients. A process model that supports neural mapping. Biomarkers that allow tracking in humans. Circuit-level understanding that tells us what is altered and how it might be amended. That gives clinicians and drug developers something concrete to work with beyond symptom checklists. It also gives patients and families more satisfying explanations to their lived experience.

This is the framework. It requires all of the pieces to work together, which helps explain why progress has been slow. I still think it is the path forward for precision psychiatry.

Psychiatry And Society

Near the end of our conversation, Ben asks what I hope clinicians take away from all this. My answer has less to do with the technical framework and more to do with what psychiatry should aspire to be.

We can think of ourselves as people who reduce symptoms. Bring down the hallucinations, quiet the delusions, stabilize the mood. I understand why that framing exists, but I don’t think it captures the full picture.

Society trusts us with its most vulnerable. People in profound distress, people whose reasoning has been altered in ways that compromise their ability to function. To simply reduce their suffering is a low bar. I’d like to think of psychiatrists (and mental health providers more generally) as professionals who can facilitate giving our patients a life that can be full and fulfilling. Get them well enough to work, to maintain relationships, to be functioning members of society. Afterall, it is their society too.

If you enjoyed this post please consider sharing it with your network. You can subscribe, follow the conversation and participate. Consider becoming a paid subscriber if you would like to support this newsletter.

Lately, there has been heightened awareness that stimulant medications like Adderall, Ritalin and Concerta require careful consideration. Beyond abuse potential, these medications have been scrutinized for their long-term risk of triggering psychotic symptoms in vulnerable patients. The data is more concerning than the traditional “rare side effect” framing suggests. A 2025 meta-analysis found that approximately seven percent of patients developed psychotic symptoms in studies lasting more than five years. Perhaps most troubling: up to a quarter of these psychosis cases persisted after stopping stimulants, challenging assumptions about reversibility. I covered this in detail in The Long Game of Stimulants and Psychosis.

The Wall Street Journal’s recent investigative reporting raises an additional set of concerns: children who start ADHD medications are at increased risk for requiring additional psychiatric medications to address side effects or treatment complications. The analysis found that one in five kids who start ADHD meds end up on other psychiatric medications with their own set of complications.

It is worth stating upfront that stimulants are efficacious medications that help millions of people with ADHD live functional lives. Indeed, a large 2018 meta-analysis indicated that for children and adolescents, amphetamines and methylphenidate showed moderate to large clinical effects. They remain first-line treatment for these reasons. But they are controlled substances with real risks that require careful monitoring.

I have often struggled when evaluating such tradeoffs as a provider. A subset of my patients who experience psychotic symptoms also come with stimulant prescriptions and many state they are helpful for their daily function. Leaving the diagnostic ambiguity aside (charts often indicate that such patients carry both the diagnosis of ADHD and schizophrenia), it is conceivable that stimulants may objectively improve function in this population. But the narrow therapeutic range and the risk of worsening psychosis make it a real challenge to weigh risks and benefits.

Given this complexity, there is an unmet need for stimulant alternatives for managing ADHD symptoms. This need is particularly relevant for patients with comorbid bipolar disorder or psychosis, but the risk extends beyond these diagnoses. Specifically, common genetic factors underlie many psychiatric conditions, suggesting that DSM labels may underestimate the magnitude of this vulnerability.

Insight From Narcolepsy

A recent neuroimaging study suggested that ADHD stimulants may work primarily through arousal (and reward) systems in the brain, not via attention networks as some may assume. These findings raise an important question: would pharmacological targeting of the arousal system improve ADHD symptoms?

While the answer is not yet known, there is an interesting possibility we will know soon, and it’s coming from an unexpected place: narcolepsy clinical trials. This makes sense when you think about it. Stimulants themselves were originally developed for narcolepsy before becoming standard ADHD treatment. Both conditions involve arousal regulation changes, suggesting potentially shared circuit vulnerabilities. Now a new class of drugs developed specifically for narcolepsy, may be following the same path. And there is at least one major company betting on its utility for ADHD.

Two Major Pharmaceutical Companies Are Developing Orexin Agonists For Clinical Use

Alkermes announced earlier this month that they plan to advance ALKS 7290, an orexin-2 receptor agonist, into adult ADHD patients. This represents a deliberate move by a major pharmaceutical company to pilot a promising narcolepsy drug development program in ADHD.

At the same time, Takeda has been publishing compelling data on oveporexton (TAK-861), their lead orexin agonist. While Takeda hasn’t explicitly announced ADHD trials, they just published findings in JAMA Neurology showing that oveporexton improves memory and executive function in narcolepsy patients.

Together, these developments suggest something quite clear:

orexin agonists are promising therapeutics and at least one major company is piloting their use in ADHD.

What Is Orexin and Why Does It Matter?

Orexin’s discovery is a remarkable story that I first heard about when I was in graduate school. It is one of those rare stories where two independent groups simultaneously identified the same system from completely different angles. In 1998, Emmanuel Mignot’s group at Stanford was studying narcoleptic dogs, trying to understand why these animals would suddenly collapse into sleep when excited (this video ~17:00 shows the compelling phenotype). They found mutations in the orexin-2 receptor. Around the same time, a Japanese group led by Masashi Yanagisawa created knockout mice lacking orexin neurons and discovered the mice exhibited narcolepsy-like symptoms.

The convergence was striking: whether you knocked out the orexin peptide or its main receptor, you got narcolepsy. This pointed to orexin as a critical regulator of arousal state stability.

Orexin neurons sit in the lateral hypothalamus and project broadly to multiple arousal systems. They connect to the locus coeruleus (norepinephrine), ventral tegmental area (dopamine), dorsal raphe (serotonin), and basal forebrain (acetylcholine). In that sense, orexin may be thought of as a master coordinator for these systems rather than driving arousal directly; hence its disruption appears to lead to state instability.

Figure from: Sakurai, 2007 Nature Reviews Neuroscience

The relevance to ADHD comes from the fact that several patients may have problems with maintaining arousal in specific contexts rather than simply being sleepy all the time. In some individuals, this may underlie the difficulty maintaining appropriate arousal states when performing tasks and pursuing goals. If orexin coordinates arousal stability, enhancing orexin signaling might address something fundamental in these individuals that stimulants only partially fix.

Orexin Agonists Enhance Human Cognitive Function

A December 2025 paper in JAMA Neurology provides direct evidence for orexin agonists improving human cognition. The study treated 112 narcolepsy type 1 patients with oveporexton or placebo for eight weeks and measured cognitive performance using several neuropsychological tests, including psychomotor vigilance task (sustained arousal), continuous paired associate learning test (memory) and One Back test/Digit Symbol Substitution Tests (executive function).

The results may extend beyond orexin replacement, as seen in narcolepsy type 1. Indeed, earlier work showed that orexin receptor agonists also help narcolepsy type 2 patients, some of whom may have normal orexin levels. A 2022 study with danavorexton found marked improvements in sleep latency in both NT1 and NT2 patients. This implies the therapeutic mechanism may extend beyond replacing a missing signal into stabilizing the system’s output. That distinction may be particularly relevant for ADHD, where orexin levels may be intact but arousal regulation impaired, at least for a subset of patients.

Returning to the historical parallel mentioned before. Stimulants were developed for narcolepsy in the 1930s and later became standard ADHD treatment once we recognized that both conditions involved arousal regulation problems. Now orexin agonists, developed specifically to address orexin deficiency in narcolepsy could be following a similar trajectory. If successful, the parallel may give further support to the idea of shared circuit substrates.

Will These Drugs Work for Everyone With ADHD?

Almost certainly not, but recent evidence suggests we can predict who might benefit.

A study by Leikauf and colleagues analyzed youth with ADHD and identified two distinct subgroups based on cognitive testing. Both groups showed changes in arousal, with the first exhibiting hyperarousal (elevated heart rate, high anxiety, impulsive responding, and elevated slow-wave brain activity) and the other slow and variable responses suggesting difficulty maintaining adequate arousal levels.

These subgroups responded differently to atomoxetine; only the first group responded while the second did not. If this subset, who only responded to stimulants, exhibit arousal instability, particularly in task-relevant contexts, then orexin agonists may become particularly relevant. Perhaps they will function more consistently than stimulants and potentially avoid the wear-off effects many patients describe.

The discussion above is consistent with what I explored in my piece on whether ADHD is a valid entity: the diagnosis aggregates multiple distinct problems that happen to produce overlapping symptoms. Some patients have hyperarousal with impulsivity and anxiety. Others have difficulty sustaining engagement. These may be different computational problems requiring different solutions.

Orexin agonists are unlikely to help everyone with ADHD, but they may work for some. And if that does happen, it will make for progress toward matching mechanisms to interventions.

What Comes Next

Alkermes expects proof-of-concept data from their ADHD trial in 2026. If the signal is there, we’ll know soon. Takeda’s oveporexton is further along in development for narcolepsy, with Phase 3 data already reported. Whether they pursue ADHD indications will likely depend on what Alkermes finds (or maybe they have their own plans, just haven’t announced them yet).

It is worth noting that we already have non-stimulant options targeting arousal systems, though they’re far from ideal. Atomoxetine, which we have already mentioned is FDA-approved but widely considered not as effective as stimulants for most patients (but see above about subtyping). Guanfacine, which may enhance executive function, shows moderate efficacy.

The broader story here is about how psychiatric treatment development is evolving. For decades, we found medications by accident and then tried to figure out what they were doing. Stimulants helped ADHD, so we studied stimulant mechanisms. Antipsychotics reduced hallucinations, so we studied dopamine receptors.

Now we’re working in another direction: identifying circuit-level dysfunctions and developing compounds to address them specifically. Orexin agonists for ADHD represent that newer approach. We understand that arousal regulation maybe impaired in at least some ADHD patients, and we know orexin coordinates arousal systems. We have compounds that enhance orexin signaling, so testing them in ADHD is the logical next step.

This may help my clinical dilemma; as far as I understand, orexin agonists have not been associated with hyperarousal or psychosis. Perhaps these compounds will be helpful for arousal regulation in my vulnerable patients who experience psychosis. Even if the initial promise pans out, we will still have to wait for longitudinal studies to understand these medications long-term risk profile. We’ve now known about stimulants for over a century and we are still learning about them.

As we map psychiatric symptoms onto computational and circuit-level mechanisms more precisely, treatment development will become increasingly targeted. Some approaches will fail. But failing in that manner is failure with a ‘small f’, and is the best ingredient for success. It tells us how to modify our thinking and ultimately succeed. Failure (with a big F), is simply not trying— and we absolutely owe it to ourselves to try!

If you found this analysis useful, please subscribe and share with others thinking about these questions. I’ll continue covering how neuroscience, clinical practice, and drug development intersect as we move toward more mechanistic approaches to psychiatry. Paid subscriptions support this work and help advance its broader aims.

In the spirit of keeping you informed of blockbuster papers relevant to psychiatry, I decided to highlight the recent headline grabber, “A multimodal sleep foundation model for disease prediction,” which appeared in Nature Medicine earlier this month. The media coverage emphasized the breadth and early prediction angle: “Sleep Lab Data Can Predict Illnesses Years Earlier,” “SleepFM AI Reads 1 Night of Sleep to Find 130 Diseases Risks,” “Scientists Decode Sleep Patterns to Forecast Your Future Health Risks.“

So what do I think about this paper and why am I covering it?

The paper is an impressive feat of technology. It uses a particular type of AI paradigm called Foundation Models (FM for short, which I will explain in some detail later) to utilize sleep recording data to predict risks for various diseases. Just based on a single night’s sleep data, the resulting Sleep Foundation Model (SleepFM) achieves high accuracy predicting various types of disease including heart failure, dementia, and Parkinson’s disease.

The paper reports two complementary ways of judging performance. One asks: Can the model correctly rank people by risk over time? This is captured by the concordance index (C-index). If two patients eventually develop the same disease, a C-index of 0.80 means the model correctly predicts which one will get sick first in 80% of such pairings. The second asks: At a given point in the future, can the model tell who will get sick and who will not? This is captured by the Area Under the Receiver Operator Curve (AUROC; illustrative animation below). It reflects how well the model separates those who will develop a disease within, say, six years from those who will remain healthy. Across key endpoints, performance was in the mid-0.8s (for example, mortality AUROC 0.84–0.85 and C-index 0.84), consistently outperforming task-specific models trained on sleep data without the FM pre-training step.

An animation to provide intuition for AUROC, distributions and classification accuracy

It’s exciting that AI can do these things. In fact, what an exciting time to be alive! Getting sleep data may be able to tell us about disease risks early enough so we can do something about them. For example, we can start statins early to mitigate cardiovascular risk, or initiate disease-modifying dementia treatments well before the damage is irreversible. Many good things will probably be feasible with this technology that we simply wouldn’t be able to do without.

But as I was marveling at what this technology enables, I could not help but wonder about what this meant for mental health, particularly serious mental illnesses such as schizophrenia and bipolar disorder. First, what does SleepFM say about their risk and, second, can Foundation Models be utilized similarly to solve psychiatry’s biggest challenges such as figuring out which medications to use for which patient, predict disease course and prevent relapse?

Foundation Models: What They Are and How They Apply to Health Care

Foundation Models are machine learning systems designed to learn general-purpose representations from large amounts of data. They are first pretrained on raw data to capture broad statistical structure (using unsupervised or self-supervised learning), and only later fine-tuned for specific applications (supervised, with labels).

This paradigm emerged in language processing, where models trained on massive text corpora learned representations adaptable to translation, summarization, and question answering. The same idea now applies to images, audio, and physiological data.

The appeal in healthcare reflects a basic insight: expert clinicians develop intuitions from years of pattern recognition across thousands of patients. An experienced cardiologist might sense that a particular EKG pattern predicts poor outcomes before formal criteria are met. A sleep physician might recognize subtle combinations suggesting neurodegenerative disease years before diagnosis. These intuitions are implicit statistical learning from accumulated experience.

Foundation models can be viewed as formalizing this at scale. SleepFM learned from 585,000 hours of polysomnography across 65,000 patients. The model uses a convolutional neural network to extract features from raw sleep signals, attention mechanisms to integrate information across different recording channels (EEG, EKG, respiratory), and a transformer to capture temporal patterns across the night. Critically, it was trained using “leave-one-out contrastive learning”: the model learns by predicting one type of signal (say, brain activity) from all the others (heart rate, breathing, muscle tone). This forces it to learn how different physiological systems interact during sleep rather than memorizing isolated patterns.

The result is a 128-dimensional representation of a night’s sleep that captures relationships between brain state, cardiac function, and respiratory dynamics. This compressed representation can then be fine-tuned with relatively small amounts of disease labels to predict these diseases in the future.

Figure 1 of Thapa et al., A multimodal sleep foundation model for disease prediction. Nature Medicine 2026

But what about psychiatry? Can this approach help differentiate risk for different psychiatric illnesses that overlap and confuse subsequent management. A provider’s problem can be whether their patient has bipolar disorder requiring a mood stabilizer or depression with impulse control problems requiring a completely different approach. This is but one of the several real-world classification issues people deal with and a psychiatric FM capable of solving them would have a huge impact.

SleepFM is not that model. First, SleepFM did not predict specific psychiatric disorders even though the headlines emphasized the ‘130 disease’ angle. The model broadly predicted a category called “mental disorders” but the signal came primarily from neurodegenerative conditions like Alzheimer’s and developmental delays. Second, let’s consider a hypothetical success story for SleepFM in psychiatry. Say we apply SleepFM to high risk psychosis patients in their teens and we end up being able to predict which people are more likely to develop full blown schizophrenia. Does that help with prevention?

While this approach would be worth trying (and I hope someone does), I am not sure it would yield the kind of real-world utility that SleepFM promises for cardiovascular or neurodegenerative disease. In psychiatry, a central problem is selecting an intervention matched to an inferred mechanism, without the kind of objective measurements that give SleepFM its predictive power.

Consider a serious mental illness like schizophrenia. Two patients who both meet DSM criteria may respond in complete opposite ways to the same medication. In that context, even a classifier that achieved perfect accuracy at predicting who will eventually carry a “schizophrenia” label would not tell me what to do for a given patient today.

This creates a genuine chicken-and-egg problem. I have to choose a treatment, but all I can directly observe are symptoms. Symptoms, in turn, do not reliably reveal the latent causes (mechanisms) that would guide rational treatment selection. I need to know whether this patient should be started on clozapine now, rather than cycling through other antipsychotics for years before we finally decide the benefit is worth the risk. I need to tell a mother whether we currently have the tools to get her child back on a path toward independence, or whether the family should begin planning for long-term support until something fundamentally new is invented.

These are real barriers, but they are specific and addressable. Let’s dissect them to understand what the field can focus on building in order to address its grand challenges.

Problem 1: Psychiatric Labels Are Likely an Impediment for Current Foundation Models

SleepFM’s success with heart failure and neurodegenerative diseases likely reflects a basic requirement of supervised fine-tuning: the stability of predicted labels. Death is quite clear, and it either happens or doesn’t. Cardiologists may argue about the detailed mechanisms leading to heart failure, but echocardiography measurements are great for predicting course and guiding treatment, even when symptom ratings differ across sites. Even Neurodegenerative Diseases, which have all sorts of debates around them, still follow courses that are far more predictable than those of serious mental illnesses like schizophrenia and bipolar disorder. This is consistent with recent success in identifying predictive biomarkers.

Let me be even more concrete. The same patient presenting with mood instability, impulsivity, and periodic depression might receive a diagnosis of major depression at age 18, bipolar at 22 after an SSRI-triggered hypomanic episode, and schizoaffective disorder at 26 when psychotic features emerge independent of mood symptoms. Of course, clinicians with different levels of experience may have learned subtle signs when interacting with the patient and may say, years later, that they saw this coming. That’s great, but it is certainly neither actionable nor scalable. It partly reflects the lack of a psychiatric EKG or Echocardiogram upon which a SleepFM-like model can operate.

This may explain why SleepFM didn’t achieve specificity for psychiatric disorders. For conditions where the labels themselves shift as the clinical picture unfolds, fine-tuning on “schizophrenia” becomes problematic when schizophrenia means different things across patients and across time in the same patient.

Problem 2: Even With Stable Labels, We Need the Right Measurements

Foundation models require high-dimensional data with exploitable structure. SleepFM operates on sleep physiology: EEG waveforms, cardiac rhythms, respiratory patterns. These are continuous, multidimensional signals that the model can learn to parse for patterns that predict outcomes.

Psychiatric labels rely on symptoms, which are far lower dimensional than what SleepFM was trained on. Even if psychiatric labels were perfectly stable, the rich structure SleepFM was exposed to is absent from psychiatry’s descriptors and their rating scales. Experts who can make good predictions and treatment choices likely rely on learning higher dimensional statistical features that are not captured by our current measurement approaches and are therefore not scalable.

Can biomarkers solve this problem? Perhaps, but collecting EEG and fMRI on their own has had limited efficacy so far. Decades of neuroimaging research have identified group-level differences between patients and controls, yet these findings have rarely translated into clinically useful predictions for individual patients. The challenge is identifying which features of those scans can be reliably used for treatment decisions. Perhaps the FM approach will solve this particular problem via its structural learning.

More importantly, psychiatric conditions are disorders of behavior and a Psychiatric Foundation Model may need behavioral data alongside neuroimaging to learn the predictive latent structure of psychiatric disorders. Algorithmic circuit psychiatry offers one framework for guiding measurement selection. Consider patients who exhibit particular reasoning styles in the form of systematic difficulties deploying counterfactuals, or characteristic ways of handling uncertainty in probabilistic learning settings. These patterns might correlate with signs and symptoms that expert clinicians recognize intuitively. Computational tasks can measure these patterns systematically, perhaps telling us that patients exhibiting certain reasoning characteristics might benefit from muscarinic agonists more than dopaminergic antagonists, while patients with unstable learning in specific paradigms might show the opposite response profile. The framework helps identify which neural and behavioral patterns to measure, in order to build the high-dimensional dataset a Foundation Model requires.

Whether certain tasks can be used for pretraining and others for fine tuning will be important questions to address. However, the broader point is that some principled approach to measurement selection seems necessary. Foundation Models trained on poorly chosen biomarkers will almost certainly fail regardless of model architecture or dataset size.

This doesn't mean all psychiatric AI work should pause until perfect measurements exist. Clinical tools using today's messy data such as EHR patterns, wearable sensors, and clinical notes will continue evolving and will offer improvements, no doubt about that. But if we are talking about SleepFM-type solutions for psychiatry’s grand challenges, we are going to need at least some solutions to problems 1 and 2.

What Building the Right Dataset Would Require

If Psychiatric Foundation Models are feasible, they would require systematic collection of objective measures alongside treatment trials across many patients. The model needs to learn which patterns across multiple measurement modalities predict response to which interventions.

Treatment response itself becomes a crucial validation signal. A Foundation Model might learn that a particular combination of EEG features, task behavior, and resting-state connectivity predicts response to one class of medication, while a different multimodal pattern predicts benefit from another.

Some work is already moving in this direction. Foundation Models pretrained on large fMRI corpora are now being fine-tuned on richly annotated clinical datasets in depression treatment. The model estimates the probability of response to specific interventions given that person’s baseline brain patterns. In early implementations, this approach outperforms conventional machine-learning methods and can differentiate expected response across SSRIs, SNRIs, neuromodulation, psychotherapy, and placebo arms.

Nice illustration parallel for GPT vs fMRI FM from this Youtube presentation

This establishes technical feasibility, which is encouraging, but it also exposes how far we still are from addressing unmet clinical needs. The models operate on whole-brain fMRI, a modality with well-known signal-to-noise problems. Perhaps targeting specific, computationally interpretable processes would provide a more principled way to allocate resources toward mechanisms that guide treatment.

More importantly, and as far as I know, the field has yet to develop the infrastructure to collect multimodal phenotypes explicitly tied to treatment sequences. Without routinely linking behavior, physiology, and circuit measurements to treatment outcomes, it will be a struggle to fine-tune Psychiatric Foundation Models.

SleepFM succeeded because sleep physiology provided the right data type. Psychiatry needs its equivalent: objective, high-dimensional measurements that capture treatment-relevant mechanisms. Algorithmic circuit psychiatry offers one systematic approach by providing a framework for what to measure and why. Whether this framework identifies the right measurements is an empirical question answerable only through systematic collection and validation work.

The work required is substantial but tractable. It’s mostly careful, systematic work that won’t generate headlines, but it’s the foundation psychiatric AI needs. Without it, we may continue to marvel at what Foundation Models can do for other fields while psychiatry’s hardest problems remain unsolved.

If this analysis was useful, I hope you'll subscribe and follow this work as it develops. I plan to keep covering how advances in AI, neuroscience, and clinical practice intersect, and what it will take for psychiatry to move beyond trial and error. If you know others who are thinking about these questions, please share this with them.