Like most serious publications, this newsletter begins with an editorial process. Conversations bring together a cross-section of voices from our bilingual organization, and that team supports the author, much as newsrooms assign editors, researchers, and fact-checkers to a draft. An LLM is now part of that process. The byline still means what it always has: the ideas, the judgement, and the voice belong to the author.

Substack has enabled the detection of AI-generated content, and some of our posts will likely trip those detectors. Soon enough, a note like this one will read as a period piece. Until then, here is how Deep Phenotype gets made: the story arc and core ideas come from a human; AI assists with research and writing; humans verify every fact; and every draft goes through several rounds of revision.

insitro CEO Daphne Koller introduced these principles in 2023 as first author of the editorial establishing NEJM AI’s policy on LLMs: Authors may use them, provided the use is acknowledged and the authors take full responsibility for the work’s correctness and originality. This newsletter is written to that standard. A machine cannot be held accountable for what it writes. A human can be, and is.

For centuries, fluent prose was evidence of thought, because producing it required thought. That correlation has faded, and readers no longer accept polish as evidence of thinking. We share the objection to slop: content no one thought hard about or had much reason to publish in the first place – produced by someone in the belief that pressing a button can replace domain expertise. In science, a confident, fluent falsehood can send important research down a long, wrong road.

In Plato’s Phaedrus, Socrates explains how King Thamus warned Theuth, credited as the inventor of writing by ancient Egyptians, that relying on the written word rather than their own memory gives people “not truth, but only the semblance of truth.” They will “appear to be omniscient and will generally know nothing.” The same argument has followed every tool since. In software, programmers were once criticized for writing in high-level languages instead of assembly, as if using a compiler was somehow “cheating.” Then they were criticized for coding with AI. Today, a programmer who refuses AI tools is not principled; they are unproductive. The craft did not disappear at any step; it moved up a level, from writing instructions to deciding what should be built and judging whether it was built correctly.

Writing is undergoing this shift. Its value was never in the mechanics of prose production, and AI broadens who can compete on the terms that matter, since brilliant ideas have never been confined to those who can construct sentences effortlessly. So rather than ask if a machine has touched text, ask: Does the thinking offer an original and valuable perspective? Is it compelling and well substantiated in data? And does a human stand behind it? On this Substack, the answer to all three is, “yes.”

As I wrote in my last article, fulfilling AI’s promise in drug discovery requires us to address the hardest problem: identifying biological mechanisms with disease-transforming clinical benefit. To do so, we need an accurate and systematic process for reducing a desired clinical outcome to a druggable mechanism. Addressing this fundamental problem has been the North Star of our work at insitro since our inception. Critically, this is a causal reasoning problem: predicting the outcome of an experiment we have never performed (drugging a given mechanism) in a very complex system that we do not fully understand (human biology). The consequences of a wrong prediction here are massive: a failed clinical trial that can cost hundreds of millions of dollars, put hundreds of patients at risk, and take years to read out. It is not surprising that many gravitate toward causal experiments as close as possible to those already performed — “me-too” drugs against known mechanisms.

Unfortunately, causal reasoning is precisely the kind of reasoning today’s AI is least equipped for. At enormous scale, across countless observations, frontier models learn what tends to co-occur — not what causes what. That is a limitation that more observational data and more parameters do not mitigate, because causality can rarely be inferred from observation alone; it has to be deliberately measured, by perturbing a system and recording the consequence. The difficulty is that interventional data, at scale, barely exists, and almost never in the system that really matters — a human being. The core of our work at insitro has been devoted to collecting and interpreting causal data that can shed light on human biology, and to constructing from it a causal AI model — our Virtual Human™ — capable of making interventional predictions on human health and disease.

The rest of this article describes how we built this capability, and provides some evidence that it works.

If inferring causality requires interventional data, and yet intervening in a human is generally the very last step in drug development, it might seem that this endeavor is doomed from the start. Fortunately, nature has given us a rare window into causality in human traits, via human genetics — randomized natural experiments, at population scale: People carry natural variants that alter gene function, some raising disease risk, some protecting against it. A significant population-scale association between a change in a gene and a change in a phenotypic trait is quite likely a causal one.

Consistent with that intuition, it has been well established that drug programs whose target has support in human genetics are 2–4x more likely to succeed in the clinic (Nelson et al. 2015; King et al. 2019; Minikel et al. 2024; Wang et al. 2024). Moreover, as the genetic evidence for a target gets stronger, the probability of success rises too. The upper end of the range comes from nature’s most unambiguous experiments: people who carry loss-of-function variants in a gene and are protected from the disease. Yet these targets are not intrinsically better. They succeed more often because the evidence leaves little doubt that we have found the right causal gene. The gain comes from correct target identification, not from higher efficacy; the same pattern appears at cohort scale, where better gene mapping alone pushes relative success toward 5x (Wang et al. 2024). That points to a principle we return to below: by raising our confidence in selecting a causal gene, we also raise the probability of success.

But the genetic window rarely provides a clear view, and as a consequence, most genetic associations are never pursued in a drug discovery program (Figure 1). Only 3.6% of genetically supported targets have been pursued for any of their supported indications. Even among genetically supported targets that are “demonstrably druggable,” i.e., the industry has developed a drug against that target, only 33% have been pursued for a supported indication (Minikel et al. 2024). Why does the industry leave such a powerful resource largely untapped?

Figure 1. Human genetics - a powerful resource, largely untapped. Each grid shows 1,000 genetically supported opportunities; teal dots are the ones ever pursued (Minikel et al. 2024).

Figure 2. The fog in the genetic window. A gene's causal impact is progressively attenuated on its way to detection, leaving a GWAS only a faint rendition of the underlying causal biology.

First, a genetic association by itself does not tell us which disease process a given gene affects, at which stage, or by how much. To make a drug against a gene, we need to know what we want the drug to do. Do we inhibit the gene or activate it? Prevent or enhance binding to a particular molecular partner? In which cell type? Without clarity on the desired mechanism, drug discovery is like groping in the dark: challenging, often impossible. It is hard to effectively treat what we do not understand, and for a novel gene, the road to mechanistic understanding has historically taken years of bespoke wet-lab work. The task is overwhelming if we need to perform it for multiple plausible targets from which we are trying to select a handful to advance.

Indeed, a map of genetic associations for a given disease typically yields dozens or even hundreds of candidate genes. Some associated genes profoundly shape a disease’s course; many others are minor risk factors. Evolution has selected against genetic variants of massive impact, so we are often left staring at dim, low-resolution renditions of a gene’s effect, trying to make out which faint signal is a muted version of something far sharper. This effect reduces our statistical power to detect the most impactful associations, because the large-effect variants are rare in the human population. Power erodes further with limited cohort size (especially in rare diseases) and the noise and biological ambiguity inherent to what are often subjective and coarse-grained annotations. Picking the best drug target from human genetics alone is far from easy, especially since it may not appear on the list of high-confidence associations at all.

If there is fog in the genetic window, how do we bring the right targets into focus? We can only effectively treat what we understand, and we can only understand what we measure. The path to better understanding, therefore, goes through better data — data that is deep, broad, high quality, and causal.

Our Two Foundational Pillars

In our work at insitro, we bring together two bodies of data (see Figure 3): health data from human cohorts and perturbational data from cells. Each alone is only half the story. The human data sheds light on the genetics that underlie clinically relevant traits in a human, but is largely silent on mechanism; the cellular data reveals biological mechanism but doesn’t directly tie back to what happens in a person at the organism scale. Together, interpreted and integrated using AI, they allow us to uncover actionable mechanisms that affect human clinical outcomes.

Figure 3: Two sources of causal data. On the human side, multi-modal cohort data at scale — pathology, radiology, omics, and clinical outcomes — linked to genetics: nature's experiments. On the cellular side, disease-relevant exposures and genetic perturbations create genome-wide experiments that link genetics to high content cellular readouts using imaging and omics.

To build this integrated view, we collect and integrate human health data, incorporating both broad population-scale cohorts such as the UK Biobank or All of Us as well as specialized (often proprietary) cohorts focusing on deep phenotypic data for particular disease states. In parallel, we have built an automated lab that can generate massive amounts of perturbational data in human cells. Specifically, our POSH platform (Pooled Optical Screening in Human cells) does this at genome-wide scale, reading out thousands of morphological features per cell across millions of perturbed cells, so a single screen captures how each gene reshapes the cell rather than moving a handful of predefined markers. This allows us to observe what a gene actually does: which biological processes respond, in which cell types, under which disease-relevant conditions. The cellular experiments that we run are guided by our human cohort AI insights, which point to the cellular systems and pathways likely to underlie a disease, and tell us which readouts to collect.

We have found that rich data is a deep well of insights, especially when read through the lens of ML. The more we measure, the more we see, the smarter our ML becomes. We aim to move beyond subjective ascertainment of disease state and human-defined cellular markers. We collect high-content modalities, such as imaging and omics, that are unbiased, quantitative, and as close as possible to the underlying biology. These measurements contain far more information than human inspection alone can extract — but ML can.

Seeing Deeper

In a process called imputation, ML is used to extract from one measurement modality, often one that is cheap or abundant, an observation usually acquired from a totally different one, typically more expensive or limited in scale. For example, in our work on the liver disease MASH (see Box 1), we used ML models to impute quantitative estimates of both liver fat and liver fibrosis. The ground truth measurements — MRI, liver biopsy and elastography — are expensive and hard to capture. We trained models to predict liver fat and fibrosis from routine labs and metabolomics panels, resulting in a 10-100x increase in the cohort of individuals with (imputed) fat and fibrosis phenotypes. This in turn enabled a much better-powered genetic analysis, mitigating the statistical power issues we discussed earlier. In this example, as in many others, the phenotypes derived from ground truth biological measurements are more quantitative, less subjective, and are available for a much larger set of patients.

Figure 4: Seeing more in data. From a raw quantitative phase image, ML imputes in silico fluorescence staining that would otherwise require a separate labeled experiment for each marker.

On the cellular side, we have built our wet-lab capabilities deliberately with the goal of measuring biology in ways that ML can exploit, even if a human cannot. Rather than imaging cells using traditional fluorescent microscopy, which is limited to a small number of labeled channels, we have invested in a technology called Quantitative Phase Imaging (QPI) that measures the phase as well as the amplitude of light. A person cannot see phase, but an ML algorithm can. And since light refracts differently depending on what parts of the cell it passes through, an ML algorithm can effectively “see” cellular composition and be trained to tag dozens of distinct cellular structures that would otherwise need to be individually tagged and imaged (see Figure 4), an experiment that is generally intractable. This tagging can also be done retrospectively, so that as our models continue to learn new features across and within cell types, we can go back to historical experiments and “measure” structures that we didn’t even think about at the time.

Seeing Outside the Box

Our ability to see into data goes beyond identifying what we already know exists. We also use unsupervised or weakly supervised ML methods to disentangle high-content data and identify distinct biologies that contribute to disease. In human health data, this can uncover patient subgroups that present similarly but where the disease etiology is different, or uncover stages in a disease cascade, each with its own biological drivers. In cellular data, we can see distinct biological processes that were unmeasurable via individual markers and shed new light on the underlying disease biology (see Figure 5 for an example in ALS).

Figure 5: Distilling new biologies from cellular measurements. Self-supervised learning disentangles high-content cellular data into distinct, interpretable biological processes. In ALS neurons, these include changes in structure (nuclear elongation, neurite morphology), function (TDP-43 levels and localization, cryptic-exon burden, loss of full-length STMN2), and viability — endophenotypes no single predefined marker would capture.

The deeper shift is in how we decide what to measure and what to look at. Most of biology today is reductionist: pick a known marker or an established phenotype, and measure that. The approach has built much of modern biology, but it has a built-in ceiling — it can only find what we already knew to name. Our approach is the opposite — call it maximalist biology: capture as much signal as possible, and let machines identify what matters, whether or not anyone has named it yet. Some of what emerges is a sharper version of a trait we knew; some of it is a signature with no name at all, at least at first — a pattern in the data that marks an important biological process before we understand what that process is. In maximalist biology, we measure as much as we feasibly can, and then distill numerous precision traits, or endophenotypes, each capturing a distinct aspect of the biology that changes in health and disease (see Figure 6). These ML endophenotypes are quantitative, objective, consistently measured, and focused on a single biology rather than an arbitrary amalgam of many. Because they are sharper, the genes that causally drive them stand out more clearly, greatly increasing our power to find those genes and to see what they actually do.

Figure 6: From cardiac images to precision traits. Deep learning converts cardiac MRI into 4D (3D plus time) personalized meshes of the four heart chambers. ML embeddings of these volumetric time series yield endophenotypes that go beyond coarse clinical binning, capturing distinct aspects of cardiac physiology related to heart failure, atrial fibrillation, cardiac fibrosis, and many other cardiovascular diseases.

With this framework in place, we can now return to the critical problems we outlined above. The first is turning a target into an actionable insight, one that allows us to effectively design and optimize a therapeutic against it. To do that, we need to understand in what cell(s) the gene acts, which pathways it perturbs, and via which mechanistic processes. A genetic hit derived via a statistical association with a high-level clinical trait like “F4 MASH” answers none of those questions.

This challenge is one that our Virtual Human platform was designed to address. From one side, high-content human data is transformed, through our AI precision phenotyping, into representations of multiple, disentangled disease processes. From the other, thousands of granular cellular readouts are distilled into interpretable summaries of what has changed biologically. Meeting in the middle, the two turn the crossing from human to cellular biology into a short step rather than a long inferential leap. Whereas “MASH” is a label that floats far above any single mechanism, a quantitative estimate of liver fat corresponds to something we can actually watch happen in a dish: fat droplets building up inside liver cells that we feed with sugar. A quantitative estimate of fibrosis likewise maps onto scar-forming collagen laid down by the liver’s stellate cells when we expose them to inflammatory signals.

This through-line from disease to cell biology gives us a starting point for the experiments that we need to perform to further credential the gene: in what system, under what stimulus, and what are the key measurements. The same mechanistic understanding carries past target discovery into the rest of the pipeline. For the drug discovery step (Mechanism → Drug), the cellular assays that validated the target become the assays we use to test and optimize molecules against it: a drug candidate is measured directly against the human-relevant readout that identified the target in the first place. For the clinical development step (Drug → Patient), the same quantitative human readouts help identify the patients most likely to respond, and detect, earlier and more precisely, whether they are responding. The causal mechanism that tells us what drug to make flows directly into how we make it and how we clinically test it. A fully worked example in the context of MASH can be found in Box 1.

Box 1 — MASH: from a clinical label to a causal mechanism

Metabolic dysfunction-associated steatohepatitis (MASH) represents a large unmet need — around 9–16 million people in the US alone (Younossi et al. 2025) — and provides a clean illustration of how our approach allows us to chart the path from outcome to mechanism. Clinically, MASH is graded in coarse biopsy stages (F2, F3, F4): subjective, invasive, available for relatively few patients, and collapsing a great deal of distinct biology into a single ordinal scale.

Rather than anchor on that label, we built quantitative, continuous phenotypes for the two processes that drive the disease: liver fat and liver fibrosis. Grounded in a small but highly accurate PDFF-MRI dataset, our imputation models based on imaging and molecular readouts reconstructed a full continuum of liver fat, from normal to pathological, across more than 250,000 people in the UK Biobank (Somineni et al. 2024), enough for the first fully powered GWAS of this prevalent disease. We corrected for extrahepatic fat depots, so the signal reflects intrinsic pathways of liver-fat deposition shared by both obese and lean MASH. The liver fibrosis phenotype was trained on clinically ascertained F4 versus healthy.

Because the phenotypes are quantitative, objective, and focused on a single biology, the genes that act on them stand out sharply: the fat and fibrosis phenotypes surfaced over thirty times as many genetic associations as standard clinical staging. Moreover, we found that targets with support across multiple phenotypes were much more likely to be concordant with genetic hits in clinically ascertained MASH. That genetic map pointed to a predominant intrinsic driver of aberrant liver fat: de novo lipogenesis (DNL), the liver’s synthesis of new fat from carbohydrate (Donnelly et al. 2005; Lambert et al. 2014). Lipidomic patterns in the same population supported the hypothesis.

We then took the DNL-associated genes into the cell, testing whether high-effect-size candidates could reverse lipid accumulation in hepatocytes and HepG2 cells under conditions that favor DNL. Of the genes whose deletion normalized lipid accumulation, one stood out: IRS-1 (insulin receptor substrate 1), which was also the strongest-effect-size gene in the corrected liver-fat genetics. A PheWAS showed that heterozygous loss of function in IRS-1 is broadly benign, suggesting that inhibiting it would be well tolerated.

To test the therapeutic hypothesis (paper under submission), we developed a potent, selective siRNA against IRS-1, delivered liver-specifically via a clinically proven GalNAc ligand to widen the safety margin the human genetics already implied. In diet-induced obese mice, this molecule (CTRO-1013) sharply lowered IRS-1 expression and protein in the liver, normalized liver triglyceride, and reduced the associated inflammation. Recapitulating the human analysis, it downregulated DNL pathways, suppressed a lipidomic signature of DNL, and, in lean mice, dose-dependently blocked experimentally induced de novo lipogenesis, completing the causal bridge from a human experiment of nature to a validated cellular mechanism, and finally to an in vivo outcome. In an Amylin model of mouse fibrosis, the molecule also reduced multiple fibrotic markers, similarly supporting the other key genetic association.

That same DNL signature is designed to carry into the clinic. Because it can be read from lipidomics in a blood sample, it gives us a pharmacodynamic biomarker for clinical development: an early, mechanistic readout of whether IRS-1 inhibition is actually suppressing de novo lipogenesis in patients, and a basis for enriching trials with, and monitoring, the patients most likely to respond. The readouts that identified the mechanism are the readouts that help us test it.

Which brings us to the second challenge: selecting high-confidence targets to advance towards an indication. The Virtual Human encompasses multiple lines of evidence regarding the causal role of any given gene — spanning data modalities, disease biologies, and physical scales. Each convergent line of evidence that ties a gene to disease-relevant traits strengthens our conviction in its causal role. A gene supported by the genetics of multiple AI precision phenotypes and confirmed by direct perturbation in human cells should be a better-credentialed target for the disease. As we noted, Mendelian associations succeed in the clinic not because they are intrinsically better targets, but because they provide higher confidence that we have selected the right causal gene. Via the convergence of different lines of evidence, even if each is not as strong, we are aiming to manufacture Mendelian-grade confidence for genetic associations that are not Mendelian.

Figure 7: Causal evidence across biological scales. We collect a large and growing library of endophenotypes across five to six orders of magnitude of physical scale, each derived from real measurements, interrogated using ML. Each lets us ask the key causal question: What happens if I change this gene? The response is a holistic view that links diverse biologies across scales and levels, allowing us to track the mechanisms that connect them.

The principle here is compelling, but how can we convince ourselves that our model is achieving that goal? To assess that question, we deploy it toward zero-shot prediction of historical clinical trial success. Our benchmark parallels the analysis of Minikel et al. (2024) for human genetic support. Specifically, we consider clinical trials that entered phase 2 (the first real test of a drug’s efficacy), and define a trial to be successful if it subsequently entered phase 3. Since our predictions are at the target level, we aggregate trials into target-indication (T-I) pairs, with a T-I pair defined as a positive if there was at least one successful trial.

The goal of this analysis is to demonstrate the value of using convergent phenotypic insights; we therefore restricted the evaluation to a subset of metabolic and cardiac indications for which our libraries contain at least one endophenotype that is relevant to the disease (a determination made programmatically, using only the known biology of the endophenotype and the disease, with no access to our genetic results or to trial outcomes). That left 72 metabolic and 29 cardiac indications, covering 301 and 339 target–indication pairs respectively.

We tested a version of our in-house models that was simplified to allow an assessment on public benchmarks. It integrates the genetics of hundreds of insitro’s ML-derived endophenotypes derived from both human data and from our proprietary cellular data to predict the strength of a gene’s causal connection to a disease. Critically, our model sees no clinical trial outcomes at any point in training. We compared our approach against a baseline built on standard genetic association studies, similar to the approach used in Minikel et al.

In this setting, it is particularly important to reduce false positives, since we can only prosecute a limited number of programs at once. We therefore asked whether we can surface a set of high-confidence target-indication pairs with a lower false-positive rate. To test this, we compared the most promising pairs provided by each method (the top decile by score). As shown in Figure 8, the top 10% of target-indication pairs from our model had a false-positive rate of just 10% in metabolic disease and 21% in cardiac, against a historical failure rate of 57% in both — a greater than fivefold reduction in metabolic disease and a roughly 2.5x reduction in cardiac. Against the genetics baseline, our model cuts the top-decile false-positive rate by nearly 75% in metabolic disease and by 30% in cardiac.

At the same time, we don’t want to reduce the false-positive rate in a way that drastically cuts down on the total number of successful programs. Our model also recovers substantially more high-confidence targets at the same success rate. Holding the false-positive rate equal to the genetic baseline’s top decile, the Virtual Human identifies more than 2.5x as many strong targets in metabolic disease — 41% of the 131 Phase II validated target-indication pairs, versus 15% for the baseline — and more than 2x as many in cardiac — 34% of the 146 Phase II validated target-indication pairs versus 16%.

Figure 8: Zero-shot prediction of historical clinical trial success. Our model of human causal biology versus a baseline built on standard genetic association evidence (following Minikel et al. 2024), on the task of predicting which phase II trials succeeded — zero-shot, with no trial outcomes in training. Top: failure rates among top-decile pairs, compared to OpenTargets genetics and the historical phase II rate. Bottom: the full precision–recall curves for both methods. Caveat: A retrospective benchmark does not guarantee prospective performance, but it is a quantitative test of whether convergent causal evidence improves target selection.

It is important to emphasize that a false-positive rate under 10% on a retrospective benchmark (even if not used in training) does not imply that we will achieve a phase II failure rate under 10% in the real world. First and foremost, this prediction accuracy was over a small subset of T-I pairs, and may not be achievable over a larger number. In fact, the predictions that we are making are restricted to historical T-I pairs that had been selected by others to go into clinical trials in the first place, and a different selection might give rise to different results. Moreover, there are many other caveats to this benchmark: A trial may have failed to advance for reasons that have nothing to do with the target biology. The mix of indications and modalities going into the clinic changes over time. Nevertheless, this result does provide supporting evidence that our approach can identify a set of targets that are more likely to see clinical success.

The core of what we have built at insitro serves the goal we opened with: collecting and interpreting causal data that can shed light on human biology, and constructing from it a causal AI model, our Virtual Human, capable of making interventional predictions on human health and disease.

The foundation is data, collected deliberately across every axis that matters: across measurement modalities, from imaging to omics; across disease states and the cohorts that carry them; and across biological scales, from organisms to organs to single cells. As our library of traits grows, it unlocks targets for diseases we could not previously interrogate. Just as importantly, it changes the nature and quality of the insights themselves: every trait in the library is another map in our atlas of human causal biology, another independent line of evidence, so our confidence in any given target, and our understanding of the role it plays, rise as the atlas expands. More diseases unlocked, more targets per disease, and more conviction per target: that is what the expanding atlas buys.

The Virtual Human is not and cannot be a single monolithic AI model. Biology is too complicated for that, and its data too heterogeneous. Instead, we have built dozens of specialized models, each designed to find the signal hidden in one kind of data — a cellular image, a molecular profile, a clinical measurement — and make it legible to the general reasoning models that work across all of them. The specialized models see; the reasoning layer connects the dots to convert a massive and growing collection of insights into mechanisms and targets.

Figure 9: The architecture of the Virtual Human. A proprietary data engine feeds dozens of specialized domain models — enhancement, representational encoding, disentanglement, pathway and cluster analysis — spanning the ClinML and CellML verticals, which make heterogeneous biological data legible to frontier general models operating within a domain-specific harness for causal reasoning.

The result is a system whose intelligence compounds with every turn of the crank, not only because our individual AI models continue to learn, but also because we built the right loop. As our library continues to grow, so does the alignment between layers — the human data tells us which biological changes matter for disease, the cellular data tells us what causes those changes, and the AI, working across both, finds the mechanisms that hold at both scales. And that loop runs through the whole of making a medicine: the mechanism that tells us what drug to make flows into how we make it and whom we test it in — the same three problems the first part outlined, addressed by a single engine instead of three disconnected ones.

In our work so far, we have demonstrated our engine’s applicability across metabolic, cardiovascular, neurological, ophthalmic, and inflammatory disease. It has already produced multiple programs, all first-in-class, and all with multi-faceted support, including from human genetics — the kind of evidence that we believe meaningfully improves the odds. We are optimistic about our first programs, while knowing that the real proof will come only when we test them in patients.

But if this works, the payoff shows up in two ways. First, in the economics of failure. As argued in the first article of this series, roughly nine in ten drugs that enter the clinic fail, most of them because the underlying biology was wrong, and the cost of every medicine that reaches a patient is really the cost of all the programs that did not. Formal analyses of R&D productivity reach the same conclusion: attrition in the late clinical phases is the single most important determinant of the cost of each new medicine (Paul et al. 2010). Phase 2 is its epicenter: the lowest success rate of any phase, sitting immediately upstream of the most expensive trials. Nothing moves the economics of this industry more than improving the odds that a mechanism survives its first real test of efficacy, which is precisely the stage our benchmark predicts. Derisking target biology addresses Phase 2 failure at its source, and the effect compounds beyond a single program. There are fewer failed trials for every success, fewer resources sunk into programs that were never going to work, and a lower cost for each medicine that does reach a patient. And there is a cost beyond money: every trial asks patients to accept real risk on an unproven hypothesis. When the biology was wrong from the start, that risk was taken for a drug that was never going to help them. Better target selection honors what patients put on the line.

An even larger prize is also measured in patients: those afflicted with the many diseases that today lack any disease-modifying therapy. The reasons differ. For some, decades of plausible attempts have ended in failure. For others, drugs exist but barely bend the disease’s course. And for a long tail of rare and neglected diseases, no target has ever looked credible enough to justify a billion-dollar gamble. Underneath, the barrier is the same: we have not known, with confidence, which mechanisms actually drive a given disease. If we can generate high-confidence, causally credentialed targets repeatably, disease after disease, the question changes — from whether the next program will fare better than the last to how many diseases we can take on at once.

This is the wager behind the Virtual Human. Not that any single insight is guaranteed to bear out, since none is. It is that interrogating human causal biology this way will consistently turn diseases we understand too poorly to treat into diseases we can act on. If it works, the payoff is not one medicine. It is a repeatable path to medicines for the many who have none today.

Subscribe now

Benchmark set

Evaluation used clinical-trial information from the PharmaProjects release as curated by Minikel et al. (2024), restricted to target–indication (T-I) pairs that had entered Phase II. A trial was labeled positive if it subsequently progressed to Phase III. Following Minikel et al., evaluation is at the level of T-I pairs, taking the maximum stage reached by any drug targeting that gene and indication; a pair is positive if at least one of its trials succeeded. The evaluation covered 301 metabolic T-I pairs (131 positive) and 339 cardiac T-I pairs (146 positive), implying the ~57% historical Phase II failure rate quoted in the essay for both areas.

Indication selection

The goal of the analysis is to demonstrate the value of using convergent phenotypic insights to inform on target biology. We therefore restricted the evaluation to two areas where insitro’s phenotyping is most expansive, metabolic and cardiovascular disease. Within each, an automated AI procedure selected the indications to which insitro endophenotypes are biologically relevant. The procedure saw only the biology of the indications and endophenotypes: no target-level information, and no aggregate data about the genetic analyses, so indications could not be selected for known model performance. This yielded 72 of 105 indications in the metabolic category and 29 of 42 in the cardiovascular category.

The insitro model

The model evaluated is a simplified version of insitro’s in-house Virtual Human models: an ML model integrating gene-level features from multiple modalities. Input features included (a) genetic results from insitro’s internal libraries, including ML-derived endophenotypes, using both genome-wide and rare-variant association statistics; (b) tissue-specific expression; (c) cellular data, including results from internal perturbation screens; and (d) druggability-related features. The model had no access to clinical trial outcomes at any point during training; the prediction task is therefore zero-shot.

Genetics baseline

The baseline represents standard genetic support: a version of the Minikel et al. model built on Open Targets locus-to-gene (L2G) scores, excluding insitro’s proprietary phenotypes and cellular data. Where Minikel et al. used a binary classifier, the baseline here is continuous, enabling comparison of precision and recall across thresholds (the precision–recall panels of Figure 8 in the essay).

Feature parity and robustness checks

To keep the comparison apples-to-apples and reduce overfitting risk, a subset of features used by Minikel et al. was omitted from both the baseline and the insitro model: OMIM Mendelian genetics, somatic mutations in oncology (not relevant to these indications), and supplementary sources such as Genebass and PICCOLO. To verify that these choices did not drive the conclusions, the omitted features (OMIM, Genebass, PICCOLO) were re-added to both models and evaluated with binary classification as in the original Minikel analysis; two versions of Open Targets scoring (“L2G score” and “L2G share” per Minikel et al.) were also evaluated. Across all versions, for both metabolic and cardiac indications, the insitro model’s improvement over standard genetics remained consistent.

Results and caveats

At the top decile by score, the insitro model’s false-positive rate was 10% (metabolic) and 21% (cardiac) against the historical 57% — reductions of more than fivefold and roughly 2.5× respectively — and cut the genetics baseline’s top-decile false-positive rate by nearly 75% (metabolic) and 30% (cardiac). Holding the false-positive rate equal to the baseline’s top decile, the model recovered 41% of the 131 Phase-II-validated metabolic pairs versus 15% for the baseline, and 34% of the 146 cardiac pairs versus 16%. The essay’s caveats should be read as part of the result: a retrospective false-positive rate under 10% does not imply a prospective Phase II failure rate under 10%; the evaluation covers a modest number of T-I pairs, restricted to pairs that others had already selected to enter trials; trials fail for reasons unrelated to target biology; and the mix of indications and modalities entering the clinic changes over time.

Reference

Minikel EV, Painter JL, Dong CC, Nelson MR. Refining the impact of genetic evidence on clinical success. Nature. 2024;629:624–629. doi:10.1038/s41586-024-07316-0

Subscribe now

insitro founder and CEO recently stopped by The Information Bottleneck podcast to chat with hosts and .

The conversation covers a lot of ground:

→ One thread: the bitter lesson worked for language and vision because scale and general methods met an internet’s worth of data. Biology has no equivalent corpus — cells grow at the speed cells grow. AI cannot manufacture the missing data.

→ Another: a drug isn’t a pattern in data you already have, it’s a prediction about an intervention nobody has run yet. This is a significant reason why 90% of drugs that reach the clinic fail, and only 22% of diseases have any approved treatment at all.

→ Plus: what it would take to build that missing corpus, and what real foundation models for biology would require.

Listen to the conversation: https://theinformationbottleneck.com/podcast/daphne-koller-insitro-ceo-on-the-information-bottleneck/

This companion lays out the sources behind the quantitative claims in Deep Phenotype's first article, Drug Discovery Has No Magic Wands, so that anyone who wants to check the argument can trace each figure to its origin. It states where the evidence is strong, where it is suggestive, and where a number is an order-of-magnitude estimate rather than a measured quantity. Where a figure is our own estimate rather than a published result, it is labeled as such and the assumptions behind it are shown in full.

1. The scale of unmet need

The essay opens with the observation that the large majority of recognized human diseases have no approved therapy. Any figure of this kind depends on how diseases are individuated. The most direct primary source is the medical knowledge graph assembled by Huang et al. (2024), which catalogs 17,080 diseases together with their known drug indications and contraindications: 92% of those diseases have no approved drug indication at all. A different estimate is Every Cure’s figure that fewer than 22% of the world’s recognized diseases have an FDA-approved treatment, computed against a universe of roughly 18,000 recognized conditions. This is why the essay states the figure as a range, from around a quarter of diseases down to single digits. Both estimates support the same conclusion: among the diseases medicine can name and catalog, the overwhelming majority have no approved treatment.

2. Why drugs fail

Taxonomy of program failures

The essay’s headline figure, that more than 90% of drugs entering clinical trials fail, is well established. Wong, Siah, and Lo (2019), analyzing over 21,000 compounds, estimate the overall probability of success from Phase I to approval at 13.8%. The more informative question is why they fail. Three independent lines of evidence establish that when drugs fail in the clinic, the dominant cause is inadequate efficacy, rather than a failure of the molecule’s construction or of safety.

• Hwang et al. (2016) tracked 640 novel therapeutics that entered pivotal trials between 1998 and 2008, with follow-up through 2015. Of these, 54% failed in clinical development. Among the failures, 57% were due to inadequate efficacy, 17% to safety, 22% to commercial or funding reasons, and 5% for reasons not recorded.

• Arrowsmith (2011), analyzing attrition by phase, found the same ordering: in Phase 3 (2007–2010), 66% of failures were for efficacy, 21% safety, and 7% commercial; in Phase 2, 51% were for efficacy.

• The AstraZeneca “5R” analysis is the most direct evidence that wrong biology is what remains once the other causes are controlled. Cook et al. (2014) analyzed the company’s 2005–2010 pipeline, the period before the framework was applied, and identified target validation, drug exposure, and patient selection as the recurring failure points. After AstraZeneca imposed explicit discipline on those dimensions, its success rate from candidate nomination to Phase 3 completion rose from 4% (2005–2010) to 19% (2012–2016) (Morgan et al. 2018). The claim about what remained is qualitative, and should be read as AstraZeneca’s own interpretation rather than a statistical result; neither paper partitions the residual failures. But the company’s account of the reformed pipeline is unambiguous: its projects then rarely failed for safety reasons or for missing proof of mechanism, and the major remaining cause of failure was a scientific hypothesis that turned out to be incorrect (Pangalos & Rees 2018).

Together these establish the essay’s core empirical claim: the binding chemistry, whether a molecule engages its intended target, is seldom what sinks a program. Most programs fail because of lack of efficacy.

A fourth line of evidence points the same way from the opposite direction: when the biology is right, success rates rise. Drug programs whose target has human genetic support, a naturally occurring variant linking the gene to the disease, are roughly twice as likely to succeed in the clinic as those without it (Nelson et al. 2015; King et al. 2019). The effect has held up as the datasets have grown, with a relative success of 2.6 overall rising to 3.7 for Mendelian evidence (Minikel et al. 2024), and about two-thirds of the drugs approved by the FDA in 2021 carried such support (Ochoa et al. 2022). What most improves the odds is stronger evidence that the target genuinely drives the disease. Why genetic evidence carries this predictive weight, and how to generate causal evidence of this kind at scale, is the subject of Part 2; we note it here as independent confirmation that the mechanism governs clinical success.

It is worth being precise about what an efficacy failure is, because the category is broader than “wrong mechanism.” A drug can fail to show efficacy for at least three distinct reasons:

• the target was wrong: the drug engaged it, but the mechanism does not govern the disease, the “scientific hypothesis that turned out to be incorrect” that Pangalos & Rees (2018) discuss;

• the drug did not adequately engage the target, or could not reach it at a tolerable dose, a molecule and tissue-exposure problem (Sun et al. 2022);

• the trial itself was flawed: underpowered, mis-dosed, or poorly enrolled (Fogel 2018).

Below we discuss the latter two.

Failures due to insufficient exposure

Kola and Landis (2004) documented that inadequate PK and bioavailability caused roughly 40% of attrition in 1991 but only about 10% by 2000, the payoff of introducing early, high-throughput ADME screening. Sun et al. (2022) report the same trajectory for drug-like properties more broadly, from 30–40% of failures in the 1990s to 10–15% today. Efficacy, the failure mode that depends on getting the biology right, did not improve over the same period.

Sun et al. (2022) also argue that a meaningful share of efficacy and toxicity failures reflect the second cause: the right mechanism pursued with a molecule that cannot achieve adequate exposure in the diseased tissue relative to healthy tissue. They do not quantify how much of the efficacy bucket is attributable to tissue exposure versus wrong mechanism. They offer illustrative case studies rather than a portfolio-level fraction, and note that the relevant tissue-exposure data are not systematically collected.

This failure mode has persisted for the same reason mechanism prediction has not improved. Binding affinity is the ideal computational target: decades of high-throughput assays have produced abundant structured data, and the in vitro readout is fast and cheap. In vitro ADME properties also lend themselves to rapid feedback. Tissue exposure and selectivity in the right human tissue, which Sun et al. argue also affects the efficacy and toxicity balance, behave the opposite way on both counts. There is little training data because it is rarely measured, and there is no fast feedback loop, because the ground truth lives in human tissue and emerges only late and expensively. Here too, computational and AI effort has concentrated where the data and the scorecard already exist, and the harder property remains under-modeled.

Molecules do matter. But progress tracks the availability of fast, cheap, quantitative feedback, and the hardest problems in drug discovery, where the least AI progress has been made, are the ones that lack it.

Trial failure versus program failure

Fogel (2018) documents the third cause in detail: recruitment, enrollment, retention, eligibility criteria, and trial design. Its figures are a useful reminder of how much trial outcome depends on execution. Of 114 trials funded by two UK agencies and recruiting between 1994 and 2002, only 31% met their original recruitment target (McDonald et al. 2006, reported in Fogel 2018), and slow accrual is consistently the single most common reason a trial is terminated early (Fogel 2018).

A trial terminated for under-enrollment usually does not kill the asset; the program is redesigned and re-run. When a drug program is actually abandoned, the dominant causes are efficacy and safety, the biology, as the failure-attribution data above establish (Hwang et al. 2016; Arrowsmith 2011; Wong et al. 2019).

The two places where operational choices most affect outcome, the endpoint and the enrolled population, reinforce the point above. When a trial fails on the wrong endpoint, or enrolls patients who do not carry the target mechanism, that is usually a symptom of insufficient disease understanding. Knowing which readout will reveal benefit, and which patients will respond, is itself a product of understanding the mechanism.

Deep Phenotype, a Substack by insitro

Scaled biology, deep causality.

By Deep Phenotype, by insitro

3. Crowding, and the retreat from novel biology

The essay’s claim that the industry concentrates on a narrow set of familiar mechanisms is drawn from L.E.K. Consulting’s 2025 analysis of the global R&D pipeline (Mancuso et al. 2025). Its findings, all for the 2024 pipeline:

• Of roughly 13,600 unique drug–target pairs in the preclinical and clinical pipeline, about one quarter are concentrated on just 37 biological targets, about 2% of active targets, each associated with 50 or more drugs. The number of novel biological targets entering the pipeline per year fell from around 100 pre-pandemic to just 30 in 2024.

• This is not a shrinking pipeline. The overall pipeline nearly doubled, from about 11,000 active programs in 2015 to about 21,000 by the end of 2024. The retreat from novel targets happened while total activity grew.

L.E.K. also reported that roughly 350 novel targets did enter the pipeline between 2020 and 2024, concentrated in oncology, immunology, metabolism, and neuroscience, with about 70% still in preclinical development. Novel biology is not being ignored entirely. The essay’s point is one of balance: the ratio of effort has tilted heavily toward crowding around known mechanisms.

4. The data chasm

Cellular data scale

The essay’s central chart compares the scale of biological data available today with the scale a causal model of human cell biology would require. The chart is drawn in bytes, which lets the human cell-biology corpus be set against other large datasets on a common axis, including the success story most often invoked, the protein-structure corpus behind AlphaFold. This section gives the published evidence on both sides of that comparison, in two units: bytes, to match the chart and to enable the cross-domain comparison; and cells, perturbations, and cell types, the scientifically meaningful measures of coverage that reveal where the true shortfall lies.

Bytes and coverage do not move together. The single-cell field has amassed a large byte-volume of data drawn from a tiny slice of the biological condition space, so a corpus can look large in bytes while sampling almost none of the perturbations and cell contexts a causal model must cover. Bytes match the chart and enable the AlphaFold comparison; coverage is what the disease-understanding argument turns on. We give both, and note where they diverge.

What has actually been measured, in cells.

• Descriptive atlases. CZ CELLxGENE Discover, the largest single aggregation of human single-cell data, held 169.3 million cells across more than 1,550 datasets as of October 2024, of which 93.6 million are unique once duplicated submissions are removed. The Human Cell Atlas consortium reported more than 100 million cells from more than 10,000 donors at its late-2024 milestone. “Hundreds of millions of cells” is right for the aggregate corpus; the deduplicated count is still under one hundred million, which is worth knowing if anyone presses on the phrase.

• The largest genetic perturbation screens. Replogle et al. (2022) ran genome-scale Perturb-seq, 2.5 million cells across 9,866 genes, in two cell lines (K562 and RPE1). Xaira Therapeutics has since surpassed this by more than an order of magnitude. Its X-Atlas/Orion (2025) profiled about 8 million cells targeting all protein-coding genes across two cell lines (HCT116 and HEK293T), sequenced deeply to over 16,000 UMIs per cell (several times the depth of prior atlases), and released as a 520 GB download. X-Atlas/Pisces (2026) reached 25.6 million perturbed single-cell transcriptomes across sixteen biologically diverse contexts spanning cell lines, iPSCs, and differentiating iPSCs, the largest genome-wide CRISPRi Perturb-seq compendium reported to date. Together these trained X-Cell, scaled to a 4.9-billion-parameter model (X-Cell-Ultra).

• The largest chemical perturbation screen. Tahoe-100M reports more than 100 million cells across roughly 1,100 small-molecule treatments in 50 cancer cell lines.

• The largest virtual-cell training set. Arc Institute’s State model reports training on roughly 170 million observational cells together with more than 100 million perturbational cells spanning about 70 cellular contexts. The public Virtual Cell Challenge benchmark built alongside it comprises roughly 300,000 cells and 300 CRISPRi perturbations in a single cell type, H1 human embryonic stem cells.

The perturbational corpus has recently become comparable to the descriptive one in raw cell count, which is what the essay means by saying this has begun to change. It still covers only a small number of cellular contexts against the hundreds of primary human cell types the atlas effort has been enumerating: two in Orion, sixteen in Pisces, fifty lines in the largest chemical screen, roughly seventy contexts in the largest training set. The frontier is starting to move beyond immortalized lines: Pisces already includes iPSCs and differentiating iPSCs, and X-Cell reports zero-shot generalization to primary human CD4+ T cells. But even the largest efforts sample a few tens of contexts, before disease state, time course, and dose are considered at all. That is the sense in which the essay’s “narrow range of cell lines” holds.

The same corpus, in bytes, next to AlphaFold
In byte terms, today’s aggregate single-cell corpus is on the order of 10 to 50 terabytes (the CZ CELLxGENE Census and the Human Cell Atlas). Orion gives a concrete per-cell anchor: 8 million cells in 520 GB is about 65 KB per deeply sequenced cell. The AlphaFold Protein Structure Database, more than 214 million predicted structures spanning essentially all sequenced life (Varadi et al. 2024), occupies only 1 to 2 terabytes. That AlphaFold is the smaller of the two corpora is the point developed below. A 1-to-2-terabyte corpus was enough to largely solve protein folding, because folding is conserved and the data could be pooled across every species at once. The same byte-scale is nowhere close to sufficient for human cell biology, because the space is far larger and because it cannot be assembled by borrowing across species.

How large the space is

• Perturbations. The human genome contains roughly 19,400 protein-coding genes (GENCODE 2025). Single-gene loss of function alone defines about 1.9 × 10⁴ perturbations; pairwise combinations define about 1.9 × 10⁸.

• Contexts. Tabula Sapiens resolved 475 distinct cell types from 483,152 cells across 24 human tissues, and the Cell Ontology carries on the order of 3,000 cell-type terms. Several hundred is the conservative working figure for distinct human cell types, before disease state, time course, and dose are considered at all.

Multiplying only those two axes, single-gene perturbations across a few hundred cell types, gives on the order of 10⁷ distinct experimental conditions, with no replication and no combinatorial, temporal, or dose dimension included. Admit pairwise perturbations and it is 10¹⁰ to 10¹¹. Set that against what exists: on the order of 10⁴ distinct perturbations, assayed in fewer than a hundred contexts, almost all of them cell lines. The shortfall is roughly three orders of magnitude for the single-perturbation case, and seven or more if one insists on covering combinations.

In bytes, the same conclusion
Today’s cell atlas holds 10 to 50 terabytes drawn from about 10⁴ conditions. A causal dictionary that covered the roughly 10⁷ single-perturbation conditions at comparable depth would run to 10 to 100 petabytes, three to four orders of magnitude beyond today’s atlas, with combinatorial, temporal, and dose dimensions pushing it higher still. This is exactly the gap the chart depicts: from a cell atlas of 10–50 TB today, through a complete cell-type dictionary, to a causal dictionary of 10–100 PB.

The argument depends only on the size of the gap, several orders of magnitude between what exists today and what a causal model would need, which is robust to reasonable variation in the assumptions. It would survive every figure above being wrong by two orders of magnitude in either direction.

None of this addresses the other half of the equation: relating cellular mechanisms to human clinical outcomes. Most human disease is a systems-level dysfunction, a temporal interaction of multiple biologies across diverse cell types, and understanding it requires systems-level measurements that are far less scalable, often requiring living organisms. That is the subject of the next section. Article continues below…

Read "Drug Discovery Has No Magic Wands"

Measuring disease-relevant biology

The essay draws a contrast between biological processes that are conserved across species, where cross-species data transfers and models trained on it generalize, and those that are intrinsically human, where it does not. The clearest poles of that spectrum are protein folding and neurodegeneration, and the contrast is what makes the AlphaFold comparison above more than an analogy.

Protein folding is near-universal.
A protein’s three-dimensional structure is determined far more by physics than by species, and is far more conserved than its sequence. Chothia and Lesk (1986) established that structure diverges much more slowly than sequence, so that proteins sharing as little as 20–25% sequence identity routinely adopt the same fold. This is why a single model trained on the Protein Data Bank, using experimental structures drawn from across the tree of life, predicts structures for any organism. The AlphaFold Protein Structure Database now holds more than 214 million predicted structures, a roughly 500-fold expansion from the 300,000 structures across 21 model-organism proteomes released in 2021, and covers almost the complete UniProt archive (Varadi et al. 2024; Jumper et al. 2021). The folding problem is conserved enough to be learned once, from all of life, and applied anywhere. This applies to folding, not to function: a protein’s interactions and physiological role need not transfer in the same way.

Neurodegeneration is not
Genome conservation is real but uneven. Around 85% of protein-coding genes are conserved between mice and humans, but conservation is not even across the genome, with higher divergence in immune (including microglial) and vascular genes (Boyanova et al. 2026), the biology that human genetics implicates in Alzheimer’s disease. The divergence is concrete at the level of the key disease proteins: rodent amyloid-β differs from human at three residues (R5G, Y10F, H13R), altering its processing and aggregation, which is why rodents do not normally develop amyloid pathology; and developmental differences in tau splicing prevent mice from recapitulating the human tau-isoform shifts central to disease (Boyanova et al. 2026).

The consequence is measurable. Across more than 100 genetic mouse models of Alzheimer’s disease, the two most widely used (5xFAD and APP knock-in) replicate only about 30% of the human protein alterations, rising to about 42% when tau and splicing pathology are added (Yarbro et al. 2025). The best available models capture under half of the human molecular disease, the quantitative form of the essay’s statement that rodents do not get Alzheimer’s and non-human primates do not recapitulate ALS.

Some processes, core metabolism for instance, are conserved across mammals and travel reasonably well. But the diseases where progress has been slowest, neurodegeneration above all, sit at the human-specific end, where cross-species data cannot substitute for measurement in human systems, and where those measurements are the most expensive, least available, and most ethically constrained to obtain. This is why the byte comparison with AlphaFold above is only part of the story. The same scale of data solved folding because folding pools across species, and cannot solve human neurodegeneration because it does not.

5. What AI can and cannot compress in the clinic

The essay’s final chart divides the clinical-development timeline into three parts and asks how much of it AI can realistically remove. The three parts are: the IND-enabling preclinical work required before a drug can enter humans; the operational overhead of running a trial (startup and site activation, patient recruitment, monitoring, data cleaning, and database lock); and the in-life biological observation window, the time patients must actually be dosed and followed for the clinical endpoint to mature.

Two features of that timeline are well documented. First, it is long and dominated by patient time. The canonical model of per-phase durations is Paul et al. (2010): 12 months of preclinical development, then 18, 30, and 30 months for Phases 1, 2, and 3, and 18 months from submission to launch. Measured durations from DiMasi et al. (2016) agree in magnitude, at a mean of 80.8 months from first-in-human to submission and 16.0 months of regulatory review. DiMasi’s individual phase means (33.1, 37.9, and 45.1 months) cannot be added together, because the phases overlap; their sum exceeds the measured first-in-human-to-submission interval by three years. Recruitment by itself has grown from an average of about 13 months in 2008–2011 to about 18 months in 2016–2019 (Brøgger-Mikkelsen et al. 2022).

The operational layer is substantial and the single largest source of unanticipated delay; slow enrollment is consistently the leading cause of trials running late (Fogel 2018). Industry and vendor projections frequently claim that AI will cut trial timelines by 30–50%. Those figures largely describe the operational layer, faster recruitment and automated data handling, and cost, rather than the total calendar time to a matured endpoint. They largely ignore the in-life observation window. To learn whether a drug changes the course of a disease, patients have to be treated and then followed until the outcome matures, an interval set by human biology, not by computation. This is the floor beneath any compression estimate: even if every operational inefficiency were eliminated, the biological waiting period remains. Netted against that floor, the share of the end-to-end timeline that AI can realistically remove is smaller. The chart puts it at roughly 11–17%, with the in-life observation window left essentially untouched.

That range is a derived estimate rather than a published benchmark, so the full arithmetic is set out here. The window divides as follows.

• End-to-end window: 108 months. Candidate nomination to launch, built from Paul et al. (2010): 12 months of preclinical development, 18 + 30 + 30 months of clinical phases, and 18 months from submission to launch.

• In-life observation: 54 months. Taken from Phases 2 and 3 together, 60 months in the Paul model, less the portion of that calendar time that is site activation and close-out rather than dosing and follow-up. Recruitment overlaps dosing, since patients enrolled early are already being followed while later ones are still being enrolled, so the in-life and operational buckets cannot simply be summed. We have assigned the overlap to in-life.

• Regulatory review: 16 months. From DiMasi et al. (2016).

• IND-enabling preclinical work: 12 months. From the Paul model.

• Operational overhead: 26 months, taken as the residual (108 − 54 − 16 − 12), covering site activation, the enrollment time not already counted as in-life, monitoring, data cleaning, and database lock. For scale, Tufts CSDD benchmarking has put median time from protocol approval to full site activation at roughly 17 months, and last-patient-last-visit to database lock at about 37 days, the latter from an industry survey rather than the peer-reviewed literature.

Against that division, we assume AI can remove:

• 40–60% of the operational overhead, or 10.4–15.6 months. This is the aggressive end of what patient identification, site selection, and automated data handling can plausibly deliver, and it is deliberately close to the 30–50% vendor claims, since those claims describe this layer specifically.

• 10–25% of the IND-enabling work, or 1.2–3.0 months.

• None of the in-life observation window, and none of regulatory review.

That totals 11.6 to 18.6 months out of 108, or 11% to 17%. If the in-life share is 40% of the window rather than 50%, the operational residual grows and the compressible fraction rises to roughly 15–23%. And if patient selection or better efficacy biomarkers genuinely shorten the follow-up interval, part of the in-life window becomes compressible after all. That is the route the essay points to as the only way through the biological floor.

References

Arc Institute. State: Arc Institute’s first virtual cell model. arcinstitute.org/news/virtual-cell-model-state (accessed 2026). Benchmark described in the Virtual Cell Challenge, Cell, 2025.

Arrowsmith J. Trial watch: Phase III and submission failures, 2007–2010; and Phase II failures, 2008–2010. Nature Reviews Drug Discovery. 2011;10(2):87; 10(5):328–329.

Boyanova S, Katsouri L, Krupic J, Hair K, Wang S-H, Wiseman FK. Considerations for the selection and phenotyping of mouse models for the study of Alzheimer’s disease. STAR Protocols. 2026;7(3):104633. doi:10.1016/j.xpro.2026.104633

Brøgger-Mikkelsen M, Zibert JR, Andersen AD, et al. Changes in key recruitment performance metrics from 2008–2019 in industry-sponsored phase III clinical trials. PLoS One. 2022;17(7):e0271819. doi:10.1371/journal.pone.0271819

Cell Ontology. The Cell Ontology in the age of single-cell omics. arXiv:2506.10037, 2025. A preprint, not peer-reviewed. arxiv.org/abs/2506.10037

Chothia C, Lesk AM. The relation between the divergence of sequence and structure in proteins. EMBO Journal. 1986;5(4):823–826. doi:10.1002/j.1460-2075.1986.tb04288.x

Cook D, Brown D, Alexander R, et al. Lessons learned from the fate of AstraZeneca’s drug pipeline: a five-dimensional framework. Nature Reviews Drug Discovery. 2014;13(6):419–431. doi:10.1038/nrd4309

CZ CELLxGENE Discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data. Nucleic Acids Research. 2025;53(D1):D886. doi:10.1093/nar/gkae1142

DiMasi JA, Grabowski HG, Hansen RW. Innovation in the pharmaceutical industry: new estimates of R&D costs. Journal of Health Economics. 2016;47:20–33. doi:10.1016/j.jhealeco.2016.01.012

Every Cure. The Problem. everycure.org/the-problem (accessed 2026).

Fogel DB. Factors associated with clinical trials that fail and opportunities for improving the likelihood of success: a review. Contemporary Clinical Trials Communications. 2018;11:156–164. doi:10.1016/j.conctc.2018.08.001

GENCODE. Reference annotation for the human and mouse genomes, 2025 release. Nucleic Acids Research. 2025;53(D1):D966.

Huang AC, Hsieh T-HS, Zhu J, Michuda J, Teng A, Kim S, et al. X-Atlas/Orion: genome-wide Perturb-seq datasets via a scalable Fix-Cryopreserve platform for training dose-dependent biological foundation models. bioRxiv. 2025. doi:10.1101/2025.06.11.659105

Huang K, Chandak P, Wang Q, et al. A foundation model for clinician-centered drug repurposing. Nature Medicine. 2024;30(12):3601–3613. doi:10.1038/s41591-024-03233-x

Human Cell Atlas. Cellular atlases are unlocking the mysteries of the human body. Nature. 2024. doi:10.1038/d41586-024-03552-6

Hwang TJ, Carpenter D, Lauffenburger JC, et al. Failure of investigational drugs in late-stage clinical development and publication of trial results. JAMA Internal Medicine. 2016;176(12):1826–1833. doi:10.1001/jamainternmed.2016.6008

Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589. doi:10.1038/s41586-021-03819-2

King EA, Davis JW, Degner JF. Are drug targets with genetic support twice as likely to be approved? PLoS Genetics. 2019;15(12):e1008489. doi:10.1371/journal.pgen.1008489

Kola I, Landis J. Can the pharmaceutical industry reduce attrition rates? Nature Reviews Drug Discovery. 2004;3(8):711–716. doi:10.1038/nrd1470

Mancuso M, Jacquet P, Brau R, Dhulesia A, Srinivasan A. Is biopharma doing enough to advance novel targets? L.E.K. Consulting Executive Insights. May 15, 2025; Figure 1b updated 29 July 2026 (targets with 50+ associated drugs changed from 38 to 37). lek.com

McDonald AM, Knight RC, Campbell MK, et al. What influences recruitment to randomised controlled trials? A review of trials funded by two UK funding agencies. Trials. 2006;7:9. doi:10.1186/1745-6215-7-9

Minikel EV, Painter JL, Dong CC, Nelson MR. Refining the impact of genetic evidence on clinical success. Nature. 2024;629:624–629. doi:10.1038/s41586-024-07316-0

Morgan P, Brown DG, Lennard S, et al. Impact of a five-dimensional framework on R&D productivity at AstraZeneca. Nature Reviews Drug Discovery. 2018;17(3):167–181. doi:10.1038/nrd.2017.244

Nelson MR, Tipney H, Painter JL, et al. The support of human genetic evidence for approved drug indications. Nature Genetics. 2015;47(8):856–860. doi:10.1038/ng.3314

Ochoa D, Karim M, Ghoussaini M, et al. Human genetic evidence supports two-thirds of the 2021 FDA-approved drugs. Nature Reviews Drug Discovery. 2022;21(8):551. doi:10.1038/d41573-022-00120-3

Pangalos MN, Rees S. AstraZeneca R&D: improving drug development productivity with better predictivity. Drug Discovery World. Summer 2018:18.

Paul SM, Mytelka DS, Dunwiddie CT, et al. How to improve R&D productivity: the pharmaceutical industry’s grand challenge. Nature Reviews Drug Discovery. 2010;9(3):203–214. doi:10.1038/nrd3078

Replogle JM, Saunders RA, Pogson AN, et al. Mapping information-rich genotype–phenotype landscapes with genome-scale Perturb-seq. Cell. 2022;185(14):2559–2575. doi:10.1016/j.cell.2022.05.013

Sun D, Gao W, Hu H, Zhou S. Why 90% of clinical drug development fails and how to improve it? Acta Pharmaceutica Sinica B. 2022;12(7):3049–3062. doi:10.1016/j.apsb.2022.02.002

Tabula Sapiens Consortium. The Tabula Sapiens: a multiple-organ, single-cell transcriptomic atlas of humans. Science. 2022;376(6594):eabl4896. doi:10.1126/science.abl4896

Tahoe-100M: a giga-scale single-cell perturbation resource. bioRxiv preprint, 2025. doi:10.1101/2025.02.20.639398

Tufts Center for the Study of Drug Development. Study start-up and site activation benchmarks; and Tufts CSDD / Veeva clinical data management industry survey. Industry benchmarking analyses, not peer-reviewed.

Varadi M, Bertoni D, Magana P, et al. AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences. Nucleic Acids Research. 2024;52(D1):D368–D375. doi:10.1093/nar/gkad1011

Wang C, Karimzadeh M, Ravindra NG, Bounds LR, Alerasool N, Huang AC, et al. X-Cell: scaling causal perturbation prediction across diverse cellular contexts via diffusion language models. bioRxiv. 2026. doi:10.64898/2026.03.18.712807

Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics. 2019;20(2):273–286. doi:10.1093/biostatistics/kxx069

Yarbro JM, Han X, Dasgupta A, Yang K, Liu D, Shrestha HK, et al. Human and mouse proteomics reveals the shared pathways in Alzheimer’s disease and delayed protein turnover in the amyloidome. Nature Communications. 2025. doi:10.1038/s41467-025-56853-3

Welcome to Deep Phenotype, insitro’s new publication exploring the ideas, evidence, and provocations shaping the future of artificial intelligence, biology and drug discovery. We’re launching with a two-part manifesto from insitro founder and CEO Daphne Koller on why the path to transformative medicines begins with causal human biology. A special thank you to our friends at a16z for their partnership and for cross-posting our debut article. Subscribe to follow along.

The tech world has latched onto an intoxicating promise: build a superintelligence, and it will cure cancer, and every other disease as well. The logic is seductive. The human body is a system we can already read from and write to, so like other knowledge problems, a powerful enough AI should be able to solve disease.

I fully believe that AI will eventually transform human health. It is why I’ve spent close to 30 years working at the intersection of AI and biology, and the last decade in drug discovery. The need is staggering: by most counts only a quarter of diseases — and by some estimates a few percent — have an approved therapy, and most of those merely slow a disease rather than stop it. For the majority of human illness, medicine still has little to offer.

But the magic-wand promise rests on an assumption that turns out to be false: that we already understand human biology well enough for a clever enough reasoner to find the cures hidden in what we know. We don’t. Hundreds of years into modern medicine, our understanding of most human disease, and much of healthy physiology, is best captured by the parable of the blind men and the elephant; in this case, a really huge elephant. AI is undoubtedly extraordinary, but aimed at a biology we have only begun to measure and barely understand, it will mostly help us generate failures faster.

This essay is about what that actually takes. The path to real cures starts at the root of the problem: measuring biology in the right way and using AI to derive novel insights from those measurements. Below, I describe some of the most prevalent AI Magic Wand narratives, explain the key pitfalls, and offer my perspective on what we actually need to deliver on this important goal.

Three Problems, One Bottleneck

To understand where AI fits, it helps to decompose drug discovery into its three essential stages:

1. Disease-to-mechanism: Identifying a biological mechanism — a pathway, a target, a molecular interaction — where therapeutic intervention will alter the course of disease in humans.

2. Mechanism-to-drug: Creating a molecular intervention in the right therapeutic modality — a small molecule, antibody, siRNA, gene therapy — that achieves the desired mechanistic effect with acceptable safety and pharmacological properties.

3. Drug-to-patient: Designing a clinical development program that identifies the right patients and assesses the molecule’s effects — beneficial as well as adverse.

The vast majority of AI work in drug discovery has focused on stage 2. This is understandable: the origin of the AI Magic Wand exuberance is the incredible achievement of AlphaFold, a field-defining tour de force. From this starting point, we have seen an explosion of AI tools capable of designing novel proteins, small molecules, RNA therapies, and even gene therapies. Given a biological mechanism we want to hit, it seems that AI can now design a molecule to hit it faster and better than ever before.

AI will certainly generate new and better molecules at an unprecedented rate, but will it generate drugs that unlock diseases for which there is currently no meaningful treatment? There are “undruggable targets” — high-confidence mechanisms that historically we have been unable to hit. The success against KRAS, the quintessential undruggable target, shows that this journey is possible. Notably, this success emerged from decades of structural biology and medicinal chemistry, not AI; as of now, I don’t know of a single example of an AI-derived insight that has led to “drugging the undruggable.” More broadly, the biggest step functions in our ability to drug the undruggable have historically come not from better molecular design tools, but from expanding our repertoire of therapeutic modalities: first biologics, then siRNA and antisense oligonucleotides, then gene editing. Each new modality opened a class of targets that was simply inaccessible before.

Article continues below…

But an even more critical point: validated yet undruggable targets are a tiny handful in the landscape of unmet need. For the vast majority of diseases without effective treatments, we simply have no idea what the right mechanism is. More than 90% of drugs that enter clinical trials fail — a dismal statistic that has barely improved in several decades. In the large majority of cases, the molecule was engineered just fine. The mechanism it targeted was wrong. We are doing a pretty good job at manufacturing keys, but they are often for the wrong locks. Even if AI lets us make better keys at an accelerating pace, that won’t improve our ability to identify the right locks. The real bottleneck in making a novel medicine is disease understanding: identifying a biological mechanism whose modification actually changes the course of disease in patients. That, far more than molecular design, is where drug discovery succeeds or fails.

This mechanistic understanding is a rare commodity. And because no-one likes to fail in the clinic, we are seeing industry trends that are truly destructive. There are currently 37 targets that have over 50 programs against each of them — slightly better keys for those few locks where we have strong conviction. How many variants of GLP-1 do we really need? Even worse than this misallocation of capital is the disservice to patients: the number of novel targets the industry advances each year fell from ~100 in 2015 to about 30 in 2024. That collapse is the far bigger cost: the inability to help the hundreds of millions of people for whom medicine currently offers nothing.

The Data Chasm of Human Biology

The disease-understanding goal requires that we bridge the chasm between high-level clinical manifestations of disease in a patient and the granular cellular mechanisms in which a drug intervenes. This has given rise to a second manifestation of the AI Magic Wand. Large language models — with their super-human reasoning capabilities — will connect the dots across the vast published literature of human biology, reasoning their way to new mechanistic hypotheses.

This optimism makes a very strong assumption: that the scientific community has collected — or will soon collect — enough data about human biology to contain the answer, and that we just need better reasoning to extract it. The challenge is that human biology is incredibly complex, spanning multiple interconnected biological layers — DNA, protein, cells, multi-cellular environments, entire organisms. Individual components respond dynamically to even subtle changes in related components or in the environment. Moreover, biology wasn’t engineered; it is the result of billions of years of messy, stochastic evolution, which produced staggering variation — countless genes, cell types, states, and contexts, each behaving in its own way. There is too much of it, too idiosyncratic, to reason about in the abstract. You have to measure it.

How many experiments would we need? The table below is an informed attempt at a back-of-the-envelope analysis.

Even if we consider only cell biology — the layer we need to interrogate biological mechanism — the space is vast. It becomes exponentially more vast when we consider that a drug is an intervention, so we need to map not only biology as it is, but also how it would respond to a perturbation. The largest cell atlases assembled to date, now spanning hundreds of millions of cells, remain orders of magnitude too small to cover this space. Until recently they also held almost no causal, perturbational data, the measurements most critical for understanding what an intervention would do in a living system. That has begun to change: several organizations have launched the monumental effort of building a “Virtual Cell,” pairing large-scale perturbation data with AI to reduce the data collection burden. But even the largest of these efforts samples only a vanishing fraction of the possible perturbations, and does so almost entirely in a narrow range of cellular contexts.

Even more importantly, the virtual cell efforts — useful as they might eventually turn out to be — do not address the other half of the equation: relating biological mechanisms to human clinical outcomes. Most human disease is a systems-level dysfunction, involving a complex, temporal interaction of multiple biologies spanning diverse cell types. Understanding these processes requires systems-level measurements that are far less scalable, often requiring living organisms.

Which brings up the greatest data challenge. While some processes are conserved across all forms of life, others are far more specific. The folding of a single protein is a self-contained process, highly conserved — closer to physics than to biology; this allows protein folding models to be trained on sequences collected across thousands of species. Metabolism involves at least a dozen distinct cell types and might be conserved across mammals. Brain function and dysfunction involves dozens of distinct cellular identities; and these processes are exquisitely specialized to humans: rodents do not get Alzheimer’s disease; non-human primates do not recapitulate ALS. The diseases where we have made the least progress tend to be precisely those that are most human-specific, and therefore those for which the data is most expensive to collect, least available, and most fraught with ethical constraints.

The Right Experiments, Not All Experiments

But we don’t need a universal causal model in order to make medicines. We can instead build a targeted causal model that makes the space navigable toward a desired outcome — uncovering the biological mechanisms underlying human diseases. Our models will not cover all biologies or all diseases, but an informed selection process will still allow us to make considerable headway.

This insight motivates the thesis behind a third AI Magic Wand: swarms of AI agents operating in a closed loop with laboratory automation, relentlessly chipping away at scientific problems. They formulate hypotheses, direct robots, analyze results, and iterate. Early successes here have been beguiling — automated systems have already excelled at tasks like optimizing cell-free protein synthesis or designing antibodies against known targets.

But agentic iterated optimization relies on a fundamental attribute: agents thrive when there is a fast, accurate, and cheap scorecard to evaluate progress. If you give a sufficiently smart model an instant feedback loop, it will grind against that benchmark until it wins. This is why coding assistants and molecular design tools advanced so rapidly — the feedback is cheap, accurate, and fast. A compiler immediately verifies whether code will run. The iterative loop with a developer provides rapid feedback on intent. The closed loop is tight, cheap, and objective.

Drug development is the exact opposite. The ultimate scorecard — whether a drug actually provides therapeutic benefit to a patient — cannot be captured well by computational models or high-throughput assays. The only true ground truth is a human clinical trial. This feedback loop currently takes years, costs millions, and is strictly bound by human ethics and living biology. It is the ultimate slow feedback loop, and no amount of compute or process optimization can change this.

The agentic lab successes listed above are solving problems that look like coding: highly quantitative objectives within a constrained search space. These problems are valuable, but largely beside the point when it comes to predicting whether a drug will actually work in a patient population. You cannot solve the human translation problem by accelerating our ability to optimize the wrong objective function. Doing so will simply scale up our process for building more keys to the wrong locks.

In order to leverage the power of agentic AI and lab automation towards the goal of improving scientific discovery, we must first have an objective function that is a proxy to human clinical benefit, and yet allows for AI insights and rapid experimentation.

The Final Mile: Patient Impact

Some have argued that the most important AI unlock in drug discovery is in the third stage — drug-to-patient — taking a drug candidate through preclinical testing and clinical trials. This is the fourth AI Magic Wand: reduce the time and cost of this very expensive phase, and drug discovery becomes faster and cheaper. Sadly, if you accelerate a pipeline full of drugs aimed at the wrong mechanisms, all you get is faster failures.

At the same time, there are real opportunities for AI in clinical development: predictive toxicology models and AI-drafted regulatory filings can shorten IND-enabling work, and patient identification from electronic health records, smarter site selection, and automated data management can trim operational overhead. These compression levers offer meaningful benefits, and we should absolutely pursue those. But these gains sit almost entirely within the first two slices of work. The remaining 50% — in-life biological observation — is gated by how fast disease unfolds in a living human. Even a comprehensive AI-driven improvement across everything it can touch leaves the majority of development time and cost structurally unchanged.

There is, however, one pathway through which AI could accelerate even this irreducible clock — and it runs directly through biological mechanism. An AI-enabled, deep mechanistic understanding of a disease enables the identification of novel clinical readouts that serve three distinct purposes: selecting the patients most likely to respond, confirming that the drug is hitting its intended target, and detecting early and reliable signals that it is actually modifying disease biology. Together, these allow trials to enroll the right patients, read out faster, and catch failures earlier — changes that transcend clinical trial operations, transforming the trial design itself. This capability is inseparable from solving the disease-understanding problem; they are one and the same. Better trials, in the end, are downstream of better biology.

Fulfilling AI’s Promise

The AI Magic Wands described above have real value. Generating molecules quickly can be a significant accelerant, as is reasoning across the vast scientific literature and increasing the efficiency of the scientific process. But these tools are just that: point solutions that address important problems without altering the fundamentals of our industry.

To fulfill the promise of AI for the millions of patients lacking any meaningful treatment, we must direct our efforts toward the problem that really matters: the identification of biological mechanisms with disease-transforming clinical benefit. This is arguably the hardest problem in drug discovery, because the only conclusive test of whether we have correctly identified a novel biological mechanism is a human clinical trial. There are multiple other paths in this space with shorter timelines and clearer near-term proof points. Those paths are shorter because the problems are more tractable: the feedback loops are faster and the benchmarks are cleaner. But a shorter path to a smaller destination is still a smaller destination — process improvements for problems we already know how to solve.

For the hundreds of millions of patients for whom no meaningful medicines exist, the difference between a wrong mechanism and the right one is the difference between another devastating clinical failure and a life-altering breakthrough. That is what we need in order to truly deliver on AI’s promise for human health.

This harder but aspirational path is the one we have elected to take at insitro, and we’ll have more to say about our approach soon.

This is Part 1 of Daphne Koller’s two-part manifesto. Read the companion piece detailing the evidence and references behind these claims. Subscribe to Deep Phenotype to receive Part 2 — and future perspectives from insitro on scaled biology, deep causality, AI, and the future of drug discovery.