White Paper · AI in drug development

Deterministic Intelligence in a Probabilistic World

AI-derived molecules have reached Phase 3 and FDA review on a clinical record smaller than its success rates suggest. Regulators are defining what an AI output must show before it carries weight. Pricing generative chemistry, physics-based design and language models as one risk is the error.

In brief

• The clinical record for AI-derived molecules is small and inconsistent. Two analyses agree that Phase 1 success is high. The studies report different Phase 2 estimates, but their comparability remains uncertain: about 40 percent in the first, 19.5 percent in a newer and larger one. Neither shows better efficacy prediction.^1,2
• Of the assets usually grouped with AI-discovered drugs, the one closest to approval, Takeda's zasocitinib, was optimized mainly with physics-based simulation, alongside machine learning, rather than generative AI. It beat deucravacitinib head to head, and an FDA decision is expected in the first quarter of 2027.^3,4,5
• Regulators are converging on one principle: the evidence and controls an AI output needs scale with what a wrong output would cost. Europe's draft GMP Annex 22 goes further and, for critical manufacturing applications, accepts only static, deterministic models.⁶
• For sponsors, the central decision is one of design: for each use, which combination of validation, monitoring, change control and human review makes an output fit for the decision it informs.
• Alacrita's diligence work on AI and computational platforms leads to the same conclusion. What matters is where a platform's outputs enter decisions and what evidence supports them, not model architecture or benchmark scores, which count only insofar as they support a defensible case for performance in the intended use.

The scorecard, late 2026

Four programs, two analyses and one regulatory filing define what is actually known about AI-derived molecules in the clinic.

Two years ago the case for AI in drug development rested on discovery-stage speed claims and on the capital flowing in. There is now randomized clinical data, a Phase 3 under way, a head-to-head win and a filing under priority review. It is still a small body of evidence, and it does not support the conclusion commonly drawn from it: that AI-derived molecules have better odds in the clinic.

Two terms need defining at the outset. A deterministic system returns the same output for the same input, provided the model version, preprocessing, configuration and relevant execution environment remain fixed, so its results can be reproduced. That is the full extent of the guarantee: a deterministic model can still be opaque, systematically wrong or unsuited to its intended use. Reproducibility is also one property among several that tend to be run together. Whether a model represents uncertainty, whether its execution involves random sampling, and whether it changes during use are separate questions, and none of them says whether a given version performs adequately. Probabilistic models, such as generative chemistry models and large language models, represent distributions over possible outcomes; where their execution includes sampling, repeated requests can return different outputs. Neither term is a verdict on quality. Monte Carlo simulation in asset valuation is probabilistic by design and is good practice, because its assumptions are explicit and its uncertainty is stated. The question that matters for any AI output is whether it is fit for the decision it informs: what evidence shows it performs well enough for that use, and what will catch the errors that would matter. Determinism makes that case easier to build, and draft European manufacturing guidance proposes making it a requirement in one important setting.

The AI label also covers three different things, each judged on different evidence. A platform has to show that it improves a discovery or development process. An asset has to show an acceptable benefit-risk profile and a credible commercial case. A deployed system has to show adequate performance and control for its intended use. Success in one does not establish success in the others. A successful molecule shows that a computationally enabled workflow can produce a useful drug; it does not show how much the computation contributed, whether the program would have succeeded without it, or whether the method beats a realistic alternative. Those distinctions run through what follows.

Rentosertib: the generative AI test case

Insilico Medicine's TNIK inhibitor for idiopathic pulmonary fibrosis (IPF) is the clearest test of generative AI applied to both target selection and molecule design. The Phase 2a GENESIS-IPF trial randomized 71 patients across 21 sites in China. Its primary endpoint was safety and tolerability; the investigators considered the findings sufficient to support further study. The efficacy signal came from secondary and exploratory analyses in small groups: the highest-dose and placebo arms held 18 and 17 patients. Twenty-two participants lacked week-12 spirometry, including eight in the highest-dose arm, and the analysis used multiple imputation under a missing-at-random assumption. Forced vital capacity (FVC) rose by a mean of 98.4 mL over 12 weeks on 60 mg once daily, against a decline of 20.3 mL on placebo.⁸ Frequent adverse events included hypokalemia, abnormal liver function, diarrhea and raised ALT, which reached 33.3 percent in the 60 mg once-daily arm; four of the seven discontinuations for liver toxicity were in patients taking background nintedanib, which warrants further drug-drug interaction work.⁸

In July 2026 Insilico started a randomized, placebo-controlled Phase 3 in about 320 patients at 47 centers, again in China, with the annual rate of FVC decline over 52 weeks as the primary endpoint.⁷ The first patient was dosed in September.¹³ That trial tests the molecule, not the method. A positive result would be a landmark for generatively designed drugs; it would not show that generative design outperforms the alternatives. With 52 weeks of dosing per patient, a readout before 2028 looks unlikely. Whether a US or European filing will need additional regional data is a program-specific regulatory question. A commercially competitive program is likely to need persuasive evidence on use alongside nintedanib or pirfenidone, the standard background therapies in those markets, including safety and potential interactions; given where the Phase 2a liver signal appeared, that will be a central question for any partner.

Zasocitinib: the physics-based answer

Takeda's TYK2 inhibitor, acquired from Nimbus Therapeutics in a deal agreed in late 2022, is the furthest along of the assets usually grouped with AI-discovered drugs. In the LATITUDE Atlas head-to-head trial (606 patients with moderate-to-severe plaque psoriasis), the primary endpoint was complete skin clearance (PASI 100) at week 16: the sponsor's reported results show 36.5 percent of patients on zasocitinib reaching it, against 13.9 percent on deucravacitinib. PASI 90 rates were 62.5 and 34.0 percent.³ The FDA accepted the new drug application under priority review, with a decision expected in the first quarter of 2027, and the EMA has accepted the marketing authorization application.⁴

Zasocitinib is routinely listed as AI-discovered. Its design history is more specific than that. Nimbus and Schrödinger optimized the molecule with free energy perturbation (FEP), a physics-based method that estimates relative binding free energies from molecular dynamics simulations run on classical force fields. The target was the pseudokinase (JH2) domain of TYK2, and the work built on chemistry Bristol Myers Squibb had published for the same binding site; BMS's deucravacitinib was validating the mechanism in the clinic while the Nimbus work was under way. Machine learning also contributed: Nimbus trained models on historical compound data to predict absorption, distribution, metabolism, excretion and toxicity, and assessed more than 13,000 compounds computationally.^5,14,15 FEP is not strictly deterministic; it samples molecular motion, and repeat runs differ within known error bars. Sampling is also only one source of error. Force-field approximations, protonation states, the assumed binding mode and the protein conformation used each affect prediction quality, so a narrow simulation error bar is not a complete statement of uncertainty. Its predictions do, however, rest on physical models a medicinal chemist can interrogate, they carry quantified uncertainty, and they can be checked against experiment. If zasocitinib is approved, the first drug widely counted as AI-discovered to reach US patients will have come largely from the physics-based, interpretable end of the computational spectrum, against a mechanism a competitor was proving in the clinic at the same time. That matters to anyone attaching a premium to the AI label. It is also consistent with a longer record. Structure-based design, in which chemists work from the experimentally determined shape of the target, produced the HIV protease inhibitors approved in the mid-1990s.¹⁶ Physics- and structure-based methods have been contributing to approved drugs for three decades. Generative design has yet to produce one.

Platforms: capital ahead of clinical data

Isomorphic Labs, the Alphabet company built on AlphaFold, raised $2.1 billion in a Series B led by Thrive Capital in May 2026.¹⁰ In January its chief executive, Demis Hassabis, said the company expects its first clinical trials by the end of 2026, a year later than earlier guidance.⁹ Isomorphic has reported no human data. The round prices platform optionality, which is a legitimate basis for a valuation; the asset-level value of whatever the platform produces will be settled in Phase 2 and Phase 3, as it is for any other drug.

Recursion is the cautionary counterpart. After acquiring Exscientia in late 2024, it ended two clinical programs in 2025 when the data did not support further investment and began looking to out-license a third.¹¹ In the second quarter of 2026 it reported a net loss of $131.0 million on $7.7 million of revenue, with cash it expects to last into early 2028.¹² Takeda, meanwhile, has made a second computational bet: a multiyear discovery collaboration with Iambic Therapeutics, signed in February 2026, with success-based payments that could exceed $1.7 billion.¹⁷

What the aggregate numbers say

The widely quoted figures come from a 2024 analysis by Jayatunga and colleagues at BCG, using data through December 2023. Of 24 AI-derived molecules that completed Phase 1, 21 succeeded, a rate of 80 to 90 percent against historical averages of roughly 40 to 65 percent. Of ten that completed Phase 2, four succeeded: about 40 percent, in line with historical norms. Of the six Phase 2 programs that stopped, two stopped on negative data; the other four were halted for business, operational or other reasons.¹

A larger analysis published in October 2026 by Haque and colleagues identified 129 AI-enabled therapeutic assets across 69 companies, with data through March 1, 2026. Its published abstract reports observed success probabilities of 77.8 percent in Phase 1 and 19.5 percent in Phase 2.² We have not been able to review the full methods. The two studies define their cohorts differently, and observed rates from a pipeline in which many programs are still in progress depend on how ongoing and discontinued programs are counted and on the mix of therapeutic areas. Until the methods can be compared directly, the newer Phase 2 figure is a reason for caution rather than a new benchmark.

• Jayatunga 2024
• Haque 2026
• Industry benchmark, Citeline 2014 to 2023

Observed phase success, percent

Analysis	Phase 1	Phase 2
Jayatunga 2024	87.5	40
Haque 2026	77.8	19.5
Industry benchmark	47	28

Observed phase success for AI-derived molecules in the two published analyses, against the Citeline 2014 to 2023 benchmark used in Table 2.^1,2,18 The analyses define their cohorts differently and the 2026 methods have not been reviewed, so the Phase 2 gap is a reason for caution, not a measured decline.

Taken together, the two studies support a narrow conclusion. AI-derived molecules appear to do well in Phase 1, which is plausible: safety, pharmacokinetics and drug-like properties are what these tools are built to optimize. Phase 2 is unresolved. Neither study shows an improvement in efficacy prediction, and the newer figure leaves open the possibility that AI-derived molecules do no better than the industry norm in Phase 2, or worse. Both cohorts probably carry selection effects, since companies advance their best-characterized molecules first regardless of AI involvement. A strong Phase 1 record is evidence about safety and pharmacokinetics, not about whether the target was right.

Two curves, one gap

Discovery-stage performance and clinical validation are improving at different rates, and for different reasons.

Discovery-stage AI is improving quickly on tasks that can be scored: structure prediction, virtual screening, molecule generation and property prediction. AlphaFold 3 extended structure prediction to complexes containing proteins, nucleic acids, small molecules, ions and modified residues in a single framework.¹⁹ Independent benchmarking shows the limits of these methods. Deep-learning docking methods frequently produce physically invalid poses and generalize poorly to proteins unlike those in their training data.²⁰ Generative chemistry exists to produce novel scaffolds, which is exactly where reliability can be expected to fall.

Clinical validation proceeds on a different timescale. Compute does not shorten a 52-week efficacy endpoint, and the hardest problems in drug development, target selection, translation from models to patients and heterogeneity within a disease, are not the problems these tools are scored on. Training data widen the gap. Historical datasets underrepresent failed programs, non-European populations and rare diseases, and they carry the assumptions of the era that generated them. A model is only as reliable as the distribution it learned from.

AI drug discovery is not one field either. Capital is spread across five technical lanes, each at a different maturity and each with its own failure mode:

• Structure foundation models (Isomorphic Labs, Chai Discovery, Boltz): strong in silico performance, no clinical validation yet; the risk lies in the step from a good structure to a good drug.
• Generative chemistry (Insilico Medicine, Iambic): Phase 2 data in hand and one Phase 3 under way; the risks are selectivity and translation.
• Phenomics and perturbation biology (Recursion, insitro): deep cellular phenotyping; the risk is whether cellular signatures predict patient outcomes.
• Translational and clinical prediction (Iambic's Enchant model and others): the least mature evidence base, and the hardest to validate prospectively.
• Protein and antibody design (Generate Biomedicines, Absci, Chai Discovery): clinical assets so far come from optimization rather than de novo design; the risks are developability and immunogenicity as much as potency.

Takeda's two transactions show the difference: one bought a Phase 3-ready asset produced by physics-based design, the other bought access to a generative and predictive platform. Diligence on any AI company should begin by establishing which lane it is in and what evidence that lane can realistically produce in the next 24 months.

Quantitative structure-activity modeling, virtual screening and high-throughput screening each improved on what came before, and the industry's attrition rate stayed stubbornly high through each of them. The likeliest outcome for AI is the same pattern at larger scale: clear step changes within defined tasks, and a smaller change in the overall probability of reaching approval.

Biologics: a different evidence ladder

Antibodies deserve separate treatment. The label means something different there, and on one measure the clinical record runs ahead of small molecules.

Generate Biomedicines' GB-0895, an anti-TSLP antibody for severe asthma, is in two replicate Phase 3 trials recruiting across six regions, and the company listed on Nasdaq in February 2026.²¹ The program moved from Phase 1 to Phase 3 without a Phase 2, on pharmacodynamic rather than efficacy data: a single 300 mg dose produced roughly six months of biomarker suppression against placebo.²² What the platform contributed was affinity engineering and a half-life extension built around twice-yearly dosing, against a mechanism an approved antibody had already validated. That is the zasocitinib pattern in another modality: computation applied to optimization inside established biology, not to finding the target. Absci's ABS-101, an anti-TL1A antibody for inflammatory bowel disease, entered Phase 1 in 2025 on a similar basis.²³

De novo antibody design sits further back. The Chai-2 report generated binders against 26 of 52 novel antigens at an average hit rate of 15.5 percent without high-throughput screening, which is a substantial advance on earlier methods.²⁴ The qualifications matter for anyone pricing it: the work covered single-chain and VHH formats rather than full antibodies, affinities were often in the tens to hundreds of nanomolar, developability and immunogenicity were assessed computationally, and the designs came from a single blind round.

For diligence, the ladder is longer than for small molecules and the failure modes differ. A predicted structure or a confirmed binding event is the start of the evidence, not the end of it: expression, function in the intended format, specificity, aggregation and stability, pharmacokinetics and immunogenicity risk each decide whether a design becomes an asset. Continuous structural coordinates do not make protein design inherently more tractable than small-molecule discovery; they change the representation and the ways a program fails. The useful denominator is the experimental funnel. How many designs expressed, how many bound, how many kept function in the intended format, how many survived developability testing. Reporting only the strongest binders conceals the cost of finding them.

"Deterministic" has become a regulatory word

Six instruments now shape what an AI output must demonstrate. Two are binding, and the sharpest one is still a draft.

Since mid-2025 regulators have begun to define the evidence and controls AI outputs need, and one draft makes determinism an explicit requirement. Table 1, at the end of this section, shows where each instrument stands and which are binding.

Draft EU GMP Annex 22: the sharpest line

The European Commission consulted on a new Annex 22 on artificial intelligence from July to October 2025, alongside revisions to Annex 11 and Chapter 4.⁶ For critical GMP applications, the draft accepts only machine learning models that are static (parameters frozen once validated) and deterministic (identical inputs produce identical outputs). Continuously learning models are excluded. The draft states that generative AI and large language models should not be used in critical GMP applications. They may be used in non-critical applications, where qualified personnel take responsibility for the output. Human review does not open critical uses to generative or probabilistic models, although the draft allows reduced testing of an in-scope model where a human operator makes the decision and that operator's performance is monitored.⁶

Industry pushed back: models, it argued, should not be excluded on the strength of a technology label alone. At an EMA workshop in mid-2026, participants discussed whether such models could be used in critical applications if supported by a fully documented, risk-based control strategy, with controls layered across prevention, detection and containment. The report treats human-in-the-loop review as one possible control within broader, risk-proportionate human oversight, and states that "oversight cannot be a substitute for inadequate validation."²⁵ These are inputs to the revision, not adopted requirements, and the operative draft still excludes generative models from critical uses. The final text is due to the Commission in the fourth quarter of 2026.²⁶ Whatever the final wording, the direction is clear: validation evidence comes first, and human accountability sits inside the control strategy rather than standing in for it.

FDA and ICH: risk-based credibility

The FDA's January 2025 draft guidance on AI to support regulatory decision-making for drugs and biologics sets out a seven-step, risk-based credibility assessment built around a defined context of use, the influence a model has on a decision and the consequence of that decision being wrong. Its scope is specific: it covers AI that produces information or data supporting regulatory decisions on safety, effectiveness or quality, and excludes drug discovery and operational uses, such as drafting a submission, that do not affect patient safety, drug quality or the reliability of study results.²⁷ The guidance was still in draft at the time of writing, 21 months after publication. The CDER guidance agenda for 2026 adds new work on computer software assurance for AI-based systems in manufacturing and clinical investigations.²⁸

ICH M15, the harmonized guideline on model-informed drug development, reached Step 4 in January 2026. It names AI and machine learning among the modeling approaches it covers and assesses models against the question of interest, context of use, model influence, consequence of a wrong decision, model risk and model impact.²⁹ M15 is final at ICH level, was issued by the FDA as final guidance in June 2026,³⁰ and is already familiar to pharmacometrics teams. For models whose outputs serve as evidence in a submission, it is the framework to apply now.

In January 2026 the FDA and EMA jointly published ten guiding principles for good AI practice in drug development, covering among other things a clearly defined context of use, validation and oversight scaled to risk, data governance, sound engineering practice, performance assessment, lifecycle management and plain-language communication.³¹ The principles are not binding, but they show the two agencies aligned on direction.

Europe's horizontal rules, and the regulator as AI user

The EU's Digital Omnibus on AI entered into force on July 27, 2026 and deferred the AI Act's high-risk obligations: December 2, 2027 for stand-alone high-risk systems and August 2, 2028 for AI embedded in regulated products such as medical devices.³² For drug developers, the AI Act matters chiefly where software is itself a medical device or informs clinical decisions; GMP and GCP expectations will arrive through the medicines framework first.

In pharmacovigilance, the CIOMS Working Group XIV report, published in December 2025, frames AI as augmenting expert judgment and treats human oversight as essential.³³

Two further developments change the sponsor's position. The FDA launched Elsa, an agency-wide generative AI tool, in June 2025 to speed protocol reviews, summarize adverse events, compare labels and target inspections.³⁴ Sponsors should assume that parts of their submissions are now read with AI assistance, and that internal inconsistencies are correspondingly easier to find. The FDA also began publishing complete response letters in July 2025³⁵ and extended the practice that September to letters for pending and withdrawn applications,³⁶ so the agency's reasoning on deficient applications is more visible than it was.

Table 1. Where the main instruments stand, October 2026.

The common thread

Three expectations recur across these instruments. The evidence and controls an output needs are proportional to the consequence of its being wrong. The context of use is defined before deployment, not after. And accountability for the decision rests with people, backed by validation, monitoring and change control suited to the use; no one control stands in for the others.

Determinism helps, but it does not settle the question. Many neural networks return identical outputs from identical inputs once their weights are fixed. Variance comes in through sampling settings, retraining, prompt changes and, for models reached through a vendor's interface, version updates the sponsor does not control. Reproducibility makes an output testable and auditable; it does not make it correct. What makes an output defensible is evidence that it performs adequately for its intended use, explainability sufficient for the decision at hand, an audit trail and change control. Annex 22's "static and deterministic" requirement secures stable, reproducible behavior as a precondition for validation. A probabilistic system can build a comparable case where its use is bounded and its performance characterized; it does not have one by default, and in critical GMP applications the current draft does not let it try.

Generate, then govern

The practical question is not which model to use, but what evidence and controls sit between an output and a decision.

What follows from this is an architecture, not a tool choice. Probabilistic systems are suited to work where breadth matters and errors can be caught before they compound: generating hypotheses, enumerating options, finding patterns in large datasets, producing first drafts. Even there, errors have a cost; a wrong target hypothesis can commit substantial spending long before any regulated decision. Where outputs enter regulated decisions, they need evidence of fitness for that use and controls proportionate to what a mistake would cost. The task is to decide, use by use, what that evidence and those controls are. Figure 1 is an operational reading of the logic in the FDA's draft guidance and ICH M15, which weigh a model's influence on a decision against the consequence of that decision being wrong.^27,29 It adds one axis: how the output is produced, which governs how much evidence it needs before it can carry weight.

Figure 1. Generate, then govern. Two questions place any AI use: how its output is produced, and what a wrong output would cost. Generative output reaches regulated decisions only through a control strategy whose evidence matches the consequence of error; in critical GMP applications, the current EU draft excludes it.

Discovery

Discovery tolerates more variance than regulated work, and generative breadth is useful here, although a confidently wrong hypothesis can still consume a program budget. The strongest results so far pair generative or predictive models with reproducible filters: physics-based scoring, as with zasocitinib, or fast experimental feedback. Platforms that close the loop between computation and the wet lab have a structural advantage over purely computational ones, because the rate at which a program learns from experimental biology is usually what limits it.

Scientific, medical and regulatory writing

At many companies, language models went into scientific, medical and regulatory writing before anywhere else, and the failure mode has shifted since. Retrieval-grounded systems, which work from documents supplied to them, were built to stop the invented references that marked the early chatbots. The risk that remains is fluent reasoning that is plausible and unverified. The published evidence lags the models: a 2024 study of open-source models working through real patient cases found they did not follow clinical guidelines consistently and were sensitive to the order in which information arrived.³⁷ Current frontier models were not tested, and we know of no published study showing that the failure mode has gone away. In a protocol or a submission, a well-written but unsupported argument is still unsupported.

A useful way to think about a frontier model is as a capable junior colleague with extraordinary recall, high output and a tendency to confabulate near the edge of what it knows, and sometimes well inside it. That colleague is valuable on well-scoped tasks, properly supervised and checked against external evidence, and a liability once leadership mistakes confident prose for reliable reasoning. In our view current models are useful for literature synthesis, summarization and first drafts where each claim can be traced to a document a reviewer can open. Traceability makes checking possible; it does not do the checking. A reviewer still has to confirm that each source says what it is cited for and that nothing material was left out. The models are not yet dependable for unsupervised reasoning about mechanism, causality or regulatory strategy, where the errors are fluent and hard to spot.

A company that has deployed language models across its regulatory writing function without a validation framework, change control or a monitoring protocol has not implemented AI. It has implemented a compliance risk with a productivity interface on top.

Safety and pharmacovigilance

Case intake, MedDRA coding and duplicate detection are among the better-established AI applications, with performance that can be measured against human coders. Signal detection can benefit from continuous rather than periodic analysis. Causality assessment, and the decision to act on a signal, remain human judgments; the CIOMS report frames AI as taking on the repetitive work so that experts can concentrate on causality assessment.³³ An alert system without validated logic, documented performance and defined review points adds risk rather than removing it.

Manufacturing and CMC

Annex 22 draws its clearest line here. AI has a place in process development, anomaly detection, predictive maintenance and capturing what experienced operators know but have not written down. Where an output drives a controlled action, such as batch disposition, deviation classification or a change to a process parameter, the current EU draft requires a static, deterministic model validated for that use, with monitoring and change control.⁶ The EMA workshop report adds that human oversight cannot make up for inadequate validation.²⁵ Generative models are confined to non-critical support, with qualified personnel answerable for the output. Where that line is crossed without explicit design, the problem tends to surface in an inspection or a deviation investigation, and regulators are unsentimental about that kind of discovery.

Agents

Agentic systems plan and execute multi-step workflows, including laboratory work. They raise governance questions that a drafting tool does not: who is accountable at each decision node, how errors are detected and contained, and what change control applies when the system modifies its own tools or instructions. A sensible starting point is to deploy agents where each step's output can be checked, keep go/no-go decisions with people, and treat claimed productivity gains in GxP settings as projections until they are measured.

Clinical development: evidenced and projected

Three applications carry regulatory or peer-reviewed support. Three are plausible and largely untested.

Clinical operations may be where AI's largest near-term economic value sits, because time in development is expensive. The quality of the evidence varies a great deal by use.

Better evidenced

Eligibility criteria. Using electronic health records from 61,094 patients with advanced non-small-cell lung cancer, the Trial Pathfinder study emulated completed trials and found that data-driven relaxation of restrictive criteria more than doubled the eligible population on average, while the overall survival hazard ratio changed little (an average decrease of 0.05).³⁸ The work is peer reviewed and retrospective; using it prospectively depends on sponsors accepting broader populations.

Patient-trial matching. The NIH's TrialGPT matched patients to trial criteria with 87.3 percent accuracy at the criterion level, close to the 88.7 to 90.0 percent range of human experts, and retrieved more than 90 percent of relevant trials while screening less than 6 percent of the candidate pool. In a small user study (two physicians, six oncology trials and semi-synthetic patient cases) it cut screening time by 42.6 percent.³⁹ Broader screening of health records can also make trial populations more representative; whether it does depends on how sponsors design and govern its use.

Prognostic covariate adjustment. In September 2022 the EMA qualified Unlearn's PROCOVA procedure, which uses machine-learning prognostic scores ("digital twins") to increase the precision of randomized Phase 2 and Phase 3 trials, allowing smaller samples or more power without introducing bias.⁴⁰ This is a clear regulatory precedent for machine learning in trial design, and its structure is instructive: the model increases the precision of the randomized comparison inside a pre-specified, statistically controlled analysis. It does not replace the trial. That is "generate, then govern" applied to trial statistics. External and synthetic control arms that replace randomization remain largely limited to settings such as rare diseases, pediatrics and some single-arm oncology studies.

Plausible but largely projected

Protocol optimization. The problem is well documented: a Tufts CSDD analysis of 950 protocols found that 76 percent required at least one substantial amendment, up from 57 percent in the previous benchmark, and that the mean number of amendments per protocol rose to 3.3 from 2.1.⁴¹ AI tools trained on historical protocols aim to flag amendment-prone design choices before finalization. Vendor claims of fewer amendments have so far come from the vendors' own analyses rather than controlled comparisons, and are best treated as hypotheses to test.

AI-governed adaptive designs. Machine learning can refine enrichment thresholds as data accumulate, and reinforcement learning promises more ambitious adaptation. The regulatory overhead (pre-specification, Type I error control, documentation of adaptation decisions) is substantial, and sponsors should engage statistical reviewers early.

Digital endpoints. Computer vision scoring, wearable-derived measures and analysis of patient-reported free text could cut measurement noise in central nervous system, dermatology and musculoskeletal trials, where subjective rating scales inflate sample sizes. Qualification is slow and uncertain, and the benefit depends on the endpoint, the indication and the validation behind it.

What this means in practice

What we ask in diligence, how we treat the probabilities, and what we recommend sponsors put in place.

For investors and business development teams

What we see in diligence

Patterns from Alacrita engagements involving AI and computational platforms, anonymized:

• The black box. In diligence on an AI-driven discovery platform, the company's pitch materials gave prospective partners no way to compare its compounds with established drug-like chemical space, and the platform itself was not disclosed. Pharma partners are wary of technology they cannot inspect; we recommended generating new discovery data and case studies before approaching them.
• Insights that were not new. Supporting a client in negotiations with an AI discovery company, an Alacrita managing partner found publications suggesting that the insights the AI partner presented were already known.
• Biology still decides. An AI-driven discovery company asked for an independent view of a dual-kinase inhibitor program. The issue was the biology rather than the design method: there were indications of class-specific toxicity linked to one of the kinases, and we judged the project very high-risk.
• Test the tool, not the pitch. For a private equity investor weighing a computational chemistry software company, an Alacrita medicinal chemist with a computational chemistry background ran the software on real ligand- and structure-based design problems alongside competitors' products. The software held up, and we concluded the company was competitively positioned.
• Prior failures travel with the asset. For a venture investor evaluating a company that used a bioinformatics platform to select clinical-stage drugs for repurposing, we examined each asset's earlier clinical record to judge whether past failures could recur in the new indications. The investor decided not to invest.

Questions we ask in diligence on an AI platform or an AI-derived asset:

• Which lane is the company in, and what evidence can that lane produce within 24 months?
• For a clinical asset, what exactly did AI contribute: the target, the molecule, both, or optimization of a known chemotype? Is the target validated independently of the platform? Zasocitinib and rentosertib carry very different risk on this question alone.
• What is the platform's prospective record on its own predictions, including the candidates it predicted would work that did not?
• Where do model outputs enter development decisions, and what evidence shows they perform adequately there?
• How is model drift detected, and who signs off on retraining or vendor model changes?
• Who owns models trained on partner data, and who owns the improvements? In platform deals, this can matter more than the headline milestone figure.
• How much revenue is milestone-driven rather than recurring, and does cash runway extend past the next meaningful clinical readout?

The third question is where platform claims often fall short. A credible prospective record contains at least the following:

• Predictions timestamped before the experiment or clinical outcome they concern.
• Prespecified success criteria.
• The complete set of eligible predictions, including failed designs and abandoned candidates.
• A defined unit of analysis: molecule, target, program or decision.
• Comparators that reflect a realistic alternative workflow, not an artificially weak baseline.
• Costs and timelines measured from equivalent start and end points.
• Documented changes to models, inputs and experimental process over the period.

The list scales with the claim. A solubility predictor and a system that selects first-in-class targets are not judged by the same endpoint.

Signals that make us look harder: Phase 1 success presented as evidence of better efficacy prediction; benchmark performance without prospective validation on novel chemistry or targets; an "AI-discovered" label on an asset whose target was validated elsewhere, without saying so (a weakness in the pitch rather than the asset); no named owner for model changes; and governance described as future work.

A worked likelihood-of-approval example

We start AI-derived assets on standard phase-based probabilities of success and depart from them only where asset-specific evidence justifies it. The arithmetic below is illustrative; it shows how sensitive the result is to that choice.

On Citeline's 2014 to 2023 benchmarks, a program entering Phase 1 has a 47 percent chance of reaching Phase 2, 28 percent of passing Phase 2, 55 percent of passing Phase 3 and 92 percent of approval once filed: a cumulative likelihood of approval of about 6.7 percent.¹⁸

Table 2. Illustrative likelihood of approval under different assumptions. Benchmarks vary by therapeutic area and modality; the second row is a sensitivity case, not a recommended adjustment.

The second row applies the 85 percent midpoint of the reported AI Phase 1 range¹ to the industry benchmark and lifts the cumulative figure to about 12 percent. Because it combines a small, heterogeneous AI cohort with a different benchmark source, it is only as credible as the asset-specific safety and pharmacokinetic evidence offered in support of it. The third row credits Phase 2 as well, by substituting the roughly 40 percent AI-derived Phase 2 rate, and reaches about 17 percent. That figure was in line with the historical comparator in its own study, so it does not demonstrate an improvement, and it is not a higher rate to apply against a lower benchmark. The newer analysis reporting 19.5 percent Phase 2 success points in the opposite direction,² which is a further reason not to adjust Phase 2 upward. The last row is the starting point for an asset entering Phase 2, before its target, indication, modality and Phase 1 findings are taken into account; at that stage the AI label adds nothing.

None of this means an AI-derived asset carries no extra value. AI can improve the economics of development without moving the phase-transition probabilities: fewer design-make-test cycles, faster candidate nomination, lower preclinical costs or shorter recruitment can each raise a program's risk-adjusted value at the same likelihood of approval. Those gains belong in the cost and timing assumptions rather than in the probabilities, and they need evidence from the asset or platform in question, such as documented cycle counts, timelines or budgets, rather than general vendor claims. They should also be counted once. If faster discovery already lowers spending and brings launch forward in the model, a separate platform premium prices the same improvement twice.

Likelihood of approval is one input to a valuation. A risk-adjusted value also depends on development costs and their timing, time to market and the commercial forecast, and the effect of any probability adjustment depends on when the costs and revenues fall relative to the uncertain phase.

For sponsors: our recommendations

The instruments described earlier set requirements in specific contexts. What follows are Alacrita's recommendations, which go beyond those requirements in places; how far to apply each should be proportionate to the use.

Build an inventory by consequence, not by department or vendor. A one-page register is enough to start: the use, the system and version, the decision it informs, the consequence if the output is wrong, and the controls that apply. Generating a literature summary and recommending a process parameter may both involve the same model; they belong in different rows.

Table 3. A sample register. Each row names the decision an AI output informs, what an error would cost and the controls applied. The entries are illustrative.

Name the controls explicitly. For each use whose consequence is regulated, record the validation evidence, monitoring, change control and accountable reviewer that apply, and write them down before an inspector or reviewer asks. A reviewer's signature is one control among these, not evidence that an output is reliable.

Use ICH M15's framework where it applies. For models whose outputs serve as evidence in a submission, M15 is final at ICH level and its elements (context of use, model influence, decision consequence) map closely onto the FDA's draft credibility assessment. The FDA draft does not cover operational uses, such as drafting, that do not affect safety, quality or study reliability; for those, lighter internal controls are proportionate.

Put vendor models under change control. For tools built on externally hosted language models, the model can change without notice. For anything that enters a GxP record or a submission, record versions and reassess on change, with the depth of revalidation matched to the use.

Record what AI contributed. For each deliverable, note the tool, what it was used for and whether confidential data was involved. For documents bound for a regulatory submission, we recommend disclosing AI-generated content and having the responsible author sign a verification statement. Neither is yet a general regulatory requirement, but both make the review trail easier to defend.

Treat redaction with suspicion. Removing a company or product name rarely anonymizes data. An indication, a development stage and a deal size together can identify a program, which matters when deciding what may be entered into an externally hosted AI tool.

Write for an AI-assisted reviewer. Consistency across modules and traceable sourcing matter more when the regulator's own tools may be comparing documents.

Keep it proportionate. A 12-person biotech does not need an AI governance office. It needs the register, named owners and a rule that nothing generated goes into a regulated document without a responsible author.

What to watch through 2028

Six dated markers that will test the argument in public.

Several of these markers arrive within 18 months. The final text of Annex 22 will show how far the EMA is prepared to open critical GMP uses to generative systems, and under what controls. The FDA's decision on zasocitinib will show what the first approval of a drug widely labeled AI-discovered looks like: physics-based design, a mechanism validated elsewhere and a conventional clinical program. Isomorphic's first trials will begin the long work of turning platform value into asset value. The final FDA credibility guidance has no published date. Further out, the AI Act's deferred obligations take effect, and the rentosertib Phase 3 readout will give the first late-stage efficacy data for a generatively designed drug.

1. Jun 2025Rentosertib Phase 2a results published
2. Jul 2025Draft EU GMP Annex 22 published for consultation
3. Jan 2026FDA-EMA guiding principles; ICH M15 reaches Step 4
4. Jul 2026Rentosertib Phase 3 starts (China)
5. Sep 2026Zasocitinib NDA accepted under priority review
6. October 2026You are here
7. Q4 2026Annex 22 final text due to the Commission; Isomorphic's first trials (company guidance)
8. Q1 2027FDA decision on zasocitinib
9. Dec 2027EU AI Act obligations for stand-alone high-risk systems
10. Aug 2028AI Act obligations for AI in regulated products
11. 2028+Rentosertib Phase 3 readout (estimate)

Figure 2. Clinical and regulatory milestones, 2025 to 2028. The near-term markers are regulatory and late-stage; the first late-stage efficacy readout for a generatively designed drug, rentosertib's Phase 3, is unlikely before 2028.

The organizations that capture durable value from AI in drug development will be the ones that know, for each output, whether it is generating options or entering a decision, and that have the evidence and controls between the two in place before a regulator asks to see them.

Questions this paper answers

Where Alacrita fits

This paper was prepared by Alacrita Consulting. Alacrita's core team of former biotech and pharma operators carries out due diligence, valuation and development strategy work for investors, biotech companies and pharmaceutical companies, including assessments of AI-derived assets and the platforms that produce them. Views expressed are the authors' independent analysis and do not constitute investment advice.

Discuss an AI-derived asset, platform diligence or development strategy

References

1. 1.Jayatunga MKP, Ayers M, Bruens L, Jayanth D, Meier C. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discov Today. 2024;29(6):104009. doi:10.1016/j.drudis.2024.104009 ↑
2. 2.Haque W, et al. Artificial intelligence-enabled drug discovery: a growing but largely early-stage clinical pipeline. Cancer Treat Rev. 2026;151:103223. Published online October 8, 2026. doi:10.1016/j.ctrv.2026.103223 ↑
3. 3.Takeda Pharmaceutical Company. Zasocitinib delivered superior efficacy versus deucravacitinib in new Phase 3 head-to-head results, reinforcing potential to deliver rapid and durable skin clearance in a convenient once-daily pill [press release]. September 30, 2026. Accessed October 27, 2026. https://www.takeda.com/newsroom/newsreleases/2026/zasocitinib-superior-efficacy/ ↑
4. 4.Takeda Pharmaceutical Company. U.S. FDA accepts new drug application under priority review for Takeda's zasocitinib in moderate-to-severe plaque psoriasis [press release]. Business Wire. September 14, 2026. Accessed October 27, 2026. https://www.businesswire.com/news/home/20260914748441/en/U.S.-FDA-Accepts-New-Drug-Application-Under-Priority-Review-for-Takedas-Zasocitinib-in-Moderate-to-Severe-Plaque-Psoriasis-with-Potential-to-Redefine-Oral-Treatment-Expectations ↑
5. 5.Leit S, Greenwood J, Carriero S, et al. Discovery of a potent and selective tyrosine kinase 2 inhibitor: TAK-279. J Med Chem. 2023;66(15):10473-10496. doi:10.1021/acs.jmedchem.3c00600 ↑
6. 6.European Commission. Stakeholders' consultation on EudraLex Volume 4: Good Manufacturing Practice guidelines, Chapter 4, Annex 11 and new Annex 22 on artificial intelligence. Open July 7 to October 7, 2025. Accessed October 27, 2026. https://health.ec.europa.eu/consultations/stakeholders-consultation-eudralex-volume-4-good-manufacturing-practice-guidelines-chapter-4-annex_en. Draft Annex 22: https://health.ec.europa.eu/document/download/5f38a92d-bb8e-4264-8898-ea076e926db6_en?filename=mp_vol4_chap4_annex22_consultation_guideline_en.pdf ↑
7. 7.Insilico Medicine. Insilico initiates Phase III clinical trial for rentosertib, its AI-empowered TNIK inhibitor for idiopathic pulmonary fibrosis [press release]. July 7, 2026. Accessed October 27, 2026. https://insilico.com/news/xmjsn4l091-insilico-initiates-phase-iii-clinical-tr ↑
8. 8.Xu Z, Ren F, Wang P, et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nat Med. 2025;31(8):2602-2610. doi:10.1038/s41591-025-03743-2 ↑
9. 9.Reuters. Google-backed Isomorphic Labs delays clinical trial timeline. January 20, 2026. Accessed October 27, 2026. https://finance.yahoo.com/news/google-backed-ai-drug-discovery-195423147.html ↑
10. 10.Isomorphic Labs. Isomorphic Labs announces Series B investment round [press release]. May 12, 2026. Accessed October 27, 2026. https://www.isomorphiclabs.com/articles/isomorphic-labs-announces-series-b-investment-round ↑
11. 11.Pharmaceutical Technology. Recursion axes drug programmes to streamline pipeline. May 2025. Accessed October 27, 2026. https://www.pharmaceutical-technology.com/news/recursion-axes-drug-programmes-to-streamline-pipeline/ ↑
12. 12.Recursion Pharmaceuticals. Recursion reports second quarter financial results; Genentech options first neuroscience target into early discovery program [press release]. August 5, 2026. Accessed October 27, 2026. https://ir.recursion.com/news-releases/news-release-details/recursion-reports-second-quarter-financial-results-genentech ↑
13. 13.Insilico Medicine. Insilico Medicine doses first patient in GENESIS-IPF-3, the world's first Phase III trial of a generative AI-driven innovative drug [press release]. September 10, 2026. Accessed October 27, 2026. https://insilico.com/news/isn1009261-insilico-medicine-doses-first-patient-genesis-ipf-3 ↑
14. 14.Schrödinger. Design of a highly selective, allosteric, picomolar TYK2 inhibitor using novel FEP+ strategies [case study]. Accessed October 27, 2026. https://www.schrodinger.com/design-of-a-highly-selective-allosteric-picomolar-tyk2-inhibitor-using-novel-fep-strategies/ ↑
15. 15.Pecci A. Nimbus used AI to help develop Takeda's $4B psoriasis bet. Here's where it's taking the tech next. PharmaVoice. September 25, 2026. Accessed October 27, 2026. https://www.pharmavoice.com/news/nimbus-takeda-zasocitinib-ai-drug-discovery/831289/ ↑
16. 16.Wlodawer A, Vondrasek J. Inhibitors of HIV-1 protease: a major success of structure-assisted drug design. Annu Rev Biophys Biomol Struct. 1998;27:249-284. doi:10.1146/annurev.biophys.27.1.249 ↑
17. 17.Iambic Therapeutics. Iambic announces collaboration with Takeda to advance AI-driven design of small molecules [press release]. February 9, 2026. Accessed October 27, 2026. https://www.iambic.ai/post/iambic-announces-collaboration-with-takeda ↑
18. 18.Norstella. Why are clinical development success rates falling? May 16, 2024. Data: Citeline Biomedtracker phase transition analysis, 2014-2023. Accessed October 27, 2026. https://www.norstella.com/why-clinical-development-success-rates-falling/ ↑
19. 19.Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493-500. doi:10.1038/s41586-024-07487-w ↑
20. 20.Buttenschoen M, Morris GM, Deane CM. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chem Sci. 2024;15(9):3130-3139. doi:10.1039/D3SC04185A ↑
21. 21.Generate Biomedicines. Generate Biomedicines, Inc. reports second quarter 2026 financial results and provides business update [press release]. August 6, 2026. Accessed November 3, 2026. https://www.biospace.com/press-releases/generate-biomedicines-inc-reports-second-quarter-2026-financial-results-and-provides-business-update ↑
22. 22.Generate Biomedicines. Corporate overview, April 2026. Accessed November 3, 2026. https://investors.generatebiomedicines.com/wp-content/uploads/2026/04/Generate-Biomedicines-Non-Confidential-Overview_April-2026-1.pdf ↑
23. 23.Drug Discovery World. Absci becomes clinical-stage company with clinical trial initiation. June 6, 2025. Accessed November 3, 2026. https://www.ddw-online.com/absci-becomes-clinical-stage-company-with-clinical-trial-initiation-35241-202506/ ↑
24. 24.Chai Discovery Team. Zero-shot antibody design in a 24-well plate. bioRxiv. Preprint posted July 2025. doi:10.1101/2025.07.05.663018 ↑
25. 25.European Medicines Agency. Good manufacturing practice (GMP): multistakeholder workshop on expert contributions to artificial intelligence guidance development (Annex 22). Workshop held June 30, 2026. EMA/156789/2026. Accessed October 27, 2026. https://www.ema.europa.eu/en/documents/report/report-multistakeholder-workshop-expert-contributions-artificial-intelligence-guidance-development-annex-22_en.pdf ↑
26. 26.Drapeau B. Europe tried to ban generative AI from critical GMP. The ban may not survive. It does not matter. Pharmaceutical Technology. August 13, 2026. Accessed October 27, 2026. https://www.pharmtech.com/view/europe-tried-to-ban-generative-ai-from-critical-gmp-the-ban-may-not-survive-it-does-not-matter- ↑
27. 27.US Food and Drug Administration. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. Draft guidance for industry. January 2025. Accessed October 27, 2026. https://www.fda.gov/media/184830/download ↑
28. 28.Eglovitch JS. CDER plans new guidance on AI, electronic safety report submissions and more. Regulatory Focus. July 16, 2026. Accessed October 27, 2026. https://www.raps.org/resource/cder-plans-new-guidance-on-ai-electronic-safety-report-submissions-and-more.html ↑
29. 29.International Council for Harmonisation. ICH M15 Guideline on General Principles for Model-Informed Drug Development. Reached ICH Step 4 January 29, 2026; EMA publication accessed October 27, 2026. https://www.ema.europa.eu/en/documents/scientific-guideline/ich-m15-guideline-general-principles-model-informed-drug-development-step-5_en.pdf ↑
30. 30.US Food and Drug Administration. M15 General Principles for Model-Informed Drug Development; International Council for Harmonisation; guidance for industry; availability. Fed Regist. 2026;91:33179. June 3, 2026. Docket FDA-2024-D-5580. ↑
31. 31.US Food and Drug Administration; European Medicines Agency. Guiding Principles of Good AI Practice in Drug Development. January 14, 2026. Accessed October 27, 2026. https://www.ema.europa.eu/en/documents/other/guiding-principles-good-ai-practice-drug-development_en.pdf ↑
32. 32.Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulation (EU) 2024/1689 (Digital Omnibus on AI). Official Journal of the European Union. July 24, 2026. Accessed October 27, 2026. https://eur-lex.europa.eu/eli/reg/2026/1744/oj ↑
33. 33.Council for International Organizations of Medical Sciences. Artificial Intelligence in Pharmacovigilance: Report of the CIOMS Working Group XIV. Geneva: CIOMS; December 4, 2025. https://cioms.ch/working_groups/working-group-xiv-artificial-intelligence-in-pharmacovigilance/ ↑
34. 34.US Food and Drug Administration. FDA launches agency-wide AI tool to optimize performance for the American people [press release]. June 2, 2025. Accessed October 27, 2026. https://www.fda.gov/news-events/press-announcements/fda-launches-agency-wide-ai-tool-optimize-performance-american-people ↑
35. 35.US Food and Drug Administration. FDA embraces radical transparency by publishing complete response letters [press release]. July 10, 2025. Accessed October 27, 2026. https://www.fda.gov/news-events/press-announcements/fda-embraces-radical-transparency-publishing-complete-response-letters ↑
36. 36.US Food and Drug Administration. FDA announces real-time release of complete response letters, posts previously unpublished batch of 89 [press release]. September 4, 2025. Accessed October 27, 2026. https://www.fda.gov/news-events/press-announcements/fda-announces-real-time-release-complete-response-letters-posts-previously-unpublished-batch-89 ↑
37. 37.Hager P, Jungmann F, Holland R, et al. Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nat Med. 2024;30(9):2613-2622. doi:10.1038/s41591-024-03097-1 ↑
38. 38.Liu R, Rizzo S, Whipple S, et al. Evaluating eligibility criteria of oncology trials using real-world data and AI. Nature. 2021;592(7855):629-633. doi:10.1038/s41586-021-03430-5 ↑
39. 39.Jin Q, Wang Z, Floudas CS, et al. Matching patients to clinical trials with large language models. Nat Commun. 2024;15:9074. doi:10.1038/s41467-024-53081-z ↑
40. 40.European Medicines Agency. Qualification opinion for prognostic covariate adjustment (PROCOVA). Adopted by CHMP September 15, 2022. Accessed October 27, 2026. https://www.ema.europa.eu/documents/regulatory-procedural-guideline/qualification-opinion-prognostic-covariate-adjustment-procovatm_en.pdf ↑
41. 41.Getz K, Smith Z, Botto E, Murphy E, Dauchy A. New benchmarks on protocol amendment practices, trends and their impact on clinical trial performance. Ther Innov Regul Sci. 2024;58(3):539-548. doi:10.1007/s43441-024-00622-9 ↑