Fifteen simple rules for using AI in drug discovery

This is a review of a long, careful and somewhat deflationary Perspective on artificial intelligence (AI) in drug discovery. The authors take the position that it is too early to judge whether AI will be effective in drug discovery and then go on to suggest that if it is ever to be shown effective, the field must change its focus to validate processes in drug discovery rather than just the models that they operate around.

This article was written with the assistance of an adversarial agentic AI system; hyperlinks beside references will open the original Bender et al. article with the supporting text highlighted. See Section 3 for more details.

You might be forgiven for thinking this is another paper about how the latest razzle-dazzle AI machines are going to automagically create the greatest drugs ever, faster than we can say initial public offering. It isn't. Bender et al. [1] take a much more sober approach to the topic, and a much more expansive view of both AI and drug development.

For one thing, they correctly and refreshingly scope AI to include everything from classical machine learning (ML) to generative models, working on anything from molecular structure data to clinical trial design — something easy to forget in these heady days of LLM innovation. For another, they scope drug discovery well past target and ligand discovery: their Table 1 tracks each modellable endpoint through to the phase in which its impact actually shows up; for example, selection of targets with genetic support in the target-selection phase can give a 2.6-fold increase in the likelihood that they will be successful in the clinic during phases II and III [1]^ab.

This scoping decision has consequences for the paper. Drug development is more than just ligand identification and not all problems that can be attacked with AI are the problems that need solving. The authors' central recommendation is that benchmarking (of AI solutions) in this field has to graduate from model validation (does it score well on a test set?) to process validation (does it improve a real decision?). Everything else in the paper is, more or less, an argument for why that distinction is consequential.

The paper is long, but worth reading for its broad insight across multiple fields, the referenced research behind them, and an honest accounting of human dynamics that you don't normally encounter in a research journal.

Below I try to compress the recommendations and advice into the style of a "10 simple rules" article that somehow grew to 15.

Two concepts are central to the paper and I list them below. I've attempted to collect definitions for some specialist terms in the Glossary of the companion Appendices post that helped make the article more accessible to me and I hope others.

The two data domains. Biological data means data from cellular and in vivo disease models, pharmacology, and clinical trials. Chemical data means compound structures and the properties measured on them, including structure–activity relationships (SAR) — how a change to a molecule changes what it does. The paper's claim is that both behave badly for AI, but for different reasons: biological data are conditional and "epistemically opaque", chemical data are local and biased. Examples are given in the paper

Ligands and drugs. A ligand binds a target in an assay. A drug additionally has to be safe and efficacious in treating some disease phenotype in a person. The gap between them is the subject of the paper's Box 1 and of Rule 2.

Advocates for AI in drug discovery are plentiful. This paper doesn't dismiss that case, but rather advises caution on how early successes are interpreted and what's needed going forward.

The technical capability of AI is real and has moved fast. Protein structure prediction went from an open problem to a solved-enough one; the paper itself concedes that diffusion-based structure and design models have attracted serious money — its example is Xaira, whose first funding round was US$1 billion [1]^a. Phenotypic screening at industrial scale is real too: Recursion's public RxRx3 release alone covers 17,063 genes profiled across 2.2 million images and 1,674 compounds [5]^a, and the company is explicit about the ambition — it describes its datasets as maps of biology, built so that the connectedness of human biology can be navigated toward new medicines [5]^a. That is not marketing fluff bolted onto a wet lab; it is a coherent scientific bet that if you generate enough consistent, high-dimensional phenotypic data you can learn relationships that no hypothesis-driven programme would have proposed.

Generative chemistry has become routine rather than exotic, and the paper concedes the achievement while bounding it: published, experimentally validated workflows have raised hit rates and found novel chemotypes, but the successes are, in the authors' assessment, still largely successes of ligand design rather than of drug design [1]^a — which is a real capability, correctly labelled.

The clinical numbers, read charitably, are not embarrassing either. AI-designed molecules have reached the clinic, and a survey of the field concludes that they are entering trials at an increasing rate while progressing through development at about the same rate as conventionally discovered compounds [6]^a — which is to say: no worse than the industry baseline, from a standing start, in about a decade.

But Bender et al. note that most projects at "AI-first" companies are still preclinical, a few dozen are in phase I or II, and very few have reached phase III [1]^a. But they also note the confound in the optimists' favourite statistic — reported phase I success rates of up to 90% — namely that many of those projects are built on well-established biology and chemistry, which lowers the odds of running into safety problems the second time around [1]^a. Both readings are unresolved by the data, and the paper says so. Its framing, which I think is exactly right, is that the field is at a stage of "absence of evidence" and not necessarily "evidence of absence" when it comes to AI's translation into clinical impact [1]^a.

That is the position this review starts from: not "AI in drug discovery has failed", but "the experiment that would tell us has not been run properly yet, and here is why that is harder than it looks." This is the paper I wish every journalist would read instead of parroting the unchallenged words of claiming that all diseases will be cured in the next <3/5/10> years.

The authors convincingly argue that the jury is still out. AI obviously has great potential and has been an integral tool far before LLM's came into popular use. However, drug discovery is a long chain of decisions (target choice, assay choice, screening-deck luck, chemist intuition, animal-model availability, etc.), so it's very difficult to isolate why any one project succeeded or failed. This feeds a common bias — failure gets attributed to "bad luck," success gets attributed to "the method" (computational or experimental) — which distorts public and internal perception of what AI actually contributed [1]^a.

The authors argue that any AI methods that can contribute to patient stratification and ensuring efficacy and safety during Phase II will have the largest impact because they address the longest and most expensive part of the process [1]^a.

For example, using biomarkers to stratify patients to the right cohort roughly halves the capitalized cost per successful drug launch (based on ~21,000 compounds in commercial databases, 2000–2015), purely because it raises clinical-phase success rates [1]^a. Genetic support for a drug target is cited as another example of an evidence type that measurably raises clinical success odds (drug targets are reported to be ~2.6× more likely to succeed clinically when supported by genetic evidence) [1]^a, a figure that traces to Minikel and colleagues [7]^a. See Figure 1 in the paper.

On the other hand, much more attention is paid to the use of AI in the preceding Discovery Phase - partially due to the presence of large, homogenous and well-curated data sets that can support this activity [1]^a. But the authors point out that the results of these early phase studies (targets and ligands) are not the crux of the problem. There are reportedly >10⁶ known bioactive ligands (ChEMBL, PubChem) versus only ~10³ marketed drugs (see Box 1 in the paper) [1]^ab.

Factors where AI can have the largest impact on decision making in drug development include the right drug (physicochemical properties, on-/off-target activity, metabolism and dosage), given to the right patient (patient endotyping and patient stratification) — see Table 1 in the paper.

The authors warn that many choices are locked in early (hit-to-lead), but their consequences aren't visible until years later in the clinic — a fundamental structural challenge for validating any AI model against the outcome that actually matters [1]^a.

Biological data differs fundamentally from the data types where AI has succeeded. The proverbial image classifier that determines "is this an image of a cat" is "easy" to train because it is possible to easily collect thousands of images from the internet that are objectively "of cats" and thousands that are "not of cats". This pattern does not hold for most biological datasets which are:

One response of the last decade has been to manufacture large, homogeneous datasets on purpose, using high-dimensional assays such as Cell Painting or single-cell RNA-seq (Box A). Whether that works is Rule 9 and, separately, the first point under Rule 7.

Chemical space is enormous the authors estimate that the number of small molecules is a trillion times greater than the square of the number of stars in the known universe (hint ~10⁶⁰) which is still wildly unhelpful in coming to terms with a number that humans aren't built to deal with. In any case, it also presents profound problems with reliably sampling it and extrapolating predictive rules based on observations of those samples [1]^a. Figure 3a in the paper uses "chemical space as the universe" as an analogy — a new project's location in that space is, by definition, unknown at the time a model is built, so test-set performance doesn't reliably extrapolate to future projects. The paper flags that "external" test sets in the literature are often not truly external — they're frequently just another subset of the same original data pool — which systematically overestimates real-world model performance [1]^a. Figure 3b in the paper shows empirically that four commonly used ADME datasets (solubility, permeability/BBB, clearance) share almost no compounds or scaffolds with each other (overlap as low as 0.1–0.7%) [1]^a, meaning each model trained on them has a different, narrow applicability domain — which complicates multi-objective optimization, since drug design always requires balancing many properties simultaneously (see Box 1 in the paper) [1]^a. Information leakage from same-source train/test splits artificially inflates reported performance; the paper notes a psychological dimension too — scientists have limited incentive to adopt harder, more realistic (e.g., time-series) splits that make their own results look worse [1]^a.

Benchmarks drove genuine progress in image and speech recognition because performance gains there transferred to real-world use. In drug discovery, the authors point out that this analogy breaks down: the conditional, ambiguous nature of biological/chemical data means many published benchmarks are built on an underspecified problem setting [1]^a, creating a gap between "we improved the benchmark score" and "we improved real-world decisions." They explicitly name this risk "SOTA-chasing" and invoke Goodhart's law (once a measure becomes a target, it stops being a good measure) — benchmarks can and do get gamed [1]^a.

The author's section on "Generating data for AI applications" reads like a lament of past modalities of AI research that are no longer suited to the problem now closer at hand: applying these methods to realizing real gains in drug development. I take their point that benchmarking contests are rife with problems that hinder their applicability in the real world. Some benchmarks are invented to "sell" new methodologies while some community benchmarks have navigated these inventions to genuinely advance legitimate relevant endpoints even while struggling with the problem of the deeper label-conditionality problem and relying on community coordination to establish genuinely prospective ground truth, which isn't always available. CASP (protein structure), SAMPL (protein-ligand modeling), CACHE (computational hit-finding) — are cited as some of the few mechanisms that fall into this class and assess performance rigorously [1]^a.

However, I cannot concede that these benchmarks are the problem; they have brought us to this point afterall and the participants are well aware of their limitations. Rather, the crux of the problem lies in the authors' maxim: "generating data is not the bottleneck in drug discovery — it is ‘reduction to practice’, using it to inform clinically relevant decision making." [1]^a. This point too is taken well. But it's the how to do this that is non-obvious. So what then?

The real problem, how do we use knowledge about biological system parts out of context to make useful predictions about interventions in the real organism, is best expressed in one of my favorite passages from the paper [1]^ab:

Having been one of those researchers that was around in the early 2000's when we were ostensibly "building maps" of how the cell works, this resonates deeply. By all means, encyclopedic biological data in all its forms are critical to the study of life, but trying to translate this piecemeal data into biological consequence is fraught with difficulty because it is context (molecular, cellular, individual) that has the final say in biological function.

This is the crux of the problem. The study of biology (unlike physics) is not heavily deductive but heavily inductive, abductive and transductive. Whilst the authors offer some direction in addressing this problem (see Rule 9), it is far from a solved problem in biology even with AI.

This is not just a personal impression: it is a long-standing position in the philosophy of science, and I have set out the sources for it under Extra materials in the Appendices — Bechtel and Abrahamsen on why life-science explanation does not run by deriving phenomena from laws, and the Stanford Encyclopedia on mechanism-hunting, abduction and the problem of extrapolation [15][16][17].

Do the authors really offer direction here? The Rule 9 recommendations are about what to measure, and in what system: human primary cells from many donors, iPSC-derived models of the vulnerable cell type, higher-dimensional readouts, co-cultures. That is advice about improving the premises, not about the form of the inference. The closest thing to an inference-mode remedy is their causal-genetics framework, in which perturbation, context and phenotype are made to line up so that disease hypotheses can be tested from cause to effect [1]^a — designing the experiment so that an abductive inference becomes a controlled one is a real answer, but it still arrives as an experimental-design recommendation. The one sentence in the paper that speaks to inference itself sits in its penultimate paragraph: folding prior biological knowledge into the analysis, through Bayesian or causal modelling or Mendelian randomization, is "likely to be a necessary component of any viable solution" [1]^b. One sentence, no elaboration, then the Conclusion. So the honest answer is: some direction, and nearly all of it about data. On modes of inference, a gesture — which is why the question comes back in Rule 14.

The paper's recommendation is to move from just model validation to model and process validation — "models need to be linked to processes they are used in" [1]^a.

The authors expand on that idea to define two decision areas: project context and decision/follow-up experiments the prediction is actually meant to inform [1]^a.

Project context, consists of the application setting for the model (disease, endotype, target, anticipated human dose, etc.). They provide the example of a toxicity model that takes into account the concentration ultimately reached in humans in the phase of development in which it will be used [1]^a. Decision context, involves training a model that takes into account when the model will be used to make a decision (say during target selection versus deselection/toxicity-screening) [1]^a.

The authors reference a gem (All models are wrong and yours are useless) that poignantly describes why this matters in real life [8]^ab. The whole article is worth reading. Here are a few snippets:

AI in popular culture is commonly treated as a single "thing" while in reality it is a collection of methods applied to a collection of problems. It's only useful to talk about AI in drug development (like "computers" in drug development or "math" in drug development) in as much as one can identify things common to its practice and development across these subfields. This paper accomplishes that.

The "Where are we?" section (Table 3 in the paper) provides a brief overview of these subfields, dividing them between the molecules they are applied to (small molecules, proteins, peptides, antibodies), application to treatments (cell and gene therapy), data types (text and image data), functional assay development (cell-painting assays, temporal/single-cell/spatial transcriptomics) and translational data (real-world evidence, LLMs for diagnosis, trial-design, patient recruitment and response prediction).

The coverage is obviously non-exhaustive, and the authors say so themselves: the state of the art inside companies may well be further along than the published record, and they decline to judge what has not been published or clinically demonstrated [1]^a. What's important is that each subfield has its own distinctive challenges, limitations, state of maturity and recent advances. It's a non-homogenous mix. The one diagnosis that cuts across all of it is Rule 6's: the authors want more attention paid to how data, models and use cases are joined together [1]^a.

A few points were of specific interest to me. First, the authors consider that High-content imaging (Cell Painting, microscopy) used to train functional models for phenotypic drug/target discovery and disease modelling hasn't yet reached clinical validation, and needs strong data-quality control and biological diversity in training data to generalize [1]^a. The authors seem to hold the same view of use of temporal, single-cell, and spatial transcriptomics. These techniques enable finer-resolution characterization of biological systems and compound effects than before but the open challenge is going from high-dimensional data to variables that are actually useful for clinical decision-making [1]^a. There are companies that champion these exact methods — Recursion, whose public RxRx3 phenomics map is built from image-based, high-dimensional data generated in its own automated labs [5]^a; insitro, whose CellPaint-POSH platform (POSH: pooled optical screening in human cells) combines pooled CRISPR screening, high-content imaging and machine learning to map gene function without a prior hypothesis [9]^a; Relation Therapeutics, whose "lab-in-the-loop" combines tissue profiling, single-cell and spatial transcriptomics with target validation [10]^ab; and Noetik, which trains foundation models on spatial transcriptomics of tumours and offers to infer spatial transcriptomes from ordinary H&E slides [11]^a. I think there is an argument that large-scale, system-wide approaches take a disproportionate share of the attention space around AI and drug development and bear some responsibility for the popular impression that AI is just about finding drug targets - something this paper attempts to correct. A detailed review of these ongoing efforts is beyond the scope of this review but something I think should be addressed in a future post. For the time being, I would concur with the authors' carefully guarded statement that we are at a stage of "absence of evidence" and not necessarily "evidence of absence" [1]^a.

Secondly, the authors are explicit that LLMs "solely process patterns in language on a statistical basis, rather than 'reason'" [1]^a. This is true of LLM's but not necessarily true of AI systems that are agentic neural-symbolic systems - I've written about this before in my blog on agentic programming in practice [12]^a, [13]^a, [14]^ab. Notably, the paper explicitly flags agentic LLMs for drug-discovery tasks — while acknowledging they've been described in the literature, the authors caution that scoring well on such agentic tasks may not translate into better real drug discovery outcomes, because the underlying tasks are often underspecified and don't represent real-world project conditions [1]^ab. And this is really all the authors have to say about this topic which seems odd (and almost dismissive) given that it is recently a very active sub-field of AI and drug development and that Bender himself is a co-author on a recent review of this area [3]^a — one which, incidentally, treats the "reasoning" capacity of the models at the heart of agentic systems as an open question rather than a settled negative [3]^a - I can only speculate that the authors decided further discussing the topic would be distracting from their present argument.

With Rule 7 the review of "where we are" ends and the paper's own prescriptions begin. The authors frame them as two phases, in order. The first job is to find out where today's data fall furthest short of predicting what happens in clinical development [1]^a — and only then, second, to refine how models are built and tested against the use they will actually be put to, productionise them, and scale. The ordering is the whole point which they underwrite with a version of a pithy quote from Tukey: "Far better an approximate answer to the right question, which is often vague, than an exact answer to the wrong question, which can always be made precise." [18]^a.

They compress the target into a value equation — Model value = improvement in clinical success rate × project applicability domain — whose second factor is the underappreciated one: a model is worth more the more projects it can be applied to [1]^a. A generic safety or PK model that applies across a portfolio is worth more than a brilliant target-specific one, all else equal. Which is an uncomfortable finding for a field whose publication incentives reward the brilliant specific model.

Again, this reads a little like an indictment of a field (or fields) that never knew they were in the competition to speed up drug delivery and were just trying to do the best they could with the tools at hand. Still, I take their point and I think the "tech push/science pull" vocabulary they introduce is useful; datasets generated because a technology made them possible ("tech push") and only later repurposed for a question, as opposed to datasets generated because a scientific question demanded them ("science pull") [1]^a.

If the bottleneck is data predictive of human outcomes, then the recommendations are about what to run, not what to fit. The paper's list, in ascending order of ambition:

The overarching instruction is the one that indicts a fair amount of the last fifteen years of omics: piling up ever-larger datasets is no longer sufficient; what is needed is data that expose mechanism and bear on clinical outcomes [1]^a.

Note the tension with Rule 3 and with the first point under Rule 7. Rule 3 says high-dimensional phenotypic data has not yet delivered clinical validation, and Rule 9 recommends generating more of it. That is not a contradiction if you read it as "the problem with phenomics to date is experimental design and use-case linkage not dimensionality or the ability to infer". Whether that reading is right is, as far as I can tell, the single most important open empirical question in the paper, and it is not settled by anything in it.

It is described in research — though about drug discovery in general, not about phenomics in particular. The strongest statement of it is Scannell and Bosley's decision-theoretic model of R&D, which asks what actually limits the rate at which good candidates are found. Their answer: "The rate of creation of valid screening and disease models may be the major constraint on R&D productivity." [19]^a Their sensitivity analysis puts a number on it: a change of 0.1 in a model's predictive validity — "small and/or unknowable" to most people in drug discovery — offsets a ten- or hundred-fold change in its throughput [19]^b. What binds is the correspondence between assay and clinical question, not the dimensionality or volume of the measurement.

The same group's later review makes predictive validity the organising concept for every decision tool in drug discovery, and adds the reason it stays unfixed: it is "under-managed in drug R&D, not least because it is so hard to measure before projects succeed or fail later in the process" [20]^a — which is Rule 10's problem, one rule early. (That review is paywalled; the sentence quoted is in the free abstract.) Two caveats, because I do not want to overclaim on my own behalf. Both papers argue from decision theory and historical interpretation rather than from an experiment; and neither tests the phenomics-specific version of the claim, which is the version Rule 9 actually needs. The reading is supported by a serious literature and still unproven for the case at hand.

Pharmaceutical companies routinely separate preclinical and clinical functions — different databases, different access rules, different people and cultures. That separation is fatal to the one thing that would tell you which preclinical assays are actually predictive, because you only find out from clinical outcomes. The paper recommends explicit iterative feedback: phase I data to train tolerability, toxicity and PK models; phase IIb proof-of-concept data to refine target validation and efficacy prediction [1]^ab.

The part I would underline is the instruction to feed the loop with negative results, and the second-order benefit claimed for it: models trained on the failures and the marginal outcomes as well as the successes should translate better, and might also blunt the confirmation bias that steers clinical pipelines [1]^a. Not just better models — a partial antidote to the confirmation bias of an entire pipeline. That is a bigger claim than it looks, and it is a recommendation that costs nothing technically and everything politically.

This is Rule 6 turned into an operating instruction. Current validation is often model-centric, with metrics untethered from use; the fix is to validate models against their ability to aid specific project decision points [1]^a — evaluated inside the workflows they are meant to support, with realistic data availability and update cycles, and metrics chosen to reflect the actual decision and its trade-offs, which differ between early discovery and late development. It also means being honest about which experimental follow-ups are actually feasible and whether the provided data will be actionable.

The authors also cite the increased use of embedding AI models in design–make–test–analyse (DMTA) cycles to facilitate iterative learning and design [1]^a. The authors again stress the importance of modelling after the actual patient situation where decisions will be applied [1]^a.

The authors address large, industry-spanning consortia as a means to address the shear enormity of the AI-in-drug-development problem. [1]^a.

The paper addresses one past practice to be avoided: treating metadata (how the data were designed and generated, and the context they came from) as an afterthought instead of adressing the data integration and harmonization problem as something that begins before data collection with industry/academic discussions around experimental design and data formats [1]^a.

The authors do not describe how wide-spread a problem they believe this to be. However, there is evidence, that points both ways. On the credit side: MIAME, the minimum-information checklist for microarray experiments, plainly worked. Its own author, writing a paper subtitled Successes, Failures, Challenges, reports that "MIAME has been replicated for many other technologies, the major data repositories are supporting MIAME, and most scientific journals have adopted MIAME guidelines as a requirement for publishing" [21]^a. The MIBBI project — Minimum Information for Biological and Biomedical Investigations — which set out to corral the checklists MIAME spawned, makes the general case: such checklists "promote transparency in experimental reporting, enhance accessibility to data and support effective quality assessment" [22]^a. And the FAIR principles — findable, accessible, interoperable, reusable — generalised the idea past checklists, with the explicit aim of "enhancing the ability of machines to automatically find and use the data, in addition to supporting its reuse by individuals" [23]^a — which is exactly the property a purpose-built AI consortium would need.

On the debit side, the same sources are candid. MIBBI's own diagnosis of the landscape it was created to fix is that "overlaps in scope and arbitrary decisions on wording and substructuring inhibit their use in combination" [22]^b — standards proliferated faster than they harmonised, which is the metadata problem one level up. And seven years after FAIR, a group working on biomedical data funded by the US National Institutes of Health listed the barriers that remain: "(1) a lack of incentives, (2) little standardization in metadata collection and access, and (3) uncoordinated efforts" [24]^a.

So the fair reading is that standards move data from unusable to reusable wherever a journal or a funder enforces them, and integrate nothing wherever nobody does. That makes it an incentive problem rather than a technical one — Rule 13's territory more than this rule's — and it supports the authors' point that harmonisation has to be agreed before the data are generated rather than curated afterwards.

The authors also cover dedicated organisations that generate purpose-built data after a first, exploratory phase has settled what the questions are. Four real examples are given from both public and private organizations: the OpenADMET consortium [1]^a, LIGAND-AI from Ginkgo Bioworks', the Virtual Cell Pharmacology Initiative, covering ligand–target interaction data and compound DRUG-seq readouts respectively [1]^a, and DRUG-seq - being the high-throughput transcriptomic profiling of compound perturbations introduced in §1.2.

In summary, the authors call for a field-wide move in focus to what it really needs in terms of data and then taking advantage of its combined experimental and computational resources to generate those data in a manner mindful of how it will be used [1]^a.

I looked for, but could not find, a full catalogue of AI-drug discovery resources. The nearest thing to a full catalogue is the Nucleic Acids Research online Molecular Biology Database Collection, curated each year alongside the journal's Database Issue: in the 2026 round "899 entries were reviewed, 96 new resources added, and 319 discontinued URLs removed, bringing the total number of databases to 2173" [25]^a. That is molecular biology rather than drug discovery, and the churn in that quotation — 319 dead URLs in one year — is itself a comment on how durable the resource base under a model is. The other registry worth knowing is FAIRsharing, which indexes standards, databases and data policies together rather than databases alone: "As of February 2019, FAIRsharing has over 2,620 records: 1,293 standards, 1,209 databases and 118 data policies" [26]^a, and it is the descendant of the MIBBI portal mentioned a moment ago. Neither is drug-discovery-specific, and I could not find a peer-reviewed catalogue that is both comprehensive and current for AI-driven drug development; the working lists are review articles and community repositories, and they go stale quickly. For the chemical core the paper itself names only ChEMBL and PubChem (Box 1); the rest of the stack a working group would actually touch — DrugBank, the PDB, UniProt, Open Targets, GEO and ArrayExpress, ClinicalTrials.gov — has to be assembled by hand from the two registries above.

The section "Human psychology and extrinsic motivators" was entertaining precisely because it calls out a number of factors that are not normally called out (with some interesting references).

From the economic pressure imposed by Eroom's law (Box B) [1]^a to the constant pressure to adopt new tech along with its extravagent promises [1]^a and even just the emotional biases and cognitive errors that follow all of us around [1]^a. And the "publish or perish" working modality of academia and its pre-coccupation with toy-models gets another mention [1]^a.

I'll take it as the "looking up from the bench/keyboard and wondering what it is that we're all really doing here anyway" approach to field improvement. Whether any of these issues are really even addressable (other than being aware of them) and whether that has a chance to move the needle is another matter.

The last two paragraphs of the "Human psychology" section once again addresses the "reduction meets real-life context" curse of biology. I think they deserve their own section because they touch on a point that the article repeatably admits to be a core problem of drug discovery but then never turns around to square off against it.

I get the problem that reductionsist thinking is fallible to fall apart when it meets the real live human [1]^a but I dont see a convincing argument that this will not always be the case.

There is published support for the weaker form of that, and — as far as I can find — none for the strong form. The weaker form is that target-level reasoning routinely fails to predict what happens in the system around the target, and it is well evidenced. Lin and colleagues took ten cancer drugs that were in clinical trials and knocked out their putative targets with CRISPR: "the efficacy of each drug that we tested was unaffected by the loss of its putative target, indicating that these compounds kill cells via off-target effects" [27]^a. That is not even a failure of translation to the whole human; it is the reductionist account failing inside a dish. Network medicine supplies a reason to expect the organism-level version to be worse — "a disease is rarely a consequence of an abnormality in a single gene, but reflects the perturbations of the complex intracellular network" [28]^a. And Scannell and Bosley's sentence from Rule 9 lands here too: "Perhaps there has also been too much enthusiasm for reductionist molecular models which have insufficient predictive validity." [19]^a

I should be careful about what that adds up to, because my sentence above says "always" and the evidence says "often". A catalogue of failures, however long, cannot show that the next reductionist inference must fail; it establishes a base rate and gives reasons for it. The strong claim is a conjecture of the kind no accumulation of instances can prove and a single counterexample would kill — and the mechanistically clean single-target successes the field does have are candidate counterexamples. So I will hold it as a working prior rather than a finding: reductionist evidence is weaker evidence about the organism than the field prices in.

Multiomics, systems biology and high-throughput phenomics can't be the answer because they have the same fallibilities [1]^a. The authors suggest that the solution will somehow involve not more data but more structure: folding prior biological knowledge into the inference — through Bayesian or causal models, or throughMendelian randomization — though this point is cut short [1]^a.

Context is everything. This bugbear of biology underlies almost every rule above and may prove the reason that whilst AI will prove a useful tool, its perceived promise to "solve" all diseases is an illusion. A more sober stance would be that AI (along with many other technologies and some good old-fashioned human cooperation) might efficientize the process of drug discovery.

I would add one last idle question to be revisited. If LLMs can be made to appear to reason more effectively over text by constraining that output with deterministic code that is relevant to the immediate context of a conversation, is there an equivalent in biology? If reductionist thinking over biological facts is akin to the LLMs working over text, what is it that can be added to this that is akin to constraining and deterministic code that is relevant to the immediate context of a biological line of thought?

I went looking, and the nearest real answers are all partial — constraint-based modelling, prior structure wired into the model, and the ontology-and-knowledge-graph layer. None of them is the compiler-and-test-suite the coding analogy wants: biology has no cheap, fast, deterministic oracle that takes a candidate claim and returns false. The oracle is an experiment, and often a clinical one, which is the whole of Rule 10. I have set the candidates out with their sources under Extra materials in the Appendices [29][30][31]. If anyone knows of a general verifier, I would genuinely like to hear about it.

The authors' closing demand is that the field measure itself by "numbers that matter" rather than by proxies contrived to generate publications instead of R&D progress [1]^a. And the numbers that matter are named — higher clinical approval rates, and genuinely new therapeutic modalities and mechanisms of action [1]^a — not modelling novelty.

The authors frame the Perspective's goal modestly, as contributing a realistic view of where AI in drug discovery currently stands so as to clarify the conditions under which it can deliver. On the evidence of the fifteen rules above, they have done that.

My primary interest in this article was as background for a larger project to survey design patterns in Agentic AI systems used for drug development and biological research. My secondary aim was to explore the use of an adversarial agentic AI system I had created to debate positions based on published literature. The system is composed of a writer and critic agent that start with a human written position and take turns to revise the position in subsequent back-and-forth rounds until some quality metric is achieved. The system was implemented in Claude code Opus 5 and VS Code.

I wrote the starting position based on my own reading of the paper and an attempt to distill its points into a set of rules. After a single round, I expanded on the paper's treatment of LLMs in drug development (the coverage is light here) to include my own questions and ideas around use of neuro-symbolic systems in drug development as well as a second paper with Bender as co-author that covers AI agents in drug discovery in far deeper detail (see Huynh, Seal and colleagues' AI agents in drug discovery: applications and case studies [3]). The final machine version was preceded and followed by a human review. That expansion outgrew this article and has been set aside as the seed of a follow-on post on generative and agentic AI in drug development.

Quotations from these two papers and other supporting references were woven into or paraphrased in the first drafts of the piece and were checked by the critic agent using a deterministic script. Every number in the prose carries a footnote naming where it came from.

This final machine draft of the piece was heavily edited by myself for tone, content, brevity and additional follow-on thoughts.

The numbered references, the glossary of terms used above, and the provenance and verification record for every citation are in a companion post: Appendices.