AI-Designed Proteins and Drugs: From Prediction to Clinic

A handling editor once sent me a review request with a one-line note attached: "The claim in the abstract is that the molecule was 'discovered by AI.' Can you tell me what that sentence actually means?"

I spent a weekend on it and came back with an uncomfortable answer: it could mean almost anything. It could mean a generative model proposed the scaffold. It could mean a classifier ranked a virtual library and a medicinal chemist picked from the top hundred. It could mean a target was nominated by a knowledge graph and the chemistry was entirely conventional. All three had been published under the same shorthand. That review changed how I read this literature. Now the first question I ask of any paper or press release in this space is not "did AI work?" but "which step did the model do, and what would have happened without it?"

That question is still the hardest one to answer, and it structures everything below.

What the field is actually claiming

There are three distinct claims bundled together, and they have very different levels of support.

The first is structure prediction: given a sequence, predict the fold. The second is design: given a functional specification, produce a sequence that folds into something that does the job. The third is therapeutic discovery: use models to pick targets, generate or triage small molecules and biologics, and shorten preclinical timelines.

Claim one is close to settled. Claim two is advancing fast with real wet-lab validation and real failure rates nobody likes to quote. Claim three is where the clinical evidence lives, and it is thin, early, and heavily confounded by selection.

The strongest evidence: prospective blind benchmarks

The reason structure prediction convinced skeptics is that it was tested in a format most of computational biology never submits to: a prospective, blind, community-run competition. CASP assessors hold the answers. Groups predict before the experimental structures are public. There is no retrospective tuning, no quiet benchmark leakage.

Since 2020, the top accuracy on single-domain targets has been close enough to experimental error that the assessors had to discuss whether the problem was still interesting. Subsequent rounds pushed the frontier to complexes, nucleic acid interactions, and ligand-bound states, which is exactly where the remaining failures concentrate. If you want to read the primary literature rather than the commentary, the PubMed record on AlphaFold and structure prediction accuracy is now large enough to see both the triumphs and the systematic caveats in one sweep.

I'd put this in the same evidential category as a well-powered, preregistered trial. Blind prospective evaluation against ground truth, repeated across years, with independent assessors. That is as good as methods evidence gets in computational biology, and the 2024 Nobel Prize in Chemistry recognized both the prediction and the design sides of it.

What it does not establish: that a predicted structure is a good substrate for drug discovery. Predicted models are typically apo-like, single-conformer, and poorly calibrated for the side-chain geometry that structure-based design depends on. A postdoc I worked with spent eight months docking into predicted models of a transporter before realizing the pocket he was optimizing against didn't exist in any experimentally observed state. Nobody had lied to him. He had simply treated a confident-looking model as a crystal structure.

Docking and affinity prediction remain the weak joint

Co-folding models that place a ligand and protein together produce visually impressive poses. Independent re-evaluations using physical plausibility filters — checking bond geometry, steric clashes, and whether the ligand is even inside the pocket — have found that a substantial share of deep-learning poses fail basic chemistry checks, and that performance drops sharply on targets with low similarity to training data. Binding affinity prediction is worse still. Correlations that look respectable across a diverse benchmark collapse within a congeneric series, which is precisely the setting where a medicinal chemist needs the number.

My practical rule, and I'd defend it: use the models for pose generation and hypothesis ranking, use free energy methods or the assay for anything you plan to spend money on.

Strong but narrower: de novo protein design with wet-lab readouts

Design papers come with an advantage that clinical AI papers don't have — the experimental test is cheap, fast, and unambiguous. You express the protein, you measure binding, you solve a structure, you compare the design model to the result.

Diffusion-based backbone generation paired with inverse-folding sequence design has moved binder design from a heroic one-off into something a competent lab can attempt. Reported success rates for de novo binders against well-behaved targets have gone from fractions of a percent to percent-level hits from modest screening sets, with crystal and cryo-EM structures matching design models to low RMSD. Enzyme design, de novo luciferases, and metal-binding sites have all been demonstrated with experimental confirmation. The literature on diffusion-based de novo binder design is the place to see how fast this has compounded.

Two honest caveats. First, success rates are target-dependent in ways that are not well characterized — rigid, hydrophobic, convex epitopes work; flat polar interfaces, glycosylated surfaces and membrane proteins are much harder, and negative results are under-reported. Second, binding is not function. A nanomolar binder is not a drug. Immunogenicity, developability, half-life, tissue distribution and manufacturability are all downstream of design and largely orthogonal to what the generative model optimized.

Designed proteins have started entering clinical development, mostly as vaccine scaffolds and engineered biologics where the design element is one component of an otherwise conventional program. That is the realistic near-term picture, and it is still a genuine achievement.

The clinical evidence: small, promising, and easy to over-read

Here is where I'd ask you to slow down.

A meaningful cohort of compounds whose discovery involved machine learning — target nomination, generative chemistry, or ranked virtual screening — has now entered human trials. The earliest widely publicized example, an obsessive-compulsive disorder candidate from a generative chemistry collaboration, entered Phase 1 and was later discontinued. Several oncology and immunology candidates from the same generation of companies have also been shelved. Those discontinuations matter, and they get a fraction of the coverage the launches did.

The most informative readout so far has been in idiopathic pulmonary fibrosis, where a TNIK inhibitor whose target and chemistry were both nominated with the help of generative models completed a placebo-controlled Phase 2a and reported dose-dependent improvement in forced vital capacity over twelve weeks. Small, short, single-country, and not a registrational design — but it is the first time the full pipeline, from AI-nominated target to AI-generated chemotype, produced a controlled efficacy signal in patients. You can follow the primary reports and subsequent analyses through published work on this TNIK inhibitor in pulmonary fibrosis, and track registered studies on ClinicalTrials.gov.

There is also a genuine repurposing success story: a knowledge-graph-derived hypothesis that a JAK inhibitor might blunt both inflammation and viral entry in COVID-19, which was subsequently tested in large randomized trials and showed a mortality benefit. The AI contribution there was hypothesis generation, not molecule creation, and the evidence that the drug works is entirely conventional trial evidence. I think that is the honest way to describe it, and it's a stronger argument for machine-assisted hypothesis generation than most of the de novo claims.

Why the "AI molecules have higher Phase 1 success" analyses don't persuade me

You will see cited analyses reporting unusually high Phase 1 completion rates for AI-derived candidates relative to historical industry baselines. I've read them carefully and I don't think they support the inference most readers draw.

• The denominators are tiny and the follow-up is short, so survivorship dominates.

• Phase 1 is largely a safety and tolerability gate. Where AI methods plausibly help — potency, selectivity, ADMET optimization — is exactly what makes Phase 1 easy. The hard attrition is Phase 2 efficacy, which is mostly target biology, and no one has shown that generative chemistry fixes bad target choice.

• Portfolio composition is confounded. These pipelines are enriched for precedented targets and well-trodden indications, which have better historical odds regardless of how the molecule was found.

• There is no counterfactual arm. Nobody ran the parallel conventional program.

A second-year PhD student emailed me last spring asking whether she should reframe her thesis around AI-driven discovery because "the success rates are so much better." I told her to write down what the comparison group was. She couldn't, because it isn't in the papers. That exercise is worth more than any reading list.

Early and preclinical work: encouraging, not yet evidence

Deep learning screens of chemical libraries have produced structurally novel antibacterials with activity against resistant organisms, validated in vitro and in murine infection models. That work is legitimately interesting because the hits are outside the chemical space medicinal chemists were exploring, which is the clearest demonstration I know of models finding something humans wouldn't have. You can scan the deep learning antibiotic discovery literature for the current state. None of it has reached human testing, and the gap between a mouse thigh infection model and a clinical antibacterial has swallowed better-supported compounds than these.

Similarly, generative antibody design, protein language models for variant effect prediction, and sequence-to-function models for enzymes all have good retrospective and some prospective wet-lab support, and essentially no clinical support yet. That's fine. It's where the field should be. The error is importing confidence from the structure prediction results — which are genuinely strong — into claims about therapeutic outcomes, which are not.

What I'd argue most commentary gets wrong

The timeline claims are the weakest part of the case and the most repeated. "Eighteen months from target to candidate" compares a best case against an industry average that includes every stalled program. Discovery is also not the rate-limiting step for most assets; clinical development is, and no model shortens a 52-week fibrosis endpoint.

The second error is treating benchmark performance as evidence of generalization. Training-set contamination is pervasive, and the only trustworthy numbers come from blind prospective evaluation or from a wet lab. When I referee these papers now, I look for a held-out set defined by sequence or scaffold dissimilarity, not random splits, and for at least one prospective experimental test. Absent both, I don't believe the headline number.

The third is under-rating the boring contributions. Model-assisted triage of a million-compound library, better ADMET filters, structure-informed epitope selection — these quietly change hit rates without producing a narrative. I'd rather fund that than another de novo binder demo.

What to watch over the next two years

1. Phase 2b and Phase 3 readouts for the AI-derived candidates now past proof-of-concept, particularly in fibrosis and oncology. One controlled efficacy signal in a 70-patient study is a hypothesis; a confirmatory trial is evidence.

2. Published failures. The single most useful thing this field could do is report discontinuations with reasons. Without the denominator, every success-rate claim is unfalsifiable.

3. Prospective benchmarks for affinity and design, run in the CASP mold with blind experimental ground truth. Early community efforts along these lines exist; they deserve more participation than they get.

4. Designed biologics with immunogenicity data in humans. This is the specific unknown that will determine whether de novo proteins become a drug class or stay a scaffolding technology.

5. Regulatory guidance on model-derived evidence, especially where in silico results substitute for preclinical experiments. Agencies have begun consulting on this; the details will shape what sponsors bother to generate.

If you're building a reading practice around this, the volume is the problem more than the quality. I keep a running workspace where searches, saved papers, my notes on each one, and the draft sections that cite them stay connected rather than scattered across folders and reference managers — that's the part of exploring literature in PACR I actually use, because the context from a search survives into the writing. If you want to try that way of working, you can set up a PACR workspace and point it at this literature as a first test case.

The question I was asked by that editor is still the right one. Next time you read "discovered by AI," find the sentence in the methods that says which step the model performed, and ask what the comparison was. If the paper can't answer, you've learned something about the paper.