The Molecule Was Designed In Silico. The Biology Still Gets a Vote

 Drug-like chemical space is usually estimated at around 10⁶⁰ molecules. The largest physical compound collections ever assembled hold a few million; DNA-encoded libraries reach billions; ultra-large virtual libraries touch the low trillions. Set those numbers beside one another and the historical practice of medicinal chemistry looks less like a search than a ritual. For fifty years the industry sampled a vanishingly small and biased corner of that space — biased toward what was easy to synthesise, already on the shelf, and had worked before — then spent a decade and roughly $2.6 billion finding out whether the sample contained anything that worked in a human being. It mostly did not. That is the fact around which everything else turns.

The old workflow, stated plainly.

Target identification came from genetics, pathology and institutional intuition, and it remains the largest source of downstream failure; most Phase II collapses are target failures, not chemistry failures. High-throughput screening followed — a million to two million compounds at hit rates of roughly 0.01 to 0.1 per cent, most of them artefacts or interference compounds. Hit-to-lead pushed the survivors through orthogonal confirmation and early ADME to two or three viable series. Then the long middle: lead optimisation, the Design–Make–Test–Analyse cycle running on human cognition and wet-lab throughput, where a chemist proposes an analogue, synthesis takes weeks, biology reports back, and the structure–activity relationship advances by a single increment. Potency, selectivity, solubility, metabolic stability and half-life all had to move in the right direction at once, and are frequently in tension. Three to five years and one to five thousand synthesised compounds per programme was ordinary. Then toxicology, then the clinic, where nine in ten candidates never reach patients. The famous cost figure is really the cost of the failures: every approved drug carries the amortised weight of the dozens that died beside it.

What replaced the ritual — six bets, not one technology.

Conflating the layers of “AI drug discovery” is the field’s most common analytical error. Target discovery now runs on multi-omics, knowledge graphs and perturbation data to propose causal disease drivers rather than correlated ones — the layer where a genuinely novel target can emerge, and where the deepest risk sits, since a novel target is an untested biological hypothesis. Structure prediction, in the AlphaFold lineage, collapsed the cost of a protein structure from a multi-year crystallography programme to an afternoon of compute, removing structure as the rate-limiting step for most targets. Generative chemistry — diffusion models, reinforcement-learning agents, chemical language models — marks the real conceptual shift, because the model is not screening a library but sampling a distribution, and can therefore propose structures no chemist has drawn and no vendor stocks. Multi-parameter prediction of ADMET, toxicity and synthetic accessibility before anything is made is where the practical savings concentrate; potency was never the hard part, killing bad molecules early was.

The naive story is that AI replaces experiments; the accurate one is that AI decides which experiments to run. The DMTA loop closes, the wet lab stops being a brute-force instrument, and becomes an oracle queried where the model’s uncertainty is highest.

How pipelines are built now — and one thing that is routinely fudged.

The working pattern is to compress discovery hard (target-to-candidate in twelve to thirty months, hundreds of compounds rather than thousands), design against predicted metabolic and safety liabilities from the first generation rather than the fourth, choose indications with fast objective endpoints — fibrosis with FVC, oncology with response rate — and run regulatory acceleration in parallel through orphan designation, Fast Track, Breakthrough Therapy and adaptive trial designs. That last element needs a correction, because commentary consistently blurs it: AI does not fast-track anything at the regulator. Expedited designations are statutory and awarded on unmet need and preliminary clinical evidence; no agency grants faster review because a molecule was computationally designed. What AI compresses is the pre-IND segment the sponsor controls. What regulators have built instead is a framework for AI used inside the submission — the FDA’s January 2025 draft guidance set out a risk-based, seven-step credibility framework organised around a model’s context of use, and the FDA and EMA jointly issued good machine-learning practice principles in January 2026. Agencies are regulating the credibility of evidence, not the discovery method. Nobody at CDER cares how you found the molecule; they care whether the model’s output can support a claim.

The counter-evidence deserves equal billing.

Of roughly 6,147 drugs in clinical development in one widely cited survey, only about 67 — near one per cent — were AI-discovered molecules, with far fewer involving genuinely novel targets. A 2026 ASCO analysis counted 117 AI-enabled assets in interventional trials, about half through Phase I and fewer than seven per cent through Phase II. The headline success rates need care: Phase I pass rates for AI-derived molecules have been reported at 80 to 90 per cent against historical benchmarks nearer 40 to 65, which sounds transformative until one notices the small sample, the survivorship bias, and the fact that Phase I measures safety rather than efficacy — precisely what property prediction should be good at. Phase II rates sit near the industry norm, and the rentosertib paper itself concedes that AI-discovered drugs have so far failed in Phase II at rates comparable to everything else. As of mid-2026 no AI-discovered drug holds FDA approval.

The synthesis is clear: AI has demonstrably compressed the design problem and has not yet demonstrably solved the biology problem. Phase II attrition is driven by target validity and translational failure, neither of which is fixed by generating better molecules faster. A beautifully optimised inhibitor of the wrong protein fails as expensively as an ugly one. The bottleneck moved; it did not disappear.

Inventorship is settled everywhere — and almost beside the point.

The unsettled question is not who invented it, but how obvious it was.

Very few real programmes involve a machine inventing unaided; almost all involve humans using AI heavily, which puts the weight on non-obviousness — judged against a legal fiction, the person of ordinary skill, presumed to hold the ordinary tools of the field. In 2026 generative platforms are ordinary tools of the field. If that notional person is equipped with a competent model, the inventive-step threshold rises for everyone, and a compound the model ranks highly on a standard target specification begins to look like a predictable result — obvious to try, with a reasonable expectation of success. A computationally predicted structure with no synthesis and no biological data is a weak filing, and computational-only utility disclosures are routinely rejected, so file after preliminary in vitro activity data rather than before. And the oldest answer to an obviousness attack becomes the most important: evidence of unexpected results, showing what the prior structure–activity relationship predicted, what the model predicted, and where observed behaviour diverged from both. All of which reduces to one unglamorous discipline — document human conception contemporaneously: who framed the problem, chose the target, set the objective functions, curated the data, and selected this molecule from the model’s output and recognised why it mattered.

Who is actually building, as of 2026.

_{^{Note: The table contents are based on public information}}

The question worth sitting with.

The interesting thing about this moment is not that AI designs molecules but what that reveals about where difficulty in medicine lives. The industry long believed the hard part was chemistry — the needle in a 10⁶⁰ haystack. Generative models have substantially solved that framing and exposed that it was never the binding constraint. We do not understand disease biology well enough to know which needle we want. AI has given us an extraordinary engine for answering questions and left untouched our ability to ask the right one.

Two questions follow.

The scientific one: does compressing discovery merely front-load the same attrition, or does designing against liabilities from the first generation genuinely raise the probability of approval? Only Phase III data answers that, and the answers begin arriving now.

The legal one is stranger. Patent law rewards the non-obvious; if the ordinary practitioner is presumed to hold tools that make excellent molecules routinely, excellence stops being non-obvious and the reward weakens precisely as the technology succeeds.

We may soon have to decide whether we protect the molecule, the model, the data that trained it, or the judgment that framed the question. The molecule was designed in silico. The biology, the regulator and the patent examiner all still get a vote.

Disclaimer

_{This article is published for general information and academic discussion only. It does not constitute legal, patent, regulatory, medical, investment or financial advice, and no attorney–client, advisory or fiduciary relationship arises from reading it. Readers should obtain qualified, jurisdiction-specific professional advice before acting on anything set out here.}

_{The patent commentary summarises decisions, office guidance and examination practice as publicly reported up to 8 September 2026. Some of the orders discussed — including the Indian Patent Office orders of 15 April 2026 — have not been tested on appeal, and controller-level orders are persuasive rather than binding. Law, guidance and practice in this area are changing rapidly and may have changed since publication.}

_{The clinical information described relates to published trials of limited size and duration. Nothing in this article is a therapeutic recommendation, or a claim or representation about the safety, efficacy, approval status or availability of any product. No product referred to here should be understood as approved for the indication discussed unless independently verified. Patients and clinicians must rely on approved product labelling and the advice of a treating physician.}

_{Company names, funding figures, valuations, listing status and pipeline positions are drawn from public reporting and third-party trackers. They are approximate, are rounded, and in several cases differ between sources; where sources conflict, the figure shown is one reasonable reading rather than a verified total. Nothing in this article is an offer, solicitation, endorsement or recommendation to buy, sell or hold any security or to enter into any transaction.}

_{All company names, product names, trade marks and case citations are the property of their respective owners and are used for identification, comment and criticism only. Their use does not imply any affiliation with, sponsorship by, or endorsement from the parties named.}