five major platforms launched within weeks of each other this spring.
GPT-Rosalind. Claude Life Sciences. AlphaGenome commercial. Amazon Bio Discovery. the Virtual Biology Initiative. hundreds of robot arms. $500M philanthropic bets. the most concentrated burst of AI for science announcements in history.
when five platforms launch simultaneously with similar claims into the same space, it means nobody has won yet and everyone is scared of being left out. this is land-grab behavior. the announcements are real. the moats are not. the one thing almost nobody has, the thing that actually determines who wins, is validated clinical outcomes at Phase 3. zasocitinib cleared it in December 2025. one drug. after decades of promises. the field should be asking precisely why that method worked rather than assuming it validates everything else.
the metrics circulating are systematically misleading and worth examining precisely.
AI-designed drugs passing Phase 1 at 80 to 90 percent sounds transformative. Phase 1 is a safety trial. you are asking whether the molecule kills people. AI is genuinely good at this, toxicity prediction, ADMET forecasting, molecule stability. real capability. but the 90 percent failure rate in drug development sits in Phase 2 and 3, where drugs fail on efficacy. the target was wrong. the disease model was wrong. the patient population was wrong. AI has not moved that needle. measuring Phase 1 success and presenting it as validation of the pipeline is the most widespread sleight of hand in the current narrative.
there is one genuine proof point worth examining carefully, and one genuine failure worth understanding.
on June 3, 2025, the industry's first proof-of-concept clinical validation of AI-driven drug discovery was published in Nature Medicine. Insilico Medicine reported promising safety and efficacy results from a Phase IIa trial of Rentosertib, a TNIK inhibitor developed using generative AI for idiopathic pulmonary fibrosis, a disease with no currently available therapies that reverse its course. patients receiving 60mg Rentosertib experienced a mean improvement in lung function of +98.4ml, compared to a mean decline of -20.3ml in the placebo group. this is real. it matters. it is also a Phase 2a trial in 71 patients, promising, not proven. the Phase 3 question is still open.
the counterpoint is Recursion. the most well-funded, most watched AI drug platform in the world. in May 2025, Recursion terminated three clinical programs, REC-994, REC-2282, and REC-3964, after accumulated data did not justify continued investment. REC-994 had reached Phase 2 in cerebral cavernous malformation. long-term extension data failed to confirm earlier signals of efficacy. four programs halted. $463M net loss in 2024. the platform is real. the clinical outcomes are not there yet.
the virtual cell is the claim that deserves the most scrutiny.
the standard framing positions it as the AlphaFold moment for cellular biology. AlphaFold solved a prediction problem with clean inputs and clean outputs. the virtual cell is a causal inference problem. you need to predict what happens when you intervene, when you add a drug, knock out a gene, change a signalling pathway. a model can predict observed cellular states perfectly and be completely wrong about what happens under perturbation. which is exactly the regime drug discovery operates in.
but the deeper problem is information completeness. for language models, all the tasks you want the model to do, summarise, reason, generate, translate, are contained in the text data itself. the signal and the task live in the same space. biology does not have that property. no single biological modality is information-complete for the tasks drug discovery actually requires. sequence data does not contain drug response. gene expression does not contain clinical outcome. transcriptomics at hour 6 post-dose is categorically different from hour 48. this is not a data coverage problem that more experiments solve. it is a structural mismatch between the modality and the task. which is why, as Anshul Kundaje observed after AlphaFold and EVO2, naive scaling is actually very challenging in biology, and models with domain-specific constraints and inductive biases still compete with foundation models even at scale.
Biohub's $500M is the most honest of the big bets precisely because Alex Rives said the quiet part publicly: we do not yet know what the slope of the scaling law is with cellular biology. that is the entire wager. scaling worked for protein structure. whether it works for cellular dynamics is unknown. $500M on an unknown scaling curve.
the data problem does not get solved by more robots either. the critical gap in building virtual cell models is not compute. it is diverse perturbation data across biological contexts. running the same class of experiments faster does not solve the coverage problem. scale is not the same as coverage.
which brings us to the signal that actually predicts clinical success and is almost entirely absent from the current narrative.
human genetic evidence. drug targets supported by genetic variants in human populations, where a natural experiment confirms that modulating this target changes this disease phenotype in actual humans, succeed in trials at two to three times the rate of targets without that evidence. BridgeBio built an entire company on this single insight. pursuing genetically validated targets increases the probability of success by more than 2x and expands the frontier of drug feasibility by 35%. three positive Phase 3 readouts in just over three months is what that looks like in practice.
AlphaGenome points directly at this. reading regulatory DNA. identifying which variants in the non-coding genome are causally linked to disease states in human populations. that capability, if it matures, gives you a human genetic filter for target selection grounded in actual human causal biology rather than animal model proxies.
so here is the honest state of the field.
the digitisation of biology is real. the tools are arriving. the infrastructure is being built. the trial architecture is genuinely being redesigned. Insilico's Rentosertib Phase 2a result is the first clinical signal that AI-discovered drugs can work on efficacy. zasocitinib clearing Phase 3 in December 2025 is the first AI-designed drug to cross that threshold, the most significant milestone the field has produced. Recursion's failures are the honest reminder that platform scale and clinical outcomes are still not the same thing.
the gap between narrative and evidence is not a reason to disengage. it is a reason to be precise about where in the stack value actually accrues.