AINEWS
Genentech R&D chief Aviv Regev told the International Conference on Machine Learning the bottleneck is closing the experimental loop, not building bigger models.
If a transformer can write a sonnet, prove a theorem, and route a packet, why can't it make a real approved drug? Aviv Regev, who runs drug R&D at Genentech, opened her ICML 2026 keynote in Seoul on July 8 with that question, and her answer is that the field has been asking the wrong one. The bottleneck in AI pharma is not a model-size problem. It is a loop-closure problem.
She is the head of research and early development at Genentech, the biotech arm of Swiss drugmaker Roche. Her argument, delivered to a machine-learning audience that already believes in models, is that drugs fail not because the models are too small but because no one has closed the experimental loop between the model's prediction, the wet lab that tests it, and the patient who receives it.
The math, she said, is what makes a brute-force solution impossible. The human body has roughly 37 trillion cells. About 100,000 genetic loci are linked to disease. Five-locus genotype combinations run on the order of 10^19. Drug-like small molecules number around 10^60. Therapeutic antibody sequences sit near 20^32. "You cannot enumerate that space," Regev said. "You have to learn it, and you have to learn it with the world in the loop."
Regev called the pattern "lab in the loop," and showed what she meant with a demo. A model reconstructed single-cell and spatial gene-expression maps from a single 150-year-old hematoxylin-and-eosin (H&E) histology slide, the routine stain pathologists have used since the 1880s. If a model can read one old stain and recover the molecular map beneath it, every archived slide in every hospital becomes a training example. The "lab" half of the loop is the wet experiment. The "clinic" half is the patient. The AI's job is to connect them.
Four Genentech drugs anchor the case. Inavolisib targets the PIK3CA mutation in HR-positive, HER2-negative breast cancer; the patient population was selected to match the drug's mechanism, which is what Regev means by translating at the patient axis. Duradestrant, an oral selective estrogen degrader, cut the risk of invasive recurrence or death by roughly 30% compared with standard care in its Phase 3 Lidera trial, according to Regev at ICML 2026. Divarasib, a KRAS G12C inhibitor for non-small cell lung cancer, is about 25 times more potent than the first-generation G12C drugs and 10 to 50 times more selective, according to Regev at ICML 2026, and posted a positive Phase 3 CRESCENDO-1 readout on July 1, a week before the talk. Fenabrutinib, a BTK inhibitor being tested in multiple sclerosis, has produced relapse intervals of 16 to 19 years on treatment in follow-up, according to Regev at ICML 2026, a number that, if it holds, would be a category change for the disease.
The dollar-stakes counterfactual is lampalizumab, Genentech's age-related macular degeneration antibody that failed Phase 3 in 2017 at a cost Regev put at roughly $500 million. She argued, on stage, that a baseline-correction model trained on historical trial data could have flagged or even reversed the failure earlier by adjusting for the imbalance between patient arms. The number is Regev's, and the analysis is retrospective, but the lesson is the one she wants the ML audience to take home: the most expensive place to learn that your model was wrong is in a 1,000-patient trial.
Regev closed with the trajectory she thinks the loop opens up: personalized cancer vaccines, designed against the unique neoantigens of one patient's tumor, what she called the "one person, one drug" endpoint. That is an exploration, not a marketed therapy, and she was careful not to oversell it. The next test of the thesis is a data point: whether the divarasib Phase 3 readout, fully unblinded later this year, holds up across subgroups, and whether the lab-in-the-loop pipeline that produced it can be reproduced at the next target.