Predictions to Proof: Translating AI Into Real Biology

Author: Alex Morgan (moderator), Anna Marie Wagner, Payal Sheth, Martin Jensen Collection: UNLOCK-2026-Sessions

Summary

This panel takes stock of what AI has actually delivered in drug discovery as of Q1 2026, diagnoses the remaining gaps, and debates whether a "ChatGPT moment" for clinical success improvement is plausible. The moderator is Alex Morgan, partner at Khosla Ventures. Panelists are Anna Marie Wagner (Co-Founder & CEO, Transfyr.ai), Payal Sheth (SVP Therapeutic Discovery Sciences, Bristol Myers Squibb), and Martin Jensen (Co-Founder & CSO, Gordian Biotechnology).

Alex Morgan opens with a survey of what has genuinely arrived: antibody discovery automation at Medra; VHH (single-domain antibody, also called nanobody) discovery platforms on automated OpenTrons systems; Xvivo lymph node systems from Prellis; Generate Bio at a $1.6 billion market cap despite some controversy about the novelty of designs; AWS antibody discovery tools; Naba's zero-shot developable biologic design; and on the small molecule side, InSilico Medicine and Recursion/XCNC in the clinic with AI-designed molecules, plus Amicro (HERU inhibitor), Isomorphic, Charm Therapeutics (pan-inhibitor), and others.

Anna Marie Wagner founded Transfyr.ai out of her experience running AI teams at GKO (5 years), where she observed that biological data training models was often revealing where the science was performed rather than what the science was actually measuring — a consequence of uncontrolled experimental variation. Transfyr.ai builds an observability platform that passively collects metadata about scientific execution, making data trustworthy enough to train AI models and robots. Her core thesis: the bottleneck is now reliable evals (evaluations), not model capability. Eval development requires physical-world experiments that take a long time (assay development), and without high-quality ground truth evals, AI cannot self-improve on biological tasks the way it has on code and chess.

Payal Sheth describes Bristol Myers Squibb's "predict first" approach — predicting molecule properties before synthesis across small molecules, large molecules, and complex modalities — and their automation workflows augmenting biological throughput. She identifies the unsolved problem as connecting the design/automation pipeline to the disease state itself: current optimization measures what is technically measurable, not what reflects actual disease complexity. She advocates for perturbation platforms and disease-relevant complex systems integrated into the AI discovery stack.

Martin Jensen co-founded Gordian Biotechnology to study aging — a process that happens slowly in living organisms over time. Gordian invented in vivo multiplex perturbation technology capable of testing putative therapeutics in a single animal, now covering >1,000 genetic targets across organs (heart, liver, lung) to generate causal data at scale for aging biology. Jensen is building toward iterative organism-in-a-loop experiments: AI generates hypotheses, Gordian's system tests them in vivo causally, results feed back into the model, repeat — accelerating understanding of diseases like heart failure. He cautions that the "ChatGPT moment" for clinical improvement will not happen uniformly across disease areas: progress will be jagged, gated by whether available data actually contains the answers needed.