Extrapolative AI to unlock first-in-class drugs
HQ: San Francisco, CA, United States | Founded: 2026 | Employees: 3 | Stage: Seed (Y Combinator S26) | Website: https://wondertx.ai/
WonderTx is an AI-native biotech company working on replacing needles with pills: it discovers new molecules that can be taken orally in place of injectable drugs such as insulin and Ozempic. Its insight is that these oral candidates act on biology already shown to work in people, rather than only in animal models, which sidesteps the single biggest risk in drug discovery. It estimates that $200B of treatable disease with proven human biology has no pill because discovery programmes have no workable chemical starting point, and calls this a zero-shot problem: with no training data for these targets, models that work by interpolation are not enough, so it builds its own models intended to reason beyond their training data. It reports validating the platform in internal and partner programmes, including selecting molecules on four biologically and structurally different targets with zero training data, confirmed experimentally from first screen results through to crystal structures. Its longer-term vision is a 'lights-out wonder drug factory' where AI runs the entire design-make-test loop, pairing expert judgement with robotic throughput instead of adding people. Founded in 2026 and based in San Francisco with 3 employees. Founder and chief executive Abraham Heifets previously co-founded Atomwise, the first company to apply deep learning to structure-based drug design, which raised over $220M and signed more than $5B in partnership value with Sanofi, Bayer and Eli Lilly.
WonderTx reframes the risk in drug discovery rather than claiming better molecules. It targets biology already proven to work in humans — insulin and Ozempic do work, they just have to be injected — so the risk that kills most drug programmes is mostly removed, leaving the chemistry problem of finding a pill that works. Its argument that this is a zero-shot problem is interesting: with no training data for these targets, models that interpolate from what they have seen cannot solve it by construction. Whether that is marketing or real engineering is the whole question, but the evidence is specific: it selected molecules with zero training data on four different targets, confirmed in the lab through to crystal structures. The founder co-founded Atomwise, which raised over $220M.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki