$4B AI Drug Founder: Left OpenAI, $1T Eli Lilly Signed On, Designing Drugs in 24 Hours

Chai Discovery lifts antibody design success to double digits

Chai Discovery uses language models to design drugs, treating proteins and DNA as the real natural language for AI. Co-founder Josh Meier, who trained early GPT models at OpenAI, says the company pushed antibody design from about a 0.1% lab success rate to double digits in roughly a year. The bottleneck in drug discovery is no longer any single thing, since compute and data both respond, and the models already built would have a large impact even if work stopped. Eli Lilly adopted the technology quickly and stress tests it at scale, which shapes Chai's research roadmap.

Takeaways

• Antibody design success rose to double digits. A year earlier about 0.1% of designed antibodies bound as expected in the lab, sometimes one in ten thousand. Chai-2 reached roughly 15 to 16%, and Chai-3 more than doubled that.
• Design takes 24 hours plus two weeks. Meier says an initial molecule now takes about a day of compute and a couple of weeks of lab testing, where it once took months or years. The goal is a zero-shot candidate needing no changes.
• Structure prediction improved tenfold. AlphaFold predicted antibody-antigen complexes correctly about 3% of the time a few years ago. Chai's latest models improved that by an order of magnitude, with cryo-electron microscopy showing errors smaller than an atom's diameter.
• Lilly stress tests models at scale. Meier says the Lilly team adopted the technology quickly and rigorously, running the right experiments before scaling. Their testing helps set Chai's research roadmap, and pharma partnerships will decide which diseases get prioritised.
• Validation spans 100 to 200 projects. A customer spent $5m to $10m and one to two years failing to design a molecule that binds both a human and a monkey protein. Chai's models produced it zero-shot.
• Compounding gains, not one secret. Meier attributes progress from Chai-1 to Chai-3 to 5 to 10% improvements stacking each week or month, not a single breakthrough. Roughly 30 people work in person in San Francisco with very few meetings.
• Competition is now its own models. Chai's real competitor was once yeast, Meier says, and now it is Chai's previous models. Chai-4 will compete with Chai-3, and data will convince skeptics as it did for ChatGPT.

From OpenAI to Chai

1. Meier studied chemistry and programmed from childhood.
2. At OpenAI he trained early language models including GPT-1 and GPT-2.
3. He came to see DNA and proteins as the real natural language for AI.
4. Chai started as an R&D project that open sourced its models.