Is The Era Of AI-Designed Drugs Actually Here? - EP 50 Josh Meier and Jack Dent
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
This episode features Josh Meier and Jack Dent, co-founders of Chai Discovery, discussing their groundbreaking work in AI-driven drug discovery. (02:00) The conversation explores how their company is using large language models to design novel antibodies and therapeutic molecules at unprecedented speed and success rates. (23:00) Founded in 2024 and already valued as a unicorn, Chai Discovery represents a new wave of AI companies applying advanced machine learning to biology, potentially transforming how medicines are discovered and developed. (38:00)
- Main theme: How AI language models are revolutionizing drug discovery by designing functional antibodies and therapeutic molecules orders of magnitude faster than traditional methods
Speakers
Josh Meier
Josh is co-founder and CEO of Chai Discovery, bringing together his dual passions for programming and biochemistry. He started his career at Stripe at age 18, worked at OpenAI during the GPT-1 era, and later led protein language modeling research at Facebook/Meta's AI Research lab. He then spent three years at biotech company AbSci before founding Chai Discovery in 2024.
Jack Dent
Jack is co-founder of Chai Discovery and former head of engineering at Stripe, where he witnessed the company grow from under 200 to 8,000 employees. He studied computer science at Harvard alongside Josh and brings extensive engineering leadership experience from building large-scale financial infrastructure systems to the biotech space.
Key Takeaways
AI Models Achieve 15% Success Rate vs. Traditional "One in a Billion" Odds
Traditional antibody discovery involves brute force screening of billions of random molecules, often yielding success rates of one in a million or worse. (38:00) Chai Discovery's models now generate molecules with a 15% success rate - meaning 15 out of 100 designed molecules actually bind to their intended targets. This represents a revolutionary improvement that makes rapid testing economically viable, as companies can now test 20-50 molecules instead of screening billions.
Drug Discovery Timeline Compressed from Years to Days
The traditional process of discovering and optimizing antibody therapeutics typically takes 6 months to a year for initial discovery, followed by additional months of optimization. (34:30) Chai's AI models can now generate candidate therapeutic molecules within 24-48 hours on a computer, with drug-like properties built in from the start. This compression allows pharmaceutical companies to focus resources on clinical trials rather than the discovery bottleneck.
Models Display "Superhuman Intelligence" in Molecular Design
Unlike chess or Go where humans can still compete, no human can design functional protein sequences or antibodies from scratch given just a target. (40:00) The AI models demonstrate genuine scientific creativity, generating molecules that are vastly different from existing drugs and taking approaches that human researchers would likely never explore - similar to AlphaGo's famous "Move 37" moment.
Previously "Undruggable" Targets Now Accessible
Many important disease targets have been considered "undruggable" because they're embedded in cell membranes or unstable outside their natural environment, making traditional screening impossible. (74:55) AI models can design molecules against these challenging targets without needing to extract and stabilize the proteins first, potentially unlocking entirely new categories of medicines for diseases that currently have no treatments.
Models Self-Assess Confidence and Guide Experimental Priority
The AI models have developed the ability to communicate their own confidence levels about different molecular designs, similar to how ChatGPT says "I don't know" when uncertain. (81:29) This allows researchers to focus laboratory resources on the most promising candidates and avoid wasting time on designs the model itself flags as uncertain, making the entire discovery process more efficient.
Statistics & Facts
- Chai Discovery achieves a 15% success rate in designing functional antibodies, compared to traditional methods that find maybe one successful molecule out of millions or billions tested. (38:00)
- The pharmaceutical industry approves only 40-60 new medicines per year globally, with roughly half being antibody-based therapeutics. (23:30)
- The average drug has approximately 5 off-target effects (unintended interactions), while aspirin has between 20-25 known off-target effects. (51:00)
