Powering AI-Driven Drug Discovery with President of Isomorphic Labs Max Jaderberg
Max Jaderberg argues drug discovery stalls because it is empirical trial and error, and that predictive AI models reaching experimental accuracy can move design onto computers, raising success rates and enabling safer, more complex medicines.
Watch on YouTube ↗Charlie's summary
- ▶ 1:59
AlphaFold was the proof that a deep learning model can generalize in biology, which prompted Isomorphic Labs
Max says he and Demis were looking for the problems deep learning and reinforcement learning should be pointed at after game-playing work like AlphaGo and AlphaStar. The early AlphaFold results were “the first proof of concept” that a model could generalize in a meaningful area of biology. That gave the team conviction to start a dedicated company around 2021, with its own culture and people from several disciplines.
- ▶ 6:35
Drug discovery is stuck because it is an empirical, trial-and-error process that treats each drug as a one-off
Max thinks medicines have always been found by guessing, testing in the real world and iterating, which gives weak signal in a huge, chaotic space. The low-hanging fruit has been picked, and it is hard to raise the complexity of what can be found. Each drug is also run as a “bespoke frontier science project” with little that transfers. An AI platform can compound learnings across programs, so each medicine helps the next.
- ▶ 15:16
Get out of the real world: aim for models at experimental accuracy
Max says the main bottlenecks are time and cost of synthesizing real atoms and running assays. Isomorphic's goal is to move design onto a computer, as other engineering fields did with CAD. AlphaFold is the example: it replaces months or years of X-ray crystallography. Internally, every leaderboard has a dotted line for experimental accuracy. Past that line, you iterate in silico and go to the lab mainly to validate, so success is far more likely.
- ▶ 17:48
Drugs fail in the clinic because problems only appear in patients, so the opportunity is understanding biology before then
Max says costs are high because so much fails, and the failures come late: toxicity nobody accounted for, a molecule that hits an off-target, or a target that proves harmful. His answer is to model more than binding. Better understanding of disease biology and how cells work lets you check whether the target is really implicated before any patient is dosed.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.