Isomorphic Labs chief executive Max Jaderberg says AI drug discovery will need several major advances working together, not one miracle model. In an interview recounted by BigGo Finance, he describes a strategy that pairs AI-generated designs with models that predict which candidates are worth testing in the lab.
Isomorphic Labs Watch analysis
What happened
Jaderberg puts the number at “maybe half a dozen plus” AlphaFold-level breakthroughs. AlphaFold showed that AI could make important progress on protein structures, he says, but it is not, on its own, a solution to making medicines. Drug development combines problems including choosing a target, designing a molecule, getting it to the right place in the body, and establishing safety.
He describes Isomorphic’s approach as using generative models to produce large sets of candidate designs, then simulation or “world” models to narrow them down before expensive laboratory work. Jaderberg says the company can make 20 or 50 molecules instead of experimentally screening hundreds of thousands, with a high proportion of “hits”. That is his account of the company’s work, not an independently reported measure of success.
Why it matters
The proposed advantage is not simply that AI can generate more candidates. It is that models might help researchers choose which few candidates merit the slow, costly test of making them and seeing what happens. Jaderberg says progress should be judged by whether better internal benchmarks translate into breakthroughs on live discovery programmes, rather than by benchmark scores alone.
He also argues that data, not algorithms or computing power, is the constraint in many areas where AI needs to predict physical outcomes. His examples include materials science and engineering, where better models of the physical world could help researchers test ideas on a computer before running real experiments.
Our read
This is a more useful pitch than “AI will discover drugs” on its own: the proposed system has to earn its keep by shrinking the experimental search and producing results that survive contact with the lab. The distinction matters. A promising design is not a medicine, and Jaderberg gives no date for an AI-designed drug reaching patients. Biology has a stubborn habit of keeping the final vote.
For readers, the useful test is whether the company can show that its predictions improve real-world experimental outcomes, not merely that its models generate plausible candidates. Jaderberg’s account of Isomorphic’s strategy is reported by BigGo Finance from a Silicon Valley Girl podcast interview.
What to watch
- Whether Isomorphic shares evidence that its candidate-selection approach improves experimental success rates.
- How the company connects internal benchmarks to results on live drug-discovery programmes.
- Whether data generation and physical-world modelling become as important as larger models in drug research.
- What independent evidence emerges from clinical trials and later stages of development.
Discussion spark: What should count as convincing evidence that AI is improving drug discovery: better benchmark scores, fewer experiments per useful candidate, or results in clinical trials?
Sources and evidence
Independent WittyWires tracker for public updates about Isomorphic Labs. Not affiliated with or endorsed by Isomorphic Labs; this is not an official account.