Max Jaderberg, president of Isomorphic Labs, says the company proposes molecule designs with generative models, filters them on the computer with 'virtual lab' models and then makes and tests only 20 or 50, where experimental screening once went through hundreds of thousands; making a drug, he says, still needs about half a dozen AlphaFold-level breakthroughs.
Models filter on the computer first, so the lab makes only 20 or 50 molecules
Jaderberg says AI can propose thousands or millions of designs for the next round of a molecule, and rather than send them all to the lab, Isomorphic runs a second set of models, which he calls simulation, world or virtual-lab models, to predict on the computer how each would behave and keeps only the ones it is most confident in. Instead of screening hundreds of thousands of designs, he says, they can make 20 or 50, with a high enough share of hits to drop large-scale screening. The experimental success rate is a core internal metric: a gain on a model benchmark should show up in it. This is not lab-in-the-loop, he says; the aim is to bring as much of the real world onto the computer as possible. In his Bloomberg interview a week earlier he described a company that makes drugs and sells no model access, with no figures of this kind.
When a benchmark rises and the programmes do not, the evals get redesigned
A drug, he says, has to clear many hard problems at once: the target, the molecule, whether it reaches the right part of the body, safety and toxicity. So it needs about half a dozen or more AlphaFold-level breakthroughs, which the company has been working through for five years. Drug discovery does not need self-driving levels of accuracy; a clear margin over what the rest of the industry has is enough to use on live programmes, and when a machine-learning benchmark improves while the programmes do not, the evals have to be redesigned. His example is a modelling task whose team target is 80% accuracy on a benchmark where, he says, the world stands at 10%.
Chemists set the design goals; agents draw the molecules
A chemist used to change a molecule atom by atom on a whiteboard; at Isomorphic, he says, the chemist now defines the design space and goals for the next round, and design agents produce the individual designs on their own. The goals often conflict, so that pushing one property up drags another down, which makes this an open search balancing exploration and exploitation rather than greedy optimisation. Because results in biology and drug discovery cannot be verified on a computer, he says, each part of the agent system needs its own evaluation of whether it beats a person, or works faster.
No hit rate, programme or data is given
The 20 or 50 molecules and the high share of hits are his description, with no programme, hit rate or data; the 80% and 10% belong to an unnamed benchmark. That animal toxicity models predict only about 30% of human toxicity is a figure he cites. Asked when an AI-designed drug will reach a pharmacy, he says he does not know; the internal pipeline is in oncology and in immunology and inflammation, and which programme reaches the clinic first is, he says, too early to say.