Jaderberg says every leaderboard inside Isomorphic carries a dotted line for experimental accuracy, and the goal for each model is to reach it: once a model is there, that step need not be done in the real world. Public data repositories, he says, are mostly the byproduct of passive collection and badly biased, so the company generates its own; animal toxicity predicts only about 30% of human toxicity, and he would rather drop animals altogether. He will not give a date for the first clinical trial.
The aim is to move drug design onto a computer and enter the real world to validate, not to learn and iterate
Jaderberg says the north star for the core AI team is to bring drug discovery, molecule design and the understanding of disease biology onto a computer: how to get out of the real world. Every internal leaderboard, he says, carries a dotted line that is experimental accuracy, and what each model is for is to reach it; once there, that experiment need not be run. Automated labs and robotic platforms have got much better and the company has worked with them for years, but he says they eventually meet the floor set by physics: synthesising a molecule means waiting for the reaction, culturing cells means waiting for the cells to reach the right state. So what he wants to cut is the number of trips into the real world to learn and iterate, not the speed of each one; the real world is still entered, because a new scientific hypothesis has in the end to be shown safe and effective in people.
Public repositories are a byproduct of making medicines, badly biased, so the company generates its own data
Asked where the data for these models comes from, Jaderberg says the company has a ridiculously comprehensive data strategy, and that what it has learned over five years is how much opportunity there is to generate data itself. The big repositories, public ones and those inside pharma companies, are mostly the byproduct of years of passive collection while making medicines, he says, and so can be very biased; if generalisation is what you care about, the distribution of data you need is a different one. Generating it themselves, he says, has shown really amazing data scaling laws. The models and the drug design engine are built entirely in house, while the experimental platforms are other people's: contract research organisations and other companies working on platforms. Proprietary data is the part he wants to own.
Animal toxicity predicts about 30% of human toxicity, so he wants to model human clinical toxicity directly
Most drugs fail in the clinic, Jaderberg says, because the problems cannot be seen until the drug is in patients: toxicity nobody accounted for, a molecule that hits what it was designed to hit and something else as well, or a target that turns out to be harmful to modulate at all. The gold standard regulators require is proof that a molecule is not toxic in rats and other animals, but, he says, choose your study and animals are about 30% predictive of human clinical toxicity; if one of his AI scientists brought him a model that was 30% predictive, he would say do a bit better. Animals have been a proxy for 50 or 70 years and he would forget them: put in vitro endpoints that are easy to measure in the lab together with AI predictive models and with newer human-tissue technologies such as organoids and organs on a chip, and model human clinical toxicity directly. The difficulty is that there is little data on that side, because only a finite number of clinical trials have been run.
Cocktails are out of reach because the molecules are not precise enough, not because of the biology
Asked why combinations are not given more often, Jaderberg says the trouble is that these molecules often carry some toxicity, or act on the liver, so taking several at once produces drug-drug interactions that can harm patients; it is because we are not good at engineering exquisite molecules that they cannot be combined safely, which limits the surface of what can be treated. If molecules could be engineered to be safe, to spare the liver and not to react badly with each other, he hopes for quite complex combinations that modulate not one target at a time but a whole vector of targets spanning complex pathways. The internal pipeline is in oncology and in immunology and inflammation, while the engine and the models, he says, are agnostic to disease area, target class and modality; internally they pick root-node problems, where one solution reaches many downstream indications.
He will not date the first clinical trial, and five years is the only deadline he gives
All of it is his own account: no programme is named, no hit rate, no dataset, and no number for experimental accuracy or for which model has reached the line. Asked when the first clinical trial will start, he says he would love to give a hard date but this is frontier science, and says only that several medicines in the works are going well and that a clinical development team is being built. The record is more specific: the timeline entry of 3 November quotes an R&D World report putting the first trials on a year-end target. The one deadline he does give is five years: he would be really disappointed if, five years on, the predictive understanding of disease biology and the ability to design much more precise and less toxic molecules had not step-changed in a massive way. He also gives a team ratio: one core domain expert to about five people building models.