Research

We want to shorten the distance between an idea for a molecule and the evidence that it works. That means building the models that propose and judge molecules, the simulations that test them, and keeping the chemistry close enough that an answer takes weeks rather than years.

Generative design

Molecules that did not exist yesterday

Screening a catalogue can only find what someone already made. We train generative models that design molecules directly in three dimensions, so that the shape and the chemistry are proposed together rather than one being fitted to the other afterwards. A proposal only counts if a chemist could actually make it.

Prediction

Deciding what is worth the effort

Models that estimate what a molecule will do before anyone spends a month on it: how strongly it acts, how it behaves in the body, and where it is likely to be toxic. We care as much about being able to interrogate a model’s reasoning as about its accuracy — a ranking nobody can question is hard to act on.

Simulation

Testing the answer with physics

Statistical models are confident in ways that are not always earned. Simulation puts a proposed molecule in its real environment and asks whether it stays where it is supposed to. It is slower and much harder to fool, which is exactly why it sits between prediction and synthesis.

Chemistry

Making it, and finding out

The point of all of the above is a real compound in a real assay. The lab designs, makes and tests molecules, and the results return to the models as training data. This loop is the reason the lab is built the way it is.

Beyond one target

Where we point it

Cancer is where the lab’s experience and data are deepest, so it is where the method is proved. Nothing about the approach is specific to it, and newer projects are already pointed elsewhere.