Approach
Symmetry as latent dynamics
Equivariant networks guarantee correct behaviour under rotation, but the cost grows badly with molecule size. Voxel models scale comfortably yet learn rotational structure only implicitly.
FLDWM takes a third route: a molecule is voxelised, compressed into a latent grid, then randomly rotated. A world model conditioned on that rotation learns to invert it — treating the latent as state and the rotation as an action. The result is approximate SO(3) equivariance with a provable error bound, at the speed of an ordinary convolutional network.
Application
Alternatives to drugs that price patients out
Seeding generation with FDA-approved drugs produces novel analogs in the lead-optimisation range. We chose our first five — Osimertinib, Imatinib, Sotorasib, Ivacaftor, Bedaquiline — because each carries a real unmet need behind it: cost, toxicity, or acquired resistance.
Research
Work
FLDWM: a world model for learning approximate SO(3) equivariance for scalable and generalizable molecule generation.