Drug discovery ML · R&D demonstration, public data · Mar 2026

Generative Design of Fragment-Like Molecules

Expanding a small fragment library into novel, synthesizable candidates using a diffusion model trained on public ZINC20 fragment data.

Challenge

Expanding a small fragment library into novel, synthesizable candidates with drug-like properties, without a chemist manually enumerating variations.

Approach

A diffusion-based generative model, conditioned on QED and synthetic accessibility, trained on public ZINC20 fragment subsets.

What we found

A majority of generated molecules passed standard drug-likeness filters. The filter thresholds and evaluation set behind that figure aren't published alongside it yet — see the caveat on the metric above.

Limits of this demonstration

This is an R&D demonstration on public ZINC20 data, not a client engagement. No repo or notebook is linked yet, which is why the metric above is flagged TODO-VERIFY. Passing a drug-likeness filter is not the same as synthesizability in practice, which this pass does not attempt to confirm.