Drug-discovery models, in context

I contribute to the Ersilia Open Source Initiative, a non-profit building free ML tools for neglected-disease research. My work there is chem-icl.

The idea

Predicting a molecule's toxicity usually means training a fresh model per task. chem-icl skips training: it groups Ersilia model outputs into representation families, selects the best features, and hands a labelled context to TabPFN, a transformer that reasons over it at inference. The context is the input; nothing is fit.

The pipeline

                    ┌─ concat → select ─┐ physchem
SMILES+labels ─group┤  (Isaura)   (≤k)  ├─ concat → TabPFN → P(active)
  into families     └─ concat → select ─┘ learned            (+ RF baseline)

Molecules resolve cheapest-first: local store → cloud lake → a one-off run, cached back. Every model must cover every molecule.

What the benchmark says

family_stack vs family_concat across 9 TDC tasks: a dead heat — Per-family stacking vs. a plain concat, over nine TDC ADME/Tox tasks × ten seeds — a dead heat. So `chem-icl` ships the simpler concat as default.

GitHub ↗Live website ↗