Binding Affinity Prediction for Kinase Inhibitors

Challenge

Ranking candidate kinase inhibitors by predicted binding affinity, before they reach a costly assay screen, only helps if the model is right often enough to change which compounds get tested first — not just loosely correlated with the true affinity.

Approach

A graph neural network (PyTorch Geometric) trained on public ChEMBL kinase bioactivity records, using RDKit for featurization and a physics-aware representation of ligand–protein contacts rather than a 2D-only fingerprint.

What we found

The model's ranking outperformed a baseline docking score on the data it was tested against. That comparison has not yet been through a documented held-out split, a fixed baseline definition, or a fixed seed — see the caveat on the metric above.

Limits of this demonstration

This is an R&D demonstration on public ChEMBL data, not a client engagement. No repo or notebook is linked yet, which is why the metric above is flagged TODO-VERIFY. Kinase selectivity and off-target effects are out of scope for this pass.