Learning inhibitor interactions from protein deep mutational scanning data and deep learning models

Ligand-aware geometric models (LigandMPNN, LASErMPNN) partially recover resistance-associated hotspots across 11 inhibitor conditions for the MET kinase domain, while current AI models remain largely ligand-invariant.

• Why it matters: Drug responses that depend on specific mutations challenge generalization because small sequence changes can yield large, inhibitor-specific fitness and binding effects.
• What they did: The study benchmarks mutational effect and drug-target affinity prediction models on deep mutational scanning datasets spanning 11 inhibitor conditions for MET, then extends checks to additional kinase-inhibitor datasets.
• The result: Affinity models underestimate mutation-driven binding changes, and the authors attribute this to mutation-invariant structural priors and limited mutation-conditioned supervision, implying data scarcity is a key bottleneck.

AI summary of the abstract · may contain errors · not medical advice · read the paper.