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

Recent advances in protein language models, geometric deep learning, and structure prediction frameworks have the potential to accelerate computational drug discovery. However, it remains unclear how well these methods generalize beyond their training sets. Mutation-dependent drug responses present a challenge to these methods because there are often large effects from relatively small perturbations. Here, we systematically benchmark diverse mutational effect and drug-target affinity prediction models using deep mutational scanning datasets spanning 11 inhibitor conditions for the MET kinase domain. Across both mutational effect prediction and drug-target affinity tasks, current AI models produce largely ligand-invariant predictions despite experimentally distinct inhibitor-dependent fitness landscapes. Ligand-aware geometric models, including LigandMPNN and LASErMPNN, partially recover experimentally validated resistance-associated hotspots, highlighting the importance of explicitly modeling protein-ligand interactions and local structural flexibility. In contrast, deep learning affinity prediction models underestimate mutation-induced changes in binding affinity. This deficiency likely reflects mutation-invariant structural priors derived from co-folded structure prediction frameworks and the absence of mutation-conditioned supervision in current affinity datasets. Finally, we further extended our analysis to several kinase-inhibitor deep mutational scanning datasets and confirm the overall trends. Our findings suggest that a primary limitation in improving performance on this task lies in the scarcity of mutation-conditioned protein-ligand training data rather than model architecture alone.