Explainability Audit Reveals Insights into Drug-Target Interaction Prediction Models.

Drug-target interaction (DTI) and affinity (DTA) prediction models are achieving high performance, but their internal workings, especially regarding sequence, fingerprint, and graph features, often remain unclear. This research conducts an interpretability audit on the BridgeDPI architecture across multiple datasets. The study employs a combination of gradient-based attribution methods, such as integrated gradients and SmoothGrad, alongside feature-wise occlusion ablation. By using a strict intersection consensus across these methods, the aim is to mitigate biases from single explainers and provide a more robust understanding of model behavior. The findings indicate that explainability serves as valuable model criticism, uncovering issues like modality dominance, artifacts from padding, and chemistry-consistent fragment motifs. While not a substitute for experimental validation, these analyses can generate testable hypotheses for computational drug discovery and highlight the potential of XAI in understanding complex DTI/DTA models.