Where Black-box Drug-Target Interaction Prediction Models Look: Cross-Method Explainability

Model explanations reveal systematic biases rather than just predictions: The study found that explainability analysis uncovers "padding and special-token artifacts" and "modality dominance" in drug-target interaction models, meaning these AI systems may rely on technical quirks in how data is formatted rather than genuine chemical properties—a finding that wouldn't be apparent from accuracy metrics alone.

No single explanation method tells the whole story: By combining five different explainability techniques (integrated gradients, saliency maps, layer-wise relevance propagation, SmoothGrad, and others), the researchers found that "single-explainer bias" can mislead. Only findings that multiple independent methods agreed on were considered reliable, suggesting organizations should use multiple explanation approaches rather than trusting one tool.

AI explanations can generate testable scientific hypotheses: The analysis identified "chemistry-consistent fragment and composition motifs" that the model uses, which the authors note "can provide testable hypotheses for downstream validation in computational drug discovery pipelines"—meaning explainability isn't just about understanding AI, but can actually accelerate real drug discovery work.

Explainability works best as model criticism, not validation: Rather than confirming that models work correctly, this audit used explanations to identify problems and limitations. The researchers emphasize these analyses "do not substitute for structural or experimental ground truth," positioning explainability as a quality-control tool that reveals what models don't understand about chemistry.