Brief

A new research paper explores the explainability of black-box drug-target interaction (DTI) prediction models, specifically auditing the BridgeDPI architecture. The study employs a combination of gradient-based attribution methods and feature-wise occlusion to understand how these models utilize sequence, fingerprint, and graph features. The findings suggest that explainability can serve as a critical tool for model evaluation, revealing issues like modality dominance, artifactual patterns, and dataset-specific behaviors, thereby generating hypotheses for further validation in drug discovery. AI

Researchers have developed WassersteinGrad, a new method for explaining neural network predictions in dynamic physical fields, particularly for autoregressive weather forecasting. Existing gradient-based methods struggle with these complex data types, as input perturbations can cause geometric displacements in attribution maps, leading to blurred explanations. WassersteinGrad addresses this by computing an entropic Wasserstein barycenter of perturbed attribution maps to achieve a geometric consensus, showing improved explainability over baseline methods on regional weather data. AI