Confidence-Aware Explanations for 3D Molecular Graphs via Energy-Based Masking

TMLR Paper6870 Authors

Abstract: Graph Neural Networks (GNNs) have become a powerful tool for modeling molecular data. To improve their reliability and interpretability, various explanation methods aim to identify key molecular substructures, typically a subset of edges, influencing the decision-making process. Early work on 2D GNNs represented molecules as graphs with atoms as nodes and bonds as edges, ignoring 3D geometric configurations. While existing explanation methods perform well for 2D GNNs, there is a growing demand for 3D explanation techniques suited to 3D GNNs, which often surpass 2D GNNs in performance. Current explanation methods struggle with 3D GNNs due to the construction of edges based on distance cutoffs, leading to an exponential increase in the number of edges a molecular graph possesses. We identify key sources of explanation errors and decompose them into two components, derived from an upper bound between optimized masks and the true explanatory subgraph. This gap is especially significant in 3D GNNs because of the dense edge structures. To improve explanation fidelity, our method assigns two energy values to each atom, representing its contribution to predictions: one for importance and one for non-importance. The explanation model becomes more confident when the distinction between importance and non-importance is clearer. Analogous to physics, lower energy values indicate greater stability, enhancing confidence in the associated scenario. By optimizing these energy values to distinguish the two cases, we minimize both components of the error bound and identify a stable subgraph with high explanation fidelity. Experiments with various 3D backbone models on widely used datasets are conducted to validate our method's effectiveness in providing accurate and reliable explanations for 3D molecular graphs.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Shinichi_Nakajima2

Submission Number: 6870