Energy-Based Discrete Mask Approximation for 3D Molecular Graph Explanation

Xufeng Liu, Wenhan Gao, Yi Liu

Keywords: 3D Graph Explanation, 3D Molecular Graphs, Energy-Based Models, Discrete Masks

Abstract:

In recent years, Graph Neural Networks (GNNs) have become a powerful tool for modeling molecular data. To enhance their reliability and interpretability, various explanation methods have been developed to identify key molecular substructures, specifically a set of edges, in the decision-making process. Early work with 2D GNNs represented molecules as graphs with atoms as nodes and bonds as edges, neglecting 3D geometric configurations. While existing explanation methods perform well on 2D GNNs, there is a pressing need for 3D explanation methods tailored for 3D GNNs, which outperform 2D GNNs in many tasks. Current explanation methods struggle with 3D GNNs due to the construction of edges based on cut-off distances in 3D GNNs, resulting in an exponentially large number of edges. We identify the sources of errors in explanations and decompose them into two components based on a derived upper bound between the optimized masks and the actual explanatory subgraph. This gap can be significant, especially for 3D GNNs because of the large number of edges. To achieve optimal explanation fidelity, our method aims to bridge this gap by assigning two energy values to each atom based on its contribution to the prediction: one energy reflects the scenario where this node is important in making the decision, while the other represents the scenario where it is unimportant. In analogy to physics, lower energy values indicate greater stability in the prediction, and thus, we are more confident about the scenario with which it is associated. Our approach strives to push up and down the energies, respectively, to distinguish these two scenarios to simultaneously minimize both components of the derived upper bound of error, enabling us to identify a stable subgraph that maintains high explanation fidelity. Experiments conducted on backbone networks and the QM9 dataset demonstrate the effectiveness of our method in providing accurate and reliable explanations for 3D graphs.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9182