3DGraphX: Explaining 3D Molecular Graph Models via Incorporating Chemical Priors

Xufeng Liu, Dongsheng Luo, Wenhan Gao, Yi Liu

Abstract:

We consider the explanation of 3D graph neural networks (GNNs)
in the field of molecular learning. Recent studies have modeled
molecules as 3D graphs, but there exist formidable challenges for 3D
graph explanation. In this work, we propose a novel and principled
paradigm, known as 3DGraphX, for 3D molecular graph explana-
tion. Unlike existing 2D GNN explanation methods, 3DGraphX fo-
cuses on 3D motifs, which are subgraphs showing great occurrence
and function significance in molecular activities. Once generated,
3D motifs are fixed in the explanation model; hence, 3DGraphX
produces more accurate and chemically plausible explanations in
an efficient manner. 3DGraphX contains two branches with several
novel methods for instance-level and geometry-level explanations,
respectively. Two novel components, known as the mask pooling
component and mask unpooling component, are developed to dis-
cover important motifs for each 3D molecule as the instance-level
explanation. Local spherical coordinate systems are built to in-
vestigate the relative positions among motifs for geometry-level
explanation. Altogether, 3DGraphX sheds light on the character-
istics of molecules as well as the behaviors of 3D GNNs in molec-
ular learning. Experimental results show that 3DGraphX signifi-
cantly outperforms baselines in instance-level explanation with
various explanation budgets. Additional experiments show that
3DGraphX reveals the important geometries taken by 3D GNNs
for accurate molecular learning. The code is publicly available at
https://github.com/xufliu/3DGraphX.