Weisfeiler-Lehman Embedding for Molecular Graph Neural Networks
This work addresses performance bottlenecks in molecular property prediction for chemistry and drug discovery, presenting an incremental improvement by integrating established feature engineering into GNNs.
The paper tackled the limited performance of graph neural networks (GNNs) for molecular property prediction due to the 'curse of depth' by expanding atom representation using Weisfeiler-Lehman (WL) embedding to capture local atomic patterns. The result showed that WL embedding consistently improved empirical performance across multiple GNN architectures and datasets, with theoretical analysis indicating it can replace the first two layers of a ReLU GNN with a smaller weight norm.
A graph neural network (GNN) is a good choice for predicting the chemical properties of molecules. Compared with other deep networks, however, the current performance of a GNN is limited owing to the "curse of depth." Inspired by long-established feature engineering in the field of chemistry, we expanded an atom representation using Weisfeiler-Lehman (WL) embedding, which is designed to capture local atomic patterns dominating the chemical properties of a molecule. In terms of representability, we show WL embedding can replace the first two layers of ReLU GNN -- a normal embedding and a hidden GNN layer -- with a smaller weight norm. We then demonstrate that WL embedding consistently improves the empirical performance over multiple GNN architectures and several molecular graph datasets.