Expressive Graph Informer Networks

Applying machine learning to molecules is challenging because of their natural representation as graphs rather than vectors. Several architectures have been recently proposed for deep learning from molecular graphs, but they suffer from information bottlenecks because they only pass information from a graph node to its direct neighbors. Here, we introduce a more expressive route-based multi-attention mechanism that incorporates features from routes between node pairs. We call the resulting method Graph Informer. A single network layer can therefore attend to nodes several steps away. We show empirically that the proposed method compares favorably against existing approaches in two prediction tasks: (1) 13C Nuclear Magnetic Resonance (NMR) spectra, improving the state-of-the-art with an MAE of 1.35 ppm and (2) predicting drug bioactivity and toxicity. Additionally, we develop a variant called injective Graph Informer that is provably more powerful than the Weisfeiler-Lehman test for graph isomorphism. We demonstrate that the route information allows the method to be informed about the non-local topology of the graph and, thus, it goes beyond the capabilities of the Weisfeiler-Lehman test. Our code is available at github.com/jaak-s/graphinformer.

Automated summary

This paper introduces a route-based multi-attention mechanism that incorporates features from routes between node pairs, aiming at addressing the information bottleneck issue in deep learning from molecular graphs. The proposed method, called Graph Informer, is able to attend to nodes several steps away, and it outperforms existing approaches in two prediction tasks. Furthermore, a variant method called injective Graph Informer is developed and proven to be more powerful than the Weisfeiler-Lehman test for graph isomorphism.