Directed Message Passing Based on Attention for Prediction of Molecular Properties

TL;DR

This paper introduces Directed Graph Attention Networks (D-GATs), a novel GNN model that uses directed bonds and attention mechanisms to improve molecular property prediction, outperforming existing models on most benchmarks.

Contribution

The paper proposes D-GATs, a new GNN architecture utilizing directed graphs and attention to better capture molecular sub-structures, advancing molecular property prediction.

Findings

D-GATs outperform state-of-the-art models on 13 out of 15 benchmarks.

Directed bonds and attention mechanisms enhance the capture of functional groups.

The approach improves molecular property prediction accuracy.

Abstract

Molecular representation learning (MRL) has long been crucial in the fields of drug discovery and materials science, and it has made significant progress due to the development of natural language processing (NLP) and graph neural networks (GNNs). NLP treats the molecules as one dimensional sequential tokens while GNNs treat them as two dimensional topology graphs. Based on different message passing algorithms, GNNs have various performance on detecting chemical environments and predicting molecular properties. Herein, we propose Directed Graph Attention Networks (D-GATs): the expressive GNNs with directed bonds. The key to the success of our strategy is to treat the molecular graph as directed graph and update the bond states and atom states by scaled dot-product attention mechanism. This allows the model to better capture the sub-structure of molecular graph, i.e., functional groups. Compared to other GNNs or Message Passing Neural Networks (MPNNs), D-GATs outperform the state-of-the-art on 13 out of 15 important molecular property prediction benchmarks.