Directed graph attention neural network utilizing 3D coordinates for molecular property prediction

The prosperity of computer vision (CV) and natural language procession (NLP) has spurred the development deep learning in many other domains. advancement machine provides us with an alternative option besides computationally expensive density functional theories (DFT). Kernel method graph neural networks have been widely studied as two mainstream methods for property prediction. promising achieved comparable accuracy to DFT specific objects recent study. However, most high precision require fully connected graphs pairwise distance distribution edge information. This work sheds light on Directed Graph Attention Neural Network (DGANN), which only takes chemical bonds edges operates atoms molecules. DGANN distinguishes from previous models those features: (1) It learns local environment encoding by attention mechanism bonds. Every initial message flows into every passing trajectory once. (2) transformer blocks aggregate global molecular representation atomic encoding. (3) position vectors coordinates are used inputs instead distances. Our model matched or outperformed baseline QM9 datasets even without thorough hyper-parameters searching. Moreover, this suggests that directly utilizing 3D can still reach accuracies molecule rotational translational invariance incorporated.