Fig. 1From: ABT-MPNN: an atom-bond transformer-based message-passing neural network for molecular property predictionIllustration of our proposed ABT-MPNN. The given network takes the SMILES as input and generates atom features, bond features, three inter-atomic matrices and molecular descriptors as local and global encodings of the molecule. The bond feature matrix is first learned via bond attention blocks and bond update functions in the message-passing layers. After the message-passing phase, the atomic representations are obtained by summing the incoming bond hidden states, followed by the concatenation of the atom feature matrix and a multi-head atom attention block. In the atom attention block, three scaled inter-atomic matrices are individually added to each attention head’s weights as a bias term. Finally, the learned atomic hidden states are aggregated to a molecular vector, concatenated with the molecular descriptors, then entered into feed-forward layers for property predictionBack to article page