Fig. 1

Illustration of the proposed FFiNet model. a Illustration of four types of potential energy terms in a molecule. b Illustration of message passing in FFiNet. Message from 1-hop nodes (pink), 2-hop nodes (green), and 3-hop nodes (blue) are used to inform the update to the embedding of the carbon atom located at the junction of two rings. c The model structure of FFiNet. The model takes 3D molecular structure as input and generates atom features and spatial information. Then the model takes the embedded atom features and spatial information to the k-hop attention module and gets k-hop outputs. Positional encoding and axial attention are applied to get the final node embeddings to distinguish the nodes from different hops. Finally, the model uses a readout function and a multi-layer perceptron (MLP) to predict molecular properties