For message-passing neural networks, a molecule is represented as a graph: atoms are nodes, bonds are edges carrying spatial-distance information, and pairwise relationships encode three-dimensional orientation. section 7.3.2
Each atom repeatedly sends learned messages to its neighbors and updates its hidden state from incoming messages. A readout pools the atom states into one vector for predicting a molecular property. section 7.3.2
From atom messages to molecular predictions
After successive rounds, each atom’s state summarizes its local neighborhood; pooling these states can predict properties such as energy, dipole, or gap. Distant information may require many rounds to travel, motivating virtual edges or master nodes. section 7.3.2
