A graph attention network (GAT) is a graph neural network that learns how strongly each neighboring node should influence a node’s representation. In molecular graphs, this can emphasize neighbors belonging to important functional groups. section 11.5.4
Compared with graph convolutional aggregation, GAT provides:
- Learned edge weights based on node features rather than fixed degree normalization. section 11.5.4
- Multi-head attention, which learns multiple neighborhood patterns and concatenates or averages their outputs. section 11.5.4
- More interpretable attention weights, with intermediate expressiveness between GCN and GIN. section 11.5.4
Molecular graph layer
A two-head GAT layer can replace a graph convolution layer while averaging the two attention outputs: self.mol_conv1 = GATConv(num_features_mol, num_features_mol, heads=2, concat=False). section 11.5.4
For drug-target affinity prediction, GAT is evaluated alongside GCN and GIN on DAVIS and KIBA subsets; full dual-graph training can require several hours on a modern GPU. section 11.6



