Table 1.
Total number of complexes in each data set.
Fig 1.
Graph representation of protein-ligand binding structure.
The 3D structure of the binding pocket is represented as a graph. The features of the atoms and the features of the atomic interactions are embedded in the feature matrix and adjacency matrix, respectively.
Table 2.
Description of the features.
Fig 2.
Overview of the GraphBAR architecture.
Multiple graph convolution blocks process node features with different adjacency matrices. All the outputs from the graph convolution blocks are concatenated and entered into the fully-connected layer and dropout layer to predict the binding affinity.
Fig 3.
The framework of the graph convolutional block.
Each graph convolutional block consists of three graph convolutional layers. The graph convolutional layers are combined with a fully connected layer and a dropout layer, except for the last graph convolutional layer, which is followed by the graph gather layer.
Table 3.
The performance of GraphBAR on the PDBbind v.2016 core set.
Table 4.
The performance of GraphBAR with the PDBbind v.2013 core set.
Table 5.
The performance of GraphBAR with the CSAR NRC-HiQ set 1 (51 complexes).
Table 6.
The performance of GraphBAR with the CSAR NRC-HiQ set 2 (36 complexes).
Table 7.
The training time of the models with the PDBbind 2016 database.