Fig. 2: Drug-Target Interaction Network.
From: Improving the generalizability of protein-ligand binding predictions with AI-Bind
a The drug-target interaction network used to train the DeepPurpose models consists of 10,416 ligands and 1391 protein targets. Ligands and proteins are represented by green and pink nodes, respectively. b Network neighborhood of the ligand Ripk1-IN-7. Solid links represent positive or binding annotations, while dashed links refer to negative or non-binding annotations. Ripk1-IN-7 has one positive and two negative annotations in the training data, implying a degree ratio ρ of 0.33. c Protein degree ratios {ρp} and DeepPurpose predictions are highly correlated with rSpearman = 0.94. We observe that the predictions for the top 100 false positive protein-ligand pairs include the proteins with large {ρp} represented by the red crosses, whereas the false negative pairs are contributed by the proteins with small {ρp} which are represented by the blue triangles. d Examples of proteins and ligands with large degree ratios, contributing to false positive predictions. Source data are provided as a Source Data file.
