AI-Bind is a pipeline for predicting protein-ligand binding which combines network science methods with unsupervised pre-training to control for the over-fitting and the annotation imbalance of existing libraries. AI-Bind leverages the notion of shortest path distance on a network to identify distant protein-ligand pairs as negative samples and uses experimentally validated non-binding protein-ligand pairs to ensure sufficient positive and negative samples for each node in the training data. Additionally, AI-Bind learns, in an unsupervised fashion, the representation of the node features, i.e., the chemical structures of ligand molecules or the amino acid sequences of protein targets, helping circumvent the model’s dependency on limited binding data. ML models characterize the likelihood of each node (proteins and ligands) to bind to other nodes according to the features and the annotations in the training data.
Related Posts
