Gene Ontology aided Compound Protein Binding Affinity Prediction Using BERT Encoding

The drug-target binding affinity(DTA) indicates the strength of the drug-target interaction; therefore, predicting DTA by computational approaches can considerably benefit drug discovery by narrowing down the searching space and pruning those drug-target pairs with low binding affinity scores. In the computational methods, feature representation of proteins is one of the most important parts due to its strong influence on the following regression task. This paper introduces the BERT-based language representation to embed the gene ontology annotations, combined with the raw sequence to characterize a protein by fusing its physical structure and human knowledge. We exploit CNN network stacked over full connected layers to learn the prediction of DTA scores in a supervised manner. This framework enhances the feature representation ability, leading to the improvement of the DTA prediction precision. The evaluation on the Davis and KIBA datasets compared to the state-of-the-art baselines demonstrates our feature representation’s superiority.