AbstractBackgroundTraditional drug development is time-consuming and expensive, while computer-aided drug repositioning can improve efficiency and productivity. In this study, we proposed a machine learning pipeline to predict the binding interaction between proteins and marketed or studied drugs. We then extended the predicted interactions to construct a protein network that could be applied to discover the potentially shared drugs between proteins and thus predict drug repositioning.MethodsBinding information between proteins and drugs from the Binding Database and the physicochemical properties of drugs from the ChEMBL database were used to build the machine learning models, i.e. support vector regression. We further measured proportionalities between proteins by the predicted binding affinity and introduced edge betweenness centrality to construct a protein similarity network for drug repositioning.ResultsAs the proof of concept, we demonstrated our machine learning approach is capable of reflecting the binding strength between drugs and the target protein. When comparing coefficients of protein models, we found proteins SYUA and TAU that may share common ligand which were not in our training data. Using the edge betweenness centrality network based on the prediction proportionality of protein models, we found a potential target, AK1C2, of aspirin and of which the binding interaction had been validated.ConclusionsOur study could not only be applied to drug repositioning by comparing protein models or searching the protein-protein network, but also to predict the binding strength once the sufficient experimental data was provided to train the protein models.
Interpreting Potts and Transformer Protein Models Through the Lens of Simplified Attention
