A Novel ML Approach to Detect Possible Protein Allosteric Sites for Drug Discovery

Abstract

In recent years, allosteric sites on proteins, located away from the primary active binding site, have gained prominence as promising drug targets because of the significant control they provide over biological pathways. The process of identifying allosteric sites through methods such as x-ray crystallography is often time-consuming, where the protein structure and the structure of potential allosteric-site binding molecules are analyzed to discover the allosteric site. Thus, machine learning models have been applied to expedite the process of allosteric drug discovery by contributing to allosteric site identification. In this study, I investigated different machine learning models and evaluated their performance and found that the best models were the random forest classifier and a 5-layer neural network. The random forest had an accuracy of 98.89%, a precision score of 98.20%, a recall score of 99.69%, an F1 score of 98.94%, and an AUC of .9991. The neural network had an accuracy of 97.70%, a precision score of 99.15%, a recall score of 96.04%, an F1 score of 97.57%, and an AUC of .9939. The high accuracy of these models along with their efficiency in running could potentially help the pharmaceutical industry identify these allosteric sites which would help with the allosteric drug manufacturing process.