Using Machine Learning to Predict Docking Scores of Compounds to Proteins based on Molecular Properties

Abstract

Docking is a computational procedure designed to find molecules that bind possibly the “best” on the protein surface and predict their “correct” bound association. Most docking programs explore the conformation space of ligand by manipulating the position in the protein active site, and the conformation of the ligand by the rotational bonds. To determine the quality of the fit and select the best ligands, most docking programs use an energy-based scoring function. Scoring functions may have a bias that gives preference to certain types of compounds over others. This paper addresses the issue of bias in scoring functions in virtual screening. To mitigate this issue, in our preliminary experiments, we show that a surprisingly accurate model can be developed for predicting docking scores based solely on the molecular properties of the ligands, which explains why there is such a high degree of correlation of scores for compounds between different targets. Our goal is to show how to use this model to “subtract” this bias out, producing a modified score that better shows that compounds dock best, and most specifically to which targets. We then show the performance of the model can be improved by extending the first set of molecular properties calculates for each compound with additional chemical features called “fingerprints.” In this thesis, we use the AutoDock docking program and explore the use of several statistical and machine learning methods to extract and characterize the biases in the VINA score.