A Small Step Toward Generalizability: Training a Machine Learning Scoring Function for Structure-Based Virtual Screening

Over the past fewyears, many machine learning-based scoring functionsfor predicting the binding of small molecules to proteins have beendeveloped. Their objective is to approximate the distribution whichtakes two molecules as input and outputs the energy of their interaction.Only a scoring function that accounts for the interatomic interactionsinvolved in binding can accurately predict binding affinity on unseenmolecules. However, many scoring functions make predictions basedon data set biases rather than an understanding of the physics ofbinding. These scoring functions perform well when tested on similartargets to those in the training set but fail to generalize to dissimilartargets. To test what a machine learning-based scoring function haslearned, input attribution, a technique for learning which featuresare important to a model when making a prediction on a particulardata point, can be applied. If a model successfully learns somethingbeyond data set biases, attribution should give insight into the importantbinding interactions that are taking place. We built a machine learning-basedscoring function that aimed to avoid the influence of bias via thoroughtrain and test data set filtering and show that it achieves comparableperformance on the Comparative Assessment of Scoring Functions, 2016(CASF-2016) benchmark to other leading methods. We then use the CASF-2016test set to perform attribution and find that the bonds identifiedas important by PointVS, unlike those extracted from other scoringfunctions, have a high correlation with those found by a distance-basedinteraction profiler. We then show that attribution can be used toextract important binding pharmacophores from a given protein targetwhen supplied with a number of bound structures. We use this informationto perform fragment elaboration and see improvements in docking scorescompared to using structural information from a traditional, data-basedapproach. This not only provides definitive proof that the scoringfunction has learned to identify some important binding interactionsbut also constitutes the first deep learning-based method for extractingstructural information from a target for molecule design.