The prediction of protein-ligand binding affinities is very crucial for drug discovery research. Molecular docking is widely used in finding initial hits in the early drug discovery stage. Though docking programs have successfully used for years, they are still not fully reliable. Many physics-based scoring functions are developed over the years. However, machine learning based scoring functions are proven to outperform the traditional scoring functions. In this study, the performance of an ensemble of fully connected neural networks, trained on the recent PDBBind data using different descriptors, will be presented. We have achieved an R2 (Pearson) of 0.82 and Spearman Rho of 0.90 for the test set. Validation of the model using ROC curves and enrichment factors along with a comparison to other scoring functions will be presented. Besides FCNN, we will also present how 3D convolutional neural networks can be implemented to perform the same task.

Deep learning based scoring function for predicting protein-ligand binding affinities

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.