Predicting Protein-Ligand Binding Using Deep Learning with Spatial Transformations

Overview

School of Medicine, University of Pittsburgh There is a pressing need for new technologies that can accelerate the drug discovery process, which is currently laborious, time-consuming and costly due to the initial pre-screening of thousands of compounds. Because most drugs bind to proteins to produce their effects, more efficient and accurate selection of drug candidates that bind to specific proteins will save time and money. Computer-aided drug discovery and deep learning convolutional neural networks (CNNs) have been applied for virtual drug screening, however, they are limited by their ability to analyze complex 3D structures of proteins and ligands. Spatial Transformer Convolution Neural Networks (STNs) can recognize spatially transformed structures, and has yet to be tested for drug discovery. In this study, novel deep learning models incorporating STNs were built to predict protein-ligand binding by recognizing inherent relationships between proteins and ligands. Models were constructed in either Caffe or PyTorch framework and were trained and tested using the refined dataset from PDBbind database. The accuracy of the model was determined by calculating the loss and Pearson’s Rcoefficient for each model. Results show that in both Caffe and PyTorch models, the STNs are able to converge following translational and rotational perturbations of the ligand with higher accuracy predicting translational perturbations than rotational perturbations. Between the two models, the Caffe model has better accuracy, and PyTorch model has much shorter run time. This is the first study using STNs to predict ligand-protein binding, and demonstrates the potential of STNs models to become efficient tools for drug screening.