SSnet - Secondary Structure based End-to-End Learning model for Protein-Ligand Interaction Prediction

Computational prediction of bioactivity has become a critical aspect of modern drug discovery as it mitigates the cost, time, and resources required to find and screen new compounds. Deep Neural Networks (DNN) have recently shown excellent performance in modeling Protein-Ligand Interaction (PLI). However, DNNs are only effective when physically sound descriptions of ligands and proteins are fed into the network for further processing. Furthermore, previous research has not incorporated the secondary structure of the protein in a meaningful manner. In this work, we utilize secondary structure information of the protein which is extracted as the curvature and torsion of the backbone of protein to predict PLI. We demonstrate how our model outperforms previous machine and non-machine learning models on three major datasets: humans, C.elegans, and DUD-E. Visualization of the intermediate layers of our model shows a potential latent space for proteins which extracts important information about the activity of the protein. We further investigate the inner workings of our model by visualizing heatmaps through Grad-CAM. This analysis is adapted to visualize the most important aspects of the protein that the algorithm has learned. We observed that the important residues highlighted by Grad-CAM are the ones responsible for non-covalent interactions with a ligand and is not just confined to the binding site as it also includes allosteric sites and other locations where a ligand interacts. Our new model opens the door in exploration of DNN based on the secondary structure which is not just confined to protein ligand interactions