Predicting Protein-Fragment Binding

Understanding the local structure surrounding the site of a protein-drug interaction can provide important insights into the mechanism of the drug and can help inform drug design and repurposing1. While structural information exists for a subset of protein-ligand interactions, it does not cover the space of all experimentally known interactions. As a result, the sites on proteins to which drugs or ligands bind are not always known. To address this problem, our eventual goal is to use existing structures of protein-ligand interactions to train machinelearning classifiers to predict locations of ligand binding. Because the chemical search space for ligands is large, we will approach this problem from a fragment perspective. Fragments are low molecular weight compounds (150-250 Da) that are parts of a given ligand. The same fragment may be present in a variety of ligands, so examining this problem from a fragment perspective is highly useful because it reduces the number of molecules to evaluate and allows data from multiple different ligands but the same fragment to be combined2;3. In order to predict the locations of fragment binding, we will focus on the sub-problem of: