Learning Representations for Biomolecular Interaction Prediction
Deep neural networks are applied to predicting protein-ligand binding affinity by learning representations of atoms, atomic interactions, and inter-molecular interfaces.
Predicting protein-ligand binding affinity, or even just binding at all, is an important part of drug design and learning biological mechanisms and biochemical pathways. Existing techniques explicitly model the physics of atomic interactions and/or reference known similar structures. We seek to augment these approaches and refine binding affinity predictions by using deep convolutional architectures to capture new representations of atoms in their local environments and their composition in larger molecular structures. We look to leveraging the large body of unlabeled data in the form of unbound protein crystal structures and small molecules and bound complexes in conjunction with the relatively smaller body of labeled data in the form of experimentally assayed bound complexes. The usefulness of the overall representation of the molecules, whether as entities in 3-D space or as nodes in a graph representation, will also be explored.In collaboration with Prof. Aleksander Madry
