The Coley group has previously worked on ShEPhERD, an SE(3)-equivariant diffusion model that performs 3D interaction-aware drug design via joint diffusion of molecular structure with other chemically relevant features, such as a shape and electrostatic point cloud. Owing to both the graph neural network (GNN) based and equivariant denoising modules, ShEPhERD is quite slow, and it becomes intractable to train and sample from larger chemical spaces. This project aims to create a non-equivariant version model admitting a scalable mechanism for interaction-aware drug design. Currently, we are working on developing this via representing shape and electrostatics as voxelized grids and augmenting the data with random SE(3) transforms to bypass the need for equivariance or GNNs.

My background has evolved from pure chemistry to include computer science while at MIT. I have had a great experience in the Coley lab and the structure offered by the SuperUROP program seems very useful to me. I intend on both becoming more versed in relevant literature and modern machine learning research methods throughout this project, which is the perfect intersection of the passion I have for both the natural sciences and computer science.