SE(3)-Convolutional neural network molecular semantic in-painting on protein structure representat-
Siddharth Bhadra-Lobo - MS, PhD Candidate Center for Computational and Integrative Biology, Dr. Guillaume Lamoureux - PhD, Associate Professor Department of Chemistry and Center for Computational and Integrative Biology
Poster # 82
Structures of biomolecules obtained from electron microscopy or generated from AI-based algorithms such as AlphaFold are usually missing many molecular components: ligands, cofactors, drug molecules, etc. Those additional components are however essential to understand the function of the biomolecule. It is therefore important to develop computer vision tools to fill in the missing details as accurately as possible. as possible. Using SE(3)-equivariant convolutional neural networks, we predict atomic densities for seven common cofactor types (ATP, FAD, FMN, HEME, NAD, SAM, ZN) and for the water molecules and inorganic ions in their vicinity, and we investigate what latent representations have been learned from the molecular in-painting task.