Presenter: Daniel Berenberg, NYU
Topic: Protein structure embeddings through contact map decomposition and convolution
Abstract: Protein folds arise from a complex system of atomistic interactions in three dimensions. While such point clouds of atoms
prescribe near exact positioning for a fold conformation, much of its characterization is retained in its contact map. However,
while contact maps are informative, they do not constitute a fixed-size representation, which could be useful from a data analysis perspective.
This talk will discuss recent work in attempting to extract self-supervised, fixed-length feature vectors of protein structure by decomposing the contact map using a variety of signal processing techniques, including both learning and designing spectral filters.