Reconstructing Shapes and Motions from Unposed Images: From Molecular Complexes to Neural Radiance Fields
Axel Levy, Stanford University

Meeting • 2023-08-31

Abstract
Understanding how biomolecules function and interact is essential to understanding the microscopic machinery of life and developing new drugs. Cryogenic electron microscopy (cryo-EM) is an imaging method that gives access to the structure of proteins, but accessing this information involves a challenging 3D reconstruction problem. In this talk, I will explain how the recent advances in cryo-EM reconstruction and inverse graphics can benefit from each other, as they provide complementary answers to the same general problem. I will show how amortized inference can help coping with a large number of images (10k to 10M), how tailored loss functions enable gradient-based optimization to solve this non-convex problem and how new pose estimation strategies can simultaneously provide robustness and scalability.