Physics-informed geometric regularization of heterogeneous reconstructions in cryo-EM

Victor Prins, Willem Diepeveen, Erik J Bekkers, Ozan Öktem

Track: Machine learning: computational method and/or computational results

Keywords: neural network, cryo-em, protein reconstruction, heterogeneous proteins, heterogeneous cryo-em, protein folding

TL;DR: This paper presents a novel loss function that operates on the structure of a protein, facilitating accurate reconstructions of structurally flexible proteins from cryo-EM data, using a neural network.

Abstract: Many proteins are flexible and occur in a continuum of 3D conformations. A protein's 3D conformation is the main determinant of its biological function, and is therefore of paramount importance to fields such as drug development and biomolecular engineering. Cryogenic electron microscopy (cryo-EM) enables the reconstruction of protein conformations. Because cryo-EM reconstruction is a fundamentally underdetermined problem, neural networks that incorporate prior knowledge such as physical constraints into the training process, can theoretically be more effective at reconstructing heterogeneous conformation distributions. We introduce a novel such prior, called the geometry degradation loss, grounded in the theory of normal mode analysis. The loss is generally applicable to all proteins and easily integrated into reconstruction algorithms that utilize geometric protein representations. We show on synthetic datasets of the flexible ADK and Nsp13 proteins that the loss greatly improves reconstruction quality and that our network is able to reconstruct both proteins accurately across the full conformation distribution. These results are further evidence that geometric cryo-EM reconstruction networks have large potential that can be tapped with the introduction of geometric priors. Code is published at https://github.com/VictorPrins/geometric-heterogeneous-cryoEM-reconstruction

Submission Number: 45