Search papers, labs, and topics across Lattice.
This paper introduces a geometry-aware method for heterogeneous cryo-EM reconstruction using a graph neural network (GNN) autodecoder to predict atomic backbone conformations from per-image latent variables. By representing the protein backbone as a graph and incorporating geometric regularization, the method leverages protein-structure priors to improve reconstruction accuracy. Experiments on synthetic datasets demonstrate that the GNN outperforms a comparable MLP, indicating the effectiveness of the geometry-informed inductive bias.
GNNs can significantly improve heterogeneous cryo-EM reconstruction by leveraging protein structure priors, outperforming MLPs in predicting atomic backbone conformations.
We present a geometry-aware method for heterogeneous single-particle cryogenic electron microscopy (cryo-EM) reconstruction that predicts atomic backbone conformations. To incorporate protein-structure priors, we represent the backbone as a graph and use a graph neural network (GNN) autodecoder that maps per-image latent variables to 3D displacements of a template conformation. The objective combines a data-discrepancy term based on a differentiable cryo-EM forward model with geometric regularization, and it supports unknown orientations via ellipsoidal support lifting (ESL) pose estimation. On synthetic datasets derived from molecular dynamics trajectories, the proposed GNN achieves higher accuracy compared to a multilayer perceptron (MLP) of comparable size, highlighting the benefits of a geometry-informed inductive bias.
