ANTIDOTE: A Metadata-Driven Neural Network for Improving CryoEM 3-D Particle Sorting

Abstract

Despite the maturation of cryogenic electron microscopy (cryoEM) methodologies, generating high-resolution three-dimensional (3-D) reconstructions from micrographs is a time-intensive process involving iterative rounds of subjective data curation and hyperparameter optimization. Current approaches to particle classification are often unable to remove all low-quality particles from particle stacks, largely due to the low signal-to-noise ratio, the high dimensionality of particle images, and the multiple degrees of freedom associated with each particles unknown rotation, orientation, and class assignment. The retention of low-quality particles negatively affects the overall quality of the final EM density and continued efforts to eliminate their deleterious contributions are warranted. Here, we present ANTIDOTE (A Neural network Trained In Deleterious Object deTection and Elimination), a neural network framework that discriminates between constructive and deleterious particles using per-particle metadata generated during 3-D classification in RELION. Using benchmark and real-world cryoEM datasets, we demonstrate that ANTIDOTE paired with RELION 3-D classification achieves higher particle classification accuracy than conventional data processing approaches alone, yielding improvements in reconstruction quality, global and local resolution, and map interpretability while reducing time-consuming hyperparameter optimization. We additionally detail practical use-case scenarios for ANTIDOTE and demonstrate its versatility in increasing particle curation accuracy for high-quality cryoEM reconstruction.