A Bayesian Framework for Cryo-EM Heterogeneity Analysis using Regularized Covariance Estimation

Proteins and the complexes they form are central to nearly all cellular processes. Their flexibility, expressed through a continuum of states, provides a window into their biological functions. Cryogenic-electron microscopy (cryo-EM) is an ideal tool to study these dynamic states as it captures specimens in non-crystalline conditions and enables high-resolution reconstructions. However, analyzing the heterogeneous distribution of conformations from cryo-EM data is challenging. Current methods face issues such as a lack of explainability, overfitting caused by lack of regularization, and a large number of parameters to tune; problems exacerbated by the lack of proper metrics to evaluate or compare heterogeneous reconstructions. To address these challenges, we present RECOVAR, a white-box method based on principal component analysis (PCA) computed via regularized covariance estimation that can resolve intricate heterogeneity with similar expressive power to neural networks with significantly lower computational demands. We extend the ubiquitous Bayesian framework used in homogeneous reconstruction to automatically regularize principal components, overcoming overfitting concerns and removing the need for most parameters. We further exploit the conservation of density and distances endowed by the embedding in PCA space, opening the door to reliable free energy computation. We leverage the predictable uncertainty of image labels to generate high-resolution reconstructions and identify high-density trajectories in latent space. We make the code freely available at https://github.com/ma-gilles/recovar.