CryoDECO: Deconstructing Compositional and Conformational Heterogeneity in Cryo-EM with Foundation Model Priors

Abstract

Resolving compositional and conformational heterogeneity remains a fundamental bottleneck in single-particle cryo-EM. This challenge stems from a circular dependency: classification requires reliable references, while reference generation requires accurate classification. Current deep learning methods often resort to blind stochastic initialization, frequently becoming trapped in local minima within complex optimization landscapes. To address this, we present CryoDECO, a framework that breaks this deadlock by integrating representation priors from the pretrained cryo-EM foundation model, Cryo-IEF. By projecting particle images onto a semantically structured manifold, CryoDECO bypasses the tabula rasa search phase of standard ab initio reconstruction. We demonstrate that this prior-informed strategy robustly disentangles extreme compositional heterogeneity, successfully resolving 100 distinct structures from a single simulated mixture while simultaneously mapping complex conformational dynamics. CryoDECO offers a robust, high-throughput solution for systems structural biology by replacing ab initio reconstruction with prior-guided optimization.