CryoNeRF: generalizable automated cryo-EM reconstruction using neural radiance field

Cryogenic electron microscopy (cryo-EM) has revolutionized structural biology by enabling high-resolution determination of large macromolecular structures. However, inherent heterogeneity in cryo-EM datasets, arising from coexisting conformational states or compositional assemblies, remains a fundamental challenge for accurate 3D density map reconstruction. To this end, we present CryoNeRF, a pioneering neural radiance fields (NeRF)-based framework that directly resolves structural heterogeneity in 3D Euclidean space. CryoNeRF incorporates a hash encoding scheme and heterogeneity-aware encoder in NeRF to disentangle the heterogeneity in latent space. Extensive benchmarks demonstrate that CryoNeRF is generalizable to determine and model diverse cryo-EM maps, including both homogeneous and heterogeneous datasets. CryoNeRF demonstrates exceptional capability in capturing continuous and discrete heterogeneity, identifying structural differences and similarities in the latent space, and the structural discoveries are consistent with previous experimental discoveries. Notably, CryoNeRF successfully distinguished the assembly state that only accounts for 2% of particles of the dataset without filtering, which is neglected by previous methods. CryoNeRF is open-source software freely available at https://github.com/UNITES-Lab/CryoNeRF.