CryoCodex: learning discrete structural representations for cryo-EM map post-processing

Abstract

Cryo-EM maps frequently suffer from weak structural signals, spatially variable quality, and noise, complicating interpretation and downstream modeling. Existing post-processing approaches treat density enhancement, quality assessment, and masking as separate tasks, often amplifying residual noise or introducing local artifacts. Here, we present CryoCodex, a unified framework that jointly performs density enhancement, local quality estimation, and molecular masking. CryoCodex introduces a multi-head, multi-scale vector-quantization module that projects encoder-derived latent features onto a shared codebook of structural prototypes. By imposing discrete constraints across feature subspaces and spatial scales, the codebook suppresses stochastic noise while preserving recurrent structural patterns. On a non-redundant test set of 126 experimental maps spanning 3–8 Å, CryoCodex outperformed four representative methods, achieving a mean map–model FSC-0.5 of 3.93 Å versus 4.62 Å for EMReady2 and 6.87 Å for the deposited maps. Downstream automated de novo model building with Phenix further confirmed the quality of the CryoCodex-enhanced maps. These gains extended to maps containing nucleic acids, where CryoCodex produced the best map–model agreement and improved model building. Half-map cross-validation indicated that these improvements are anchored to experimental data rather than driven by half-map-specific noise or structural hallucination. CryoCodex also delivered reliable local quality estimates and molecular masks within the same framework. Together, these results establish CryoCodex as a practical tool for cryo-EM map post-processing and structural determination in challenging cases.