CryoAtom2: inverse folding-inspired deep learning enables accurate model building of protein-nucleic acid complexes in cryo-EM

Abstract

Automated model building of protein–nucleic acid complexes in cryo-electron microscopy (cryo-EM) maps is hindered by the highly similar density signatures of chemically related bases at moderate resolutions. Here, we present CryoAtom2, an inverse folding-inspired multimodal deep learning framework for de novo model building of proteins, nucleic acids, and their complexes within a single pipeline. CryoAtom2 jointly integrates three complementary modalities: density, sequence, and—crucially—a diverse set of plausible 3D structural conformations. For each residue, rather than generating a single coordinate set, it produces 28 candidate coordinate sets (20 amino acids, 4 RNA nucleotides, and 4 DNA nucleotides) and employs a dedicated Structure Encoder to capture the rich structural context necessary for accurate sequence inference, following ProteinMPNN. By reasoning over this expanded structural space, the model effectively discriminates between similar bases, yielding globally optimal sequences that best fit the density map. This approach is analogous to the principle of inverse folding, where structural context guides accurate sequence assignment. Benchmarking on 175 non-redundant cryo-EM maps demonstrates that CryoAtom2 substantially outperforms existing state-of-the-art tools. It delivers more complete models with a superior fit to the density and improved geometric quality across proteins, nucleic acids, and their complexes. For example, under the most stringent criterion—completeness—which considers both atomic position and sequence identity—the CryoAtom2 model achieves an average of 60.0% for nucleic acids, significantly outperforming other methods: EM2NA (40.6%), ModelAngelo (32.6%), CryoREAD (32.4%), and Phenix (11.7%). We successfully applied CryoAtom2 to three challenging, unseen experimental structures: an in situ human 80S ribosome, an archaeal 70S ribosome, and a large RNA-only nanocage. In each case, CryoAtom2 achieved near-expert levels of model completeness, exceptional sequence accuracy, and stereochemically sound geometry. Together, these results position CryoAtom2 as a transformative tool for interpreting complex biological machinery in cryo-EM.