De novo all-atom model building from cryo-EM maps for proteins, RNA, DNA and their complexes, guided by ESM-2 and RiNALMo sequence embeddings.
No providers recorded yet. Browse all providers
A cryo-EM map of an RNA or a protein–RNA complex can be good enough to trace every backbone and still leave the builder guessing at the bases. Purines look alike at 3–4 Å, and pyrimidines look alike, so a map-only method that traces a nucleic acid chain correctly will often thread the wrong sequence onto it. Deep-learning builders tuned for proteins inherit this weakness: they cover most of an RNA chain but assign only a minority of its nucleotides correctly. EM3DFold attacks the sequence-assignment step directly by letting pretrained sequence language models inform residue-type prediction alongside the density.
EM3DFold was developed by Tao Li, Hong Cao and Sheng-You Huang at the School of Physics, Huazhong University of Science and Technology, and posted to bioRxiv in September 2026. It is a single framework for de novo building of proteins, RNA, DNA and mixed protein–nucleic acid complexes, and it extends the same lab's protein-only (EMProt) and nucleic-acid-only (EM2NA) builders. The lab also develops the map post-processing tool EMReady2.
The pipeline has two learned stages. A Swin-Conv-UNet (SCUNet) predicts a probability map of Cα (protein) and C4′ (nucleotide) positions, which a mean-shift step converts into atom coordinates. A density-aware Three-Track Attention (TTA) network then refines those positions and predicts backbone frames, torsion angles and residue types, drawing protein sequence embeddings from ESM-2 and nucleotide embeddings from RiNALMo.
hmm_search step matches the traced fragments against a sequence database to propose the identity of unknown chains.Training used PDB/EMDB entries with resolution better than 4 Å released before 1 January 2025, filtered on model-to-map correlation and clustered at 80% nucleic acid or 30% protein sequence identity. The TTA network was trained on 6,870 structures (2,528 containing nucleic acids, 4,342 protein-only); the SCUNet atom predictor used 1,376 cluster representatives. TTA training follows a ModelAngelo-style denoising scheme, perturbing Cα/C4′ positions and frames, randomly deleting or fragmenting up to 20% of residues, and optimizing RMSD losses on Cα/C4′, backbone and all atoms. It ran for 150k steps on four A100 GPUs, the last 30k adding a structure-violation loss.
The test set holds 298 maps released from 2025 onward and filtered to at most 10% protein or 80% nucleic acid sequence identity to any training entry: 54 RNA-only structures, 124 protein–nucleic acid complexes and 120 protein-only structures. Across the 178 nucleic-acid-containing maps, EM3DFold reached a median completeness of 75% (88% coverage, 94% sequence accuracy), against 45% for EM2NA, 26% for CryoREAD and 30% for ModelAngelo. The gain comes mainly from sequence accuracy, which is 43–56% for those baselines. Completeness was 90% versus 79% for ModelAngelo on complexes, and 95% versus 88% (EMProt) and 87% (ModelAngelo) on protein-only targets. Median-sized structures build in about 10 minutes on one A100, and an 11,000-nucleotide RNA assembly in 54 minutes.
EM3DFold is aimed at structural biologists building initial atomic models into single-particle cryo-EM maps, particularly ribonucleoprotein complexes, protein–DNA assemblies and large RNA structures, where correct nucleotide assignment is hardest to obtain automatically. Its sequence-free mode and database search help when a map contains density for a chain whose identity is unknown. The authors recommend refining outputs in Phenix or Coot, since weak density, symmetric assemblies and repetitive sequences can still cause misalignment.
EM3DFold's results show that residue-type assignment, not backbone tracing, is the main bottleneck in automated nucleic acid model building, and it answers that bottleneck by pairing local density with nucleotide language-model embeddings. That places it beside other catalog builders such as CryoAtom, Cryo2Struct and DeepCryoRNA, with coverage of all three polymer types in one model. The work is a preprint that has not been peer reviewed, and its benchmarks were run by the developers. Code and pretrained weights are downloadable, though the repository has no top-level license and its bundled EM3DFit fitting package carries an all-rights-reserved proprietary notice.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.