High-resolution de novo structure prediction from primary sequence
Ruidong WuiD, Fan DingiD, Rui Wang, Rui SheniD, Xiwen ZhangiD, Shitong LuoiD, Chenpeng Su, Zuofan Wu, Qi XieiD, Bonnie BergeriD, Jianzhu MaiD, Jian PengiD
Abstract
Recent breakthroughs have used deep learning to exploit evolutionary information in multiple sequence alignments (MSAs) to accurately predict protein structures. However, MSAs of homologous proteins are not always available, such as with orphan proteins or fast-evolving proteins like antibodies, and a protein typically folds in a natural setting from its primary amino acid sequence into its three-dimensional structure, suggesting that evolutionary information and MSAs should not be necessary to predict a proteins folded form. Here, we introduce OmegaFold, the first computational method to successfully predict high-resolution protein structure from a single primary sequence alone. Using a new combination of a protein language model that allows us to make predictions from single sequences and a geometry-inspired transformer model trained on protein structures, OmegaFold outperforms RoseTTAFold and achieves similar prediction accuracy to AlphaFold2 on recently released structures. OmegaFold enables accurate predictions on orphan proteins that do not belong to any functionally characterized protein family and antibodies that tend to have noisy MSAs due to fast evolution. Our study fills a much-encountered gap in structure prediction and brings us a step closer to understanding protein folding in nature.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 392
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (22 Jul 2022) · CC BY-NC-ND 4.0
From a single sequence, OmegaFold reaches mean LDDT 0.82 on 146 CAMEO targets, between MSA-based RoseTTAFold (0.75) and AlphaFold2 (0.86), and runs about ten times faster.
AlphaFold2 and RoseTTAFold depend on multiple sequence alignments (MSAs), and their accuracy drops sharply for proteins with few homologs, such as orphan proteins, or with noisy alignments, such as antibodies. OmegaFold replaces the alignment with a roughly 670-million-parameter protein language model feeding a geometry-aware transformer, which the authors present as the first method to predict high-resolution structure from one sequence alone.
Key results
ShownGiven only a single sequence, OmegaFold comes close to MSA-based AlphaFold2 and RoseTTAFold: it beats RoseTTAFold and trails AlphaFold2 on recent CAMEO targets, and matches AlphaFold2 but slightly trails RoseTTAFold on hard CASP free-modelling targets.
CAMEO, 146 single-chain proteins from the first six months of 2022: mean LDDT 0.82 vs 0.75 for RoseTTAFold and 0.86 for AlphaFold2, both run with MSAs in default mode. CASP13/CASP14 free-modelling set, 29 proteins: mean TM-score 0.79 vs 0.81 for RoseTTAFold and 0.79 for AlphaFold2. · OmegaFold performs well on CASP and CAMEO…; 2.6.1 Validation Data
ShownMSA-based methods lose much of their accuracy without alignments, and retraining AlphaFold2 on single sequences does not close the gap; ablations indicate that both OmegaPLM and Geoformer are needed.
Single-sequence AlphaFold2 and RoseTTAFold are described as statistically highly inferior to OmegaFold on both benchmark sets. CAMEO ablations compared ensembled and single OmegaFold, single-sequence and retrained AlphaFold2, and OmegaPLM plus structure module alone. The metric values are not in this text. · OmegaFold outperforms RoseTTAFold and AlphaFold2…; Towards interpretability of OmegaFold; 2.9 Ablation Study
ShownOmegaFold is more accurate than AlphaFold2 on recently released orphan proteins and on antibody CDR loops, the cases where alignments are missing or noisy.
Reported as much higher accuracy than AlphaFold2, which gives very low predicted LDDT (pLDDT) on heavy-chain CDR3 loops because their MSAs are noisy. Set sizes and accuracy values are not in this text. · OmegaFold enables accurate predictions on orphan…
ShownBecause no MSA search is needed, OmegaFold runs roughly an order of magnitude faster than AlphaFold2 and is also faster than the optimised ColabFold-AF2.
Stated as about ten times faster than MSA-based methods with comparable or better accuracy; the runtime comparison is in a figure whose values are not in this text. · OmegaFold improves runtimes by an order of…; Main Text
ShownOmegaPLM embeddings alone already give reasonable contacts; the first 20 Geoformer layers sharply cut triangle-inequality violations and raise contact accuracy, and later layers mostly keep reducing inconsistency while accuracy levels off.
Contact accuracy and geometric inconsistency were tracked layer by layer from OmegaPLM through the 50 Geoformer layers; values are in a figure not in this text. OmegaPLM is also reported to beat vanilla transformer language models on language modelling and contact prediction while using less memory. · Towards interpretability of OmegaFold; OmegaPLM outperforms vanilla language models
Methods, data and reproducibility
OmegaPLM, a 66-layer, ~670-million-parameter protein language model pre-trained on UniRef50, supplies residue and pair embeddings to Geoformer, a 50-layer geometry-consistency transformer, and an AlphaFold2-style structure module that outputs heavy-atom coordinates; the folding part is trained on PDB single chains with contact, frame-aligned point error (FAPE), torsion and violation losses.
- UniRef50 (2021/04 release) sequences for OmegaPLM pre-training, 99.99% used for training and the rest for validation
- PDB single-chain structures deposited before 2021 plus SCOP v1.75 single domains, at most 40% sequence identity, with test-set homologs removed · n = ~110,000 single-chain structures
- CASP13 and CASP14 free-modelling targets (target IDs listed in the text) · n = 29
- Recent single-chain CAMEO targets from the first six months of 2022 (PDB chain IDs listed in the text) · n = 146
- Recently released antibody and orphan protein structures from the PDB
- Methods & tools
- OmegaFold, OmegaPLM, Geoformer, AlphaFold2, RoseTTAFold, ColabFold-AF2, MMseqs2, Biopython PDBparser, PyTorch, PowerSGD
- Reagents & environment
- PyTorch with the AdamW optimiser and TensorFloat-32 matrix precision on A100 GPUs; MMseqs2 clustering and Biopython PDBparser for data preparation. No package versions, containers or requirement files are given.
- Assay conditions & compute
- OmegaPLM pre-training took 2,560 Nvidia A100 80G GPU-days; PowerSGD gradient compression (rank 32) sped training by 30%. Compute for training the folding model is not stated.
- Reproducibility gaps
- No code, model-weight or data availability statement and no links appear in this text.
Limitations (4)
- Single-sequence accuracy still trails MSA-based AlphaFold2 on CAMEO (mean LDDT 0.82 vs 0.86) and RoseTTAFold on the CASP free-modelling set (mean TM-score 0.79 vs 0.81). · stated by the authors
- The authors note that many protein chains are hard to predict because they need RNA or protein partners to fold or have sparse co-evolution signal; fragments that seemed impossible to fit were down-weighted during training. · stated by the authors
- The CASP and CAMEO sets used as benchmarks are also called validation datasets, and later-released structures were used for hyperparameter selection, so it is unclear whether they were fully held out. · noted in this summary
- The orphan-protein and antibody advantage is reported only against AlphaFold2, with no set sizes or numbers in this text; the abstract's claim of beating RoseTTAFold holds on CAMEO but not on the CASP set. · noted in this summary
Who should read:Read this if you need structures for proteins with few or noisy homologs, such as orphans or antibodies, or want an MSA-free and faster alternative to AlphaFold2-style prediction.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Single-sequence protein structure prediction using language models from deep learning
Ratul Chowdhury, Nazim Bouatta, … Mohammed AlQuraishi·Nature BiotechnologyPeer-reviewed·Oct 3, 2022
- So ManyFolds, So Little Time: Efficient Protein Structure Prediction With pLMs and MSAs
Thomas D. Barrett, Amelia Villegas-Morcillo, … Arthur Flajolet·bioRxiv·Oct 15, 2022
- GhostFold: Accurate protein structure prediction using structure-constrained synthetic coevolutionary signals
Nitesh Mishra, Bryan S. Briney·bioRxiv·Oct 13, 2025
- RaptorX-Single: single-sequence protein structure prediction by integrating protein language models
Xiaoyang Jing, Fandi Wu, … Jinbo Xu·bioRxiv·Apr 24, 2023
- Highly accurate protein structure prediction with AlphaFold
John Jumper, Richard Evans, … Demis Hassabis·NaturePeer-reviewed·Jul 15, 2021
- Effective High-Accuracy Prediction of Protein Structures from Easily Obtainable Artificial Homologous Sequences by Structure-Stability-Based Selection
Jinle Tang, Zhe Zhang, … Yaoqi Zhou·bioRxiv·Nov 22, 2023