henhao
Sign in
BiophysicsNature
Peer-reviewed journal article

Highly accurate protein structure prediction with AlphaFold

John JumperiD, Richard EvansiD, Alexander PritzeliD, Tim GreeniD, Michael FigurnoviD, Olaf RonnebergeriD, Kathryn TunyasuvunakooliD, Russ BatesiD, Augustin ŽídekiD, Anna PotapenkoiD, Alex Bridgland, Clemens MeyeriD, Simon KöhliD, Andrew J. BallardiD, Andrew CowieiD, Bernardino Romera‐ParedesiD, Stanislav NikoloviD, Rishub JainiD, Jonas AdleriD, Trevor BackiD, Stig PeterseniD, David ReimaniD, Ellen ClancyiD, Michał ZielińskiiD, Martin SteineggeriD, , Demis HassabisiD

Abstract

Abstract Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous experimental effort 1–4 , the structures of around 100,000 unique proteins have been determined 5 , but this represents a small fraction of the billions of known protein sequences 6,7 . Structural coverage is bottlenecked by the months to years of painstaking effort required to determine a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence—the structure prediction component of the ‘protein folding problem’ 8 —has been an important open research problem for more than 50 years 9 . Despite recent progress 10–14 , existing methods fall far short of atomic accuracy, especially when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14) 15 , demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates physical and biological knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
43,764
FWCI
2028.17
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text (15 Jul 2021) · CC BY 4.0

AlphaFold, an attention network over sequence alignments and residue pairs, rivaled experimental accuracy in most CASP14 cases: median backbone r.m.s.d.95 0.96 Å versus 2.8 Å for the runner-up.

Experimental structures exist for only around 100,000 unique proteins against billions of known sequences, and prior physics- and evolution-based predictors fell far short of atomic accuracy when no close homologue had been solved. AlphaFold is presented as the first method to regularly reach near-experimental accuracy, including for proteins without similar known structures, opening a route to proteome-scale structural bioinformatics.

Key results

  1. ShownIn the blind CASP14 assessment (Critical Assessment of protein Structure Prediction), AlphaFold domain models far exceeded competing entries in backbone and all-atom accuracy, with median backbone error below the roughly 1.4 Å width of a carbon atom.

    n = 87 CASP14 domains: median backbone r.m.s.d.95 0.96 Å (95% CI 0.85 to 1.16 Å) vs 2.8 Å (95% CI 2.7 to 4.0 Å) for the next-best method; all-atom r.m.s.d.95 1.5 Å (95% CI 1.2 to 1.6 Å) vs 3.5 Å (95% CI 3.1 to 4.2 Å); compared against the top-15 of 146 entries. · Main; Figure legends

  2. ShownAccuracy carried over to an uncurated, redundancy-reduced set of Protein Data Bank (PDB) chains released after the training cut-off and scored as full chains, including chains lacking close training templates; side-chain accuracy was high wherever the backbone was accurate.

    10,795 chains released after 30 April 2018; on 3,144 chains with no training-set template above 40% identity over more than 1% of the chain, median full-chain backbone r.m.s.d.95 was 1.46 Å (95% CI 1.40 to 1.56 Å). Rotamer accuracy (torsion within 40°) increased with backbone local-distance difference test (lDDT-Cα) across 5,317 chains at <2.5 Å resolution. · Main; Test set of recent PDB sequences; Figure legends

  3. ShownBuilt-in confidence estimates track true accuracy on recent PDB chains: per-residue predicted lDDT (pLDDT) follows lDDT-Cα, and the predicted template modelling score (pTM) follows full-chain TM-score.

    n = 10,795 chains: least-squares fit lDDT-Cα = 0.997 × pLDDT - 1.17 (Pearson's r = 0.76) and TM-score = 0.98 × pTM + 0.07 (Pearson's r = 0.85), with 95% bands from 10,000 bootstrap samples; pTM came from models fine-tuned after CASP14. · Main; End-to-end structure prediction; Training regimen; Figure legends

  4. ShownAccuracy depends on multiple sequence alignment (MSA) depth, dropping sharply below a median effective depth of about 30 sequences with small gains past about 100; metagenomic databases are critical for targets poorly represented in UniRef.

    Depth curves on chains with ≤25% heteromeric long-range contacts, split by template coverage (n = 6,743 chains with >60% coverage; n = 1,596 with <30%). Re-running CASP14 without the Big Fantastic Database (BFD) cut mean global distance test (GDT) by 0.4, without MGnify by 0.7, without both by 6.1; a few outliers shifted by 20+ GDT. · MSA depth and cross-chain contacts; Inputs and data sources; Test set of recent PDB sequences; Figure legends

  5. ShownBecause coordinates are output directly, inference runs in GPU minutes to hours depending on length and scales to long chains; the 2,180-residue target T1044, lacking structural homologues, was predicted after CASP with correct domain packing.

    Single model on a V100 GPU with ensembling: 4.8 min (256 residues), 9.2 min (384), 18 h (2,500); without ensembling, 8× faster: 0.6 min, 1.1 min and 2.1 h. A 16 GB V100 fits about 1,300 residues without ensembling. Network-only timings; MSA search and relaxation add CPU time. · Main; Discussion; Inference regimen; Figure legends

Methods, data and reproducibility

Sequence, multiple sequence alignment and optional templates pass through 48 Evoformer blocks mixing alignment and residue-pair information via axial and triangle updates; a structure module refines per-residue rigid frames with invariant point attention. Training uses frame-aligned point error and masked-alignment losses, recycling and self-distillation.

  • CASP14 blind-assessment target domains (May to July 2020), benchmarked against the top-15 of 146 entries · n = 87 protein domains
  • Recent-PDB evaluation set: redundancy-reduced chains released after the 30 April 2018 training cut-off (PDB copy of 15 February 2021) · n = 10,795 chains; 3,144 without >40%-identity training templates
  • Recent-PDB subsets: ablation set with ≤30% template coverage at 30% identity, and MSA-depth sets with >60% or <30% template coverage · n = 2,261; 6,743; 1,596 chains
  • PDB training structures released up to 30 April 2018 (copy downloaded 28 August 2019), sampled inversely to 40%-identity cluster size
  • Uniclust30 v.2018_08 sequences predicted by an initial model and filtered to high confidence for noisy-student self-distillation (sampled 75% of the time in final training) · n = 355,993 sequences predicted; size of high-confidence subset not stated
  • Big Fantastic Database (BFD): custom protein-family MSAs and hidden Markov models from UniProt, soil and marine eukaryotic catalogues and Metaclust NR · n = 65,983,866 families; 2,204,359,010 sequences
  • MSA and template databases: UniRef90 v.2020_01, MGnify clusters v.2018_12, PDB copy of 14 May 2020 and PDB70 of 13 May 2020
  • Experimental PDB structures shown for comparison, including CASP14 targets T1049, T1056 and T1044 and an intertwined homotrimer · PDB 6Y4F, 6YJ1, 6VR4, 6SK0, 6FES, 6W6W, 6T1Z, 7JTL
Methods & tools
AlphaFold (Evoformer trunk plus structure module with invariant point attention), HHBlits and HHSearch (hh-suite v.3.0-beta.3) for BFD/Uniclust30 MSAs and PDB70 templates, jackhmmer (HMMER v.3.3) for UniRef90 and MGnify MSA search, Hmmsearch for template-identity filtering of the recent-PDB test set, OpenMM v.7.3.1 with Amber99sb force field (post-prediction relaxation), MMseqs2 v.925AF/Linclust and FAMSA v.1.2.5 (BFD construction), JAX v.0.1.69 with XLA, Haiku v.0.0.4, TensorFlow v.1, Sonnet v.1, TM-align v.20190822 (TM-score computation), Accuracy metrics: lDDT, lDDT-Cα, GDT, TM-score, r.m.s.d.95, CASP14 blind structure-prediction assessment, Python v.3.6 stack: NumPy, SciPy, pandas, statsmodels, seaborn, plotnine, Colab, Pymol v.2.3.0 (structure visualization)
Code
https://github.com/deepmind/alphafoldhttps://github.com/google/jax/https://github.com/deepmind/dm-haikuhttps://github.com/soedinglab/hh-suitehttp://eddylab.org/software/hmmer/https://github.com/openmm/openmm
Data & accessions
https://www.wwpdb.org/ftp/pdb-ftp-siteshttps://wwwuser.gwdg.de/~compbiol/data/hhsuite/databases/hhsuite_dbs/https://ftp.ebi.ac.uk/pub/databases/uniprot/previous_releases/release-2020_01/uniref/https://bfd.mmseqs.com/https://wwwuser.gwdg.de/~compbiol/uniclust/2018_08/https://ftp.ebi.ac.uk/pub/databases/metagenomics/peptide_database/2018_12/
Software & environment
JAX v.0.1.69, Haiku v.0.0.4 (XLA bundled), TensorFlow v.1, Sonnet v.1; hh-suite v.3.0-beta.3, HMMER v.3.3, OpenMM v.7.3.1 with Amber99sb; MMseqs2 v.925AF, FAMSA v.1.2.5; Python v.3.6, NumPy v.1.16.4, SciPy v.1.2.1; TM-align v.20190822.
Assay conditions & compute
Training: 128 TPU v3 cores, batch 128, 256-residue crops, about 1 week, then about 4 days fine-tuning on 384-residue crops. Inference: single V100 GPU (16 GB), minutes to hours by length; 2,500-residue chains use unified memory (four GPUs requested).
Reproducibility gaps
Released: model code, trained weights, inference script; training code and the self-distillation predicted-structure set are not stated as released. Hyperparameters, pseudocode and ablation details are only in the Supplementary Information.
Limitations (6)
  • Accuracy falls substantially when median MSA depth is below roughly 30 sequences, so proteins with few detectable homologues remain hard; extra depth beyond about 100 sequences brings only small gains. · stated by the authors
  • Predictions are much weaker for chains whose shape is set mainly by heterotypic contacts, such as bridging domains in large complexes; full hetero-complex prediction is left to a future system. · stated by the authors
  • The r.m.s.d.95 metric discards the worst-aligned 5% of Cα atoms, which the authors note can sometimes contain genuine modelling errors; full-chain r.m.s.d. is also highly sensitive to domain-packing errors. · stated by the authors
  • Memory scales roughly quadratically with length: a 16 GB V100 handles about 1,300 residues without ensembling, and 2,500-residue chains need unified memory with a multi-GPU allocation. · stated by the authors
  • The recent-PDB benchmark excludes NMR structures, chains over 1,400 residues and chains with fewer than 80 resolved Cα atoms, so accuracy on such proteins is not measured systematically; very long chains appear only as single examples such as T1044. · noted in this summary
  • Each chain is scored as one static model against one deposited structure; alternative conformational states, ligand- or partner-induced rearrangements and dynamics are not evaluated. · noted in this summary

Who should read:Read this if you build or apply protein structure predictors and need AlphaFold's architecture, CASP14 and recent-PDB accuracy, confidence calibration, compute costs and stated failure modes.

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

Similarity Map
SelectedSame fieldCross-field
THISHeo '19Terwilliger '22Finkelstein '22Roney '22López‐Sagas '25Moffat '21
77% semantic matchBiophysics
L. Heo, M. Feig·bioRxiv·Aug 10, 2019