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.
first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known
blind CASP14 benchmark shows 0.96 Å median backbone accuracy versus 2.8 Å next best
demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods
CASP14 and all-atom results show large margins over competing entries
the high accuracy that AlphaFold demonstrated in CASP14 extends to a large sample of recently released PDB structures
post-cutoff PDB validation is broad, with filtering and documented residual failure modes
pLDDT, reliably predicts the Cα local-distance difference test (lDDT-Cα) accuracy of the corresponding prediction
correlation is substantial at r=0.76, though residual variance remains
accuracy decreases substantially when the median alignment depth is less than around 30 sequences
explicit limitation is quantified in the MSA-depth analysis
Derived from the full evaluation — not a separate score.
Strengths
CASP14 is described as the gold-standard assessment for structure prediction and was unavailable to developers during model training. AlphaFold’s median backbone accuracy of 0.96 Å r.m.s.d.95, compared with 2.8 Å for the next-best method, directly supports the headline performance claim.
↳ Results p. 584; Fig. 1a
The paper evaluates recent PDB structures deposited after the training data cutoff and applies redundancy and template-similarity filtering. This provides evidence that the CASP14 result is not confined to a single benchmark setting.
↳ Fig. 2a; Methods, Test set of recent PDB sequences
The Code availability section provides source code, trained weights, and inference scripts under an open-source license. The Methods also report software versions, compute requirements, and inference timings, making practical adoption unusually well supported.
↳ Code availability; Methods, Training regimen and Inference regimen
Limitations
The paper reports that accuracy decreases substantially when median alignment depth falls below around 30 sequences. This limits transferability for targets with sparse evolutionary information.
↳ MSA depth and cross-chain contacts; Fig. 5a
AlphaFold is described as weaker for proteins with few intra-chain or homotypic contacts compared with heterotypic contacts. This leaves a meaningful gap for chains whose folded shape is largely determined by interactions in larger complexes.
↳ MSA depth and cross-chain contacts, p. 588
The pLDDT confidence measure correlates with observed lDDT-Cα accuracy, but the reported Pearson correlation of 0.76 leaves non-trivial residual variance. Users still need caution when relying on confidence scores for difficult downstream decisions.
↳ Fig. 2c
The score is driven by the unusually direct match between the paper’s main claim and its evidence: CASP14 provides a blind field benchmark, and Fig. 1a shows a large margin over competing methods. The recent PDB analysis, ablations, confidence calibration, and open release of code and weights support both methodological quality and practical translation. The paper’s highest-impact pathway is immediate use in structural biology workflows and proteome-scale prediction. Generalizability is scored slightly lower than the other impact components because Fig. 5a and the cross-chain-contact discussion identify target classes where accuracy is materially less reliable.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Exemplary4.7
Confidence highThe score reflects a major advance over prior protein-structure prediction methods, anchored in a blind CASP14 benchmark and a clear claim of atomic-accuracy prediction. The contribution is strongly scoped to the structure-prediction component of the protein-folding problem rather than all protein biophysics.
first computational method that can regularly predict protein structures with atomic accuracy
The design uses CASP14 as a blind benchmark, a post-training-cutoff PDB evaluation, explicit metrics, uncertainty intervals, ablations, confidence calibration, and public code and weights. The score is slightly below the maximum because full architectural and training details are partly distributed into supplementary methods and because performance limits depend on MSA depth and complex-contact context.
Data are median and the 95% confidence interval
The paper presents a coherent sequence from problem framing to model architecture, benchmark evaluation, broader validation, and limitations. Some metric distinctions, especially domain-level versus full-chain reporting, require careful reading but are explained in figure legends and Methods.
A description of the most important ideas and components is provided below.
The paper situates AlphaFold within physical simulation, evolutionary-history approaches, CASP history, deep-learning contact prediction, and direct-coordinate prediction. It also states remaining limits, including MSA depth and heterotypic contacts, rather than presenting the method as universally solved.
has proceeded along two complementary paths
No concerns4 of 4 checks
The central performance figures are consistent across the abstract, results, and figures, including CASP14 median r.m.s.d.95 and recent PDB validation metrics. Methods, reported outcomes, bootstrap intervals, and confidence-calibration analyses align without detected internal contradictions.
The study is computational and declares human, animal, and clinical ethics review as not applicable. Patent applications are disclosed, and the Data availability and Code availability sections provide public data sources, trained weights, source code, and inference scripts.
Flags: 4 declared / 5 total
All 85 cited references resolved to real publications via PubMed, Crossref, or OpenAlex.
References not verified automatically are resolved manually within 24 hours.
No retraction or expression of concern found in Crossref (includes Retraction Watch).
Sources: Crossref (incl. Retraction Watch) ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
The method is close to operational use because it reports practical inference timings and releases source code, trained weights, and inference scripts. The paper also cites use in molecular replacement and cryo-EM map interpretation.
GPU minutes to GPU hours
AI-generated, human-governed. Something look off? Contact us to request a review.
Keep reading
Other papers Nabu has evaluated blind, starting with the same field.
Han, Bruls, Soyman, et al. · 2019 · Neuroscience
rats exchange information about danger in both directions: how the observer reacts to the demonstrator's distress also influences how the demonstrator responds to the danger
SupportedIveland, Hagen, Sharma, et al. · 2020 · Basic medicine
UNG2 degradation appeared to be ubiquitous and was observed across several mammalian cell lines of different origin and with several HDACis.
SupportedKavli, Iveland, Buchinger, et al. · 2021 · Biological sciences
we demonstrate that uracil-DNA glycosylase (UNG), but not SMUG1 efficiently excises uracil from replication protein A (RPA)-coated ssDNA
Supported