CoChat
Sign Up
Cited by 155Protein Structure and DynamicsEnzyme Structure and FunctionMachine Learning in Bioinformatics

The power and pitfalls of AlphaFold2 for structure prediction beyond rigid globular proteins

Open access
Agarwal, Vinayak, McShan, Andrew C.·2024·Nature Chemical Biology
DOI 10.1038/s41589-024-01638-wPMC PMC11956457
Open this paper in CoChat — free accountOpen PDF

Abstract

Artificial intelligence-driven advances in protein structure prediction in recent years have raised the question: has the protein structure-prediction problem been solved? Here, with a focus on nonglobular proteins, we highlight the many strengths and potential weaknesses of DeepMind's AlphaFold2 in the context of its biological and therapeutic applications. We summarize the subtleties associated with evaluation of AlphaFold2 model quality and reliability using the predicted local distance difference test (pLDDT) and predicted aligned error (PAE) values. We highlight various classes of proteins that AlphaFold2 can be applied to and the caveats involved. Concrete examples of how AlphaFold2 models can be integrated with experimental data in the form of small-angle X-ray scattering (SAXS), solution NMR, cryo-electron microscopy (cryo-EM) and X-ray diffraction are discussed. Finally, we highlight the need to move beyond structure prediction of rigid, static structural snapshots toward conformational ensembles and alternate biologically relevant states. The overarching theme is that careful consideration is due when using AlphaFold2-generated models to generate testable hypotheses and structural models, rather than treating predicted models as de facto ground truth structures.

Verified 2026-10-07 · 1 reference not checked yet

AI snapshot

Written from the paper’s full text.

TL;DR

This review examines AlphaFold2’s strengths and failure modes for proteins beyond rigid globular folds, stressing that confidence scores do not guarantee agreement with experimental structures. It argues AF2 is best used as a hypothesis-generating complement to experimental data rather than as definitive structural truth.

Key findings

  • AF2 has broad coverage—modeling 98.5% of the human proteome, with over 200 million precomputed predictions in the AlphaFold Protein Structure Database—but accuracy varies by target class.
  • pLDDT and PAE are useful but imperfect: low pLDDT (<70) or high PAE (>5 Å) signal caution, yet high-confidence models can still deviate from experiments. On average, AF2 predicts about 50% of residues with high confidence, and one cited estimate suggests ~10% of high-confidence residues differ from native backbone by more than 2 Å.
  • Non-globular and dynamic systems are especially challenging: peptides were excluded from training, IDPs/IDRs are often conformational ensembles, membrane proteins are underrepresented in the PDB (<3%), and AF2 struggles with alternative conformations, ligand/cofactor effects, and domain placement. Heterodimeric complex success rates were reported at 51% and 63% in two benchmarks, with antibody complexes particularly difficult.
  • Integrating AF2 with SAXS, NMR, X-ray, and cryo-EM can improve models; AF-guided molecular replacement solved 187 of 215 structures in one benchmark, though cryo-EM maps worse than 6 Å are not recommended for fitting.

Methods

  • Narrative review/perspective, not a primary research study; no new experimental data or systematic meta-analysis are presented.
  • Synthesizes published benchmarks, case examples, and prior literature on AF2 workflow, pLDDT/PAE interpretation, protein classes, experimental integration, and conformational ensembles.
  • Uses selected examples—such as OSBP1, insulin, ChRmine, SecA, and beta-endorphin amyloid—to illustrate strengths and failure modes.

Limitations

  • Stated AF2 limitations include no training or validation for mutation effects, training bias toward soluble/PDB folds, peptide exclusion, membrane-protein underrepresentation, poor handling of IDPs/IDRs and alternative conformations, and cases where high pLDDT/PAE still mislead; inter-domain PAE accuracy is less benchmarked.
  • Evident from the text, not stated by the authors: as a narrative review, the selection of examples and benchmarks is not systematic, and the conclusions depend on the cited literature rather than new quantitative assessment.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41589-024-01638-w

Cites

110 of 111 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 132 cited sources checked · 0 supported · 0 flagged

  • Premisecitation

    AF2 has enabled modeling of 98.5% of the human proteome.

    “modeling of 98.5% of the human proteome5”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [5] Tunyasuvunakool, K. et al. Highly accurate protein structure prediction for the human proteome. Nature 596, 590–596 (2021). source

      Not checked in the free run.

  • Premisecitation

    Machine-learning tools such as AF2, RoseTTAFold, and ESMFold are currently the most accurate answer to the structure prediction problem.

    “are the best and most accurate answer to the structure prediction problem to date”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [1] Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). source

      Not checked in the free run.

    • cites [17] Akdel, M. et al. A structural biology community assessment of AlphaFold2 applications. Nat. Struct. Mol. Biol. 29, 1056–1067 (2022). source

      Not checked in the free run.

    • cites [18] Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871–876 (2021). source

      Not checked in the free run.

    • cites [19] Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123–1130 (2023). source

      Not checked in the free run.

  • Premisecitation

    It is not immediately clear from pLDDT or PAE whether an AF2 model resembles the apo or holo form.

    “whether the modeled structure resembles the apo or holo form of the protein is not immediately clear from analysis of pLDDT scores or PAE graphs32”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [32] Saldaño, T. et al. Impact of protein conformational diversity on AlphaFold predictions. Bioinformatics 38, 2742–2748 (2022). source

      Not checked in the free run.

  • Premisecitation

    Benchmarking AF2 on membrane proteins is difficult because current AlphaFold versions do not directly account for the membrane environment.

    “Benchmarking AF2 against membrane proteins represents a challenge since the membrane environment, which includes lipids and other proteins, is not directly considered by current versions of AlphaFold38”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [38] Hegedűs, T., Geisler, M., Lukács, G. L. & Farkas, B. Ins and outs of AlphaFold2 transmembrane protein structure predictions. Cell. Mol. Life Sci. 79, 73 (2022). source

      Not checked in the free run.

  • Premisecitation

    NMR structural ensembles provide a unique validation metric for AF2 because AF2's training PDB subset excluded NMR data.

    “NMR structural ensembles offer a unique validation metric to assess the accuracy of predicted AlphaFold2 models since AlphaFold2 was trained on a subset of the PDB that excluded NMR data1,27–30”
    Quote in the paperSubstantial claim0 of 5 sources checked
    • cites [1] Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). source

      Not checked in the free run.

    • cites [27] Alderson, T. R., Pritišanac, I., Kolarić, Đ., Moses, A. M. & Forman-Kay, J. D. Systematic identification of conditionally folded intrinsically disordered regions by AlphaFold2. Proc. Natl Acad. Sci. USA 120, e2304302120 (2023). source

      Not checked in the free run.

    • cites [28] Fowler, N. J. & Williamson, M. P. The accuracy of protein structures in solution determined by AlphaFold and NMR. Structure 30, 925–933 (2022). source

      Not checked in the free run.

    • cites [29] Zweckstetter, M. NMR hawk‐eyed view of AlphaFold2 structures. Protein Sci. 30, 2333–2337 (2021). source

      Not checked in the free run.

    • cites [30] Tejero, R., Huang, Y. J., Ramelot, T. A. & Montelione, G. T. AlphaFold models of small proteins rival the accuracy of solution NMR Structures. Front. Mol. Biosci. 9, 877000 (2022). source

      Not checked in the free run.

  • May combine two claims. 28
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 15

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    In the OSBP1 AF2 model, the PH, CC, and ORD domains have high predicted confidence, the FFAT domain has very low confidence, and the PAE graph shows low confidence in relative domain placement.

    “the PH, CC, and ORD domain structure are assigned very high to high predicted confidence, while the FFAT domain is predicted with very low confidence”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Current AF2 versions do not employ energy functions that seek native-like conformations, unlike Rosetta.

    “current versions of AF2 do not use energy functions that seek to identify native-like protein conformations, unlike its competitor Rosetta13”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [13] Alford, R. F. et al. The Rosetta all-atom energy function for macromolecular modeling and design. J. Chem. Theory Comput. 13, 3031–3048 (2017). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    ColabFold and OpenFold let users run a modified AF2 protocol on open-access servers within minutes.

    “allow users to run a modified AF2 protocol on open access servers in minutes”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [8] Mirdita, M. et al. ColabFold: making protein folding accessible to all. Nat. Methods 19, 679–682 (2022). source

      Not checked in the free run.

    • cites [9] Ahdritz, G. et al. OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization. Nat. Methods https://doi.org/10.1038/s41592-024-02272-z (2024). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    Running AF2 locally demands substantial computational resources, including up to 3 terabytes of disk and modern NVIDIA GPUs.

    “requires significant computational resources to run the software locally (up to 3 terabytes of disk space and a modern NVIDIA GPUs with gigabytes of memory)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The AlphaFold Protein Structure Database contains more than 200 million precomputed AF2 predictions.

    “houses over 200 million pre-run AF2 predictions”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [6] Varadi, M. et al. AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucleic Acids Res. 50, D439–D444 (2022). source

      Not checked in the free run.

    • cites [7] Varadi, M. et al. AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences. Nucleic Acids Res. 52, D368–D375 (2024). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    AlphaFold2 can predict three-dimensional protein structures accurately from sequence alone, which may solve the long-standing structure prediction problem.

    “enables accurate prediction of three-dimensional protein structures from only the target amino acid sequence”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [1] Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). source

      Not checked in the free run.

    • cites [2] Bertoline, L. M. F., Lima, A. N., Krieger, J. E. & Teixeira, S. K. Before and after AlphaFold2: an overview of protein structure prediction. Front. Bioinform. 3, 1120370 (2023).

      Not checked in the free run.

    • cites [3] Perrakis, A. & Sixma, T. K. AI revolutions in biology. EMBO Rep. 22, e54046 (2021). source

      Not checked in the free run.

    • cites [4] Bouatta, N., Sorger, P. & AlQuraishi, M. Protein structure prediction by AlphaFold2: are attention and symmetries all you need? Acta Crystallogr. D 77, 982–991 (2021). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Tsaban et al showed AF2 can be adapted to model peptide/protein complexes accurately regardless of peptide length, though results were biased toward helical peptides and those not rearranging greatly upon binding.

    “Tsaban et al showed that AF2 can be adapted to accurately model peptide/protein complexes irrespective of peptide length, although the results seemed biased towards helical structures and peptides that do not undergo large structural rearrangements upon binding24”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24] Tsaban, T. et al. Harnessing protein folding neural networks for peptide–protein docking. Nat. Commun. 13, 176 (2022). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    AF2 outputs a per-residue pLDDT confidence score from 0 to 100, where higher values mean greater confidence.

    “the predicted Local Distance Difference Test (pLDDT) score, which ranges from 0 to 100 with higher values assigned higher confidence in the model”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [1] Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). source

      Not checked in the free run.

    • cites [22] Mariani, V., Biasini, M., Barbato, A. & Schwede, T. lDDT: a local superposition-free score for comparing protein structures and models using distance difference tests. Bioinformatics 29, 2722–2728 (2013). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    For categories i and iv, low pLDDT and PAE warn users to interpret cautiously, but this warning is not obvious for categories ii and iii.

    “For cases i and iv, low confidence in pLDDT scores and PAE graphs alerts users to interpret structures with caution, which is not immediately clear in cases ii and iii.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Membrane proteins make up less than 3% of PDB structures, so AF2's training set was strongly biased toward soluble proteins.

    “membrane proteins represent less than 3% of total structures in the PDB15, meaning that the training set used by AF2 was highly biased towards soluble proteins39”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [15] Laurents, D. V. AlphaFold 2 and NMR spectroscopy: partners to understand protein structure, dynamics and function. Front. Mol. Biosci. 9, 906437 (2022). source

      Not checked in the free run.

    • cites [39] Topitsch, A., Schwede, T. & Pereira, J. Outer membrane β-barrel structure prediction through the lens of AlphaFold2. Proteins 92, 3–14 (2024). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The top-ranked AF2 peptide models chosen by high pLDDT often did not have the lowest Cα RMSD to experimental structures, indicating pLDDT is not ideal for classifying peptide conformations.

    “the best ranked AF2 models (selected on the basis of high pLDDT score) often did not exhibit the lowest Cα root-mean-square deviation (RMSD) relative to the experimental structure, suggesting the pLDDT metric used by AF2 to assess protein models is not optimal for classification of peptide conformations25”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [25] McDonald, E. F., Jones, T., Plate, L., Meiler, J. & Gulsevin, A. Benchmarking AlphaFold2 on peptide structure prediction. Structure 31, 111–119 (2023). source

      Not checked in the free run.

  • May combine two claims.

    Premiseown data

    For Complement C6, AF2 models individual domains well but misplaces them relative to each other.

    “AF2 predicts the structure of individual domains well but deviates in placement of domains relative to each other (Fig. 2e)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Peptide structure prediction is difficult because AF2's training set omitted peptides, short sequences yield poor MSAs, and many peptides form conformational ensembles rather than one static structure.

    “Peptide structure prediction poses additional challenges given that the benchmark set used to train AF2 excluded peptides, the difficulty in generating robust MSAs for short sequences, and observations that many peptides exist in solution as conformational ensembles rather than a single static conformation1,24,25”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [1] Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). source

      Not checked in the free run.

    • cites [24] Tsaban, T. et al. Harnessing protein folding neural networks for peptide–protein docking. Nat. Commun. 13, 176 (2022). source

      Not checked in the free run.

    • cites [25] McDonald, E. F., Jones, T., Plate, L., Meiler, J. & Gulsevin, A. Benchmarking AlphaFold2 on peptide structure prediction. Structure 31, 111–119 (2023). source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    Several groups suggest the default AF2 pipeline struggles to model alternative conformations.

    “Several groups have suggested that the default AF2 pipeline has difficulty in modeling alternative conformations45”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [45] Lane, T. J. Protein structure prediction has reached the single-structure frontier. Nat. Methods 20, 170–173 (2023). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    AF2 also produces a PAE matrix that assesses the relative orientation and position of different model parts or domains.

    “AF2 also generates a Predicted Aligned Error (PAE) matrix, which evaluates the relative orientation and position of different parts (i.e., domains) of the model6”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [6] Varadi, M. et al. AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucleic Acids Res. 50, D439–D444 (2022). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    AF2 performs poorly on targets embedded in membrane thickness outside 15-35 Å and on targets with novel features uncommon in the PDB.

    “AF2 performs poorly for targets embedded in membrane thickness outside the 15–35 Å range as well as targets with novel features not commonly present in the PDB”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [38] Hegedűs, T., Geisler, M., Lukács, G. L. & Farkas, B. Ins and outs of AlphaFold2 transmembrane protein structure predictions. Cell. Mol. Life Sci. 79, 73 (2022). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The AF2 ChRmine model captures the overall fold, but its N-terminal region and extracellular loops differ from the cryo-EM structure, likely because of ChRmine's unique covalent Schiff base.

    “while the AF2 model of the channelrhodopsin ChRmine captures its overall fold, the modeled N-terminal region and extracellular loops deviate from its experimental high-resolution cryogenic electron microscopy (cryo-EM) structure (Fig. 2f), likely due to ChRmine’s unique covalent Schiff base feature41”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [41] Kishi, K. E. et al. Structural basis for channel conduction in the pump-like channelrhodopsin ChRmine. Cell 185, 672–689 (2022). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    AF2 models with transmembrane lengths matching non-physiological membrane thickness can have very high pLDDT, so pLDDT alone cannot select native membrane protein conformations.

    “AF2 models with trans-membrane region lengths corresponding to non-physiological membrane thickness values can exhibit very high pLDDT scores (high model confidence), suggesting pLDDT scores alone are not sufficient to select native membrane protein conformations”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [38] Hegedűs, T., Geisler, M., Lukács, G. L. & Farkas, B. Ins and outs of AlphaFold2 transmembrane protein structure predictions. Cell. Mol. Life Sci. 79, 73 (2022). source

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    The NMR structure of Ca2+-bound polycystin 2 differs from the AF2 model, possibly due to conformational changes from calcium binding.

    “the NMR structure of Ca2+ bound polycystin 2 deviates from the AF2 model, potentially due to conformational changes upon calcium binding (Fig. 2c)”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Hegedűs et al benchmarked membrane proteins absent from AF2's original training set and concluded AF2 predicts transmembrane proteins as well as soluble proteins on average.

    “Hegedűs et al benchmarked several membrane proteins not included in the original AF2 training set and concluded that on average AF2 predicts transmembrane proteins as well as soluble proteins38”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [38] Hegedűs, T., Geisler, M., Lukács, G. L. & Farkas, B. Ins and outs of AlphaFold2 transmembrane protein structure predictions. Cell. Mol. Life Sci. 79, 73 (2022). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    PAE values are only confidence estimates, and PAE accuracy for inter-domain prediction has been benchmarked less extensively than for intra-domain contacts.

    “it is important to remember that the PAE values are only confidence estimates. Furthermore, the accuracy of PAE graphs for inter-domain prediction has not been as extensively benchmarked as for intra-domain contacts36”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [36] David, A., Islam, S., Tankhilevich, E. & Sternberg, M. J. E. The AlphaFold database of protein structures: a biologist’s guide. J. Mol. Biol. 434, 167336 (2022). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The AF2 insulin model deviates greatly from its experimental NMR structure, possibly because AF2 cannot orient cysteine pairs for disulfide formation.

    “the AF2 model of insulin deviates significantly from its experimental NMR structure (Fig. 2b), potentially due to the inability of AF2 to orient cysteine pairs for disulfide bond formation31”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [31] Thornton, J. M., Laskowski, R. A. & Borkakoti, N. AlphaFold heralds a data-driven revolution in biology and medicine. Nat. Med. 27, 1666–1669 (2021). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    High pLDDT or low PAE does not guarantee that a model matches native conformations; these metrics instead estimate confidence in local and global coordinate positions or orientations.

    “high pLDDT or low PAE metrics, indicating high confidence in the prediction, do not promise agreement with native protein conformations, but instead estimate a likelihood for local and global coordinate positions and/or orientations”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    AF2 deviations from experimental structures fall into four categories: wrong secondary/tertiary/quaternary structure in low-to-moderate confidence regions; wrong structure in high-confidence regions; correct backbone but wrong side-chain rotamers; and correct individual domain backbones but wrong relative domain placement.

    “These deviations can be broadly categorized into cases with i) inaccurate secondary, tertiary, and/or quaternary structure in regions where AF2 predicts low to moderate confidence, ii) inaccurate structure in regions where AF2 predicts high model confidence, iii) correct backbone structure but incorrect fine details (i.e., side chain rotamer placements), and iv) correct backbone”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Bryant et al reported a 63% success rate for heterodimeric complexes.

    “Bryant et al reported a 63% success rate for heterodimeric complexes56”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [56] Bryant, P., Pozzati, G. & Elofsson, A. Improved prediction of protein–protein interactions using AlphaFold2. Nat. Commun. 13, 1265 (2022).

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Modifications of AF2 can produce models that differ substantially from each other, enabling sampling of conformational landscapes.

    “several groups have shown that modifications of AF2 have the potential to generate models that significantly deviate from each other, allowing for sampling of conformational landscapes”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Other reports compared pLDDT with crystallographic B-factors and suggested AF2 confidence metrics cannot directly report local flexibility.

    “other reports have compared pLDDT scores with crystallographic B-factors to suggest AF2 confidence metrics are unable to provide direct information on local flexibility66”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [66] Carugo, O. pLDDT values in AlphaFold2 protein models are unrelated to globular protein local flexibility. Crystals 13, 1560 (2023). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    If the best AF2 model has many residues with low pLDDT (<70) or high PAE (>5 Å), its backbone is very unlikely to match native and the model cannot be reliably interpreted.

    “if the “best” AF2 model exhibits many residues with low pLDDT (< 70) or high PAE values (> 5 Å), the likelihood that the backbone structure matches the native conformation is very low and the model cannot be reliably interpreted”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    More robust evaluation of AF2-SAXS fits has been proposed using combinations of Vc, Qr, X2free, and Rsas metrics.

    “additional strategies such as the combination of Vc, Qr, X2free, and Rsas metrics have been proposed as more robust evaluation metrics75.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [75] Rambo, R. P. & Tainer, J. A. Accurate assessment of mass, models and resolution by small-angle scattering. Nature 496, 477–481 (2013). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    AlphaMissense does not explicitly determine a mutation's structural effects but gives the probability that a missense variant is pathogenic.

    “AlphaMissense, does not explicitly determine the structural effects of a mutation on a protein but provides the probability of a missense variant being pathogenic51”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [51] Cheng, J. et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science 381, eadg7492 (2023). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Standard molecular replacement requires a template or homologous structure and performs best when the template is within 2 Å Cα RMSD of the target.

    “Standard molecular replacement strategies require 3D coordinates of a template/homologous structure and work best when the template is < 2 Å Cα RMSD from the target structure80.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [80] Abergel, C. Molecular replacement: tricks and treats. Acta Crystallogr. D 69, 2167–2173 (2013). source

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    This failure may stem from training bias toward stable structures or inability to extract signal from small mutations via MSAs.

    “which may be due to a training bias on stable structures or an inability to extract signal from small mutations through MSAs”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Some groups used MD simulations to suggest pLDDT and PAE provide information about dynamics and disorder.

    “Some groups have used MD simulations to suggest that pLDDT and PAE metrics provide information on dynamics and disorder42,65”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [42] Ruff, K. M. & Pappu, R. V. AlphaFold and implications for intrinsically disordered proteins. J. Mol. Biol. 433, 167208 (2021). source

      Not checked in the free run.

    • cites [65] Guo, H.-B. et al. AlphaFold2 models indicate that protein sequence determines both structure and dynamics. Sci. Rep. 12, 10696 (2022).

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Preliminary reports suggest AF2 may predict oligomeric states of proteins and amyloids.

    “Preliminary reports also suggest that AF2 may also be able to predict oligomeric states of proteins and amyloids17,57”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [17] Akdel, M. et al. A structural biology community assessment of AlphaFold2 applications. Nat. Struct. Mol. Biol. 29, 1056–1067 (2022). source

      Not checked in the free run.

    • cites [57] Jeppesen, M. & André, I. Accurate prediction of protein assembly structure by combining AlphaFold and symmetrical docking. Nat. Commun. 14, 8283 (2023). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    For NT-9, slight tertiary structure deviations occur, but the overall structure is well described.

    “Slight deviations in tertiary structure are noted, such as for the nuclear transport factor 2 derived de novo designed protein NT-9, but the overall structure is well described (Fig. 2j)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The standard AlphaFold implementation detects only a single structural snapshot, the ground state, likely because its training set lacks a large set of redundant protein conformers.

    “The standard implementation of AlphaFold performs well only in detecting a single structural snapshot (the “ground state” structure), likely due to the lack of a large set of redundant protein conformers in the training set16,32.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [16] Chakravarty, D. & Porter, L. L. AlphaFold2 fails to predict protein fold switching. Protein Sci. 31, e4353 (2022). source

      Not checked in the free run.

    • cites [32] Saldaño, T. et al. Impact of protein conformational diversity on AlphaFold predictions. Bioinformatics 38, 2742–2748 (2022). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Methyl side-chain labeling increases signal and resolution in NMR studies of large biological assemblies.

    “This is especially helpful for large biological assemblies where methyl side-chain labeling affords an increase in signal and resolution91,92.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [91] McShan, A. C. Utility of methyl side chain probes for solution NMR studies of large proteins. J. Magn. Reson. Open 14–15, 100087 (2023).

      Not checked in the free run.

    • cites [92] Ruschak, A. M. & Kay, L. E. Methyl groups as probes of supra-molecular structure, dynamics and function. J. Biomol. NMR 46, 75–87 (2009).

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Current AF2 implementations can provide highly accurate working models for most rigid, well-folded globular proteins but may have issues predicting other protein classes.

    “Current implementations of AF2 can provide highly accurate working models for most rigid, well-folded globular proteins, but may have issues predicting other classes of proteins.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    AF2 models are not suitable for fitting into cryo-EM maps with resolution worse than 6 Å.

    “Reports suggest AF2 models should not be fit into cryo-EM maps with resolution greater than 6Å98,99.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [98] Dai, X., Wu, L., Yoo, S. & Liu, Q. Integrating AlphaFold and deep learning for atomistic interpretation of cryo-EM maps. Brief. Bioinform. 24, bbad405 (2023).

      Not checked in the free run.

    • cites [99] Alshammari, M., He, J. & Wriggers, W. Appraisal of AlphaFold2-predicted models in cryo-EM map interpretation. Microsc. Microanal. 29, 977–978 (2023). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Experimental data, especially SAXS, NMR, and cryo-EM, have been described as guides for modeling ensembles and alternative conformations.

    “The use of experimental data, especially SAXS, NMR, and cryo-EM, has also been described to guide modeling of ensembles and alternative conformations71,82,108.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [71] Brookes, E., Rocco, M., Vachette, P. & Trewhella, J. AlphaFold-predicted protein structures and small-angle X-ray scattering: insights from an extended examination of selected data in the Small-Angle Scattering Biological Data Bank. J. Appl. Crystallogr. 56, 910–926 (2023). source

      Not checked in the free run.

    • cites [82] Abdollahi, H., Prestegard, J. H. & Valafar, H. Computational modeling multiple conformational states of proteins with residual dipolar coupling data. Curr. Opin. Struct. Biol. 82, 102655 (2023). source

      Not checked in the free run.

    • cites [108] Ramelot, T. A., Tejero, R. & Montelione, G. T. Representing structures of the multiple conformational states of proteins. Curr. Opin. Struct. Biol. 83, 102703 (2023). source

      Not checked in the free run.

40 of 40

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Cites

Flagged first, then by citations

1 of the 112 references couldn’t be matched to a known paper.

Research works printed but not linked: 6

  1. Bertoline, L. M. F., Lima, A. N., Krieger, J. E. & Teixeira, S. K. Before and after AlphaFold2: an overview of protein structure prediction. Front. Bioinform. 3, 1120370 (2023).
  2. Zhao, B., Ghadermarzi, S. & Kurgan, L. Comparative evaluation of AlphaFold2 and disorder predictors for prediction of intrinsic disorder, disorder content and fully disordered proteins. Comput. Struct. Biotechnol. J. 21, 3248–3258 (2023).
  3. McShan, A. C. Utility of methyl side chain probes for solution NMR studies of large proteins. J. Magn. Reson. Open 14–15, 100087 (2023).
  4. Ruschak, A. M. & Kay, L. E. Methyl groups as probes of supra-molecular structure, dynamics and function. J. Biomol. NMR 46, 75–87 (2009).
  5. Dai, X., Wu, L., Yoo, S. & Liu, Q. Integrating AlphaFold and deep learning for atomistic interpretation of cryo-EM maps. Brief. Bioinform. 24, bbad405 (2023).
  6. Johansson-Åkhe, I. & Wallner, B. Improving peptide–protein docking with AlphaFold-Multimer using forced sampling. Front. Bioinform. 2, 959160 (2022).

Cited by

Sorted by citations

Related

Authors
Agarwal, V., & McShan, A. C.
Affiliations
Georgia Institute of Technology (US)
Corresponding author
  • Agarwal, Vinayak
  • McShan, Andrew C.
ORCID
Venue
Nature Chemical Biology
Publisher
Springer Science and Business Media LLC
Published
2024-06-21
Volume
20
Issue
8
Pages
950-959
ISSN
  • 1552-4450 (ISSN-L)
  • 1552-4469
DOI
10.1038/s41589-024-01638-w
PMC
PMC11956457
OpenAlex
W4399893066
Cited by
155
References
112
Citations by year
2026: 79 · 2025: 63 · 2024: 13
FWCI
6.31
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Chemical Biology (publishedVersion)
  • PubMed (publishedVersion)
  • Europe PMC (PubMed Central) (acceptedVersion, other-oa)
  • PubMed Central (acceptedVersion)
Funders
  • U.S. Department of Health & Human Services | NIH | National Institute of General Medical Sciences (R35GM142882)
  • National Science Foundation (CHE-2238650)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Physical Sciences › Materials Science › Materials Chemistry › Enzyme Structure and Function
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
Keywords
AlphaFold2, protein structure prediction, intrinsically disordered proteins, conformational ensembles, model reliability, pLDDT, small-angle X-ray scattering, solution NMR, cryo-electron microscopy, X-ray diffraction
MeSH terms
Artificial Intelligence, Humans, Models, Molecular, Protein Conformation, Proteins, X-Ray Diffraction, Protein Folding, Cryoelectron Microscopy, Scattering, Small Angle

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review