---
title: How AlphaFold results compare to the bench
description: >-
  What an AlphaFold Database model has been shown to agree with
  experimentally, which AlphaFold version you are holding, and how to read
  pLDDT and PAE.
tags:
  - alphafold
  - validation
  - translatability
---

**You have an AlphaFold result. Here is what it supports.**

- **You are holding a database record, not a prediction run.** The platform's AlphaFold fetch is a lookup against the EBI AlphaFold Database. The database names its own producing pipeline in the metadata it returns for every entry — `AlphaFold Monomer v2.0 pipeline` — so the model is an **AlphaFold 2 monomer**. It is not AlphaFold 3, not AlphaFold-Multimer, and not ColabFold. Figures for those three do not describe what you have.
- **On single chains it is the most accurate method in every independent benchmark surveyed here, and only AlphaFold 3 matches it.** At CASP14 its models had a median backbone accuracy of **0.96 Å RMSD<sub>95</sub>** against the next best method's 2.8 Å, author-reported on 87 CASP14 domains. On 1,337 PDB chains deposited after every tested method's training cutoff, an independent group measured median **TM-score 0.96, RMSD 1.30 Å and GDT-TS 94%**, ahead of ESMFold and OmegaFold on all three. The comparison set in these studies is ESMFold, OmegaFold, RoseTTAFold and AlphaFold 3 — not every predictor in existence.
- **pLDDT and PAE are the two fields you actually hold, and they answer different questions.** pLDDT is per-residue and local; PAE is pairwise and tells you whether two domains are placed correctly relative to each other. A model can be high-pLDDT everywhere and still have its domains hinged wrong — [reading pLDDT and PAE](#reading-plddt-and-pae) has what each band is and is not backed by.
- **pLDDT ≥ 70 is highly sensitive and poorly specific.** On 1,666 template-poor PDB monomers, 1,067 of AlphaFold 2's 1,069 accurate models carried mean pLDDT ≥ 70 — but so did **150 of its 200 outright-wrong ones (75%)**. A confident score is not a correctness claim.
- **Docking into a fetched receptor costs roughly half your pose accuracy, and nothing warns you.** Top-1 pose accuracy fell from **41% on co-crystallised receptors to 17% on AlphaFold Database models** across 2,474 PDBBind complexes, with the share of poses landing over 5 Å away rising from 34% to 60%. Neither pLDDT nor binding-pocket RMSD predicts which way a given target will go.
- **A stock AlphaFold 2 run is not the CASP14 headline.** At CASP15, the control run of DeepMind's own AlphaFold 2 protocol reached correct topology on **85.3% of 109 evaluation units** and crystal-form-equivalent accuracy (GDT_TS ≥ 90) on only **34.9%**. The groups that beat it did so by building deeper and more diverse MSAs around the same model.
- **Not every entry in the database got the same treatment.** Swiss-Prot and reference-proteome entries are the most confident of five models; bulk UniProt entries are **the output of a single model run**. The original human-proteome release also used a deliberately reduced pipeline. Neither fact is visible in the file you download.
- **It is one chain, alone.** A fetched entry is a single chain modelled by itself, with no partners, ligands, ions or cofactors. Beware the version trap when you read the complex literature: AlphaFold release v2.2 is the release that ships the Multimer model, so published "AlphaFold2" figures on dimers are usually **Multimer** figures and describe something you cannot fetch.
- **This model is genuinely useful at the bench — for phasing.** Across four independent benchmarks, AlphaFold 2 search models solved **86–93%** of crystal structures by molecular replacement — on differing map-correlation criteria, and on sets deliberately chosen to be hard for MR. On the largest of those sets, the authors' own control — search models drawn only from existing PDB structures — solved **35.6%**. That contrast is the best-evidenced wet-lab use of a predicted structure anywhere in this documentation.
- **Confirm a structure with** X-ray crystallography, cryo-EM, or NMR. Nothing computational substitutes.

Jump to [agreement with wet lab](#agreement-with-wet-lab) or [against other in silico methods](#against-other-in-silico-methods). The general framing for every tool is in [reading an in silico result against the bench](/docs/translatability).

## How in silico is used currently

Predicted structures are used to **supply a model where no experimental one exists**, and for AlphaFold specifically the field's answer to "how much did that help" has been measured directly against the structural coverage of a proteome.

- **Covering the gap, quantified.** An independent analysis of the human proteome found that structural coverage of all human protein residues rose from **48% to 76%** once AlphaFold models were counted alongside experimental structures and template-based homology models. Coverage by *high-quality* models (≥ 50% sequence identity to a PDB chain) rose from **31% to 50%**, and the number of human proteins with no structural coverage at all fell from **5,027 to 29**. The dark proteome shrank from 26% to 10%. That is the largest single jump in structural coverage the field has had.
- **But the gap closed least where it was widest.** In the same analysis, AlphaFold produced high-confidence structure (pLDDT > 90) for only **37% of residues in Pfam domains lacking any experimental structure and 21% of dark-proteome residues**, while **21% and 39%** of those same two sets came back below pLDDT 50. The coverage headline and the confidence breakdown point in opposite directions, and both are real.
- **Triage before an experiment.** Rank candidates for crystallography, pick a construct boundary, decide which domain to express. A wrong model here costs a failed plate rather than a wrong conclusion.
- **Phasing a crystal structure.** This is the one use where a predicted structure is consumed by a wet-lab workflow and the outcome is measured crystallographically rather than by a similarity metric. It has the best evidence on this page and it has its own section below.
- **An input to a structure-based step** — pocket finding ([P2Rank](/docs/p2rank)), docking ([DiffDock](/docs/diffdock)), protonation ([PDB2PQR](/docs/pdb2pqr)). This is where prediction error propagates, and for docking the cost has been measured on database models directly: top-1 pose accuracy roughly halves against a co-crystal receptor. It is still the common practice, because for most targets there is no co-crystal to use instead.
- **Variant interpretation.** The database also ships AlphaMissense substitution annotations per entry. The platform does not surface them, and they are a different model with its own validation — do not read a structural pLDDT as a variant-effect claim.

What the database's own maintainers say, and what the field broadly accepts, is that these are models. The AlphaFold Database's user guidance ends on the point that the files "provide coordinates for all regions, and it is incumbent upon the user to interpret the model prudently" — coordinates exist even where confidence does not.

## Against other in silico methods

**At CASP15, with no DeepMind entry, the stock AlphaFold 2 protocol was the baseline that better groups beat.** 132 groups submitted for 112 evaluation units from 77 single-chain targets; 118 groups that attempted at least 10 units were evaluated. Among them were two "control" runs: `NBIS-af2-standard`, implementing DeepMind's original AlphaFold 2 protocol as of March 2022, and `ColabFold`. On `model_1` submissions by GDT_TS:

| Group | GDT_TS ≥ 45 (correct topology) | GDT_TS ≥ 90 | Median GDT_TS |
|---|---|---|---|
| Yang (MSA-based) | 97.2% (105/108) | 47.2% | 89.26 |
| UM-TBM | 96.3% (105/109) | 42.2% | 87.36 |
| PEZYFoldings | 94.4% (101/107) | 49.5% | 89.65 |
| ColabFold | 85.3% (93/109) | 38.5% | 86.67 |
| **NBIS-af2-standard** (DeepMind protocol) | **85.3% (93/109)** | **34.9% (38/109)** | **85.88** |
| ESM-single-sequence (language model only) | 77.4% (72/93) | 20.4% | 77.71 |

Read the denominators: the AlphaFold 2 controls attempted 109 of the 112 evaluation units while `ESM-single-sequence` attempted 93, so the percentages are over different populations.

The assessors' own reading: the AlphaFold 2 controls "are a little way off the best performance possible, though each produces many high-quality models," and **"the best performing method that was entirely independent of AF2 was that of the BAKER group, 28th place, using RoseTTAFold2."** What separated the leaders was not a better model but **deeper and more diverse MSAs** — collecting from more databases and sampling across sub-clustered alignments. Read against the CASP14 result, this is the most useful single comparison on the page: **the headline "AlphaFold solved it" figures came from CASP14's hardest-target set graded on best-of-five submissions, and a plain AlphaFold 2 run on a fresh target set two years later got correct topology 85% of the time and crystal-equivalent accuracy a third of the time.**

**Against AlphaFold 3, the difference on monomers is within noise.** On the 1,666 template-poor monomers, AlphaFold 3 reached 1,076 accurate / 388 domain movement / 191 incorrect against AlphaFold 2's 1,069 / 386 / 200 — a 7-structure difference in the accurate class out of 1,666. AlphaFold 3's advantages are in ligands and complexes, not in monomer backbones, and the database does not ship AlphaFold 3 models. In the GH11 case study the two were indistinguishable at 0.30 Å and 0.30–0.42 Å.

**Against single-sequence predictors, AlphaFold 2 wins on accuracy and loses on everything else.** On the 1,337-chain set it beat ESMFold on every metric (TM-score 0.96 vs 0.95, RMSD 1.30 vs 1.74 Å, GDT-TS 94% vs 90%) and on the 1,666-monomer set by a wide margin (1,069 accurate vs 832). But ESMFold needs no alignment and runs in 30 seconds to a few minutes where a local AlphaFold 2 run takes half an hour to two hours plus the MSA search. For a sequence already in the database, the AlphaFold fetch wins on both axes at once — seconds *and* higher accuracy — because the inference was already paid for. For a sequence that is not, [ESMFold](/docs/esmfold/translatability) is the trade to evaluate.

**Where the single-sequence methods are reported to win, it is narrow and specific.** On the de novo designed subset (n = 102) of the 1,337-chain benchmark, ESMFold and OmegaFold achieved significantly lower RMSD for designs than for natural proteins, while **AlphaFold 2 showed a weakness unique to it — significantly lower TM-scores for designs than for monomers or complexes**. Direction and significance are published; the per-group medians sit in a figure panel and were not extracted. If what you folded is a design rather than a natural sequence, AlphaFold 2's advantage is the one case where it is reported to invert.

### On complexes, and the version trap

The database ships **single chains** — a UniProt F1 entry is one chain, reports `isComplex: false`, and carries no ligands, ions or cofactors. Modelling an assembly needs AlphaFold-Multimer or a co-folding model, and the AlphaFold fetch reaches neither. The figures below are therefore about what you *cannot* get from this tool, and they exist on this page mainly so that a Multimer number is never mistaken for a reason to trust a fetched monomer.

**First, the version trap, because it has already corrupted the secondary literature.** AlphaFold release **v2.2 is the release that ships the Multimer model**. A paper that reports "AlphaFold2 v2.2.x" on dimers is reporting an **AlphaFold-Multimer** figure unless it states its `model_preset`, because predicting a dimer in that release requires the multimer preset. Several published complex figures labelled "AlphaFold2" are Multimer figures. Check the preset, not the label.

**AlphaFold-Multimer, author-reported.** On 2,609 homology-reduced in-contact chain pairs drawn from 4,446 recent-PDB complexes — filtered so neither chain exceeds 40% template overlap with the training set — the system reached DockQ ≥ 0.23 (acceptable or better) on **70% of 1,323 heteromeric interfaces and 72% of 1,286 homomeric interfaces**, selecting the top model by its own confidence score (0.8·ipTM + 0.2·pTM). This work was never journal-published; it remains a preprint.

**The plain AlphaFold 2 monomer model, used on complexes.** AlphaFold v2.0's monomer model, given per-chain MSAs concatenated and padded into one unpaired alignment and no structural templates, reached **51% acceptable-or-better, 43% medium-or-better and 21% high accuracy** on 149 of the 152 heterodimers in Docking Benchmark 5.5, ranked top-1 by pTM (54 / 44 / 23% taking the best of five). Rigid-body docking on the same set reached 13 / 9 / 1%. **Read that with its caveat, which the authors state:** this set predates the training cutoff. The interfaces were not trained on, but the constituent chains very likely were, so these are not blind-prediction figures.

**At CASP15, blind, on assemblies.** The organizers ran stock public AlphaFold-Multimer v2.2 at default settings as a baseline group. It produced an acceptable model for **54% of the 41 assembly targets**, where the union over all 87 predictor groups reached **90% (37 of 41)** — up from 31% at CASP14. ⚠︎ That assessment's "acceptable" means **ICS, IPS, TM-score and lDDT all above 0.50**, not a DockQ class, so the 54% is not the same kind of quantity as the 70% above and the two must not be tabled together.

**Antibody–antigen is the measured floor.** AlphaFold v2.2.0's multimer model, on 427 non-redundant antibody–antigen complexes at ≤ 3.0 Å released after its own training cutoff, reached **26% acceptable-or-better, 18% medium-or-better and 5% high** as the top-ranked prediction — rising only to 37 / 22 / 6% when all 25 predictions per complex were considered. On a 41-case subset released after v2.3's cutoff, v2.3 reached **36% medium-or-better against v2.2's 23%**. At CASP15 nobody submitted a correct model for two of the three blind antibody–antigen targets, and the baseline Multimer run reached medium quality on the one target where a template with the correct binding mode existed. The Multimer authors name the limitation themselves: "we observe anecdotally that AlphaFold-Multimer is generally not able to predict binding of antibodies." The consistently given mechanism is that antibody–antigen interfaces carry no co-evolutionary signal between the partners, so an MSA-driven method has nothing to exploit.

For a complex, use a co-folding model — [Boltz-2](/docs/boltz2/translatability) or [Chai-1](/docs/chai/translatability) — rather than any monomer fetched here.

> Evans R, O'Neill M, Pritzel A, Antropova N, Senior AW, Green T, et al. Protein complex prediction with AlphaFold-Multimer. *bioRxiv*. 2022:2021.10.04.463034. doi: [10.1101/2021.10.04.463034](https://doi.org/10.1101/2021.10.04.463034). Preprint: not peer reviewed.
>
> Yin R, Feng BY, Varshney A, Pierce BG. Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants. *Protein Science*. 2022;31(8):e4379. doi: [10.1002/pro.4379](https://doi.org/10.1002/pro.4379)
>
> Yin R, Pierce BG. Evaluation of AlphaFold antibody-antigen modeling with implications for improving predictive accuracy. *Protein Science*. 2024;33(1):e4865. doi: [10.1002/pro.4865](https://doi.org/10.1002/pro.4865)
>
> Ozden B, Kryshtafovych A, Karaca E. The impact of AI-based modeling on the accuracy of protein assembly prediction: insights from CASP15. *Proteins*. 2023;91(12):1636-1657. doi: [10.1002/prot.26598](https://doi.org/10.1002/prot.26598)
>
> Lensink MF, Brysbaert G, Raouraoua N, Bates PA, Giulini M, Honorato RV, et al. Impact of AlphaFold on structure prediction of protein complexes: the CASP15-CAPRI experiment. *Proteins*. 2023;91(12):1658-1683. doi: [10.1002/prot.26609](https://doi.org/10.1002/prot.26609)

**Confidence calibration, head to head.** On the 1,337-chain set, pLDDT's rank correlation with agreement to the experimental structure was **strongest for ESMFold and OmegaFold**, and the authors concluded their "confidence scores more accurately reflect prediction quality than those of AlphaFold2." That is not the same as ESMFold being more accurate — it is less accurate on every metric in the same table — and it cuts the same way as the 75%-of-wrong-models-above-pLDDT-70 figure above. **AlphaFold 2 is more accurate and somewhat more overconfident when it is wrong.** CASP15's assessors, grading by a different method on a different set, placed the AlphaFold 2 confidence estimates among the most accurate available. Both readings are published and they are measuring different things.

## In silico vs. wet lab

The standard wet-lab approach for the question AlphaFold answers is **experimental structure determination** — X-ray crystallography, cryo-EM, or solution NMR. These produce the structure against which every prediction on this page is graded.

Agreement has settled definitions. Global agreement is reported as **GDT-TS** (percentage of residues within distance cutoffs after superposition; ≥ 45 is taken as correct topology, and CASP treats ≥ 90 as roughly the difference between two crystal forms of the same protein) or **TM-score** (0–1; above 0.5 is mostly the same fold). Local agreement is **lDDT**, a superposition-free per-residue distance-difference score. Positional error is **RMSD**, or **RMSD<sub>95</sub>** — the Cα RMSD at 95% residue coverage, which discards the worst-fitting 5% and so is insensitive to one badly placed tail. For complexes the metric is **DockQ**. CASP14's assessors published the mapping between the two scales most often confused: **GDT_TS 90 corresponds to approximately 1.5 Å RMSD, and GDT_TS 80 to about 2.5 Å**.

| | Time to answer | Cost basis | Accuracy, and its metric |
|---|---|---|---|
| **Standard wet lab** — X-ray crystallography, cryo-EM, or NMR | Weeks to months; may never crystallize or resolve | €2,000 data collection + €1,000/day refinement at a core facility's commercial tier; cryo-EM has no public price at all. [cost basis](/docs/translatability#what-the-wet-lab-cost-figures-mean) | Ground truth, limited by resolution and by what crystallizes |
| **Conventional / other in silico** — a fresh prediction run: [ESMFold](/docs/esmfold), RoseTTAFold, AlphaFold 3 | ESMFold 10–60 s for 50–200 residues on an A100, no alignment step. A local AlphaFold 2 run is **0.5–2 h per chain on an A100**, over 3–5 h above 1,000 residues, **plus** the MSA search. AlphaFold 3 server ~1–5 min | No per-run price is published for this class. [cost basis](/docs/translatability#what-the-wet-lab-cost-figures-mean) | ESMFold median **TM-score 0.95**, **RMSD 1.74 Å**, **GDT-TS 90%** on 1,337 recent PDB chains. On 1,666 template-poor monomers, **AlphaFold 3** reached 1,076 accurate / 388 domain-movement / 191 incorrect, essentially matching AlphaFold 2 |
| **Combined tooling here** — AlphaFold fetch (database lookup), optionally → [P2Rank](/docs/p2rank) → [DiffDock](/docs/diffdock) | **2–5 s, no GPU** — the inference hours were already spent by DeepMind and the result cached. Nothing to queue, nothing to pay a GPU for | No GPU job — the fetch is an HTTP lookup, so it is not billed as a compute run | AlphaFold 2: median backbone **0.96 Å RMSD<sub>95</sub>** on 87 CASP14 domains and **1.46 Å** on 3,144 recent PDB chains, both author-reported; median **TM-score 0.96 / RMSD 1.30 Å / GDT-TS 94%** on 1,337 recent chains, independent. **Only one published study graded an actual database download**, and got **0.30 Å** against a newly solved crystal structure (n = 1) |

**Read the accuracy column by cell, never down the column.** Experimental determination's "accuracy" is definitional — it is the thing the other two rows are graded against. The two prediction rows are only partly comparable: the 1,337-chain and 1,666-monomer figures *are* head-to-head on shared sets with shared metrics, and those comparisons are the point of the table. The CASP14 figures are not comparable to anything else here.

On that reading, the honest position for the AlphaFold fetch is: **seconds instead of months, and seconds instead of the alignment-plus-inference hours a fresh AlphaFold 2 run costs, for the output of the same model.** Mostly you are not trading accuracy — you are collecting compute someone else already paid for.

But "the same model" is not quite "the same run", and the difference cuts one way. The database's models came from a pipeline tuned for proteome scale: a reduced metagenomics search, reduced ensembling, and — for entries outside Swiss-Prot and the reference proteomes — **a single model run rather than the best of five**. A fresh full-pipeline AlphaFold 2 run with a hand-built MSA can therefore be *better* than the database entry for the same sequence, and the CASP15 result below is the measurement of how much better a deeper MSA makes a stock run. What you also give up is control: *one* model of *one* chain, at a version you do not choose, with no ability to vary the alignment, and only if the protein is in the database at all.

Three asymmetries matter beyond the table. Experimental determination can fail outright for a given protein and no effort fixes it, which is most of why structure prediction exists. An experiment reports what *is*, while AlphaFold reports what is *likely* — pLDDT 95 and a solved structure are not interchangeable claims. And a crystal structure is one conformation captured under one set of conditions, which is why a model-versus-crystal disagreement is not automatically the model's error: CASP14's assessors attributed part of the residual gap to experimental data quality and to crystal lattice contacts, examining seven targets whose best models differed from experiment in small regions at GDT-TS 72–93.

## Agreement with wet lab

All figures below are against experimentally determined structures, with the metric named each time.

### The authors' own benchmarks

**On CASP14** (n = 87 protein domains, median with a 95% confidence interval bootstrapped over 10,000 samples, against the top-15 of 146 entries), author-reported:

- Median backbone accuracy **0.96 Å RMSD<sub>95</sub>** (95% CI 0.85–1.16 Å). The next best performing method: **2.8 Å** (95% CI 2.7–4.0 Å).
- All-atom accuracy **1.5 Å RMSD<sub>95</sub>** (95% CI 1.2–1.6 Å) against the best alternative method's 3.5 Å (95% CI 3.1–4.2 Å).

**On a held-out set of recent PDB chains** — every structure deposited after the training cutoff, filtered to remove any with a pre-cutoff template above 40% identity covering more than 1% of the chain (n = 3,144 chains): median backbone **RMSD<sub>95</sub> 1.46 Å** (95% CI 1.40–1.56 Å). The authors note this measure is highly sensitive to domain packing, so a high value is expected for some chains with packing uncertainty rather than fold error.

**The independent CASP14 assessment reaches the same place by a different metric.** CASP's own assessors reported that the best-model trend line for CASP14 "starts at a GDT_TS of about 95, and finishes at about 85 for difficult targets," that **about two-thirds of the 96 targets** exceeded the GDT_TS ≈ 90 ceiling previous CASPs had implied and "are considered competitive with experiment in backbone accuracy," and that **for only four targets did another group obtain a higher GDT_TS** than AlphaFold 2. In the free-modelling category, where no homologous structure is available, **about 30% of targets reached GDT_TS above 90 and 75% reached 80 or greater**.

> Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, et al. Highly accurate protein structure prediction with AlphaFold. *Nature*. 2021;596(7873):583-589. doi: [10.1038/s41586-021-03819-2](https://doi.org/10.1038/s41586-021-03819-2)
>
> Kryshtafovych A, Schwede T, Topf M, Fidelis K, Moult J. Critical assessment of methods of protein structure prediction (CASP)-Round XIV. *Proteins*. 2021;89(12):1607-1617. doi: [10.1002/prot.26237](https://doi.org/10.1002/prot.26237)

### Independent benchmarks after the training cutoff

**On 1,337 PDB chains deposited July 2022 – July 2024** — after the training cutoff of every tested method, 20–400 residues, filtered to at most 70% pairwise identity within the monomer and de novo groups, comprising 980 monomers, 245 small complexes and 102 de novo designs — medians were:

| Metric (median) | AlphaFold 2 | ESMFold | OmegaFold |
|---|---|---|---|
| TM-score | **0.96** | 0.95 | 0.93 |
| RMSD | **1.30 Å** | 1.74 Å | 1.98 Å |
| GDT-TS | **94%** | 90% | 89% |
| pLDDT | 92.65 | 87.40 | 89.00 |

All four differed significantly across tools (Kruskal–Wallis, post-hoc Dunn with Bonferroni). Split by experimental method, AlphaFold 2's median RMSD was **1.24 Å against X-ray structures and 2.31 Å against NMR** structures. The AlphaFold 2 in this study was **v2.1.1 with the monomer model and reduced database settings**, MSAs from UniRef90, BFD and MGnify, five models generated and the highest-confidence one (`ranked_0`) evaluated.

**On 1,666 monomers deposited 2022–2024 and filtered to lack close pre-2022 homologs** (sequence identity < 40% and query coverage < 70% against anything released before 1 January 2022) — a stricter test. Monomers were classified by RMSD and lDDT jointly: accurate = RMSD ≤ 3.0 Å and lDDT ≥ 0.7; domain movement = RMSD > 3.0 Å with lDDT ≥ 0.7; incorrect = both failing.

| Monomers (n = 1,666) | Accurate | Domain movement | Incorrect | Special cases |
|---|---|---|---|---|
| **AlphaFold 2** | **1,069 (64.2%)** | **386 (23.2%)** | **200 (12.0%)** | 11 |
| AlphaFold 3 | 1,076 | 388 | 191 | 11 |
| ESMFold | 832 | 430 | 391 | 13 |

Counting accurate plus domain movement as correct, that is **1,455 of 1,666 = 87.3% for AlphaFold 2**. Restricted to X-ray and cryo-EM targets only — the fairer comparison if your reference is a crystal or a map — Table 1's per-method breakdown gives 1,037 accurate and 365 domain-movement of 1,475, or **95.1% correct**. Most of AlphaFold 2's outright failures were NMR targets: of the 200 incorrect models, **128 were NMR**, 47 X-ray and 25 cryo-EM. This study used **AlphaFold 2 v2.2.3** installed locally, five structures per sequence, top-ranked model (`ranked_0`) assessed.

The same table lets you read the effect of template availability, which matters because the set was *already* filtered to lack good templates. Grouping by the study's template-quality score (sequence identity × query coverage), AlphaFold 2 was accurate on 399 of the 652 entries with no template at all (61.2%, or 77.6% counting domain movement) against 637 of 936 for entries with a weak template (68.1%, or 94.2%). **Removing templates does not break AlphaFold 2, but it measurably costs it.**

### The one study that graded an actual database download

Every benchmark above re-runs AlphaFold 2 locally with a full MSA pipeline. Only one published study surveyed here graded the artifact the AlphaFold fetch actually returns — **a model downloaded from the AlphaFold Protein Structure Database** — against a structure that did not exist when the model was made.

A GH11 xylanase from *Hypocrea virens* was solved and then compared against predictions for the same sequence. The database's AlphaFold 2 model (mean pLDDT 95.25) superimposed on the experimental structure at **RMSD 0.30 Å**, against 0.30–0.42 Å for five AlphaFold 3 server models, 0.77–0.88 Å for RoseTTAFold and 0.87 Å for ESMFold. The N-terminal region Ala20–Arg30 scored a mean pLDDT of 58.48 and was disordered in the electron density map — the model's low-confidence call and the experiment's missing density agreed.

Then the useful part. Docking xylohexaose into the database model reproduced the experimental structure's result exactly: **pose 1 occupied the same five subsites (+2 to −3) at the same −12.5 kcal/mol** as the docking run against the experimental structure. ESMFold's model, by contrast, put the ligand across all six subsites at −12.1 kcal/mol — the same cleft, a different binding mode.

**And yet the pocket geometry was not identical.** The database model's substrate-binding cavity volume was **812 Å<sup>3</sup> against the experimental structure's 909 Å<sup>3</sup>, about 10% smaller**, with a narrower cleft and a domain angle of 46.3° against 44.4°. A 0.30 Å global RMSD did not guarantee the pocket. This is one case and it cannot carry a rate, but it is the clearest published illustration that **a near-perfect global superposition and a correct pocket volume are different claims.**

> Hýsková A, Maršálková E, Šimeček P. Balancing speed and precision in protein folding: a comparison of AlphaFold2, ESMFold, and OmegaFold. *Frontiers in Genetics*. 2025;16:1715037. doi: [10.3389/fgene.2025.1715037](https://doi.org/10.3389/fgene.2025.1715037)
>
> Mahtha SK, Venkadesan S, Mohanty D. Comparative evaluation of the prediction accuracy of AlphaFold and ESMFold for monomeric and dimeric proteins. *NAR Genomics and Bioinformatics*. 2026;8(1):lqag002. doi: [10.1093/nargab/lqag002](https://doi.org/10.1093/nargab/lqag002)
>
> Nam KH. Evaluation of AI-predicted GH11 xylanase models against a previously unreported experimental structure: implications for conformational accuracy and ligand binding. *International Journal of Molecular Sciences*. 2026;27(3):1370. doi: [10.3390/ijms27031370](https://doi.org/10.3390/ijms27031370)

### Molecular replacement: the one wet-lab workflow that consumes these models

Crystallographic phasing by molecular replacement needs a search model close enough to the true structure for Phaser to place it. This is the only published use where a predicted structure is graded by a **crystallographic** outcome — map correlation, log-likelihood gain, R<sub>free</sub> — rather than by similarity to an answer you already had. The evidence is strong and it is all AlphaFold 2.

| Success rate | Criterion | n / test set | Processing |
|---|---|---|---|
| **87% (187/215)** | map CC ≥ 0.5 vs deposited | 215 SAD-phased PDB entries, Dec 2021 – Jun 2022 | AlphaFold 2, iterative `PredictAndBuild`, sequence + diffraction data only |
| **86% (349/406)** | map CC ≥ 0.25 | 406 SAD-phased PDB entries, Jul 2022 – Oct 2023 | **DeepMind AlphaFold 2 v2.3.2 alone**, unmodified or pLDDT < 70 trimmed, pLDDT → pseudo-B |
| **87.7% (356/406)** | map CC ≥ 0.25 | same 406 | AlphaFold 2 **∪ ColabFold 1.5.1**, first pass |
| **93.5% (201/215)** | map–model CC > 0.5 | the 215 set above | AlphaFold 2 via ColabFold, revised AlphaFold-guided MR script |
| **92.4% (146/158)** | map–model CC > 0.5 | 158 entries **selected for Cα RMSD ≥ 5 Å** between the default prediction and the deposited structure | same |
| **91% (31/34)** | σ<sub>A</sub> map CC > 0.3 | 34 CASP14 crystal structures | the five submitted CASP14 models as an ensemble, trimmed on predicted per-residue error |

Four things about this table are load-bearing.

**The selection rules differ enormously and the rates are not interchangeable.** The two 406-entry rows are **SAD-phased** structures — deliberately enriched for cases where MR was not the method of choice. In the same deposition window 10,084 structures were deposited as MR-solved, so the 3% that remained unsolved amount to **0.1% of the 10,492 total depositions**. The 92.4% row is the opposite kind of bias: that set was filtered *to* cases where the default AlphaFold prediction is more than 5 Å from the deposited structure, so it measures recovery from a wrong starting conformation, not a PDB-wide rate.

**The control makes the case.** On the same 406 SAD-phased entries, search models drawn only from pre-existing PDB structures solved **99 by MrParse/HHsearch and 87 by GESAMT — 145 in union, 35.6%**. The other 261 were not tractable with classical MR at all. That is the measurement of what AlphaFold 2 added, and it is the largest wet-lab effect any figure on this page reports.

**pLDDT is the field that makes it work, and it has to be converted.** pLDDT runs the opposite way to a B-factor: left unconverted, the *most* confident residues get the *least* weight in the MR calculation. The standard treatment is to rescale pLDDT to a coordinate-error estimate (pLDDT 0.7 ≈ 1.5 Å RMSD ≈ B ≈ 60 Å<sup>2</sup>; pLDDT 1.0 ≈ 0.45 Å) and to delete residues below pLDDT 70 outright, because overweighted low-confidence residues clash with crystal symmetry mates and degrade the log-likelihood gain. In the one published cutoff sweep, on a single target, not trimming **failed outright** while trimming at pLDDT 70 found both copies in the asymmetric unit (LLG 54/127, TFZ 7.7/11.2). The same sweep found that a *lower* cutoff of 40–55 built better models downstream even though 70 was best for the search itself. Correlationally, successful MR cases averaged **90% of residues at pLDDT ≥ 70 against 50% in the 28 failures**.

**At CASP15, MR success was not reported as a rate and should not be quoted as one.** The assessors ran 17 targets with diffraction data and published per-target log-likelihood-gain heatmaps, not percentages. What they do state: with search models trimmed below pLDDT 70 and pLDDT converted to B-factors, **no group produced a successful search for two of the 17 targets, and for the remaining 15, many groups produced models that could be successfully placed**. Splitting models into domains rescued one further target. The best-scoring groups were those using AlphaFold 2.

> McCoy AJ, Sammito MD, Read RJ. Implications of AlphaFold2 for crystallographic phasing by molecular replacement. *Acta Crystallographica Section D*. 2022;78(Pt 1):1-13. doi: [10.1107/S2059798321012122](https://doi.org/10.1107/S2059798321012122)
>
> Oeffner RD, Croll TI, Millán C, Poon BK, Schlicksup CJ, Read RJ, Terwilliger TC. Putting AlphaFold models to work with phenix.process_predicted_model and ISOLDE. *Acta Crystallographica Section D*. 2022;78(Pt 11):1303-1314. doi: [10.1107/S2059798322010026](https://doi.org/10.1107/S2059798322010026)
>
> Terwilliger TC, Afonine PV, Liebschner D, Croll TI, McCoy AJ, Oeffner RD, et al. Accelerating crystal structure determination with iterative AlphaFold prediction. *Acta Crystallographica Section D*. 2023;79(Pt 3):234-244. doi: [10.1107/S205979832300102X](https://doi.org/10.1107/S205979832300102X)
>
> Keegan RM, Simpkin AJ, Rigden DJ. The success rate of processed predicted models in molecular replacement: implications for experimental phasing in the AlphaFold era. *Acta Crystallographica Section D*. 2024;80(Pt 11):766-779. doi: [10.1107/S2059798324009380](https://doi.org/10.1107/S2059798324009380)
>
> Wang W, Gong Z, Hendrickson WA. AlphaFold-guided molecular replacement for solving challenging crystal structures. *Acta Crystallographica Section D*. 2025;81(Pt 1):4-21. doi: [10.1107/S2059798324011999](https://doi.org/10.1107/S2059798324011999)
>
> Millán C, Keegan RM, Pereira J, Sammito MD, Simpkin AJ, McCoy AJ, et al. Assessing the utility of CASP14 models for molecular replacement. *Proteins*. 2021;89(12):1752-1769. doi: [10.1002/prot.26214](https://doi.org/10.1002/prot.26214)

### Reading pLDDT and PAE

These are the two fields you hold, so they deserve their own accounting.

**Mind the scale and the provenance.** pLDDT is per-residue, on a **0–100** scale, stored in the B-factor column of the PDB file and under `_ma_qa_metric_local` in the mmCIF. The four bands the 3D viewer colours by are the AlphaFold Database's own: **very high (> 90), confident (70–90), low (50–70), very low (< 50)**, plus a "no score" grey for atoms without one. Unlike ESMFold — which inherited AlphaFold 2's prediction head and the band convention with it — **these bands originate with this model and were measured on it.** Here is what each one actually rests on:

- **pLDDT > 90.** Taken as the high-accuracy cutoff because above it "AlphaFold χ<sup>1</sup> rotamers are 80% correct" on a recent-PDB test set — restricted to structures at resolution < 2.5 Å (n = 5,983 chains) and residues with B-factor < 30 Å<sup>2</sup> (n = 609,623 residues), with a χ<sup>1</sup> angle counted correct within 40°. The database's guidance is that this band is "suitable for accuracy-dependent uses such as characterising binding sites."
- **pLDDT > 70.** Stated as corresponding to "a generally correct backbone prediction," with the per-band lDDT-Cα distribution given in an extended-data table over 10,215 recent-PDB chains at resolution < 3.5 Å.
- **pLDDT < 50.** "Long regions with pLDDT < 50 adopt a readily identifiable ribbon-like appearance, and should not be interpreted as structures but rather as a prediction of disorder." This is the one band with a real benchmark behind it, and also the one where the benchmark has since been qualified. Evaluated as a disorder predictor on the CAID DisProt-PDB set (n = 178,124 residues), AlphaFold 2's pLDDT reached **AUC 0.897**, competitive with the then state of the art, and the model's separate experimentally-resolved head reached 0.921. An independent replication on 475 CAID targets got **AUC 0.905** on the same DisProt-PDB reference — **but only 0.731 on the harder full DisProt reference, where a dedicated sequence predictor reached 0.794.** DisProt-PDB masks every residue without experimental evidence either way, so it is the easier of the two. By CAID's third round, solvent accessibility derived from the AlphaFold model ranked third of all entries at **AUC 0.950** while **pLDDT itself ranked 11th**. Read a low-pLDDT region as "no model here," which is always true. Reading it as a *disorder call* is supported on the evidence-restricted benchmark and measurably weaker on the full one.
- **Underlying fit.** Per-residue pLDDT against lDDT-Cα gives **Pearson r = 0.73** (n = 10,215 chains, 2,756,569 residues, resolution < 3.5 Å), with the least-squares fit y = (0.967 ± 0.001)x + (1.9 ± 0.1). At the chain level the fit is **lDDT-Cα = 0.997 × pLDDT − 1.17, Pearson r = 0.76** (n = 10,795 chains).

**pLDDT ≥ 70 is a sensitive filter and a weak guarantee.** On the 1,666-monomer set, **1,067 of 1,069 accurate AlphaFold 2 models (99.8%)** carried mean pLDDT ≥ 70, so almost nothing accurate is below the threshold. But **150 of the 200 incorrect models (75%)** were also above it. Read the per-residue plot over the residues your question depends on, not the mean.

**PAE is the field that answers a question pLDDT cannot.** PAE is a pairwise matrix of predicted error in the *relative* position of two residues, and it is the only output that speaks to whether two domains are placed correctly with respect to each other. The database ships it as a PNG and as a JSON of all residue pairs; the AlphaFold fetch returns both URLs under `qualityMetrics`. A block-diagonal PAE with high off-diagonal error means "each domain is fine, their relative orientation is not" — which is precisely the 23.2% "domain movement" class in the benchmark above. **Two high-pLDDT domains plus a bad inter-domain PAE block is a normal, correct reading of a model, and pLDDT alone will not show it to you.**

**CASP15's assessors graded AlphaFold's confidence estimates and named this database.** Ranking all 118 evaluated groups by median accuracy-self-estimate Z-score, they found that "the confidence estimates produced by the control ColabFold and DeepMind AF2 entries are among the most accurate currently available, which should be seen as reassuring for users of the ColabFold pages and the AlphaFold Protein Structure DataBase, respectively." That is an independent endorsement of this specific artifact's confidence field — the strongest one on this page — and it is a statement about *relative* calibration among available methods, not a claim that a given pLDDT value implies a given accuracy.

> Tunyasuvunakool K, Adler J, Wu Z, Green T, Zielinski M, Žídek A, et al. Highly accurate protein structure prediction for the human proteome. *Nature*. 2021;596(7873):590-596. doi: [10.1038/s41586-021-03828-1](https://doi.org/10.1038/s41586-021-03828-1)
>
> Varadi M, Bertoni D, Magana P, Paramval U, Pidruchna I, Radhakrishnan M, et al. AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences. *Nucleic Acids Research*. 2024;52(D1):D368-D375. doi: [10.1093/nar/gkad1011](https://doi.org/10.1093/nar/gkad1011)
>
> Simpkin AJ, Mesdaghi S, Sánchez Rodríguez F, Elliott L, Murphy DL, Kryshtafovych A, Keegan RM, Rigden DJ. Tertiary structure assessment at CASP15. *Proteins*. 2023;91(12):1616-1635. doi: [10.1002/prot.26593](https://doi.org/10.1002/prot.26593)
>
> 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*. 2013;29(21):2722-2728. doi: [10.1093/bioinformatics/btt473](https://doi.org/10.1093/bioinformatics/btt473)

## What the figure doesn't cover

- **Which AlphaFold you have is not which AlphaFold was benchmarked.** Every independent figure on this page comes from a *fresh local run*: v2.1.1 with reduced databases, v2.2.3 with five models, v2.3.2 for the phasing work, ColabFold for others. The database's models were produced by a different pipeline, and nobody has benchmarked the database as shipped at scale. The one study that graded a real download got an excellent result on one protein. Treat the benchmark figures as informative about the model and not as measurements of the artifact.
- **The database's own pipeline was deliberately reduced, and not every entry got the same treatment.** The original human-proteome release replaced the full BFD metagenomics search with a "Reduced BFD" and cut ensembling by a factor of eight, to keep a proteome-scale run tractable. More consequentially today: **Swiss-Prot and reference-proteome entries are the most confident of five models trained with different random seeds, while bulk UniProt entries are "outputs of a single model run."** Model ranking by mean pLDDT is a real quality step, and for most of the database's 214 million entries it did not happen. Nothing in the downloaded file tells you which kind you have.
- **The version you fetch can change under you, and one version was wrong.** The database has shipped six model versions, deletes the old file keys on a bump, and does not expose a version parameter through this tool. A November 2022 update also rectified structures hit by a temporary numerical bug in that July's release, which "led to low accuracy predictions with correspondingly low pLDDT for ∼4% of the total structure predictions in the database." Record `modelVersion` from the result with any figure you derive, and do not compare a model you fetched today against a figure from a paper that used v3 or v4.
- **Training overlap is the first thing to check, and the benchmarks here were built to avoid it — which is why they are lower than the headlines.** The 1,337-chain and 1,666-monomer sets were filtered by deposition date and by template identity precisely to exclude lookups. The 87.3% and the 0.96 TM-score are what AlphaFold 2 does on novel-ish targets; a figure from an unfiltered set describes something else.
- **Shallow MSAs are the measured failure mode, and no amount of confidence fixes it.** The authors state plainly that "accuracy decreases substantially when the median alignment depth is less than around 30 sequences," with a threshold effect above which more depth buys little — improvements beyond roughly 100 sequences "lead to small gains." Removing both metagenomics databases from the MSA search cost 6.1 GDT on average at CASP14, with a few targets losing 20+ GDT. CASP15's assessors found the same thing from the outside: "the most obvious characteristic of difficult targets is the availability of only a shallow MSA," and 17 of the 18 worst-predicted evaluation units were free-modelling targets. The dependence shows up from the outside too: on the 1,337-chain benchmark, **AlphaFold 2 was significantly less accurate for proteins lacking a Pfam family annotation (odds ratio 0.67)** while ESMFold and OmegaFold held steady regardless — ESMFold not degrading, which is not the same as ESMFold winning, since it starts lower. **If your protein has no homologs, the database entry for it is likely to be low-confidence, and that is the model telling you the truth.**
- **Heterotypic contacts are the authors' own stated second limitation.** AlphaFold "is much weaker for proteins that have few intra-chain or homotypic contacts compared to the number of heterotypic contacts," which "typically occurs for bridging domains within larger complexes in which the shape of the protein is created almost entirely by interactions with other chains." The human-proteome paper measured the same effect: a high ratio of heterotypic contacts is associated with lower accuracy. A database entry is a single chain modelled alone, so a protein whose shape is made by its partners will be modelled badly and there is no flag for it.
- **Multi-domain orientation is the measured bottleneck, and PAE flags it without a published hit rate.** 23.2% of AlphaFold 2's models on the 1,666-monomer set were locally correct (lDDT ≥ 0.7) but globally displaced (RMSD > 3.0 Å) — roughly one model in four. A dedicated benchmark of 219 non-redundant multi-domain proteins puts AlphaFold 2 at full-chain **TM-score 0.900 and RMSD 3.58 Å**, and its authors conclude that "the main bottleneck for AlphaFold2 structure prediction on multi-domain proteins is capturing the correct inter-domain orientations." The worked cases are stark: soluble epoxide hydrolase's two domains came out individually at **TM-score 0.990 and 0.995** while the full chain scored **0.689 at 6.13 Å**. On 164 human AlphaFold DB entries *selected* for full-chain TM-score below 0.8, the mean was 0.610 — a set biased by construction, but drawn from the database itself. DeepMind's own figure, on 151 long recent-PDB chains with no close template, is that **70% of predictions reached TM-score above 0.7**, with their own caveats attached: CASP14's inter-domain assessment "was based on a small target set," and "we expect the inter-domain accuracy of AlphaFold to be lower than its within-domain accuracy." **PAE does flag these cases** — published examples show a low-error block per domain with uniformly high error across the inter-domain region, and across 128 autoinhibited proteins the across-domain PAE is significantly higher than the within-domain PAE. But **no study publishes PAE's sensitivity, specificity or AUC for catching a mis-oriented pair.** The nearest calibration figure is pTM against realised TM-score at Pearson r = 0.84, which is global rather than per-domain-pair. Read the PAE block; just do not treat a clean one as a guarantee.
- **Point mutations: do not use this for variant effects, and the model's authors say so.** DeepMind's own documentation, quoted in the literature, states that AlphaFold "has not been validated for predicting the effect of mutations" and "is not expected to produce an unfolded protein structure given a sequence containing a destabilising point mutation." The measurements bear that out. On four mutants across three domains, each experimentally known to destroy folding, the predicted mutant differed from the predicted wild type by only **0.1–0.6 Å mean Cα RMSD** — less than the 0.7 Å between two wild-type copies of the same domain — and mean pLDDT barely moved (84 vs 84; 95 vs 94), with one mutant scoring *higher* than its wild type. At scale, ΔpLDDT at the mutated residue correlates with experimental ΔΔG at **Pearson r = −0.17** over 1,154 single mutations in 73 proteins, and whole-chain ΔpLDDT at **r = 0.02**. A separate line of work does extract signal from the *coordinates* rather than the confidence — a local "effective strain" measure reaches Spearman **ρ = 0.36** against ΔΔG pooled over 2,499 mutants and about 0.54 within a single protein — but that work reports no consistent correlation with RMSD and found that selecting the top-pLDDT model made results *worse*. **The confidence score carries essentially no mutation-effect signal.**
- **Induced fit and alternative conformations: you get one state, and nothing tells you which.** On 86 curated apo–holo pairs (mean apo–holo RMSD 3.5 Å), the top-confidence model was closest to the **holo** form for **69.8%** of proteins and the apo form for 30.2%, with the unfavoured conformer always 1.5–1.8 Å worse — and pLDDT medians for the two groups were 95.47 and 94.53, so confidence does not tell you which state you received. On 93 fold-switching proteins with both folds solved, **94% of predictions sat closer to one fold than the other**, and AlphaFold 2 failed to produce both folds within its top five models for **75 of 93 (81%)**. A follow-up pooling over 282,000 predictions from every AlphaFold sampling variant plus AlphaFold 3 captured both conformations for 35% (32/92) of fold switchers that were likely in the training set — and for **1 of 7** confirmed after it. For autoinhibited, allosterically switching proteins, over 75% of individual domains came within 3 Å while **about half the inhibitory modules were mis-placed** relative to their functional domain. ⚠︎ Fold-switch accuracy must be scored over the switching region, not the whole chain; a whole-chain TM-score can clear 0.8 against *both* conformers simply because the switching region is short.
- **pLDDT is confident about things that are not proteins, and more so than ESMFold.** On 644 spurious AntiFam sequences (average length 120), AlphaFold 2 weights run through ColabFold 2.3.6 gave an average mean pLDDT of **57.36**, rising to **71.18 on sequences ≤ 50 residues** — a 24% increase over its own mean, purely from shortness, and across the "confident" threshold. AlphaFold 3 inflates further (65.15 and 80.74); ESMFold inflates less in absolute terms (39.82 and 56.3) while rising more in relative terms. On the matched set of 644 real Swiss-Prot proteins, AlphaFold 2 reached confident mean pLDDT on 90.8%. Read both with their sample: those 644 sequences were binned across 1–200 residues to match the spurious set, so they are short proteins, not a representative proteome. **A short sequence buys pLDDT it has not earned.**
- **A near-perfect global superposition does not give you the pocket.** The GH11 database model matched the crystal structure at 0.30 Å and still had a binding cavity 10% smaller, a narrower cleft and a 2° domain-angle difference. If your question is a cavity volume, a rotamer, or an active-site distance, the global RMSD figures overstate what you have — and the one published check of this kind found the error in exactly the direction that matters for docking.
- **A predicted receptor costs pose accuracy, and the largest measurement was made on this exact artifact.** On 2,474 PDBBind refined-set complexes re-docked with AutoDock-GPU, top-1 pose accuracy (RMSD < 2 Å) fell from **41% against the co-crystallised receptor to 17% against the AlphaFold model** — and those receptors were downloaded from the AlphaFold Database by UniProt ID, which is exactly what the AlphaFold fetch returns. The failure is not a mild degradation: the fraction of top poses landing **more than 5 Å** from the right place rose from **34% to 60%**. Two independent studies on narrower sets agree on direction and magnitude — **44% → 15%** on GPCR cross-docking (54 structures of 18 receptors, self-docking excluded) and **67% → 46%** on 24 kinases. Budget for roughly a halving before treating a pose from a fetched receptor as a binding mode, and read [DiffDock's translatability page](/docs/diffdock/translatability) before ranking anything.
- **⊘ Nothing you can compute from the model tells you whether it will dock well.** Two of those studies independently report that **pLDDT does not predict docking success**. The GPCR study goes further and closes the loophole: AlphaFold 2's *binding-pocket* geometry was excellent — median pocket RMSD **1.3 Å**, which the authors describe as "typically nearly as low as the RMSD between two experimentally determined structures of the same protein with different ligands bound," against 3.3 Å for traditional homology models — and pose accuracy still did not follow. At *matched* pocket RMSD the AlphaFold receptor still underperformed the experimental one, significantly so below 1.1 Å. **High pLDDT, low pocket RMSD and a good docking result are three separate things.**
- **Retrospective enrichment halves as well, and every one of those comparisons is inflated.** Across DUD-E-style screens, EF1% fell from **24.2 on holo crystal receptors to 13.0** on AlphaFold models over 27 targets, from **20.5/24.1 to 8.8/8.9** under two consensus scores over 21 targets, and from a median of **4.1 to 1.6** over 10 GPCRs. ⚠︎ But these benchmarks *selected* their experimental structures for high ligand enrichment in the first place, as one of those papers warns directly, so the holo-versus-model gap is an overestimate of what you would see prospectively. Induced-fit refinement recovers much of it (13.0 → 18.9 against holo's 24.2).
- **The fair comparator matters: AlphaFold beats an *apo* crystal structure.** On the overlapping unliganded subset, top-1 pose accuracy was **16% for the AlphaFold model against 10% for the apo crystal**, and enrichment ran 13.0 against 11.4. AlphaFold loses badly to a liganded structure and wins against an unliganded one. If your real alternative is an apo crystal, the model is the better receptor — which is the opposite of what the 41% → 17% headline suggests on its own.
- **Two published null results exist and they are not good news.** One study found docking into AlphaFold models statistically indistinguishable from docking into experimental structures (mean auROC 0.48 vs 0.46) against measured enzyme inhibition, and another found no significant difference across eight docking protocols on 16 protein–protein interfaces. In both cases absolute performance sat at or near random, so the instrument had almost no dynamic range — and **both sets of authors say so themselves**, attributing the weak performance to the docking method rather than the receptor. A null from an instrument that cannot resolve the effect is not evidence that the effect is absent.
- **A complex figure is not a monomer figure, and a selection rule changes a number more than a method does.** Nothing on this page about Multimer, dimers or antibodies describes a fetched monomer. And across the complex literature the same system reports wildly different rates depending on whether the reported model was picked by the method's own confidence, by best-of-N, or by comparing against the answer. When you meet a success rate anywhere, find the selection rule before you compare it to another one.
- **No one has benchmarked the AlphaMissense annotations the database ships.** They appear in the entry metadata and this platform does not surface them. They are a separate model and nothing on this page describes their accuracy.

> Orr AK, Bateman A. Folding the unfoldable 2: using AlphaFold and ESMFold to explore spurious proteins. *Bioinformatics Advances*. 2026;6(1):vbag160. doi: [10.1093/bioadv/vbag160](https://doi.org/10.1093/bioadv/vbag160)
>
> Porta-Pardo E, Ruiz-Serra V, Valentini S, Valencia A. The structural coverage of the human proteome before and after AlphaFold. *PLoS Computational Biology*. 2022;18(1):e1009818. doi: [10.1371/journal.pcbi.1009818](https://doi.org/10.1371/journal.pcbi.1009818)
>
> Xia Y, Zhao K, Liu D, Zhou X, Zhang G. Multi-domain and complex protein structure prediction using inter-domain interactions from deep learning. *Communications Biology*. 2023;6(1):1221. doi: [10.1038/s42003-023-05610-7](https://doi.org/10.1038/s42003-023-05610-7)
>
> Buel GR, Walters KJ. Can AlphaFold2 predict the impact of missense mutations on structure? *Nature Structural and Molecular Biology*. 2022;29(1):1-2. doi: [10.1038/s41594-021-00714-2](https://doi.org/10.1038/s41594-021-00714-2)
>
> Pak MA, Markhieva KA, Novikova MS, Petrov DS, Vorobyev IS, Maksimova ES, Kondrashov FA, Ivankov DN. Using AlphaFold to predict the impact of single mutations on protein stability and function. *PLoS One*. 2023;18(3):e0282689. doi: [10.1371/journal.pone.0282689](https://doi.org/10.1371/journal.pone.0282689)
>
> McBride JM, Tlusty T. AI-predicted protein deformation encodes energy landscape perturbation. *Physical Review Letters*. 2024;133(9):098401. doi: [10.1103/PhysRevLett.133.098401](https://doi.org/10.1103/PhysRevLett.133.098401)
>
> Saldaño T, Escobedo N, Marchetti J, Zea DJ, Mac Donagh J, Velez Rueda AJ, et al. Impact of protein conformational diversity on AlphaFold predictions. *Bioinformatics*. 2022;38(10):2742-2748. doi: [10.1093/bioinformatics/btac202](https://doi.org/10.1093/bioinformatics/btac202)
>
> Chakravarty D, Porter LL. AlphaFold2 fails to predict protein fold switching. *Protein Science*. 2022;31(6):e4353. doi: [10.1002/pro.4353](https://doi.org/10.1002/pro.4353)
>
> Chakravarty D, Schafer JW, Chen EA, Thole JF, Ronish LA, Lee M, Porter LL. AlphaFold predictions of fold-switched conformations are driven by structure memorization. *Nature Communications*. 2024;15(1):7296. doi: [10.1038/s41467-024-51801-z](https://doi.org/10.1038/s41467-024-51801-z)
>
> Perkins-Jechow BH, Iglesias Ahualli JP, Nhu HT, Omidi A, Li C, Holguin-Cruz JA, et al. Challenging AlphaFold in predicting proteins with large-scale allosteric transitions. *Communications Chemistry*. 2025;8(1):378. doi: [10.1038/s42004-025-01763-0](https://doi.org/10.1038/s42004-025-01763-0)
>
> Wilson CJ, Choy WY, Karttunen M. AlphaFold2: a role for disordered protein/region prediction? *International Journal of Molecular Sciences*. 2022;23(9):4591. doi: [10.3390/ijms23094591](https://doi.org/10.3390/ijms23094591)
>
> Holcomb M, Chang YT, Goodsell DS, Forli S. Evaluation of AlphaFold2 structures as docking targets. *Protein Science*. 2023;32(1):e4530. doi: [10.1002/pro.4530](https://doi.org/10.1002/pro.4530)
>
> Karelina M, Noh JJ, Dror RO. How accurately can one predict drug binding modes using AlphaFold models? *eLife*. 2023;12:e89386. doi: [10.7554/eLife.89386](https://doi.org/10.7554/eLife.89386)
>
> Song J, Ha J, Lee J, Ko J, Shin WH. Improving docking and virtual screening performance using AlphaFold2 multi-state modeling for kinases. *Scientific Reports*. 2024;14:25167. doi: [10.1038/s41598-024-75400-6](https://doi.org/10.1038/s41598-024-75400-6)
>
> Scardino V, Di Filippo JI, Cavasotto CN. How good are AlphaFold models for docking-based virtual screening? *iScience*. 2023;26(1):105920. doi: [10.1016/j.isci.2022.105920](https://doi.org/10.1016/j.isci.2022.105920)
>
> Mitjavila-Domènech N, Díaz-Holguín A, Hu H, Kahlous NA, Cabeza de Vaca I, Wallner B, Carlsson J. Improving AlphaFold2 performance in virtual screens targeting GPCRs by enhancing binding-site conformational sampling. *Journal of Chemical Information and Modeling*. 2026;66(10):6045-6056. doi: [10.1021/acs.jcim.6c00034](https://doi.org/10.1021/acs.jcim.6c00034)
>
> Wong F, Krishnan A, Zheng EJ, Stärk H, Manson AL, Earl AM, Jaakkola T, Collins JJ. Benchmarking AlphaFold-enabled molecular docking predictions for antibiotic discovery. *Molecular Systems Biology*. 2022;18(9):e11081. doi: [10.15252/msb.202211081](https://doi.org/10.15252/msb.202211081)
>
> Gómez Borrego J, Torrent Burgas M. Evaluating ligand docking methods for drugging protein-protein interfaces: insights from AlphaFold2 and molecular dynamics refinement. *Journal of Cheminformatics*. 2025;17:144. doi: [10.1186/s13321-025-01067-4](https://doi.org/10.1186/s13321-025-01067-4)

## Go / no-go

**Go — act on the result when:**

- The protein is in the database and you want a structure to reason about: a fold, a domain architecture, the rough shape of a pocket. This is the fastest accurate structure you can get, and it costs no GPU.
- Mean pLDDT is above 70 **and** the residues your question depends on are individually above 70. Check the per-residue plot, not the mean.
- Your question spans two domains **and** you have read the PAE matrix for the block between them.
- You are phasing a crystal structure. This is the best-evidenced use of a predicted model anywhere in this documentation. Convert pLDDT to B-factors and trim below 70 before the search, and split into domains on the PAE if the first attempt fails.
- You will use the model to design an experiment — a construct boundary, a crystallography target, a mutagenesis panel — rather than to conclude one.

**No-go — do not act on the result when:**

- An experimental structure exists. [Download it from the PDB](/docs/pdb) instead; a crystal structure also brings ligands, ions and cofactors that an AlphaFold monomer does not have.
- The protein has few homologs. A shallow MSA is the measured failure mode, the database entry will usually say so through low pLDDT, and no re-fetch improves it.
- The region you care about is below pLDDT 70, or the sequence is short enough (under ~100 residues) that a high pLDDT may be a length artefact.
- The answer is a cavity volume, a rotamer, or an active-site geometry. A 0.30 Å model had a 10%-smaller pocket in the one case where somebody checked.
- Your protein's shape is made by its binding partners. The authors name this as a limitation, the database ships single chains, and nothing flags it.
- You are modelling a complex, a multimer, or an assembly. A fetched entry is one chain with no partners, no ligands and no cofactors; use [Boltz-2](/docs/boltz2) or [Chai-1](/docs/chai). If the complex is an antibody–antigen pair, the published top-1 success rates are 26% acceptable and 5% high even for AlphaFold-Multimer, so do not expect a usable interface from any AlphaFold route.
- You intend to treat a docked pose from this receptor as a binding mode. Expect roughly a halving of top-1 pose accuracy relative to a co-crystal, and no computable signal telling you whether this target is one of the bad ones. ⚠︎ The exception is worth knowing: if your real alternative is an **apo** crystal structure, the model is measurably the better receptor (16% vs 10%).
- You want the structural effect of a point mutation. The model's own authors state it has not been validated for this, predicted mutant and wild-type structures differ by a few tenths of an Ångström even for folding-destroying mutations, and ΔpLDDT correlates with experimental ΔΔG at r = −0.17.
- You need a *particular* conformational state — apo versus holo, active versus inactive, one fold of a fold switcher. You get one state, it skews holo, and pLDDT does not tell you which one you got.
- You need a specific database version for reproducibility against a published figure. The tool does not expose one and the version can change between runs.
- You need the structure established rather than proposed. That requires crystallography, cryo-EM, or NMR, and no pLDDT value substitutes for it.

**Caveats on reading this page:** every figure above is quoted from the cited source with its original metric and sample. Figures from different studies are not directly comparable, because test sets, cutoff dates, filtering criteria, AlphaFold versions and metrics differ between them. Figures labelled author-reported come from the method's own paper and have not been independently reproduced on those sets.