How ESMFold results compare to the bench

You have an ESMFold result. Here is what it supports.

• A high-pLDDT model is a fold hypothesis at near-experimental resolution. On CAMEO, ESMFold predictions with pLDDT > 90 had a median all-atom RMSD₉₅ of 1.42 Å to the experimental structure (backbone 0.94 Å), author-reported. That is close enough to reason about secondary structure, domain architecture, and a pocket's rough shape.
• Read pLDDT as a per-residue forecast, not a global guarantee. It is the best single quality signal ESMFold gives you, and it is wrong a measurable fraction of the time: on 1,666 template-poor PDB monomers, 35% of ESMFold's incorrect models still carried a mean pLDDT ≥ 70. A confident core beside a disordered tail is the normal case, and the mean hides both. Colour by pLDDT and look at the residues your question actually depends on — reading pLDDT has what each band is and is not backed by.
• The pLDDT bands everyone quotes were measured on AlphaFold2, not ESMFold. ESMFold's own paper defines no band scheme. Treat >90 / 70–90 / 50–70 / <50 as a borrowed convention and lean on the two ESMFold-specific figures instead: pLDDT > 90 → median all-atom RMSD₉₅ 1.42 Å, and pLDDT > 70 → lDDT 0.83, both on CAMEO.
• Only the > 90 band is trustworthy. The largest template-poor benchmark concluded that ESMFold's pLDDT "is well aligned with accuracy only in the uppermost confidence bin (pLDDT > 90)" and that 70–90 "contain a mixture of correct and incorrect structures." ESMFold also scores optimistically in absolute terms — its pLDDT-to-measured-lDDT fit has a flatter slope and a higher intercept than AlphaFold2's. Read 70–90 as unresolved, not confident.
• ESMFold is not the most accurate structure predictor; it is the fastest one that needs no alignment. Its authors report an average TM-score of 0.83 on CAMEO and 0.68 on CASP14, against AlphaFold2's 0.88 and 0.85. At CASP15 the ESMFold entry placed 74th of 118 evaluated groups. If your protein has homologs, fetch the AlphaFold model instead.
• The accuracy gap widens exactly where you hoped it would close. On ESMFold's own CAMEO split, the deficit against AlphaFold2 grows from 0.03 TM-score on easy, alignment-rich targets to 0.17 on hard, shallow-alignment targets. CASP15's assessors reached the same conclusion independently.
• Docking into a predicted receptor costs pose accuracy. DiffDock's top-1 success fell from 35.0% on crystal receptors to 21.7% on esmfold_v1 receptors; search-based docking fell to 2–10%. Budget for that before treating a pose from a predicted structure as a result.
• Confirm a structure with X-ray crystallography, cryo-EM, or NMR. Nothing computational substitutes.

Jump to agreement with wet lab or against other in silico methods. The general framing for every tool is in reading an in silico result against the bench.

How in silico is used currently

Predicted structures are used to supply a model where no experimental one exists, and the field treats them as models throughout. In practice that is three jobs:

• Covering the gap. The PDB holds experimental structures for a small fraction of known sequences. ESMFold's authors folded 617 million MGnify90 metagenomic sequences to build the ESM Metagenomic Atlas; about 365 million (≈59%) reached mean pLDDT > 0.5 and pTM > 0.5, and about 225 million (≈36%) reached mean pLDDT > 0.7 and pTM > 0.7. The remaining ~64% is the honest denominator for what single-sequence folding can characterize at scale.
• Triage before an experiment. Rank candidates for crystallography, pick a construct boundary, decide which domain to express. Here a model is used to choose an experiment rather than to replace one, and a wrong model costs a failed plate rather than a wrong conclusion. CASP15's assessors graded exactly this use case for molecular replacement, and reported that search models from protein-language-model methods scored less well than MSA-based ones but still cleared the LLG = 60 threshold for many targets.

• An input to a structure-based step — pocket finding (P2Rank), docking (DiffDock), electrostatics. This is where prediction error propagates, and it is the use that needs the figures below.

How often you get a usable model at all is itself measurable. On 644 real Swiss-Prot proteins, 65.7% of ESMFold predictions reached a confident mean pLDDT (> 70), against 90.8% for AlphaFold2 and 94.1% for AlphaFold3 on the same sequences. Read that with its sample: those 644 sequences were deliberately length-matched to a spurious-protein set, so they are short proteins binned across 1–200 residues, not a representative proteome.

What nobody does, with validation they report honestly, is treat a predicted structure as a determined one.

Against other in silico methods

At CASP15, the independent blind assessment. ESMFold competed as the group ESM-single-sequence and placed 74th of 118 evaluated groups (groups submitting for at least 10 of 112 evaluation units across 77 single-chain targets, cumulatively ranked over 109 units). The assessors' wording: protein-language-model methods "are not currently competitive with MSA-based methods like AF2." On model_1 submissions by GDT-TS:

ESMFold attempted 93 evaluation units against 107–109 for the MSA-based groups, so the denominators differ. The same assessment found language-model methods lag on side-chain placement for models of equal backbone quality, and that "the ESMFold quality estimates are a little less accurate [than ColabFold and DeepMind's AlphaFold2], something users of the new ESM Metagenomic Atlas should bear in mind."

Against MSA-based methods stripped of their MSAs, ESMFold wins outright and this is its only clean win. Running AlphaFold2 and RoseTTAFold on single sequences by ablating the alignment drops them to TM-score 0.41 and 0.47 on CAMEO and 0.38 and 0.37 on CASP14, far below ESMFold's 0.83 and 0.68. The authors note this is artificial, since neither model was trained for single-sequence input — but it is the setting that applies to a designed or orphan sequence, and it is the reason to reach for ESMFold at all.

On short peptides it is genuinely level with the MSA-based methods. Over 56 post-cutoff NMR peptide structures, mean GDT_TS was ESMFold 68.59, AlphaFold2 69.75, RoseTTAFold2 66.52, with no significant pairwise difference (Wilcoxon signed-rank). Worth noting why that is less flattering than it looks: on the pre-cutoff set the MSA-based methods gained significantly from having seen the data (AlphaFold2 +9.24, p = 0.014) while ESMFold's gain was not significant (p = 0.105) — so the parity here is partly AlphaFold2 having less to exploit on short sequences. ESMFold was significantly worse than both on β-sheet-only and mixed helix/β-sheet peptides.

Against other single-sequence predictors. On the 1,337-chain independent benchmark, ESMFold beat OmegaFold on every metric (median TM-score 0.95 vs 0.93, RMSD 1.74 Å vs 1.98 Å, GDT-TS 90% vs 89%). On nanobody CDR3 loops, OmegaFold beat it (median RMSD 2.5 Å against ESMFold's reported 3.3 Å). At CASP15, ESMFold was the better of the two language-model-only groups entered.

The language model is doing the work. With an 8-block folding trunk, ESMFold reached 0.74 lDDT on CAMEO. Removing the language model dropped that to 0.58 lDDT; removing the folding trunk and keeping the language model dropped it to 0.66. Recycling, triangular updates, and AlphaFold2 distillation targets each moved lDDT by only 0.01–0.04.

On complexes, it is out of distribution. ESMFold v1 trains on single chains. Against AlphaFold-Multimer on 2,978 recent multimeric PDB complexes, it produced the same qualitative DockQ category for 53.2% of chain pairs — notable for a model never trained on complexes, and not a basis for modelling an assembly. On the stricter 994-dimer set above it reached DockQ ≥ 0.23 on 41.1% against AlphaFold2's 76.8%. Use a co-folding model (Boltz-2, Chai-1) instead.

Where ESMFold is reported to win on accuracy, not just speed. Two results, both narrow. 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 AlphaFold2 showed a weakness unique to it — significantly lower TM-scores for designs than for monomers or complexes. Direction and significance are published (Kruskal–Wallis, Dunn, Bonferroni); the per-group medians sit in a figure panel and were not extracted. And on 614 experimentally characterized de novo monomeric designs from 11 studies (269 successful, 43%; AUCs over n = 570 after excluding NMR-resolved cases), ESMFold's mean pLDDT was the best single computational discriminator of design success — expressed, soluble, monomeric and correctly folded — at ROC AUC 0.72 ± 0.05, the highest of every confidence metric tested, where AlphaFold3's pLDDT was lowest at 0.60 ± 0.06. A logistic regression over all metrics reached 0.72 ± 0.14, no better than ESMFold's pLDDT alone. But there is no usable threshold: pLDDT distributions are topology-dependent, and failed designs from easy topologies out-score successful designs from hard ones. Taking the top half by both AlphaFold2 and ESMFold pLDDT lifted the success rate from 39% to 57%, which is the realistic size of the win.

Speed, measured. On a single V100, ESMFold predicted a 384-residue protein in 14.2 s, about 6× faster than a single AlphaFold2 model, rising to roughly 60× on shorter sequences. Independent benchmarks put the alignment-free methods at 10–30× faster end to end, or ~30 s to a few minutes per chain on an A100 against 0.5–2 h for AlphaFold2. The alignment search alone takes over 10 minutes per sequence with published high-sensitivity protocols. The speed claim is real and it is the reason to use this tool.

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

Garcia M, Dixit SM, Rocklin GJ. Evaluating zero-shot prediction of monomeric protein design success by AlphaFold, ESMFold, and ProteinMPNN. Protein Science. 2026;35(2):e70453. doi: 10.1002/pro.70453

In silico vs. wet lab

The standard wet-lab approach for the question ESMFold 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, which is why this section can carry real numbers. Global agreement is reported as TM-score (0–1; pairs above 0.5 are mostly the same fold, below 0.5 mostly not) or GDT-TS (percentage of residues within distance cutoffs after superposition; ≥ 45 is taken as correct topology and ≥ 90 as roughly the difference between two crystal forms of the same protein). Local agreement is lDDT, a superposition-free per-residue distance-difference score. Positional error is RMSD to the experimental coordinates.

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 comparable to each other, because the figures quoted for them come from the same test sets scored with the same metrics, and that comparison is the point of this table.

On that reading, the honest position for ESMFold is: seconds instead of months, and seconds instead of the hours-plus-alignment an MSA-based run takes, at a cost of roughly 0.01–0.05 TM-score on routine targets and 0.17 on hard ones. That is a good trade when no alignment exists, when you need throughput, or when 0.83 is enough for the decision in front of you. It is a bad trade when a homolog-rich target already has an AlphaFold model you could simply fetch.

Two 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. And an experiment reports what is, while ESMFold reports what is likely — pLDDT 95 and a solved structure are not interchangeable claims.

Agreement with wet lab

All figures below are against experimentally determined structures, with the metric named each time. ESMFold's authors evaluated on structures held out by a May 2020 cutoff, which is also AlphaFold2's and RoseTTAFold's cutoff, so the three are comparable on those sets.

The authors' own benchmark

On the CAMEO test set (194 structures assessed April–June 2022), author-reported:

• Average TM-score 0.83. AlphaFold2 with MSAs and templates: 0.88. RoseTTAFold: 0.82.
• Median all-atom RMSD₉₅ 1.91 Å; backbone RMSD₉₅ 1.33 Å.
• Restricted to pLDDT > 0.9: median all-atom RMSD₉₅ 1.42 Å, backbone 0.94 Å.
• Restricted to pLDDT > 0.7: lDDT 0.83, against AlphaFold2's 0.85.
• ESMFold matched AlphaFold2 to within 0.05 lDDT on more than half the proteins.

On the CASP14 test set (51 publicly released structures, selected for difficulty): average TM-score 0.68, against AlphaFold2's 0.85. This is the figure that should set expectations for a hard target, and it is the one most often omitted when ESMFold is called competitive.

Broken down by alignment depth, which is the most useful table in the paper and the one that contradicts the usual intuition about where ESMFold helps. Metric is mean TM-score; depth is from AlphaFold2's MSA pipeline:

The paper's own caption says it: differentiation is greater in the medium and hard classes, which have lower MSA depth and are better predicted by AlphaFold2. ESMFold's advantage is that it does not need an alignment, not that it performs better when one is unavailable.

Accuracy tracks language-model perplexity. Pearson correlation between ESM-2 perplexity and TM-score to the experimental structure was −0.55 on CAMEO and −0.67 on CASP14. Sequences the language model finds unsurprising are predicted about as well as AlphaFold2 predicts them; sequences it finds surprising are not.

Independent benchmarks

On 1,337 PDB chains deposited July 2022 – July 2024 — after the training cutoff of all three methods, 20–400 residues, comprising 980 monomers, 245 small complexes and 102 de novo designs — medians were:

All four differed significantly across tools (Kruskal–Wallis p < 0.001, post-hoc Dunn with Bonferroni). Split by experimental method, ESMFold's median RMSD was 1.65 Å against X-ray structures and 2.89 Å against NMR structures (AlphaFold2: 1.24 Å and 2.31 Å). The same study found AlphaFold2 significantly degraded on proteins lacking Pfam family annotation (odds ratio 0.67, p < 0.01) while ESMFold and OmegaFold held steady — ESMFold not degrading, which is not the same as ESMFold winning.

On 1,666 monomers and 994 dimers deposited 2022–2024 and filtered to lack close pre-2022 homologs, 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:

Counting accurate plus domain movement as correct, that is 76% for ESMFold against 88% for AlphaFold2 and AlphaFold3. Most of the incorrect predictions were NMR targets; restricted to X-ray and cryo-EM monomers, the same study reports 83% for ESMFold against 95% for AlphaFold — the fairer comparison if your reference structure is a crystal or a map. On the 994 dimers, scored by DockQ, ESMFold reached acceptable-or-better (DockQ ≥ 0.23) on 409 (41.1%) against AlphaFold2's 763 (76.8%) — expected, since ESMFold v1 is a single-chain model being used out of distribution.

Reading pLDDT

pLDDT is the field you are actually holding, so it deserves its own accounting. Four things about it are load-bearing and three of them are usually left out.

Mind the scale. The esmfold_v1 runtime emits pLDDT on a 0–100 scale, which is what the platform shows you. Lin et al.'s own analyses state every threshold as a 0–1 fraction (pLDDT > 0.5, > 0.7, > 0.9). Those are the same thresholds — 70 here is 0.7 there — but a figure quoted without its scale is unreadable.

The bands are AlphaFold2's. ESMFold computes pLDDT using AlphaFold2's prediction head applied to its own internal representations, and the familiar >90 / 70–90 / 50–70 / <50 scheme comes from AlphaFold2 work, not from ESMFold's paper, which defines no bands at all and uses only "very high (pLDDT > 0.9)". The bands' evidence is:

• pLDDT > 90 — AlphaFold's χ¹ rotamers were 80% correct on a recent-PDB test set (n = 5,983 chains / 609,623 residues, resolution < 2.5 Å, B-factor < 30 Ų, χ¹ correct within 40°).
• pLDDT > 70 — "a generally correct backbone prediction."
• pLDDT < 50 — long regions in this band "should not be interpreted as structures but rather as a prediction of disorder." AlphaFold2's own paper reports its pLDDT reaching AUC 0.897 as a disorder predictor on the CAID DisProt-PDB benchmark (n = 178,124 residues); an independent evaluation on the CAID main test set (n = 646 proteins / 336,595 residues, DisProt annotations) put AlphaFold2 pLDDT at AUC 0.722, MCC 0.137, F1 0.278, below both AlphaFold2's own relative-solvent-accessibility signal (0.768) and a dedicated disorder predictor (flDPnn, 0.814). The gap between 0.897 and 0.722 is the usual author-reported-versus-independent gap, and all of it is AlphaFold2.
• Underlying fit: per-residue pLDDT against lDDT-Cα, Pearson r = 0.73 (n = 10,215 chains / 2,756,569 residues), and in AlphaFold2's own paper lDDT-Cα = 0.997 × pLDDT − 1.17, r = 0.76 (n = 10,795 chains).

No primary study has validated those bands for ESMFold. Every ESMFold paper surveyed here that applies a ≥ 70 threshold either cites the AlphaFold2 work or asserts cross-method equivalence from Lin et al. One of them does both at once, stating in its Methods that ESMFold's score "is calibrated differently" and then that the ≥ 70 threshold is "equivalent across methods." A shared prediction head is not shared calibration.

The ESMFold-specific evidence splits, and the split is the useful part. Rank ordering is good: on 1,337 recent PDB chains, mean pLDDT correlated with agreement to the experimental structure at Spearman ρ = 0.66 against TM-score, −0.61 against RMSD and 0.60 against GDT-TS — slightly tighter than AlphaFold2's 0.60, −0.56 and 0.55 on the same set. On the 1,666-monomer set, 98.7% of ESMFold's accurate models carried mean pLDDT ≥ 70, so the threshold is highly sensitive. Absolute calibration is worse than AlphaFold2's: on 56 post-cutoff NMR peptide structures, the least-squares fit of mean pLDDT to measured lDDT-Cα gave ESMFold a slope of 0.769 with an intercept of +16.317, against AlphaFold2's 0.916 and +3.850 (RoseTTAFold2 0.836 / +14.788). A flatter line sitting higher means pLDDT overstates quality, and most at the low-quality end — exactly where you need it to warn you.

So pLDDT ranks ESMFold's own predictions well and scores them optimistically. Two consequences follow. The ≥ 70 threshold is sensitive but not specific — 35% of ESMFold's outright-wrong monomer models still cleared it, and 57% of its wrong dimers did. And the bands are not uniformly trustworthy: the 1,666-monomer study concluded that pLDDT "is well aligned with accuracy only in the uppermost confidence bin (pLDDT > 90)" while "intermediate scores (70–90) contain a mixture of correct and incorrect structures." Treat 70–90 as unresolved, not as confident.

Two further cautions. CASP15's assessors, grading self-assessment by median Z_ASE, found ColabFold's and DeepMind's AlphaFold2 confidence estimates among the most accurate available and noted that "the ESMFold quality estimates are a little less accurate, something users of the new ESM Metagenomic Atlas should bear in mind." And Lin et al. report no pLDDT-versus-experimental-lDDT correlation coefficient; the r = 0.79 sometimes quoted from that paper is pLDDT against agreement with AlphaFold2's predictions on ~4,000 metagenomic sequences, not against an experimental structure. Nor do they state how many of the 194 CAMEO targets fall in the pLDDT > 0.7 or > 0.9 subsets, so those two figures have no citable n.

Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-589. doi: 10.1038/s41586-021-03819-2

Tunyasuvunakool K, Adler J, Wu Z, et al. Highly accurate protein structure prediction for the human proteome. Nature. 2021;596(7873):590-596. doi: 10.1038/s41586-021-03828-1

Sauvestre C, Zagury JF, Langenfeld F. Benchmarking AlphaFold2, RoseTTAFold2 and ESMFold on peptide structure prediction. Proteins. 2026;94(3):675-691. doi: 10.1002/prot.70049

Case studies against individual structures

A newly solved crystal structure. GH11 xylanase from Hypocrea virens was solved and then compared against predictions for the same sequence. The ESMFold model (mean pLDDT 88.0) superimposed at RMSD 0.87 Å, against 0.30 Å for AlphaFold2, 0.30–0.42 Å for AlphaFold3 and 0.77–0.88 Å for RoseTTAFold. The canonical β-jelly-roll fold was correct in every model; the thumb domain and the N-terminus were not. Docking xylohexaose into the ESMFold model put the ligand across all six subsites at −12.1 kcal/mol, where the experimental structure gave five subsites at −12.5 kcal/mol — the same cleft, a different binding mode.

Nanobodies. On 75 non-redundant nanobodies (median resolution 2.59 Å, none in any tested model's training set), ESMFold placed the framework region at a median RMSD of 0.6–0.7 Å, statistically indistinguishable from OmegaFold, AlphaFold2, IgFold and NanoNet. CDR3 was a different story for every method: medians ran 2.5 Å to 4.7 Å across the six programs, with about half or more of all structures above 2.5 Å. ESMFold's CDR3 median is 3.3 Å, and it modelled 35% of CDR3 loops below 2.5 Å — behind OmegaFold (52%) and AlphaFold2 (44%), ahead of IgFold (29%) and NanoNet (25%). Overall ranking across the benchmark: OmegaFold > AlphaFold2 > ESMFold > IgFold > NanoNet. The authors' own conclusion is that global TM-score and GDT-TS are unsuitable here — all five of the leading tools exceed 0.91 — because CDR3 error is diluted by a conserved framework. If a CDR loop is the answer you need, the global score is actively misleading.

Ion channels. A three-channel case study against cryo-EM structures (human Na_V1.8, Ca_V1.1, K_V1.3) found ESMFold's pore domains under 2 Å Cα RMSD — comparable to AlphaFold2 and RoseTTAFold2, all three within 0.7–1.8 Å on Na_V1.8 and at 1.09 Å and 1.5 Å on the other two — but its voltage-sensing domains consistently worse than AlphaFold2's (VSD-III 1.5 Å vs 0.6–0.8 Å; VSD-IV 1.8–1.9 Å vs 1.2–1.3 Å), and its K_V1.3 model mis-assembled the T1 tetramer — each subunit correctly folded to under 1 Å while the inter-subunit contacts seen in AlphaFold2, RoseTTAFold2 and the experimental structures were broken. One domain (VSD-I) matched no experimental conformation in any method despite very high pLDDT.

Lin Z, Akin H, Rao R, Hie B, Zhu Z, Lu W, Smetanin N, Verkuil R, Kabeli O, Shmueli Y, dos Santos Costa A, Fazel-Zarandi M, Sercu T, Candido S, Rives A. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023;379(6637):1123-1130. doi: 10.1126/science.ade2574

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

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

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

Valdés-Tresanco MS, Valdés-Tresanco ME, Jiménez-Gutiérrez DE, Moreno E. Structural modeling of nanobodies: a benchmark of state-of-the-art artificial intelligence programs. Molecules. 2023;28(10):3991. doi: 10.3390/molecules28103991

Nguyen PT, Harris BJ, Mateos DL, González AH, Murray AM, Yarov-Yarovoy V. Structural modeling of ion channels using AlphaFold2, RoseTTAFold2, and ESMFold. Channels. 2024;18(1):2325032. doi: 10.1080/19336950.2024.2325032

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

Xu J, Zhang Y. How significant is a protein structure similarity with TM-score = 0.5? Bioinformatics. 2010;26(7):889-895. doi: 10.1093/bioinformatics/btq066

What the figure doesn't cover

• The gap widens as alignments thin, which is the opposite of the usual claim — but read the two regimes separately. On hard natural targets, where AlphaFold2 still finds a shallow alignment, ESMFold's deficit grows from 0.03 TM-score (easy, median UniRef depth 17,627) to 0.17 (hard, depth 56), and CASP15's assessors reached the same conclusion independently: "the performance deficit actually increases on the right as available MSAs become shallower," and "claimed benefits for targets with no or few homologues were not apparent on the CASP15 targets." On true orphans, where nobody has an alignment, the relative gap is reportedly flat at roughly 0.05 across depth-1, ≤10 and ≤100 bins, because AlphaFold2 loses its MSA too — but both methods sit near TM-score 0.65 there, so the problem is the absolute level, not the ranking. The two regimes are not in conflict, and neither supports reading ESMFold's MSA independence as an accuracy advantage. It is a workflow advantage.

• The orphan case has no concordance number, and the evidence that exists is discouraging. ESMFold's authors built an orphan benchmark for exactly this — 104 PDB-derived sequences with MSA depth ≤ 100, 70 with depth ≤ 10, 22 with depth = 1, all temporally held out — but reported the comparison as a scatter plot (Fig. S8), not a statistic. Separately, on an expert-curated set of 8,974 orphan orthogroups (48,681 proteins) from plant-parasitic Meloidogyne nematodes, ESMFold, OmegaFold and AlphaFold2 all returned median and average pLDDT in the low (50–70) or very low (< 50) bands, with ESMFold systematically lowest, against just above or below 70 for the same species' non-orphan proteome. Disorder did not explain it: two non-transformer disorder predictors found no significant orphan/non-orphan difference. ESMFold was not even self-consistent across homologs — within an orthogroup, only 28% of predictions were structurally identical and 50% had a median TM-score below 0.5. That study has no experimental ground truth (no orphan structures are deposited), so it measures confidence and inter-predictor agreement, not accuracy. Its authors are explicit that "this limited performance cannot be fully attributed to intrinsic disorder," having checked with three non-transformer disorder predictors. Net position: on a true orphan, expect a fold-level answer at best — around TM-score 0.65 in Lin et al.'s own orphan bins, above the 0.5 same-fold threshold but well short of detail — accompanied by a low pLDDT that is telling you the truth.
• pLDDT is a forecast that misses. Of ESMFold's 391 incorrect monomer models on the 1,666-chain set, 137 (35%) still had mean pLDDT ≥ 70. The comparison cuts the other way too: of AlphaFold2's 200 incorrect models, 150 (75%) had pLDDT ≥ 70, so ESMFold's confidence is the better calibrated of the two in absolute terms — it is simply wrong more often overall. On dimers, 57% of ESMFold's incorrect predictions still scored pLDDT ≥ 70.
• "Low pLDDT means disorder" has never been tested on ESMFold. The claim traces to AlphaFold2 work, and ESMFold has never been assessed in CAID — the CAID3 paper, which re-assesses every CAID2 predictor, mentions AlphaFold 40 times and ESMFold not once; CAID1 predates ESMFold. ESMFold's own paper makes no disorder claim at all, and calibrates pLDDT against lDDT only. ⚠︎ Two name traps: CAID3's ESMDisPred entries and its four ESM2-embedding methods use ESM embeddings, not ESMFold, and must never stand in for it. One study on Drosophila proteins reported the relationship reversing sign — a positive correlation between predicted disorder and confidence for de novo and random proteins, against the expected negative correlation for conserved ones — though its comparator was predicted rather than curated disorder, and only its abstract was readable. Treat a low-pLDDT region as "no model here," which is always true, rather than as a disorder call, which is borrowed.
• pLDDT is confident about things that are not proteins. On 644 spurious AntiFam sequences, ESMFold's average mean pLDDT was 39.82, rising to 56.3 on sequences ≤ 50 residues — a 41% increase over its own mean, purely from shortness. ESMFold inflates less than AlphaFold2 (57.36 overall, 71.18 short) or AlphaFold3 (65.15, 80.74), so it is the better-behaved of the three, but the bias runs the same way: a short sequence buys pLDDT it has not earned. Discrimination between real and spurious improves beyond 100 residues.
• Confidence is not a model-selection oracle, and the published evidence for that is about AlphaFold, not ESMFold. A systematic adversarial evaluation found predicted structures invariant to mutation of up to 40% of residues and to 10% deletions, with confidence selecting the most accurate of five models at most 35% of the time. Those are AlphaFold 2 and 3 figures, and the metric is the ranking score (pTM/ipTM plus a clash penalty), not pLDDT — that paper does not analyse pLDDT at all. ESMFold enters it only as a comparator showing greater, though still imperfect, mutational sensitivity, with no number given. The transferable caution is the general one: do not use a confidence score to pick between two near-identical mutant models.
• Side chains lag the backbone. CASP15 found language-model methods place side chains worse than MSA-based methods for models of the same low backbone score. If your question is a rotamer, an active-site geometry, or a hydrogen-bond network, the backbone RMSD figures above overstate what you have.
• Domain orientation is the weak axis within an otherwise correct model. 25.8% of ESMFold's models on the 1,666-chain set were classified as correct locally (lDDT ≥ 0.7) but displaced globally (RMSD > 3.0 Å) — the "domain movement" class. The GH11 case had the fold right and the thumb domain wrong, and the K_V1.3 case had every subunit right and the tetramer wrong. Read the PAE matrix, not just pLDDT, before trusting the relative placement of two domains — see outputs and confidence.
• A predicted receptor costs pose accuracy downstream, measurably — and costs classical docking far more than it costs DiffDock. On the PDBBind date-split test set re-folded with esmfold_v1 (349 complexes; 12 of 361 discarded after out-of-memory failures), top-1 RMSD < 2 Å rates fell as follows: DiffDock 35.0% → 21.7%, GNINA 22.9% → 2.0%, SMINA 18.7% → 3.4%, P2Rank+GNINA 28.8% → 8.6%, TANKBind 20.4% → 10.4%. GNINA's median top-1 RMSD went from 7.7 Å to 22.3 Å. Adding side-chain flexibility did not recover it. So the honest reading is not "DiffDock degrades" but "search-based docking loses 85–90% of its success rate on an ESMFold receptor, and DiffDock loses about a third." The authors' own summary: existing methods "are not capable of docking against these approximate apo-structures." This was measured on the exact model this platform runs. ⚠︎ The DiffDock-L paper never revisits ESMFold receptors, so there is no updated figure for the newer docking model.
• No virtual-screening enrichment study on ESMFold receptors exists. Searches for ESMFold paired with enrichment factor, logAUC or docking power return nothing usable. Every published enrichment number for predicted receptors is AlphaFold2's, and those must not be relabelled. For orientation only, and labelled: redocking into AlphaFold2 models dropped top-pose accuracy from 41% to 17% over 2,474 complexes; mean EF_1% across 22 targets fell from roughly 20–24 on PDB structures to about 9 on AlphaFold2 models; and on 18 GPCRs, cross-docking pose accuracy was 15% into AlphaFold2 models against 44% into experimental structures even though the AlphaFold2 binding-pocket median RMSD was 1.3 Å — structural accuracy in the pocket did not transfer to pose prediction.
• The one counterweight on downstream cost is also AlphaFold2's, and it matters. When large libraries were docked prospectively against unrefined AlphaFold2 models, hit rates were comparable to experimental structures: σ2 receptor 51% (70/138) on the crystal versus 54% (64/119) on the model, and 5-HT2A 23% (51/223) on cryo-EM versus 26% (42/161) — neither difference significant — while retrospective enrichment on the same targets was much worse for the models (logAUC 39 versus 16 for σ2). Retrospective enrichment and prospective hit rate are not the same question, and the predicted-structure penalty is clearest in the former. Whether that generalizes from AlphaFold2 to ESMFold is unknown and should not be assumed, particularly given the pose-accuracy figures above.
• Version matters. This platform runs facebook/esmfold_v1 — the model in Lin et al. 2023, and the model used in the DiffDock apo-receptor benchmark above. Figures for ESM-3, AlphaFold 3, or any later single-sequence predictor do not describe it.
• Anything with an interface is a different question, and ESMFold loses it badly. ESMFold v1 models one chain, so every interface result below is it being used out of distribution — but these are the tasks people actually ask of it. On 200 post-cutoff antibody–antigen complexes it produced zero successful predictions by either CAPRI or DockQ criteria, where AlphaFold2 v2.3.2 exceeded 60% success over its top 25 models. On 67 GPCR–peptide complexes, mean DockQ 0.025 with 3 of 67 acceptable. On 51 post-2023 TCR complexes it was the best method for 3, against AlphaFold3's 17. Note the pattern in the GPCR case: the receptor backbone came out at 1.55 Å mean Cα RMSD, better than RoseTTAFold-AA without templates — a good monomer and a failed complex at the same time. Quoting the 1.55 Å without the DockQ 0.025 would be a serious misread.
• NMR reference structures are where the monomer numbers go bad. Of 190 monomeric NMR targets in the 1,666-chain set, ESMFold failed to predict 139 (73%) accurately, against AlphaFold2's 128 (67%). That is most of what separates the headline 76% from the 83% on X-ray and cryo-EM only. If your reference is an NMR ensemble, or your protein is flexible enough to have needed one, discount accordingly.
• A high-confidence model is not a design that works. On de novo β-barrels, ESMFold cleanly separated designable from non-designable backbones (all designable models above pLDDT 0.75 and within 2.0 Å of the design model) but made "almost no distinction between experimentally folded and misfolded designs," handing high pLDDT and low RMSD to nearly every sequence tested across 42 water-soluble barrels (15 folded) and 90 weak-β-sheet transmembrane barrels. The mechanism is the useful part: ESM-2's long-range contact detection decayed as barrel diameter grew from 16.0 Å to 25.8 Å, and for the largest barrel no contacts were predicted between the first and last strands — yet the structure module closed the barrel anyway. That is the clearest published account of how ESMFold manufactures a confident wrong answer. (In the same work it was AlphaFold2 whose best scores went to designs that failed to express.)

Corso G, Stärk H, Jing B, Barzilay R, Jaakkola T. DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking. ICLR. 2023. doi: 10.48550/arXiv.2210.01776

Seçkin E, Colinet D, Danchin EGJ, Sarti E. Evaluating transformer-based models for structural characterization of orphan proteins. Bioinformatics Advances. 2026;6(1):vbag215. doi: 10.1093/bioadv/vbag215

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

Feldman J, Brogi M, Skolnick J. Adversarial sequence mutations in AlphaFold and ESMFold reveal nonphysical structural invariance, confidence failures, and concerns for protein design. Computational and Structural Biotechnology Journal. 2026;35(1):0142. doi: 10.34133/csbj.0142

Middendorf L, Eicholt LA, Taubert A, Bornberg-Bauer E, Weisman CM. Random, de novo, and conserved proteins: how structure and disorder predictors perform differently. Proteins. 2024;92(6):757-767. doi: 10.1002/prot.26652

Ünsal S, Holland B, Sardag I, Timucin E. Benchmarking antibody-antigen complex structure prediction. Briefings in Bioinformatics. 2026;27(2):bbag137. doi: 10.1093/bib/bbag137

Hoegen Dijkhof LR, Rönkkö TKE, von Vegesack HC, Lenzing J, Hauser AS. Deep learning in GPCR drug discovery: benchmarking the path to accurate peptide binding. Briefings in Bioinformatics. 2025;26(2):bbaf186. doi: 10.1093/bib/bbaf186

Huang Q, et al. A comparative and exploratory analysis of computational methods for TCR structural prediction and antigen-specific TCR discovery. Briefings in Bioinformatics. 2026;27(4):bbag382. doi: 10.1093/bib/bbag382

Hermosilla AM, Berner C, Ovchinnikov S, Vorobieva AA. Validation of de novo designed water-soluble and transmembrane β-barrels by in silico folding and melting. Protein Science. 2024;33(7):e5033. doi: 10.1002/pro.5033

Sauvestre C, Zagury JF, Langenfeld F. Benchmarking AlphaFold2, RoseTTAFold2 and ESMFold on peptide structure prediction. Proteins. 2026;94(3):675-691. doi: 10.1002/prot.70049

Zhao B, Ghadermarzi S, Kurgan L. Comparative evaluation of AlphaFold2 and disorder predictors for prediction of intrinsic disorder, disorder content and fully disordered proteins. Computational and Structural Biotechnology Journal. 2023;21:3248-3258. doi: 10.1016/j.csbj.2023.06.001

Mehdiabadi M, Del Conte A, Nugnes MV, Aspromonte MC, Tosatto SCE, Piovesan D. Critical assessment of protein intrinsic disorder prediction (CAID) — results of round 3. Proteins. 2026;94(1):414-424. doi: 10.1002/prot.70045

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

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

Karelina M, Noh JJ, Dror RO. How accurately can one predict drug binding modes using AlphaFold models? eLife. 2023;12:RP89386. doi: 10.7554/eLife.89386

Lyu J, Kapolka N, Gumpper R, et al. AlphaFold2 structures guide prospective ligand discovery. Science. 2024;384(6702):eadn6354. doi: 10.1126/science.adn6354

Go / no-go

Go — act on the result when:

• No experimental structure and no AlphaFold DB entry exists for your sequence. That is the case ESMFold was built for.
• Your sequence is engineered, designed, or mutated, so an MSA either does not exist or does not describe what you folded. On de novo designs ESMFold is reported to do better than on natural proteins, which is the reverse of AlphaFold2's pattern.
• You need throughput — hundreds or thousands of sequences — and a 0.01–0.05 TM-score deficit is cheaper than 10 minutes of alignment search plus an hour of inference per chain.
• The residues your question depends on are individually above 90, not merely above 70. Check the per-residue plot, not the mean. A 70–90 region is a region you have not resolved yet.
• You will use the model to design an experiment — a construct boundary, a crystallography target, a mutagenesis panel, a molecular-replacement search model — rather than to conclude one.

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

• Your protein has homologs and an AlphaFold DB entry. Fetch it; it is faster than folding and more accurate.
• The target is genuinely orphan and you expect that to be ESMFold's strength. The published evidence says the gap to AlphaFold2 is widest there, and that all methods return low confidence on true orphans. A low-pLDDT orphan model is a "no model available" result, not a structure.
• You are modelling a complex, a multimer, or an assembly. ESMFold v1 is single-chain and reaches acceptable DockQ on only 41% of dimers; use Boltz-2 or Chai-1.
• The answer is a loop, a rotamer, or an active-site geometry. Nanobody CDR3 medians sit at 3.3 Å with half the structures above 2.5 Å, and CASP15 found language-model side-chain placement lags the backbone.
• The region you care about is below pLDDT 70, sits in the unresolved 70–90 band, or the sequence is short enough (under ~100 residues) that high pLDDT may be a length artefact.
• You are modelling an antibody–antigen, receptor–peptide, or TCR complex. ESMFold scored zero successes on 200 antibody–antigen complexes and 3 of 67 on GPCR–peptide pairs. A good monomer does not give you an interface.
• You intend to treat a docked pose from this receptor as a binding mode. Expect roughly a 13-point drop in top-1 pose accuracy relative to a crystal structure, and read DiffDock's translatability page before ranking anything.
• 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, 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.