How Chai-1 results compare to the bench

You have a Chai-1 result. Here is what it supports.

• A predicted complex is a geometric hypothesis, and for a protein–protein assembly the published agreement measure is DockQ against an experimentally determined complex. In independent testing on antibody–antigen complexes that no method had seen, the mean DockQ of Chai-1's top-ranked model was 0.12–0.14 — below DockQ's conventional "acceptable" floor of 0.23.
• The platform's configuration is the weak one. It runs Chai-1 with ESM embeddings and no MSAs, no templates, no restraints, a fixed seed, and one returned structure. Every published lever that raises Chai-1's antibody–antigen accuracy — MSAs, many seeds, epitope restraints — is unavailable here. See what the figure doesn't cover.
• iptm estimates how accurately the interface was modelled, not whether the chains bind. It has real predictive power for that (AUC 74.6 for pose correctness on a 489-ligand prospective set) and only a weak relationship to measured potency.
• Two co-folding models agreeing is weaker evidence than it looks. Co-folding methods' errors correlate with each other far more than with physics-based methods, so Chai-1 and Boltz-2 concurring is not two independent votes.
• Confirm a predicted interface with X-ray crystallography or cryo-EM of the complex, or a biophysical binding measurement (SPR, MST, ITC) if you only need to establish engagement.

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

Co-folding is used to propose an assembly you cannot yet see, and to triage which assemblies are worth an experiment. In a live program that means three things:

• Complex hypothesis from sequence. You have chains and no structure, and you want a model of how they sit together — which residues form the interface, which mutations to test, what a cryo-EM or crystallography request should be looking for.
• Screening and ranking by confidence. Confidence metrics are used as a cheap pre-filter over many candidate pairings. This is the weakest published use: evaluating four co-folding models on nanobody–antigen ranking across 17 antigens, no single confidence measure performed best across tasks and all models showed limited generalization to previously unseen antigens.
• Orthogonal validation of another method's result. Co-folding the same chains after a docking run, to see whether an independent architecture places the interface in the same region.

What co-folding is not used for by anyone reporting their validation honestly is deciding affinity. Abriata and Dal Peraro's practical-outcomes review of CASP16 put it flatly for the ligand case: "Ligand affinity prediction is totally unreliable." Chai-1 on this platform returns no affinity value at all.

Against other in silico methods

Against AlphaFold 3, on the platform's own use case, Chai-1 loses. The 110-complex antibody–antigen benchmark above is the only independent head-to-head that isolates this: AF3's top-ranked mean DockQ of 0.37 against Chai-1's 0.14, with all methods given the same MSAs and 200 samples per target. A separate independent evaluation of Chai-1 on 1628 antibody–antigen complexes found that adding even shallow MMseqs2-generated MSAs "nearly doubled its success rate… making it comparable to AlphaFold-3 (MSA)", while in the same study Chai-1 "performed relatively well even without MSA, likely due to its use of ESM2 language model embeddings." Those two findings pull in opposite directions from the Chai-1 report's own single-sequence claim, and the disagreement is unresolved.

Against Boltz-2, there is no clean winner and the choice is task-dependent. On Mac1 ligand poses Chai-1 reached ~70% against Boltz-2's 53%; on nanobody–antigen binder ranking, "Boltz-2-derived confidence measures achieved the highest median performance for identifying true binders." On unseen antibody–antigen complexes the two are indistinguishable, both at ~0.13 top-ranked DockQ. Both accept a small-molecule ligand as wired here; Boltz-2 is the only one of the two that accepts DNA and RNA chains, and the only one with an affinity head.

Against conventional docking, for the ligand case. On Mac1, Chai-1's iptm was a better predictor of pose correctness than DOCK3.7's energy (AUC 74.6 vs 67.1), and the authors concluded that co-folding and physics-based errors are less correlated with each other than two co-folding methods' errors are — which is the argument for combining them rather than treating one as a check on the other. For covalent protein–ligand complexes, a 218-complex benchmark comparing AutoDock-GPU, CovDock and GNINA against AlphaFold 3, Chai-1 and Boltz-1x found co-folding methods achieved superior ligand RMSD accuracy but that "their performance markedly declines for novel pocket-ligand pairs", while classical docking was stable but modest.

Chai-1 was not in CASP16, and no CASP-level blind assessment of it exists. The user-facing CASP16 summary says so directly: "servers and software like RoseTTAFold-AllAtoms, Chai and Boltz did not participate and were not tested as baselines in CASP16, so any performance claims lack the rigor of a CASP-level blind study." The same assessment notes that for antibody–antigen targets a ClusPro-based pipeline "significantly outperformed other groups, including extensively sampled AF3", and that antibody–antigen docking remains an area "where some traditional docking methods still hold an edge."

The widely cited cross-docking comparison does not include Chai-1. A 2026 study across three Alzheimer's targets (acetylcholinesterase, BACE1, GSK-3β) reported cross-docking success rates of 93.1% for AlphaFold 3, 89.6% for Boltz-2 and 85.7% for DiffDock, where GNINA 1.3 rigid fell from 87.7% on redocking to 13.5% on cross-docking. Chai-1 was not among the evaluated methods, so no figure from that paper describes it — and it is a cross-docking study in any case, which is not what co-folding does: there is no receptor structure input here.

Vogt Y, Roßberg R, Habermann J, et al. Analysing open-source protein folding models for nanobody binding prediction. Frontiers in Bioinformatics. 2026;6:1858891. doi: 10.3389/fbinf.2026.1858891

Zhang TH, Zhu JT, Huang ZX, Xie J, Pei JF, Lai LH. Benchmarking co-folding methods to predict the structures of covalent protein-ligand complexes. Acta Pharmacologica Sinica. 2026;47(5):1374-1384. doi: 10.1038/s41401-025-01721-5

Abriata LA, Dal Peraro M. Practical outcomes from CASP16 for users in need of biomolecular structure prediction. Proteins. 2026;94(1):435-446. doi: 10.1002/prot.70078

Suri K, Yadav A, Tripathi A, Murugan NA. Cross-docking and redocking reveal distinct determinants of success in physics-based and AI-driven binding pose prediction in protein-ligand complexes. RSC Advances. 2026;16(43):53425-53444. doi: 10.1039/d6ra05440d

In silico vs. wet lab

The standard wet-lab approach for the question Chai-1 answers is experimental structure determination of the complex — X-ray crystallography or cryo-EM. That resolves the assembly directly: which chains contact which, at which residues, in which conformation.

The comparison is well defined, which is why this section can carry real numbers. For a protein–protein or antibody–antigen interface, a predicted and an experimental complex are compared by DockQ, a 0–1 score combining interface RMSD, ligand RMSD and the fraction of native contacts. The CAPRI thresholds are acceptable ≥ 0.23, medium ≥ 0.49, high ≥ 0.80. For a small-molecule pose inside a co-folded complex the measure is ligand heavy-atom RMSD, conventionally ≤ 2 Å — which applies to a complex submitted here with a ligand entry. See inputs.

Read the accuracy column by cell, never down the column. The experimental row's "accuracy" is definitional — it is the thing the other two are graded against. The two in silico rows carry two different DockQ statistics (a success rate above a threshold versus a mean score), measured on different sets, one author-reported and one independent. Nothing meaningful comes from comparing them to each other; what the table supports is comparing time and cost at a stated accuracy.

On that reading, the honest position for this chain is: minutes instead of months, for an assembly hypothesis that is reliable on targets resembling the PDB and unreliable on novel interfaces — and least reliable precisely on the antibody–antigen case the tool is most often reached for.

The two asymmetries from the docking case apply here too, and one more. Structure determination can fail outright for a given complex and no amount of effort fixes it. It reports what is, while Chai-1 reports what is likely. And a co-folded complex comes with no RMSD and no DockQ: there is no reference-structure input, so if you want an agreement number for your own target you have to compute it yourself against a deposited structure. See results.

Agreement with wet lab

Chai-1's agreement with experiment has been measured on three different tasks, and the figures are not interchangeable.

Protein–protein interfaces, author-reported. On a set of structures released after Chai-1's training cutoff and filtered for low homology to its training set — 2362 protein–protein interfaces across 1054 PDB structures, reduced to 929 interface clusters — Chai-1's authors report a mean DockQ > 0.23 success rate of 0.751, against 0.677 for AlphaFold-Multimer 2.3 (two-sided Wilcoxon, p = 6.24 × 10⁻¹⁰). Without templates this falls to 0.743, not significantly. In single-sequence mode without MSAs — the mode the platform runs — it is 0.698, which the authors describe as comparable to AF2.3 given MSAs.

Antibody–protein interfaces, author-reported. On a 268-interface subset (129 structures, 122 redundancy-reduced clusters), the same report states Chai-1 outperforms AF2.3 by a significant margin on DockQ > 0.23 (p = 3.25 × 10⁻⁵), and that single-sequence Chai-1 performs similarly to Chai-1 given full MSAs on these interfaces.

Antibody–antigen interfaces, independent. This is where the picture changes. On 110 antibody–antigen complexes released after the 30 September 2021 training cutoff shared by all methods tested, determined by X-ray or EM at 3.5 Å or better, with 200 models generated per method per target, the mean DockQ of the top-ranked model was 0.12 at one sample, rising only to 0.14 at the largest sample size for both Chai-1 and Boltz-1. AlphaFold 3 on the same set went from 0.29 to 0.37, against a best-of-200 oracle of 0.52. The authors state plainly that "AlphaFold3 outperforms AlphaFold2, Boltz-1, and Chai-1", and that at a single model AlphaFold2, Chai-1 and Boltz-1 perform similarly. Sampling more does not rescue Chai-1 here, because its own confidence cannot identify its best model.

Ligand poses inside a co-folded complex, independent and prospective. This is the strongest published Chai-1 result, and it is a task you can run here with a ligand entry. Testing 551 ligands — 489 of sufficient quality to evaluate — bound to the SARS-CoV-2 NSP3 macrodomain (Mac1) and crystallographically determined after every method's training cutoff: "Using the field standard, 2 Å cutoff for correct predictions, AF3 and Chai-1 predicted roughly 70% of the ligands correctly. Boltz-2 performed less well… an overall predictive accuracy of 53%." Chai-1's iptm discriminated correct from incorrect poses with AUC 74.6, against 75.6 for AlphaFold 3's ligand pLDDT, 67.1 for DOCK3.7 energies and 53.8 for Boltz-2's predicted pIC₅₀. Confidence tracked measured potency only weakly for both AF3 and Chai-1.

That progression — 0.751 author-reported on low-homology protein–protein, 0.698 without MSAs, ~0.13 top-ranked DockQ on unseen antibody–antigen, and ~70% ligand-pose recovery on a prospective single-target set — is one model measured four ways. It is the most important thing to understand about a Chai-1 number.

Boitreaud J, Dent J, McPartlon M, et al. Chai-1: Decoding the molecular interactions of life. bioRxiv. 2024. doi: 10.1101/2024.10.10.615955

Fromm S, Ludaic M, Elofsson A. Evaluating deep learning based structure prediction methods on antibody-antigen complexes. Bioinformatics. 2026;42(4):btag136. doi: 10.1093/bioinformatics/btag136

Kim J, Correy GJ, Hall BW, et al. Large-scale prospective evaluation of co-folding across hundreds of Mac1-ligand complexes and three virtual screens. eLife. 2026;15:RP110475. doi: 10.7554/eLife.110475

Clifford JN, Nielsen M. Pocket restraints guided by B-cell epitope prediction improve Chai-1 antibody-antigen structure modeling. Protein Science. 2026;35(8):e70730. doi: 10.1002/pro.70730

What the figure doesn't cover

• The platform runs the configuration every paper identifies as the weak one. A chai_predict job calls Chai-1 with 3 trunk recycles, 200 diffusion timesteps, ESM embeddings, no MSAs, no templates, no restraints, a single fixed seed, and one returned structure. The published levers that improve Chai-1's antibody–antigen accuracy are: MSAs ("nearly doubled its success rate" on 1628 complexes), many independent runs (most of the gain within the first 10 of 30), and epitope restraints ("6%–8% more AbAgs achieving acceptable or medium DockQ" over 30-seed Chai-1, and in Chai-1's own report a single antibody–antigen distance restraint raised the DockQ-acceptable rate "from the baseline of 35% to 57%"). None of the three is reachable through this tool.
• Training overlap dominates the headline numbers, and Chai-1's cutoff is early. Chai-1's training cutoff is 12 January 2021. On complexes released after the cutoff, one independent study found accuracy "nearly halved on the independent 'After' test set" compared to pre-cutoff structures. Revealingly, 49% of the post-cutoff failures (21 of 43) still modelled the antigen below 3 Å and the antibody below 1.5 Å individually — the chains were fine and the antibody was simply placed in the wrong spot. A good per-chain plddt_mean is not evidence the interface is right.
• DockQ-acceptable is a low bar, and the platform reports neither DockQ nor RMSD. 0.23 is the CAPRI floor for an "acceptable" interface, not a usable model. There is no reference-structure input, so nothing on the platform tells you where your own result falls.
• Confidence cannot identify the model's own best answer. On the 110-complex antibody–antigen set, the gap between the top-ranked model and the best-of-200 model was the dominant error for every method; for Chai-1 and Boltz-1 more sampling bought essentially nothing. With a fixed seed and one returned structure, this platform does not sample at all.
• Two co-folding models agreeing is correlated evidence. On Mac1, Chai-1 and AlphaFold 3 pose RMSDs correlated at r = 0.72, while AF3 vs DOCK3.7 was r = 0.45 and AF3 vs Boltz-2 r = 0.52. The published conclusion was that co-folding and physics-based methods are complementary because their errors are less correlated — which is an argument against treating Chai-1 and Boltz-2 as two independent opinions.
• Conformational state is biased toward whatever the PDB holds. Benchmarking AlphaFold 3, Boltz-2, Chai-1 and BioEmu on four multi-state proteins found that models "frequently default to a dominant state represented in the PDB", with small-molecule ligands having weak or inconsistent effects. On Mac1, common conformational rearrangements including a peptide flip and a large loop opening were not recapitulated by any co-folding method.
• One ligand, and no DNA or RNA on this platform. A complex here takes protein chains plus at most one small molecule, and an entry typed dna, rna or glycan is refused rather than folded as a protein. Use Boltz-2 for those chains. The ligand-pose figures on this page also come from runs with MSAs, which this configuration does not use.
• No affinity, at all. Chai-1 returns no binding energy. Abriata and Dal Peraro's CASP16 practical-outcomes review called ligand affinity prediction generally "totally unreliable." For a model-predicted affinity see Boltz-2; for physics-based rescoring of a pose see MM-GBSA.
• No prospective confirmation of a Chai-1 protein–protein interface. No study has reported a Chai-1 interface predicted in advance and then resolved by a newly determined experimental structure. The Mac1 work is prospective but is a ligand-pose study on a single target.
• Do not borrow Chai-2 figures. Chai-2 is a separate, not-open-sourced generative model for de novo antibody and miniprotein design (16% de novo antibody hit rate across 52 targets). Its report credits its own folding module with "predicting 34% of antibody–antigen complexes with DockQ > 0.8 (compared to 17% for its predecessor, Chai-1)" — a DockQ > 0.80 "high" threshold, not the 0.23 floor used elsewhere on this page, and not comparable to any figure above. The platform runs chai_lab 0.6.1, which serves Chai-1.

Ye M, Wang YH, Brogi M, Parks JM, Kuo KM, Gumbart JC. Benchmarking AI protein structure predictors reveals a persistent bias in multi-state proteins. bioRxiv. 2026. doi: 10.64898/2026.07.10.737860

Chai Discovery Team. Zero-shot antibody design in a 24-well plate. bioRxiv. 2025. doi: 10.1101/2025.07.05.663018

Go / no-go

Go — act on the result when:

• Your complex is a protein–protein assembly whose chains have close structural relatives in the PDB, and you want an interface hypothesis to guide mutagenesis, epitope mapping or a structure-determination request.
• You will read iptm as an estimate of interface modelling accuracy and per-chain plddt_mean as an estimate of each chain's own fold, and nothing more. See results.
• You are using the predicted complex as one input among several, with a physics-based method rather than a second co-folding model as the cross-check.
• You will inspect the interface geometry before building on it. There is no physical-validity check in the result.

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

• Your target is an antibody–antigen or nanobody–antigen complex with no close PDB relative, and you intend to treat the returned structure as the binding mode. This is the case with the weakest published figures and the configuration here has none of the mitigations.
• You are ranking candidate binders by confidence. On nanobody–antigen ranking, no single confidence measure generalized to unseen antigens, and Boltz-2's confidence outperformed Chai-1's.
• You need an affinity, or any statement about binding strength. Nothing in the result carries one.
• Your complex contains DNA or RNA, or more than one small molecule. Those requests are refused rather than approximated. Use Boltz-2.
• You are relying on agreement between Chai-1 and Boltz-2 as independent confirmation. Their errors are correlated.
• You need the interface established rather than proposed. That requires the experimental structure.

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, sampling budgets, MSA availability and success criteria differ between them. Author-reported figures are labelled as such.