How Boltz-2 results compare to the bench

You have a Boltz-2 result. Here is what it supports.

• The predicted complex is a geometric hypothesis. For the small-molecule part the published agreement measure is ligand heavy-atom RMSD against a deposited structure, conventionally ≤ 2 Å. On the largest independent prospective test — 489 ligand complexes crystallographically determined after every method's training cutoff — Boltz-2 reached 53%, against roughly 70% for AlphaFold 3 and Chai-1 and 40% for DOCK3.7 on the same set.
• The affinity number is graded against real measured potency, which is unusual — but it ranks, it does not measure. Against measured IC50 it reaches Pearson r = 0.6 on a 202-compound prospective set and ρ = 0.68 inside a single congeneric series, yet falls to ρ = 0.31 when four series of one target are pooled, and to no correlation at all against physics-based free energies for a top-100 shortlist. Treat it as a triage score, never as a ΔG.
• The platform runs the thin configuration. No MSA unless you ask for it, one diffusion sample, no steering potentials, no templates, no pocket constraint, no experimental-method conditioning. Every published figure on this page was produced with more than that. See what the figure doesn't cover.
• The confidence fields estimate pose accuracy, not binding. Boltz-2's ligand ipTM separated correct from incorrect poses at AUC 75.2; its own predicted pIC50 separated them at AUC 53.8 — near random — and the two are architecturally decoupled (Pearson r = −0.001).
• Confirm the structure with X-ray crystallography or cryo-EM of the complex. Confirm engagement with a biophysical binding measurement (SPR, MST, ITC). Confirm affinity with MM-GBSA rescoring and then an assay.

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 a complex you cannot yet see, and to decide which complexes are worth an experiment. In a live program that means four things:

• Complex hypothesis from sequence. You have chains and no structure you trust, and you want a model of the assembly to reason about — which residues form the interface, where to grow a substituent, what a crystallography or cryo-EM request should be looking for.
• Library triage by predicted affinity. The accepted use of a Boltz-2 affinity value is as a cheap pre-filter in front of something rigorous. A study that recommends it inside a workflow is explicit about the shape: "use an efficient AI approach like Boltz-2 to prefilter large compound libraries (on the order of 10⁵ ligands), followed by rigorous alchemical ABFE calculations on a few hundred top-ranked hits."
• Rescoring a docking hit list. Co-folding scores are run over molecules a docking campaign already ranked, on the argument that their errors are independent of the docking score's. Measured on three prospective screens, Boltz-2's predicted pIC50 did separate actives from docking false positives (AUC 83.8, ~71 and 68.1 on the σ2 receptor, dopamine D4 and AmpC β-lactamase), but "showed little advantage over docking, a method that is about four orders of magnitude faster."
• Orthogonal check on another method's pose. Co-folding the same receptor sequence plus ligand after a DiffDock run, to see whether a different architecture places the ligand in the same region.

What co-folding is not used for, by anyone reporting their validation honestly, is settling a binding mode or quoting an absolute binding free energy. CASP16's user-facing assessment said of the whole ligand track that "predicting binding affinities proved extremely challenging, with very poor correlation with the experimentally measured values", and added that handing predictors the real experimental complex in a second stage did not improve affinity accuracy at all — the limitation is the scoring, not the pose.

Against other in silico methods

Against AlphaFold 3 and Chai-1, head to head. Three independent comparisons put Boltz-2 behind AlphaFold 3 on structural accuracy, and two of the three put it behind Chai-1 as well. On the 489 prospective Mac1 complexes, Boltz-2 reached 53% ≤ 2 Å against roughly 70% for both AlphaFold 3 and Chai-1. On cross-docking across three Alzheimer's targets, AlphaFold 3 beat Boltz-2 in every scenario when the top-ranked pose was scored, and the authors attribute the gap to confidence calibration rather than data: "Boltz-2 frequently sampled native-like conformations but often failed to rank them as its highest-confidence prediction", and "the most striking observation was the consistent superiority of AlphaFold 3 over Boltz-2 despite the latter having access to a larger and more recent structural corpus." On an allosteric-site benchmark Boltz-2 was the weakest of five models, with mean ligand-pose RMSD of 3.8–4.1 Å even on orthosteric sites where AlphaFold 3, Protenix and Chai-1 reached 2.3–2.8 Å.

It is worth noting one place Boltz-2 came out ahead on geometry: across 12 wild-type kinase–inhibitor complexes, median ligand RMSD was 0.80 Å for Boltz-2, 0.95 Å for AlphaFold 3 and 1.87 Å for Chai-1, with Boltz-2 and AlphaFold 3 statistically indistinguishable (Wilcoxon, p = 0.791). Those are heavily represented PDB kinases, which is exactly the regime where co-folding is strongest.

Against docking. Boltz-2 clearly beats docking at pose prediction when no receptor conformation is on hand: 53% against DOCK3.7's 40% on the prospective Mac1 set, and 82.3% against GNINA 1.3 rigid's 5.5% under cross-docking. It does not clearly beat docking at telling ligands from non-binders. Rescoring three prospective screens, docking out-performed Boltz-2 on AmpC (AUC 76.5 vs 68.1), the two were level on dopamine D4 (both ~71), and Boltz-2 won on the σ2 receptor (83.8 vs 78.8) — with the caveat, stated by the authors, that "the hit lists were composed of molecules prioritized by docking in the first place, giving it an advantage on these particular sets." For affinity ranking on an ordinary medicinal-chemistry campaign, the empirical Vina scoring function beat Boltz-2 on three of four series and on the pooled set (Vina ρ = 0.45 / 0.55 / 0.57 / 0.67 against Boltz-2's 0.31 / 0.68 / 0.17 / 0.55 for pooled, Series 1–2, Series 3 and Series 4) — though at n = 12–18 per series the bootstrap confidence intervals are wide enough that only the pooled comparison with FEP+ reached significance.

Against free-energy methods. Boltz-2's authors position it as "the first AI model to approach the performance of free-energy perturbation (FEP) methods in estimating small molecule-protein binding affinity" while being "at least 1000× more computationally efficient than FEP." Independent testing supports the speed claim and qualifies the accuracy claim sharply. Within one congeneric series of conservative modifications it did match FEP+ — ρ = 0.68 against 0.64, statistically indistinguishable at n = 18 — which is consistent with the "approach FEP" framing: on the congeneric benchmarks curated for FEP methods, Boltz-2's reported average per-target Pearson r is 0.62 against FEP+ at 0.72, within about 0.10. Across series and on novel chemistry it does not: pooled ρ = 0.31 against FEP+ at 0.70, no correlation with ensemble MM-PBSA-class free energies on a top-100 shortlist, and 81% true-positive rate against 86–100% for alchemical absolute binding free energy.

Against Boltz-1. The headline comparison here is not an improvement. Benchmarking against systems released after each model's training cutoff, stratified by similarity to the training set, "the additional data used for training (more than 25,000 PDB entries) increases the number of cases in the higher similarity bins but does not seem to help with generalization to unseen data", and in the discussion: "the generalizability of Boltz-2 did not discernibly improve compared to Boltz-1, despite being trained on more data."

Škrinjar P, Eberhardt J, Studer G, Tauriello G, Schwede T, Durairaj J. Evaluating generalization in protein-ligand cofolding methods. Nature Structural & Molecular Biology. 2026;33(5):782-794. doi: 10.1038/s41594-026-01797-5

Sun K, Head-Gordon T. A curated benchmark for cofolding models on kinase conformational states. npj Drug Discovery. 2026;3(1). doi: 10.1038/s44386-026-00068-z

Parikh V, Foley B, Gatlin W, Ludwick M, Turano L, Verkhivker GM. Explainable AI reveals the allosteric blind spot in protein-ligand binding predictions. Cell Reports Physical Science. 2026;7(8):103448. doi: 10.1016/j.xcrp.2026.103448

Passaro S, Corso G, Wohlwend J, et al. Boltz-2: towards accurate and efficient binding affinity prediction. bioRxiv. 2025. doi: 10.1101/2025.06.14.659707

In silico vs. wet lab

The standard wet-lab approach for the question Boltz-2 answers is experimental structure determination of the complex — X-ray crystallography or cryo-EM — plus a separate biophysical or biochemical measurement (ITC, SPR, or an IC50 assay) for the affinity half. Structure determination resolves the assembly directly: which chains contact which, at which residues, in which conformation.

The comparison is well defined for the structural half, which is why this section can carry real numbers. A predicted and an experimental complex are compared by ligand heavy-atom RMSD for the small molecule, conventionally with a ≤ 2 Å success threshold, often paired with lDDT-PLI ≥ 0.8 for the local protein–ligand interaction environment; for a protein–protein or antibody–antigen interface the measure is DockQ (acceptable ≥ 0.23). For the affinity half there is no equivalent geometric threshold — agreement is reported as a correlation or a mean absolute error in log units or kcal/mol against measured IC50 or Ki.

Read the accuracy column by cell, never down the column. The experimental row's "accuracy" is definitional — it is the thing the other three are graded against. The three in silico rows carry success rates from different sets with different training-overlap properties, and the affinity figure in the last row is a correlation, not a geometric success rate. Nothing meaningful comes from comparing those numbers 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 a complex hypothesis that is reliable where the PDB is dense and degrades where it is thin — and an affinity value good enough to order compounds within a series and not good enough to price one.

The spread between the two structural figures in the bottom row is the most useful thing on this page. The 82.3% was measured on three heavily studied targets whose complexes Boltz-2's training set "effectively cover[s] nearly all" of; the 53% was measured on structures no method had seen. Both are real. The second describes a novel target.

Three asymmetries matter beyond the table. Structure determination can fail outright for a given complex and no amount of effort fixes it, which is a large part of why co-folding exists. It reports what is, while Boltz-2 reports what is likely. And a Boltz-2 result carries no RMSD and no reference: there is no experimental-structure input, so if you want an agreement number for your own target you must compute it yourself against a deposited structure. See results.

Agreement with wet lab

Boltz-2's agreement with experiment has been measured on two different questions — structure and affinity — and the figures are not interchangeable.

Structure, against deposited experimental structures

The strongest independent prospective test. As part of a real drug-discovery campaign against the SARS-CoV-2 NSP3 macrodomain (Mac1), 551 ligand complexes were crystallographically determined — typically near 1 Å resolution — and never previously disclosed, so the structures themselves could not have been in any method's training set; 489 were of sufficient quality to evaluate. Quoting the source directly:

"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, with more structures shifting to the 2 Å distribution, and an overall predictive accuracy of 53%. This result was surprising because the later training cutoff date for Boltz-2 meant that it could potentially benefit from additional Mac1 ligand-bound structures."

On the coarser question of whether the ligand landed in the right part of the site — ligand centre of mass within 2.5 Å after alignment — Boltz-2 placed 444 of 489 (91%), against 449 (92%) for AlphaFold 3 and 440 (90%) for Chai-1. So the common failure is orientation inside the right pocket, not gross mislocation.

Cross-docking against crystal structures, with near-total training overlap. Across acetylcholinesterase, BACE1 and GSK-3β, under paired non-parametric analysis with Holm correction, cross-docking success at HA-RMSD ≤ 2 Å was AlphaFold 3 90.0% (95% CI 87.0–92.3), Boltz-2 82.3% (78.4–85.6), DiffDock 72.9% (69.8–75.9) and GNINA 1.3 rigid 5.5% (4.2–7.2). Under each method's own top-ranked pose rather than the best sampled pose, Boltz-2 falls to 82.1% and AlphaFold 3 to 88.2%; taking the single best of the sampled poses, Boltz-2 reaches 89.5% at a mean HA-RMSD of 1.07 ± 1.59 Å (n = 95). Per target, Boltz-2's top-ranked pose failed for 54.5% of 22 acetylcholinesterase complexes, 16.0% of 337 BACE1 complexes and 14.1% of 64 GSK-3β complexes. The authors state that Boltz-2's training through 30 June 2023 "effectively cover[s] nearly all benchmark complexes evaluated here", making this "a stringent upper-bound scenario."

Protein–peptide complexes. On a protein–peptide benchmark, Boltz-2 produced a high-quality pose (CAPRI criteria) in its top-ranked prediction 68.4% of the time, rising to 73.7% within the top 20 — the highest proportion of high-quality models among the methods examined. On a complex-level non-redundant subset, however, Boltz-2 reached only 10%, against 30% for AlphaFold-Multimer.

Nanobody–antigen interfaces. Ranking true nanobody–antigen complexes above non-binding bait pairs across 17 antigens, Boltz-2's confidence measures achieved the highest median ROC AUC of the four models tested, though the three single best-performing measures came from IntFold and AlphaFold 3 (median ROC AUC 0.90, 0.88 and 0.86). On a real camelid immunisation campaign against CD33 that ranking did not transfer: Boltz-2's measures "perform poorly" there, and all four models showed limited generalisation.

No CASP assessment covers Boltz-2. CASP16's pharmaceutical protein–ligand track used Boltz-1, not Boltz-2, as a post-hoc deep-learning baseline, and CASP16's own practical-outcomes paper records that "the latest Boltz-2 model purportedly capable of modeling complexes and predicting affinities, was not well represented among CASP16 predictors." Any Boltz-2 figure attributed to CASP is a misattribution.

Affinity, against measured potency

This is the rarer comparison, and the one Boltz-2 is distinctive for making. It is a comparison against an assay readout, not against a structure.

Prospectively, against measured IC50. On the 202 Mac1 compounds with full eight-point dose–response HTRF IC50 values, Boltz-2's predicted pIC50 correlated at Pearson r = 0.6, significantly better than DOCK3.7 energies (r = −0.225, Friedman and Conover with Bonferroni, p = 0.0057) and better than AlphaFold 3's confidence (r = 0.314, not significantly). Before any correction the mean absolute error was ~1.6 kcal/mol (1.19 pIC units), with most compounds predicted too potent; an empirical linear fit reduced it to ~0.7 kcal/mol (0.54 pIC units), against a baseline of 0.93 kcal/mol for simply predicting the dataset mean for every compound.

Across resistance mutations and a kinase panel. Over 57 kinase–inhibitor pairs (eight human kinases, nine clinical inhibitors, representative resistance variants), Boltz-2's predicted log10(IC50) agreed with measurement at Pearson r = 0.820, Spearman ρ = 0.816; in leave-one-pair-out testing r = 0.806–0.821 with RMSE 0.799–0.834 log units, against the strongest simple baseline (inhibitor-mean potency) at r = 0.531 and RMSE 1.168 log units. The authors are careful about what this establishes: "relative kinase-inhibitor interaction patterns were reproduced more consistently than absolute potency values", with a compressed dynamic range throughout.

Inside and across a real lead-optimisation campaign. On four related series against 15-hydroxyprostaglandin dehydrogenase (43 compounds, Boltz-2 at default settings with the default MSA server), rank correlation with measured potency was ρ = 0.68 for the 18 compounds closest to published inhibitors — statistically indistinguishable from FEP+ at ρ = 0.64 — and ρ = 0.55 for a 12-compound series. For the 13-compound series that explored a novel subpocket by displacing structured water it was ρ = 0.17, against FEP+ at 0.78, and Boltz-2 predicted every member to be less potent than the published parent despite experimental potency gains of up to 500×. The authors' conclusion is the operative one: "Used prospectively to evaluate these specific changes, none would have been synthesized." Pooling all four series, Boltz-2 fell to ρ = 0.31 against FEP+ at 0.70 (padj < 0.05) — a Simpson's-paradox effect from systematic between-series bias.

Against a curated affinity set, and against physics. On a 632-complex PDBbind set, Boltz-2's affinity predictions correlated with experiment at Pearson r = 0.58, Spearman ρ = 0.58; discriminating DUD-E actives from decoys across nine targets it reached a true-positive rate of 81%, the best of seven machine-learning methods tested (RF-Score-VS 79%, GNINA 38%, the rest 20–38%), though alchemical ABFE reached 86–100%. Benchmarked instead against ensemble physics-based free energies over 38,482 compounds, correlation was r = 0.45 / ρ = 0.46 for tankyrase 2 and r = 0.24 / ρ = 0.25 for the SARS-CoV-2 main protease; restricted to each target's Boltz-2 top 100, correlation collapsed to approximately zero, and the top-100 lists from the two methods overlapped in 2 and 1 compounds respectively.

The same study reports Boltz-2 at r = 0.77, ρ = 0.78 against 1,137 BindingDB measurements for tankyrase 2, and then argues against reading it as generalisation: compounds deposited after Boltz-2's training cutoff "cluster tightly within the exact same narrow islands already defined by the historical" data, and a bootstrap over 271 compounds with replicate measurements put the correlation between two independent experimental values of the same compound at only r = 0.56 ± 0.25. An r of 0.77 against data whose own internal reproducibility is 0.56 is a figure about the label set as much as about the model.

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. doi: 10.7554/eLife.110475

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

Suenaga A, Okimoto N. Benchmarking Boltz-2 for relative kinase-inhibitor interaction patterns across conformational states and resistance mutations. npj Drug Discovery. 2026;3(1). doi: 10.1038/s44386-026-00086-x

Rich AS, Dodda LS, Price DJ. Evaluating machine learning affinity prediction methods on novel lead optimization data: a 15-PGDH case study. ChemMedChem. 2026;21(20):e70531. doi: 10.1002/cmdc.70531

Wan S, Zhang X, Xue X, Coveney PV. Reliability of AI methods in drug discovery: evaluation of Boltz-2 for structure and binding affinity prediction. Journal of Chemical Theory and Computation. 2026;22(15):7811-7824. doi: 10.1021/acs.jctc.6c01334

Wang X, Hentabli H, Mettu A, Gautam S, Kireev D. The last mile problem: a critical assessment of physics-based and AI tools for small molecule binding prediction in virtual screening. Journal of Chemical Information and Modeling. 2026;66(10):5964-5977. doi: 10.1021/acs.jcim.6c00942

Fonteyne M, Badr N, Kuppen PJK, Vahrmeijer AL, van Westen GJP, Jespers W. Molecular docking versus co-folding for protein-peptide pose prediction: effects of protein flexibility and training set overlap. Digital Discovery. 2026;5(10):4471-4488. doi: 10.1039/d6dd00242k

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

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

What the figure doesn't cover

• The platform's configuration is thinner than every published figure's. Boltz-2 runs here as boltz 2.2.1 with the released boltz2_conf.ckpt and boltz2_aff.ckpt weights, 3 recycling steps and 200 sampling steps — and with one diffusion sample, no steering potentials, no templates, no pocket constraint, no experimental-method conditioning, and no MSA unless you set use_msa_server. The studies above ran it with MSAs on and, in several cases, 25 structures per system (five seeds × five samples) or ten independent runs per complex, then reported the best or the median. You get one sample and whatever its confidence happens to be. See inputs.
• Training overlap dominates the headline numbers. The spread 53% → 82% is one model measured against two very different degrees of exposure to its own training set, on structures it had never seen versus targets its training "effectively cover[s] nearly all" of; the further step 82% → 89.5% is not exposure at all but selection rule, the difference between the pose Boltz-2 ranked first and the best pose it sampled — and on this platform you only get the first. Benchmarking stratified by training-set similarity found that successes in the low-similarity bins come mostly from ligands with many analogous occurrences in the PDB, and that "data scarcity is one of the drivers of memorization." If your target has no close relative in the PDB, the 53% figure describes your situation better than the 82%.
• Protein conformational change is not modelled, even when the pose is right. On the prospective Mac1 set, global C-alpha RMSD was 0.1–0.4 Å but "rare local and high magnitude conformational changes observed experimentally were not captured by co-folding methods" — a peptide flip and an ~8 Å loop opening were both missed. On a 2,225-chain kinase benchmark, Boltz-2 recovered the correct conformational state (all eight KinCoRe labels) for 67.9% of geometrically successful predictions, showed over 40% "apo-drift" — predicting the ligand-free state for a holo complex — even for complexes in its training set, and collapsed to near-zero ensemble diversity (~0.1 Å in distance metrics, under 5° in torsions across 20 inferences, with roughly two-thirds of systems all-correct or all-failed). So the structure you get is not a conformational hypothesis about your ligand's induced fit.
• Allosteric sites are a documented blind spot. On two independent allosteric benchmarks, Boltz-2's success at 2 Å was 23–24% under best-of-25 selection and 9–11% under mean-of-25, and median pocket RMSD up to 6.2 Å. On the same paper's 2,275-complex orthosteric reference, AlphaFold 3 and Protenix both exceeded 80% success at the same 2 Å threshold. The gap persisted across all five models tested and across every selection strategy, so it is not a Boltz-2 defect specifically — but it is the regime where a co-folded pose is most likely to be confidently wrong.
• Ligand chemistry can come back altered. For the top-100 shortlists on two targets, 28 of 100 and 30 of 100 predicted ligands had hydrogen counts and indices of hydrogen deficiency that disagreed with their own input SMILES — rings predicted more unsaturated than specified, aliphatic chains more saturated. The platform runs no physical-validity check, so check saturation, stereochemistry, bond geometry and clashes against your SMILES before building anything on a pose.
• Confidence is compressed and over-stated. Across two libraries of 16,780 and 21,702 compounds, every ligand scored above 0.80 confidence and the means were 0.946 and 0.944, so "thresholds below 0.9 provide little discriminatory power." Only above 0.95 does the score separate anything, and the separation is target-dependent.
• The affinity value is not reproducible to better than it needs to be. Two independent inference runs over ~800 ligands per target agreed at Pearson r = 0.913 and 0.962, yet individual predictions differed by up to ~1.5 kcal/mol — against a prediction range of roughly −4 to −9 kcal/mol for the great majority of ligands. That is enough to reorder a shortlist.
• Affinity and structure are architecturally decoupled. Boltz-2's predicted affinity had Pearson r = −0.001 with its own pose confidence on the Mac1 set, and the one study that looked directly found "larger ligand RMSD values were not associated with larger prediction errors in magnitude." A good pose does not validate the affinity number, and a good affinity number does not validate the pose.
• The binder/non-binder probability and the affinity value are separate answers. Boltz-2 emits a binary binding probability alongside the affinity value, and classification is the task the literature finds it most reliable at. This platform returns both: affinity_probability_binary (0–1, no cut-off applied) next to affinity_score from affinity_pred_value. The probability head is internally inconsistent with the affinity head: correlations between the two were only 0.21 and 0.25 across two large libraries, so do not read one as confirming the other.
• No in silico to in vivo comparison exists. No published study measures how often a Boltz-2 prediction predicts an animal or clinical outcome, and sequential workflows that co-fold and then run biology are not concordance measurements. The real axes are the two above.
• No prospective co-crystal confirmation of a Boltz-2 pose. The Mac1 study is prospective in the sense that the structures were determined after the training cutoff and withheld — which is the strongest design available — but no study reports a Boltz-2 pose predicted in advance and then resolved by a newly solved structure of a previously unknown complex.

Go / no-go

Go — act on the result when:

• You need a complex hypothesis to guide mutagenesis, SAR, or a structure-determination request, and your target or a close homolog is well represented in the PDB.
• You are ordering compounds within one congeneric series on predicted affinity, as a filter in front of FEP, MM-GBSA or an assay — and you accept that the ranking is good and the absolute value is not.
• You are co-folding as a second opinion on a DiffDock pose or a docking hit list, treating agreement as weak corroboration rather than confirmation.
• You will inspect the returned ligand against your input SMILES before anything downstream.
• You turned use_msa_server on, or you have read single_sequence_mode in the result and discounted the confidence accordingly.

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

• You are comparing predicted affinities across series or scaffolds. The documented failure is a systematic per-series offset, and pooling destroyed the correlation in the one campaign-scale test published.
• You need an absolute binding free energy, or a ΔΔG for a design decision. Use MM-GBSA and then FEP, and then an assay.
• The site is allosteric, cryptic, or defined by a ligand-induced conformational change. Success at 2 Å falls to 9–24% there, and confidence does not flag it.
• Your target is structurally novel with no close PDB relative, and you intend to treat the single returned structure as correct. One sample from a model with documented mode collapse is not an ensemble.
• You are relying on the pose to validate the affinity, or the affinity to validate the pose. The two heads are decoupled by design.
• You need the binding mode 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, selection rule and sample size. Figures from different studies are not directly comparable, because test sets, training-overlap properties, sampling budgets and MSA settings differ between them — and several of the published figures were obtained under settings this platform does not expose.