Choose structural hypotheses to guide experiments

• Scored target populations differ; Table 1 assessable counts are retained as coverage context, not verified metric denominators. Some Table 3 rates do not reconcile with integer counts after rounding. No complete-cohort estimate is inferred.
• No intervals in Table 3; no prospective experimental utility or matched conventional mutation/construct baseline.
• Automated source review only; independent human scientific review remains outstanding.

Protocol: FoldBench protein-protein published full-set assessment

What each column measures:

• Lipid poses on 36 post-cutoff complexes (LiPP): Share of 36 post-cutoff lipid-protein complexes whose predicted lipid pose counts as a success (lipid pose all-atom RMSD below 2 Angstrom)
• Lipid poses on 36 post-cutoff complexes (LiPP): Share of 36 post-cutoff lipid-protein complexes whose predicted lipid pose counts as a success (RMSD below 2 Angstrom and PoseBusters valid)

• Author-run predictions, not an independent reproduction by Rewire.
• The PoseBusters columns in the same tables are literature values compiled by the LiPP authors and are recorded as paper_compilation.
• Pose accuracy only; the paper's confidence-score analysis is in figures and was not extracted.
• Several printed test-set percentages are truncated rather than rounded (for example 11 of 36 printed as 30.5); the values are kept as printed.
• The full LiPP set and the PoseBusters columns are not linked to the use case: the full set includes complexes inside some tools' training windows, and the PoseBusters values are literature compilations on drug-like ligands.

Protocols: LiPP Table 2: lipid pose all-atom RMSD below 2 Angstrom, LiPP test set, LiPP Table 3: RMSD below 2 Angstrom and PoseBusters valid, LiPP test set

What each column measures:

• Accuracy of five tools (FoldBench): FoldBench success rate, LDDT, interface RMSD and ligand RMSD for AlphaFold 3, Boltz-1, Chai-1, HelixFold 3 and Protenix on antibody-antigen targets with low homology to earlier PDB entries
• Picking AlphaFold3 models by confidence (independent): Mean DockQ of the AlphaFold3 model picked by each of five scores a user can compute (ipTM, pTM, ranking confidence, pDockQ2 and ipSAE) among 200 models per target, and Spearman correlation with DockQ over all models and per target, on 110 antibody-antigen complexes released after the training cutoff

• Benchmark authors ran all five tools; per-tool assessable counts differ, so the scored populations are not matched.
• Structural success against the deposited complex, not binding, affinity or the outcome of an experiment.
• No uncertainty is printed.
• Antibody-antigen targets only.
• The mapped evaluations are stored records reviewed in an earlier batch; this review checked their status, their links and their fit, not their values against FoldBench.
• Antibody-antigen complexes only; results do not transfer to other complex classes.
• Structural accuracy against the deposited complex, not binding, affinity or the outcome of an experiment.
• Only AlphaFold3 models are ranked in Table 1; the source does not print the same table for AlphaFold2.3, Boltz-1 or Chai-1.
• Table 1 does not state the number of models ranked per target. Its DockQ row (0.544) matches the best-of-200 value given in the text (0.54), so the table is read as selection among all 200 models; this is an inference.
• The DockQ, aeTM, aeiTM and aeRankConf rows use the experimental structure, so they are upper bounds, not usable selection rules.
• Figure 6 and the text give a per-target ranking-confidence correlation of 0.28, while Table 1 prints 0.214 for RankConf under a Spearman caption; the figure's correlation type is not stated.
• The four oracle rows of Table 1 (DockQ, aeTM, aeiTM, aeRankConf) are excluded: each needs the experimental structure, so none is a rule a user could apply. The gap between the best achievable pick and the best computable score is kept as a descriptive claim.
• Section 3.5 gives the mean per-target correlation for ranking confidence as 0.28 while Table 1 prints 0.214; the stored value is the table's and carries the conflict as a source warning.

Protocols: FoldBench antibody-antigen published full-set assessment, Picking the top AlphaFold3 model per antibody-antigen target with a confidence score

• Benchmark authors ran all five tools; per-tool assessable counts differ, so the scored populations are not matched.
• Structural success against the deposited complex, not binding, affinity or the outcome of an experiment.
• No uncertainty is printed.
• Protein-ligand targets only.
• The mapped evaluations are stored records reviewed in an earlier batch; this review checked their status, their links and their fit, not their values against FoldBench.

Protocol: FoldBench protein-ligand published full-set assessment

What each column measures:

• Best of N samples: Median best DockQ among 1 and among 100 samples for AlphaFold3, Boltz-2, Boltz-1 and Chai-1 on 106 nanobody-antigen complexes
• Confidence against accuracy: Pearson correlation of ipTM with DockQ, share of confident failures and unconfident successes, and correlation of ipTM change with DockQ change under sampling, for AlphaFold3, Boltz-2 and Chai-1 on 106 nanobody-antigen complexes
• Confidence as evidence of binding: Average precision for ranking cognate above shuffled nanobody-antigen pairings by ipTM, for AlphaFold3, Chai-1 and Boltz-2, against a random baseline of about 0.011

• Best-of-N DockQ needs the experimental structure to pick the best sample, so it is an upper bound on what a user could select.
• Each sampling depth is a separate run with its own seed (seeds 1 to 5 for N = 1, 10, 25, 50 and 100), unlike the single 50-sample runs of the other protocols.
• For Chai-1 the smallest depth is 5 samples (5 trunk samples x 1 diffusion sample), so its N = 1 value is the best of 5.
• Boltz-1 appears only in this analysis.
• Nanobody (VHH)-antigen complexes only; results do not transfer to conventional antibodies or other complex classes.
• The set mixes systems inside and outside each tool's training data (AF3 30 of 106 in training, Chai-1 25, Boltz-2 64); the printed values are over all systems, so they are not post-cutoff results, and Boltz-2's are the most exposed.
• Preprint, not peer reviewed. Supplementary tables were not read.
• No uncertainty is printed for these values.
• The dataset paragraph (P62) says post-October 2021 depositions were kept, 'corresponding to the earliest training cutoff among the evaluated tools (Boltz-2)'. The methods (P76) give Chai-1 about 12 January 2021, AF3 about 30 September 2021 and Boltz-2 about 1 June 2023, so Chai-1 is the earliest and Boltz-2 the latest, and the October 2021 date matches AF3. The per-tool train and test counts follow the methods, and no printed value depends on the mis-stated sentence.
• Preprint, not peer reviewed.
• Calibration is computed on cognate complexes only; it says nothing about whether ipTM separates binders from non-binders.
• Quadrant thresholds (DockQ 0.23, ipTM 0.5) are the source's choice, and the text does not say which sample (best, first or worst) the quadrant shares refer to.
• Not every tool has every value printed in the text.
• Negatives are shuffled pairings assumed to be non-binders; none was tested, and some may bind.
• Ranks pairings by ipTM; it measures whether confidence separates observed from assumed-absent interactions, not structural accuracy.

Protocols: Best DockQ against sampling depth on cognate nanobody-antigen complexes, How well ipTM tracks DockQ on cognate nanobody-antigen complexes, Telling cognate from shuffled nanobody-antigen pairs by ipTM

• Cannot compare numerical differences directly to FoldBench because datasets, structures and protocols differ.
• A successful docking pose does not establish binding, affinity, mechanism or cellular interaction.
• One of102 others targets was unevaluable; scores use101.
• No model execution or human domain review; exact software revision unextracted.

Protocol: ClusPro BM5 Others top10

• Cannot compare numerical differences directly to FoldBench because datasets, structures and protocols differ.
• A successful docking pose does not establish binding, affinity, mechanism or cellular interaction.
• All88 enzyme targets evaluated.
• No model execution or human domain review; exact software revision unextracted.

Protocol: ClusPro BM5 Enzyme top30

• Cannot compare numerical differences directly to FoldBench because datasets, structures and protocols differ.
• A successful docking pose does not establish binding, affinity, mechanism or cellular interaction.
• All88 enzyme targets evaluated.
• No model execution or human domain review; exact software revision unextracted.

Protocol: ClusPro BM5 Enzyme top10

• Cannot compare numerical differences directly to FoldBench because datasets, structures and protocols differ.
• A successful docking pose does not establish binding, affinity, mechanism or cellular interaction.
• One of102 others targets was unevaluable; scores use101.
• No model execution or human domain review; exact software revision unextracted.

Protocol: ClusPro BM5 Others top30