Find hits for a target by virtual screening

What each column measures:

• Round 1: each workflow's own selection, synthesised and assayed: Compounds selected and experimentally tested per workflow, confirmed hits and chemically novel hits among them, and the challenge's activity and novelty scores
• Round 1: every workflow applied to the same merged 1,739-compound library: Compounds predicted active and predicted weak on the merged library per workflow, confirmed Tier 1 and Tier 2 hits among them, and the challenge's normalised values

• Each workflow selected its own compounds, so differences between workflows mix the ranking method with the library and chemistry each team chose. The merged-library arm is the comparison on one library.
• The source prints no per-workflow hit rate; this batch records the numerator and the denominator as separate printed counts and does not divide them.
• The denominators are small and unequal, so a one-hit difference moves a workflow's apparent rate by several percentage points.
• The activity and novelty scores are committee judgements, not measurements.
• Transcribed from the source; no reproduction by Rewire and no human scientific review.
• No ligand-similarity or pharmacophore baseline is scored on any protocol in this batch, although issue #177 asks for one as the pre-machine-learning comparator alongside docking.
• The merged library is the union of what the workflows themselves selected, so it is not an independent library and each workflow's own picks are in it.
• Four of the 23 workflows (1694, 1695, 1716, 1717) have no values printed in these columns; they are absent, not zero.
• Only 29 confirmed hits scored above 15, four of them chemically novel, so the Tier 1 counts rest on four compounds in total.
• The two normalised columns are labelled as hit counts but defined in the article's figure caption as an aggregated score divided by the number of compounds predicted active, and the final column has no printed header.
• One target, one round and one organiser's library: nothing here shows how the ordering would hold on another target.

Protocols: CACHE #3 Round 1 own-selection outcome per workflow (Herasymenko et al. 2026 Table S8), CACHE #3 merged Round 1 library prediction per workflow (Herasymenko et al. 2026 Table S8)

• Only high-ranking docked molecules were synthesised, so the list is a biased sample of the 490-million library and no value from it estimates performance over a whole library.
• The comparison is a retrospective re-ranking of a prospectively tested set, not a new prospective screen.
• No training-overlap control is stated for this arm.
• Transcribed from the source; no reproduction by Rewire and no human scientific review.
• No ligand-similarity or pharmacophore baseline is scored on any protocol in this batch, although issue #177 asks for one as the pre-machine-learning comparator alongside docking.
• The source prints a single approximate value, 'both with AUC ~ 71', for DOCK3.7 and Boltz-2 pIC50; no separate figure is given, so neither result carries a numeric value.
• Only high-ranking docked molecules were synthesised, so the list is a biased sample of the 138-million library.
• The 1,293 tested compounds were sampled from scoring bins within the top 1% of 1.7 billion docked molecules, so the list is a biased sample of the library.
• The activity threshold for a true hit on this target is an apparent Ki below 400 uM, far weaker than the thresholds used for the two receptors.

Protocols: Re-ranking the Sigma-2 receptor docking hit list by co-folding and docking scores (Kim, Correy, Hall et al. 2026), Re-ranking the Dopamine D4 receptor docking hit list by co-folding and docking scores (Kim, Correy, Hall et al. 2026), Re-ranking the AmpC beta-lactamase docking hit list by co-folding and docking scores (Kim, Correy, Hall et al. 2026)

What each column measures:

• DEKOIS 2.0, built-in co-folding scores against docking tools (24 arms): AUROC, BEDROC (alpha = 80.5) and enrichment factors at 0.5%, 1% and 5%, mean and median over 79 DEKOIS 2.0 targets
• DEKOIS 2.0, rescoring of Protenix-predicted complexes (13 arms): AUROC, BEDROC (alpha = 80.5) and enrichment factors at 0.5%, 1% and 5%, mean and median over 79 DEKOIS 2.0 targets
• DEKOIS 2.0, rescoring of AF3-predicted complexes (10 arms): AUROC, BEDROC (alpha = 80.5) and enrichment factors at 0.5%, 1% and 5%, mean and median over 79 DEKOIS 2.0 targets
• GPCRrecent, 16 targets with post-2022 structures (16 arms): AUROC, BEDROC (alpha = 80.5) and enrichment factors at 0.5%, 1% and 5%, mean and median over 16 GPCR targets
• GPCRrecent, six targets after the Boltz-2 training cutoff (2 arms): AUROC, BEDROC (alpha = 80.5) and enrichment factors at 0.5%, 1% and 5%, mean and median over six GPCR targets

• Eight of the 24 arms are values quoted from five other papers, so they were not produced under one common pipeline.
• The two Boltz-2 affinity arms are flagged by the authors themselves as likely over-estimates from training-data overlap with the databases that source DEKOIS 2.0.
• Transcribed from the source; no reproduction by Rewire and no human scientific review.
• No ligand-similarity or pharmacophore baseline is scored on any protocol in this batch, although issue #177 asks for one as the pre-machine-learning comparator alongside docking.
• Structures the Protein Preparation Wizard could not process were given an extremely low score rather than excluded.
• The temporal control covers crystal structures only; the authors exclude Boltz-2's affinity module from this analysis because its binding-data training sources overlap the benchmark.
• AF3 was not run on this set.
• No docking or rescoring arm was run on these six targets, so there is no baseline to compare against within this protocol.

Protocols: DEKOIS 2.0 screening with co-folding confidence scores against docking tools (Shen et al. 2025 Table 1), DEKOIS 2.0 screening by rescoring Protenix-predicted complexes (Shen et al. 2025 Table 2, Protenix block), DEKOIS 2.0 screening by rescoring AF3-predicted complexes (Shen et al. 2025 Table 2, AF3 block), GPCRrecent screening (16 GPCR targets) (Shen et al. 2025 Table 3), GPCRrecent screening (the six targets whose first structures postdate the Boltz-2 training cutoff) (Shen et al. 2025 Table 3)

• MSA construction is excluded: pre-computed MSAs were fed as input, so these are inference times only.
• No docking tool was timed, so the table does not compare co-folding cost with docking cost.
• Single measurements on five targets with no repeats or uncertainty; the authors note that runtime varies substantially with token length and other factors.
• Transcribed from the source; no reproduction by Rewire and no human scientific review.
• No ligand-similarity or pharmacophore baseline is scored on any protocol in this batch, although issue #177 asks for one as the pre-machine-learning comparator alongside docking.

Protocol: Co-folding time per compound on five DEKOIS 2.0 targets (Shen et al. 2025 Table 5)