DEKOIS 2.0 screening with co-folding confidence scores against docking tools (Shen et al. 2025 Table 1)

Stable ID: hitdisc-20261022-protocol-shen-dekois-builtin

areas

molecular-interactions

contexts

research

task

Protein-ligand virtual screening (library ranking)

publication status

peer_reviewed

review

method: scope-review; automated-source-review; reviewer: claude; reviewer note: Claude review agent, independent of the extraction agent; no human review claimed; date: 2026-10-22; method note: Read the record against the pinned sources and against docs/use-cases.md and docs/record-contract.md.; note: Record checked against the pinned source and the record contract. Transcription checked; not reproduced and not reviewed by a human scientist.; artifact sha256: 11a4cee97057d9e72f9393024d22fdd169e3134f493367eee2f7869facf9ccf8; retrieval url: https://www.ebi.ac.uk/europepmc/webservices/rest/PMC12704515/fullTextXML; receipt: data/omics/use-case-coverage-hitdisc-20261022/review.json

metric

enrichment-factor

metric direction

higher

unit

unitless

metric qualifier

mean and median over the DEKOIS 2.0 targets

metric definition

AUROC, BEDROC with alpha = 80.5 (the top 2% of the ranked list contributes 80% of the score), and enrichment factors at 0.5%, 1% and 5%, each reported as the mean and the median over targets. EFx% is the ratio of true positives found in the top x% of the ranked compounds relative to random selection.

protocol

Co-folding from protein sequence and ligand SMILES with pre-computed MSAs, then ranking by the model's own confidence score; docking tools rank the same compounds for the same targets.

denominator

79

version

Shen et al. 2025, Table 1

source locator

Table 1 and its footnotes a to f

selection bias

The quoted baseline rows were produced by other groups under their own protocols; only the co-folding rows and the three Gnina rows were run by these authors.

limitations

Enrichment factors depend on the actives-to-inactives ratio and on how the negatives were built, which differ by more than an order of magnitude between the benchmarks in this store (1:30 for DEKOIS 2.0, 1:50 for GPCRrecent, and from about 1:43 to about 1:200 across the five LIT-PCBA targets scored here, reaching about 1:18000 for LIT-PCBA ADRB2). Values from different protocols are not comparable.; The decoys are property-matched synthetic molecules, not assayed inactives.; Eight of the 24 rows are quoted from other papers (footnotes b to f: refs 34, 35, 36, 38 and 39), so they were not run under one common pipeline; none of those references was retrieved in this pass.; The SurfDock row's indicators were calculated on all 81 targets rather than the 79 used elsewhere in the table, because the per-target labels were unavailable (footnote f).; The authors warn that the Boltz-2 Affinity and Probability rows are likely to be significantly over-estimated because Boltz-2 fetched affinity data from the public databases that also source DEKOIS 2.0, and they exclude those two rows from their own primary comparison (footnote a and Methods 'Affinity score and probability').; DEKOIS 2.0 was released over ten years before this evaluation and its template protein structures may overlap the co-folding models' training data; the ligand-similarity analysis of Figure 5 is the authors' partial control and was not extracted.

scope note

Retrospective: the method ranks a curated set of actives and negatives that were assembled before the evaluation, not compounds selected and then tested.

• evaluates task: Protein–ligand virtual screening
• uses data: DEKOIS 2.0, 79 targets as scored by Shen et al. 2025
• part of: DEKOIS 2.0
• assessment: AF3 (ipTM) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: AF3 (min-iPAE) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: AF3 (pTM) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: AF3 (ranking score) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Boltz-2 (Affinity) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Boltz-2 (ipTM) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Boltz-2 (min-iPAE) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Boltz-2 (Probability) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Boltz-2 (pTM) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Boltz-2 (ranking score) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Gnina (CNNaffinity) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Gnina (CNNscore) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Gnina (Vina) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Protenix (ipTM) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Protenix (pTM) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Protenix (ranking score) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: CarsiDock on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Glide_SP on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: GOLD_CHEMPLP on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: KarmaDock (Aligned) on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: LeDock on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: SurfDock on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: Surflex-Dock on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessment: TankBin on DEKOIS 2.0 (Shen et al. 2025 Table 1)
• assessed by: Find hits for a target by virtual screening