Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634); Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14)

What this source contributed

Records in this release that cite this source. A record can cite several sources.

5 methods
11 configurations
1 benchmark
6 protocols
8 dataset subsets
28 evaluations
25 results
2 evidence claims

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Claims

Statements this source supports, each with where in the source it was checked.

Geneformer V2-316M layer 13 pre-specified as primary layer (Kendiukhov 2026): Layer selection Layer 13 was pre-specified as the primary attention layer from DLPFC brain data (TRRUST recovery AUROC 0.694) before any K562 perturbation data were examined; other layers are secondary analyses. (Page 7, 'Perturbation-first validation reveals gene-level dominance', paragraph 2, sentence 2: 'L13 was pre-specified as the primary layer based on independent data')
Systematic evaluation of single-cell foundation model interpretability: attentio: attributes.evidence_concerns: Evidence concerns [] (moved to attributes.limitations; the two CRISPRa results are disputed) (Page 7 paragraphs 1-3; page 10; Table 1; reference 25)

Models, benchmarks and other records

Methods (5)

Configurations (11)

Benchmarks (1)

Perturbation-target prediction from attention-derived edge scores (Kendiukhov 2026, Objective B)

Protocols (6)

Dataset subsets (8)

Evaluations (28)

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.

Evidence table

Inspect claims, sources and review details

Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.

One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.

28 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-02-a7f03c9e55a5
Property and statement	Original source and location	Review and provenance
Artifact SHA-256 4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.artifact_sha256` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact
Artifact url https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/s12864-026-12965-8.pdf Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.artifact_url` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact
Doi 10.1186/s12864-026-12965-8 Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.doi` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact
Licence CC-BY-4.0 Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.licence` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact
Limitations 5 values Single-author study; the author developed none of the evaluated foundation models but did develop CSSI and the evaluation framework. Objective B results are printed in running text only; Figures 3, 4, 6 and 7 are not read. Primary layer L13 was pre-specified on DLPFC brain data; L15 and L6 values come from an 18-layer profile (L15 selected by nested cross-validation). K562 CRISPRa is cited as reference 25 (Adamson et al. 2016), a CRISPRi study; the two CRISPRa results are disputed for that reason. K562 CRISPRi results come from two perturbation sets (n = 151 for correlation and gene-level baselines, Fig. 3; n = 280 for attention versus correlation) and RPE1 from two (n = 1,251 and n = 1,167). The article does not say how the sets relate; each is a separate dataset subset. Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.limitations` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact
Media type application/pdf Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.media_type` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact
Publication status peer_reviewed Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.publication_status` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact
Retrieved at 2026-10-14T12:01:07Z Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.retrieved_at` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact
Review: artifact sha256 4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.review.artifact_sha256` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact
Review: date 2026-10-14 Source metadata	Systematic evaluation of single-cell foundation model interpretability: attention-derived edge scores add no incremental value over gene-level features for perturbation-target prediction (Kendiukhov 2026, BMC Genomics 27:634) Original source ↗ No field-specific location recorded Version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14) Retrieved: 2026-10-14T12:01:07Z	catalogued No individual claim review recorded Audit details Field: `attributes.review.date` Source artifact SHA-256: `4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe` Hash scope: Hash scope not separately documented; inspect source record Inspected artifact

How to read the evidence

Sources and history

Release 2026-11-02-a7f03c9e55a5 · Record review: source checked

Publisher PDF of the article, retrieved and hashed for the 2026-10-14 genomic foundation model guide fixes (issue rewire-benchmark-data#129).

Release history

No supporting source is linked yet.

Read original source

Download this release (gzip)

Technical metadata and extraction receipts

Stable ID: gfm-guides-20261014-kendiukhov2026-source-article

areas: single-cell; biological-networks
contexts: research
url: https://doi.org/10.1186/s12864-026-12965-8
artifact url: https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/s12864-026-12965-8.pdf
version: Version of record, BMC Genomics 27:634, published 2026-07-22 (received 2026-03-10, accepted 2026-05-14)
retrieved at: 2026-10-14T12:01:07Z
artifact sha256: 4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe
doi: 10.1186/s12864-026-12965-8
publication status: peer_reviewed
licence: CC-BY-4.0
media type: application/pdf
venue: BMC Genomics
year: 2026
limitations: Single-author study; the author developed none of the evaluated foundation models but did develop CSSI and the evaluation framework.; Objective B results are printed in running text only; Figures 3, 4, 6 and 7 are not read.; Primary layer L13 was pre-specified on DLPFC brain data; L15 and L6 values come from an 18-layer profile (L15 selected by nested cross-validation).; K562 CRISPRa is cited as reference 25 (Adamson et al. 2016), a CRISPRi study; the two CRISPRa results are disputed for that reason.; K562 CRISPRi results come from two perturbation sets (n = 151 for correlation and gene-level baselines, Fig. 3; n = 280 for attention versus correlation) and RPE1 from two (n = 1,251 and n = 1,167). The article does not say how the sets relate; each is a separate dataset subset.
review: method: source-hash-verification; method note: Fresh download on 2026-10-14 returned a bot-check HTML page; the archived copy in artifacts/ hashes to the pinned SHA-256 and was used.; reviewer: claude; reviewer note: Claude (Opus 5.5) research agent, independent of the extractor; no human review claimed; date: 2026-10-14; artifact sha256: 4c1699cb964ae588117a6737f5e9e189eb9a5ef5ff78568d2a99e7a13d6608fe; retrieval url: https://bmcgenomics.biomedcentral.com/counter/pdf/10.1186/s12864-026-12965-8.pdf; note: Independent review 2026-10-14.

Related records

subject: Systematic evaluation of single-cell foundation model interpretability: attentio: attributes.evidence_concerns

Model or configuration	Metric	Value	Dataset	Evaluation
Absolute Spearman correlation edges (Kendiukhov 2026)	AUROC (AUROC over propensity-matched perturbation-gene pairs (k = 5 negatives per positive))	0.561 unitless	Replogle CRISPRi RPE1, propensity-matched pairs (Kendiukhov 2026)	Absolute Spearman correlation edges on Replogle CRISPRi RPE1, propensity-matched pairs
Absolute Spearman correlation edges (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets (LFC > 0.5), summarised over perturbations; K562 CRISPRi)	0.696 unitless	Replogle genome-scale CRISPRi, K562, 151 perturbations (Kendiukhov 2026)	Absolute Spearman correlation edges on Replogle genome-scale CRISPRi, K562, 151 perturbations
Absolute Spearman correlation edges (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets (LFC > 0.5), summarised over perturbations; K562 CRISPRi)	0.703 unitless	Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations (Kendiukhov 2026)	Absolute Spearman correlation edges on Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations
Absolute Spearman correlation edges (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets, summarised over perturbations; RPE1 CRISPRi)	0.658 unitless	Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery (Kendiukhov 2026)	Absolute Spearman correlation edges on Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery
Absolute Spearman correlation edges (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets, summarised over perturbations; RPE1 CRISPRi)	0.658 unitless	Replogle genome-scale CRISPRi, RPE1, 1,251 evaluable perturbations (Kendiukhov 2026)	Absolute Spearman correlation edges on Replogle genome-scale CRISPRi, RPE1, 1,251 evaluable perturbations
Gene-level logistic regression (Kendiukhov 2026)	AUROC (cross-validated AUROC over perturbation-gene pairs, 5-fold GroupKFold by perturbation)	0.895 unitless	Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations (Kendiukhov 2026)	Gene-level logistic regression on Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations
Gene-level logistic regression (Kendiukhov 2026)	AUROC (cross-validated AUROC over perturbation-gene pairs, 5-fold GroupKFold by perturbation)	0.942 unitless	Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery (Kendiukhov 2026)	Gene-level logistic regression on Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery
Gene-level logistic regression plus attention edge (Kendiukhov 2026)	AUROC difference (Gene-level features plus attention edge minus gene-level features only (logistic regression); AUROC over propensity-matched perturbation-gene pairs (k = 5 negatives per positive))	−0.000 unitless	Replogle CRISPRi K562, propensity-matched pairs (Kendiukhov 2026)	Gene-level logistic regression plus attention edge on Replogle CRISPRi K562, propensity-matched pairs
Gene-level logistic regression plus attention edge (Kendiukhov 2026)	AUROC difference (Gene-level features plus attention edge minus gene-level features only (gene-level logistic regression); cross-validated AUROC over perturbation-gene pairs, 5-fold GroupKFold by perturbation)	−0.0004 unitless	Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations (Kendiukhov 2026)	Gene-level logistic regression plus attention edge on Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations
Gene-level logistic regression plus attention edge (Kendiukhov 2026)	AUROC difference (Gene-level features plus attention edge minus gene-level features only (gene-level logistic regression); cross-validated AUROC over perturbation-gene pairs, 5-fold GroupKFold by perturbation)	+0.0001 unitless	Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery (Kendiukhov 2026)	Gene-level logistic regression plus attention edge on Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery
Gene-level logistic regression plus correlation edge (Kendiukhov 2026)	AUROC difference (Gene-level features plus correlation edge minus gene-level features only (gene-level logistic regression); cross-validated AUROC over perturbation-gene pairs, 5-fold GroupKFold by perturbation)	−0.002 unitless	Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations (Kendiukhov 2026)	Gene-level logistic regression plus correlation edge on Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations
Gene-level logistic regression plus correlation edge (Kendiukhov 2026)	AUROC difference (Gene-level features plus correlation edge minus gene-level features only (gene-level logistic regression); cross-validated AUROC over perturbation-gene pairs, 5-fold GroupKFold by perturbation)	−0.0006 unitless	Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery (Kendiukhov 2026)	Gene-level logistic regression plus correlation edge on Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery
Geneformer V2-316M attention edges, layer 6 (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets, summarised over perturbations; RPE1 CRISPRi)	0.776 unitless	Replogle genome-scale CRISPRi, RPE1, 1,251 evaluable perturbations (Kendiukhov 2026)	Geneformer V2-316M attention edges, layer 6 on Replogle genome-scale CRISPRi, RPE1, 1,251 evaluable perturbations
Geneformer V2-316M attention edges, layer 13 (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets (LFC > 0.5), summarised over perturbations; K562 CRISPRi)	0.704 unitless	Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations (Kendiukhov 2026)	Geneformer V2-316M attention edges, layer 13 on Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations
Geneformer V2-316M attention edges, layer 15 (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets (LFC > 0.5), summarised over perturbations; K562 CRISPRi)	0.743 unitless	Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations (Kendiukhov 2026)	Geneformer V2-316M attention edges, layer 15 on Replogle genome-scale CRISPRi, K562, 280 evaluable perturbations
Geneformer V2-316M attention edges, layer 15 (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets, summarised over perturbations; RPE1 CRISPRi)	0.748 unitless	Replogle genome-scale CRISPRi, RPE1, 1,251 evaluable perturbations (Kendiukhov 2026)	Geneformer V2-316M attention edges, layer 15 on Replogle genome-scale CRISPRi, RPE1, 1,251 evaluable perturbations
Geneformer V2-316M attention edges, layer not stated (Kendiukhov 2026)	AUROC (AUROC over propensity-matched perturbation-gene pairs (k = 5 negatives per positive))	0.609 unitless	Replogle CRISPRi K562, propensity-matched pairs (Kendiukhov 2026)	Geneformer V2-316M attention edges, layer not stated on Replogle CRISPRi K562, propensity-matched pairs
Geneformer V2-316M attention edges, layer not stated (Kendiukhov 2026)	AUROC (AUROC over propensity-matched perturbation-gene pairs (k = 5 negatives per positive))	0.568 unitless	Replogle CRISPRi RPE1, propensity-matched pairs (Kendiukhov 2026)	Geneformer V2-316M attention edges, layer not stated on Replogle CRISPRi RPE1, propensity-matched pairs
Geneformer V2-316M attention edges, layer not stated (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets, summarised over perturbations; RPE1 CRISPRi)	0.747 unitless	Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery (Kendiukhov 2026)	Geneformer V2-316M attention edges, layer not stated on Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery
Geneformer V2-316M attention edges, layer not stated (Kendiukhov 2026)	AUROC difference (Geneformer V2-316M attention edge minus absolute Spearman correlation edge (per-perturbation attention advantage); per-perturbation AUROC for DE targets, summary over perturbations not stated; RPE1 CRISPRi)	+0.090 unitless	Replogle genome-scale CRISPRi, RPE1, 1,251 evaluable perturbations (Kendiukhov 2026)	Geneformer V2-316M attention edges, layer not stated on Replogle genome-scale CRISPRi, RPE1, 1,251 evaluable perturbations
Geneformer V2-316M attention edges, layer not stated (Kendiukhov 2026)	AUROC difference (Geneformer V2-316M attention edge minus absolute Spearman correlation edge (per-perturbation attention advantage); per-perturbation AUROC for DE targets, summary over perturbations not stated; RPE1 CRISPRi restricted to the top 2,000 HVGs)	+0.168 unitless	Replogle genome-scale CRISPRi, RPE1, 418 perturbations with the top 2,000 HVGs (Kendiukhov 2026)	Geneformer V2-316M attention edges, layer not stated on Replogle genome-scale CRISPRi, RPE1, 418 perturbations with the top 2,000 HVGs
Target-gene dropout rate baseline (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets, summarised over perturbations; RPE1 CRISPRi)	0.797 unitless	Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery (Kendiukhov 2026)	Target-gene dropout rate baseline on Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery
Target-gene mean expression baseline (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets, summarised over perturbations; RPE1 CRISPRi)	0.851 unitless	Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery (Kendiukhov 2026)	Target-gene mean expression baseline on Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery
Target-gene variance baseline (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets (LFC > 0.5), summarised over perturbations; K562 CRISPRi)	0.881 unitless	Replogle genome-scale CRISPRi, K562, 151 perturbations (Kendiukhov 2026)	Target-gene variance baseline on Replogle genome-scale CRISPRi, K562, 151 perturbations
Target-gene variance baseline (Kendiukhov 2026)	AUROC (per-perturbation AUROC for DE targets, summarised over perturbations; RPE1 CRISPRi)	0.866 unitless	Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery (Kendiukhov 2026)	Target-gene variance baseline on Replogle genome-scale CRISPRi, RPE1, 1,167 perturbations, confound battery