Systematic evaluation of single-cell foundation model interpretability reveals attention captures co-expression rather than unique regulatory signal

Systematic evaluation of single-cell foundation model interpretability reveals attention captures co-expression rather than unique regulatory signal | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Systematic evaluation of single-cell foundation model interpretability reveals attention captures co-expression rather than unique regulatory signal Ihor Kendiukhov This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9082476/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 13 You are reading this latest preprint version Abstract Background : Single-cell foundation models such as scGPT and Geneformer are increasingly used for gene regulatory network (GRN) inference, with attention-derived edge scores routinely interpreted as regulatory proxies. However, whether attention patterns capture causal regulatory relationships—rather than statistical associations already present in expression data—has not been systematically tested. This gap is critical because the NLP interpretability literature has established that attention weights do not reliably indicate feature importance, yet this finding has not been rigorously evaluated in biological foundation models. Results : We present a systematic evaluation framework comprising thirty-seven analyses, 153 statistical tests, four cell types (K562, RPE1, T cells, iPSC neurons), and two perturbation modalities (CRISPRi, CRISPRa). Attention patterns encode layer-specific biological structure—protein–protein interactions in early layers, transcriptional regulation in late layers—but this structure provides no incremental value for perturbation prediction: trivial gene-level baselines outperform both attention and correlation edges (AUROC 0.81–0.88 versus 0.70), pairwise edge scores add zero predictive contribution beyond gene-level features (∆AUROC = −0.0004 to −0.002; 559,720 observations), and causal ablation of regulatory heads produces no degradation across three independent intervention channels. The attention–correlation relationship is context-dependent (equal in K562, worse in CRISPRa, better in RPE1), but gene-level dominance is universal. Cell-State Stratified Interpretability (CSSI) addresses an attention-specific scaling failure, improving GRN recovery up to 1.85×. Conclusions : Attention patterns in single-cell foundation models encode structured biological information but not the causal regulatory signal they are commonly interpreted as capturing. The evaluation framework establishes reusable quality-control standards for the field, and CSSI provides an immediately deployable tool for improved edge recovery from heterogeneous populations. single-cell foundation models mechanistic interpretability gene regulatory networks attention mechanisms scGPT Geneformer perturbation prediction CRISPR screens benchmarking Full Text Additional Declarations No competing interests reported. Supplementary Files supplementarygb.pdf Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 10 Apr, 2026 Reviews received at journal 09 Apr, 2026 Reviews received at journal 08 Apr, 2026 Reviews received at journal 02 Apr, 2026 Reviewers agreed at journal 27 Mar, 2026 Reviewers agreed at journal 25 Mar, 2026 Reviewers agreed at journal 24 Mar, 2026 Reviewers agreed at journal 24 Mar, 2026 Reviewers invited by journal 24 Mar, 2026 Editor invited by journal 13 Mar, 2026 Editor assigned by journal 11 Mar, 2026 Submission checks completed at journal 11 Mar, 2026 First submitted to journal 10 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9082476","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":611724933,"identity":"2e774ee9-eef7-442f-9773-a87390889019","order_by":0,"name":"Ihor Kendiukhov","email":"data:image/png;base64,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","orcid":"","institution":"University of Tübingen","correspondingAuthor":true,"prefix":"","firstName":"Ihor","middleName":"","lastName":"Kendiukhov","suffix":""}],"badges":[],"createdAt":"2026-03-10 10:09:44","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9082476/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9082476/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105565481,"identity":"bba4eca1-b4ed-4d8b-af93-f0892f2c5a68","added_by":"auto","created_at":"2026-03-27 12:53:23","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3041691,"visible":true,"origin":"","legend":"","description":"","filename":"genomebiologysubmission.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9082476/v1_covered_8c3da3e9-8c49-4b3b-a9ee-cab0c21c66d1.pdf"},{"id":105438489,"identity":"74c00a49-195d-4e27-9e74-b250732642c0","added_by":"auto","created_at":"2026-03-26 04:56:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":11184971,"visible":true,"origin":"","legend":"","description":"","filename":"supplementarygb.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9082476/v1/00df28bcc921c29d01c2c7f7.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Systematic evaluation of single-cell foundation model interpretability reveals attention captures co-expression rather than unique regulatory signal","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-genomics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"gics","sideBox":"Learn more about [BMC Genomics](http://bmcgenomics.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/gics","title":"BMC Genomics","twitterHandle":"#BMCGenomics","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"single-cell foundation models, mechanistic interpretability, gene regulatory networks, attention mechanisms, scGPT, Geneformer, perturbation prediction, CRISPR screens, benchmarking","lastPublishedDoi":"10.21203/rs.3.rs-9082476/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9082476/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e: Single-cell foundation models such as scGPT and Geneformer are increasingly used for gene regulatory network (GRN) inference, with attention-derived edge scores routinely interpreted as regulatory proxies. However, whether attention patterns capture causal regulatory relationships—rather than statistical associations already present in expression data—has not been systematically tested. This gap is critical because the NLP interpretability literature has established that attention weights do not reliably indicate feature importance, yet this finding has not been rigorously evaluated in biological foundation models.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e: We present a systematic evaluation framework comprising thirty-seven analyses, 153 statistical tests, four cell types (K562, RPE1, T cells, iPSC neurons), and two perturbation modalities (CRISPRi, CRISPRa). Attention patterns encode layer-specific biological structure—protein–protein interactions in early layers, transcriptional regulation in late layers—but this structure provides no incremental value for perturbation prediction: trivial gene-level baselines outperform both attention and correlation edges (AUROC 0.81–0.88 versus 0.70), pairwise edge scores add zero predictive contribution beyond gene-level features (∆AUROC = −0.0004 to −0.002; 559,720 observations), and causal ablation of regulatory heads produces no degradation across three independent intervention channels. The attention–correlation relationship is context-dependent (equal in K562, worse in CRISPRa, better in RPE1), but gene-level dominance is universal. Cell-State Stratified Interpretability (CSSI) addresses an attention-specific scaling failure, improving GRN recovery up to 1.85×.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e: Attention patterns in single-cell foundation models encode structured biological information but not the causal regulatory signal they are commonly interpreted as capturing. The evaluation framework establishes reusable quality-control standards for the field, and CSSI provides an immediately deployable tool for improved edge recovery from heterogeneous populations.\u003c/p\u003e","manuscriptTitle":"Systematic evaluation of single-cell foundation model interpretability reveals attention captures co-expression rather than unique regulatory signal","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-26 04:56:41","doi":"10.21203/rs.3.rs-9082476/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-04-10T07:53:50+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-09T20:57:10+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-09T02:39:31+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-02T23:46:13+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"333252682666716700875675877125736194180","date":"2026-03-27T13:03:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"53613228871143556512127799280047235725","date":"2026-03-25T04:33:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"186250924718236580338709779935798415034","date":"2026-03-25T03:54:22+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"163277837268827699258648346171884672811","date":"2026-03-25T02:39:46+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-25T02:32:13+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-03-13T09:28:00+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-12T00:06:17+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-12T00:06:12+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Genomics","date":"2026-03-10T10:06:28+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-genomics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"gics","sideBox":"Learn more about [BMC Genomics](http://bmcgenomics.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/gics","title":"BMC Genomics","twitterHandle":"#BMCGenomics","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e38d42b3-d409-4b58-932b-0836fc32d075","owner":[],"postedDate":"March 26th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-14T14:55:05+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-26 04:56:41","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9082476","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9082476","identity":"rs-9082476","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}