{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,20]],"date-time":"2026-10-20T14:29:30Z","timestamp":1792506570572,"version":"4.5.0"},"reference-count":60,"publisher":"Oxford University Press (OUP)","issue":"7","license":[{"start":{"date-parts":[[2026,6,15]],"date-time":"2026-06-15T00:00:00Z","timestamp":1781481600000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"name":"Department of Computer Science, University of T\u00fcbingen"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"published-print":{"date-parts":[[2026,7,2]]},"abstract":"<jats:title>Abstract<\/jats:title>\n                  <jats:sec>\n                    <jats:title>Motivation<\/jats:title>\n                    <jats:p>Sparse autoencoders (SAEs) decompose foundation-model activations into interpretable features, but the model-internal causal interactions between those features (i.e. what ablating one feature does to the others, as distinct from the biological causal structure of the underlying cells)\u2014and how those model-internal relationships relate to biological structure\u2014are uncharacterized in single-cell foundation models.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Results<\/jats:title>\n                    <jats:p>We introduce model-internal causal circuit tracing\u2014zeroing one SAE feature at a source layer and measuring the resulting change in all downstream SAE features, for each of 120 source features\u2014and apply it to Geneformer V2-316M and scGPT whole-human across four conditions (96\u00a0892 ablation-derived edges, 80\u00a0191 forward passes). On annotation-selected source features, edges share GO\/KEGG\/Reactome\/STRING\/TRRUST ontology terms at 50.9%\u201368.5%, a 2.9\u20136.2\u00d7 enrichment over a configuration-preserving permutation null (P&amp;lt;.002); on 20 randomly sampled source features this attenuates to 21.5%\u201326.3%\u2014still 2.5\u20133.1\u00d7 above null\u2014quantifying the annotation-selection contribution. Inhibitory dominance (fraction of ablation edges with d&amp;lt;0, i.e. source activation supports downstream target) is 65.5%\u201389.4%. scGPT produces larger raw per-edge effects (mean |d|=1.40 versus 1.05); after feature-share normalization, Geneformer is stronger (paired gene-pair ratio 0.64 on 33\u00a0301 shared pairs). Cross-model consensus yields 1142 architecture-invariant domain pairs (ordered pairs of GO biological-process categories \u201cA\u2192B\u201d each connected by at least one ablation edge in both models; 10.6\u00d7 enrichment over permutation null; P&amp;lt;.001). Circuit edge magnitude explains &amp;lt;1% of the variance in marginal driver-gene coexpression on the same cells (R2=0.010, n=31\u00a0176): the graph encodes structure beyond bivariate correlation. Against a matched-cell-type ENCODE ChIP-seq prior, circuit-predicted transcription factor (TF)\u2192target pairs are enriched 2.06\u00d7 (Fisher OR 5.84), markedly higher than 1.12\u00d7 against TRRUST; direct ChIP-seq-supported target pairs show 10\u201330\u00d7 larger CRISPRi sign-bias-corrected excess than indirect pairs. Gene-level CRISPRi validation on Replogle K562 and the noncancer RPE1 arm (and a true primary-T-cell control from Shifrut E, Carnevale J, Tobin V et\u00a0al. Genome-wide CRISPR screens in primary human T cells reveal key regulators of immune function. Cell 2018; 175: 1958\u201371.e15) after sign-bias correction shows excess over baseline of +0.03 and +0.35 percentage points on K562 and RPE1, respectively (baseline already 52%\u201356% from sign marginals); effect-magnitude Spearman correlations \u03c1\u22480. Bootstrap and per-cell-type stability (N\u2208{50,100,200}; B cell, CD4\u00a0+ T, macrophage) give Pearson r\u22650.97 on shared edges with 100% sign agreement; edge Jaccard grows monotonically with sample size. The circuit graph is therefore highly reproducible as an effect-size map, cell type specific in edge identity, consistent with coexpression encoding, and weakly but detectably enriched for ChIP-seq-supported direct regulatory edges.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Availability and implementation<\/jats:title>\n                    <jats:p>https:\/\/github.com\/Biodyn-AI\/bio-sae-circuits (Python). Archival DOI: 10.5281\/zenodo.19,633,166 (Zenodo).<\/jats:p>\n                  <\/jats:sec>","DOI":"10.1093\/bioinformatics\/btag379","type":"journal-article","created":{"date-parts":[[2026,6,11]],"date-time":"2026-06-11T11:50:05Z","timestamp":1781178605000},"source":"Crossref","is-referenced-by-count":1,"title":["Causal circuit tracing reveals distinct computational architectures in single-cell foundation models: inhibitory dominance, biological coherence, and cross-model convergence"],"prefix":"10.1093","volume":"42","author":[{"given":"Ihor","family":"Kendiukhov","sequence":"first","affiliation":[{"name":"Department of Computer Science, University of T\u00fcbingen , T\u00fcbingen,","place":["72076, Germany"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"286","published-online":{"date-parts":[[2026,6,15]]},"reference":[{"key":"2026071101161333300_btag379-B1","doi-asserted-by":"crossref","first-page":"1867","DOI":"10.1016\/j.cell.2016.11.048","article-title":"A multiplexed single-cell crispr screening platform enables systematic dissection of the unfolded protein response","volume":"167","author":"Adamson","year":"2016","journal-title":"Cell"},{"key":"2026071101161333300_btag379-B2","doi-asserted-by":"crossref","first-page":"25","DOI":"10.1038\/75556","article-title":"Gene ontology: tool for the unification of biology","volume":"25","author":"Ashburner","year":"2000","journal-title":"Nat Genet"},{"key":"2026071101161333300_btag379-B3","doi-asserted-by":"crossref","first-page":"289","DOI":"10.1111\/j.2517-6161.1995.tb02031.x","article-title":"Controlling the false discovery rate: a practical and powerful approach to multiple testing","volume":"57","author":"Benjamini","year":"1995","journal-title":"J R Stat Soc Ser B (Methodol)"},{"key":"2026071101161333300_btag379-B4","author":"Bereska"},{"key":"2026071101161333300_btag379-B5","doi-asserted-by":"crossref","first-page":"216","DOI":"10.1016\/j.molcel.2010.09.024","article-title":"The nucleolus under stress","volume":"40","author":"Boulon","year":"2010","journal-title":"Mol Cell"},{"key":"2026071101161333300_btag379-B6"},{"key":"2026071101161333300_btag379-B7","doi-asserted-by":"crossref","first-page":"a040543","DOI":"10.1101\/cshperspect.a040543","article-title":"Mitochondria","volume":"13","author":"Chandel","year":"2021","journal-title":"Cold Spring Harb Perspect Biol"},{"key":"2026071101161333300_btag379-B8","doi-asserted-by":"crossref","first-page":"179","DOI":"10.1016\/j.molcel.2010.09.019","article-title":"The DNA damage response: making it safe to play with knives","volume":"40","author":"Ciccia","year":"2010","journal-title":"Mol Cell"},{"key":"2026071101161333300_btag379-B9","volume-title":"Statistical Power Analysis for the Behavioral Sciences","author":"Cohen","year":"1988","edition":"2nd"},{"key":"2026071101161333300_btag379-B10","first-page":"16318","volume":"36"},{"key":"2026071101161333300_btag379-B11","doi-asserted-by":"crossref","first-page":"1470","DOI":"10.1038\/s41592-024-02201-0","article-title":"Scgpt: toward building a foundation model for single-cell multi-omics using generative ai","volume":"21","author":"Cui","year":"2024","journal-title":"Nat Methods"},{"key":"2026071101161333300_btag379-B12"},{"key":"2026071101161333300_btag379-B13","doi-asserted-by":"crossref","first-page":"297","DOI":"10.1038\/nmeth.4177","article-title":"Pooled crispr screening with single-cell transcriptome readout","volume":"14","author":"Datlinger","year":"2017","journal-title":"Nat Methods"},{"key":"2026071101161333300_btag379-B14"},{"key":"2026071101161333300_btag379-B15","doi-asserted-by":"crossref","first-page":"3279","DOI":"10.1038\/sj.onc.1210421","article-title":"MAP kinase signalling pathways in cancer","volume":"26","author":"Dhillon","year":"2007","journal-title":"Oncogene"},{"key":"2026071101161333300_btag379-B16","doi-asserted-by":"crossref","first-page":"1853","DOI":"10.1016\/j.cell.2016.11.038","article-title":"Perturb-seq: dissecting molecular circuits with scalable single-cell RNA profiling of pooled genetic screens","volume":"167","author":"Dixit","year":"2016","journal-title":"Cell"},{"key":"2026071101161333300_btag379-B17"},{"key":"2026071101161333300_btag379-B18"},{"key":"2026071101161333300_btag379-B19","doi-asserted-by":"crossref","first-page":"57","DOI":"10.1038\/nature11247","article-title":"An integrated encyclopedia of DNA elements in the human genome","volume":"489","author":"ENCODE Project Consortium","year":"2012","journal-title":"Nature"},{"key":"2026071101161333300_btag379-B20"},{"key":"2026071101161333300_btag379-B21","doi-asserted-by":"crossref","first-page":"1363","DOI":"10.1101\/gr.240663.118","article-title":"Benchmark and integration of resources for the estimation of human transcription factor activities","volume":"29","author":"Garcia-Alonso","year":"2019","journal-title":"Genome Res"},{"key":"2026071101161333300_btag379-B22","first-page":"9574","article-title":"Causal abstractions of neural networks","volume":"34","author":"Geiger","year":"2021","journal-title":"Adv Neural Inf Process Syst"},{"key":"2026071101161333300_btag379-B23","doi-asserted-by":"crossref","first-page":"845","DOI":"10.1038\/nmeth.3971","article-title":"Diffusion pseudotime robustly reconstructs lineage branching","volume":"13","author":"Haghverdi","year":"2016","journal-title":"Nat Methods"},{"key":"2026071101161333300_btag379-B24","doi-asserted-by":"crossref","first-page":"D380","DOI":"10.1093\/nar\/gkx1013","article-title":"TRRUST v2: an expanded reference database of human and mouse transcriptional regulatory interactions","volume":"46","author":"Han","year":"2018","journal-title":"Nucleic Acids Res"},{"key":"2026071101161333300_btag379-B25","doi-asserted-by":"crossref","first-page":"1481","DOI":"10.1038\/s41592-024-02305-7","article-title":"Large-scale foundation model on single-cell transcriptomics","volume":"21","author":"Hao","year":"2024","journal-title":"Nat Methods"},{"key":"2026071101161333300_btag379-B26","doi-asserted-by":"crossref","first-page":"89","DOI":"10.1038\/nrm3270","article-title":"The unfolded protein response: controlling cell fate decisions under ER stress and beyond","volume":"13","author":"Hetz","year":"2012","journal-title":"Nat Rev Mol Cell Biol"},{"key":"2026071101161333300_btag379-B27","doi-asserted-by":"crossref","first-page":"1071","DOI":"10.1038\/nature08467","article-title":"The DNA-damage response in human biology and disease","volume":"461","author":"Jackson","year":"2009","journal-title":"Nature"},{"key":"2026071101161333300_btag379-B28","first-page":"D498","article-title":"The reactome pathway knowledgebase","volume":"48","author":"Jassal","year":"2020","journal-title":"Nucleic Acids Res"},{"key":"2026071101161333300_btag379-B29","doi-asserted-by":"crossref","first-page":"27","DOI":"10.1093\/nar\/28.1.27","article-title":"KEGG: Kyoto encyclopedia of genes and genomes","volume":"28","author":"Kanehisa","year":"2000","journal-title":"Nucleic Acids Res"},{"key":"2026071101161333300_btag379-B60284803"},{"key":"2026071101161333300_btag379-B32","doi-asserted-by":"crossref","first-page":"435","DOI":"10.1146\/annurev-biochem-060614-033955","article-title":"The biology of proteostasis in aging and disease","volume":"84","author":"Labbadia","year":"2015","journal-title":"Annu Rev Biochem"},{"key":"2026071101161333300_btag379-B33","doi-asserted-by":"crossref","first-page":"417","DOI":"10.1016\/j.cels.2015.12.004","article-title":"The molecular signatures database (msigdb) hallmark gene set collection","volume":"1","author":"Liberzon","year":"2015","journal-title":"Cell Syst"},{"key":"2026071101161333300_btag379-B34"},{"key":"2026071101161333300_btag379-B35","doi-asserted-by":"crossref","first-page":"153","DOI":"10.1038\/nrc2602","article-title":"Cell cycle, CDKs and cancer: a changing paradigm","volume":"9","author":"Malumbres","year":"2009","journal-title":"Nat Rev Cancer"},{"key":"2026071101161333300_btag379-B36","volume-title":"Foundations of Statistical Natural Language Processing","author":"Manning","year":"1999"},{"key":"2026071101161333300_btag379-B37"},{"key":"2026071101161333300_btag379-B38","first-page":"17359","volume":"35"},{"key":"2026071101161333300_btag379-B39","doi-asserted-by":"crossref","first-page":"379","DOI":"10.1038\/nrm2163","article-title":"The spindle-assembly checkpoint in space and time","volume":"8","author":"Musacchio","year":"2007","journal-title":"Nat Rev Mol Cell Biol"},{"key":"2026071101161333300_btag379-B40"},{"key":"2026071101161333300_btag379-B41","doi-asserted-by":"crossref","first-page":"985","DOI":"10.1016\/j.cell.2017.05.016","article-title":"Wnt\/\u03b2-catenin signaling, disease, and emerging therapeutic modalities","volume":"169","author":"Nusse","year":"2017","journal-title":"Cell"},{"key":"2026071101161333300_btag379-B42","doi-asserted-by":"crossref","first-page":"3311","DOI":"10.1016\/S0042-6989(97)00169-7","article-title":"Sparse coding with an overcomplete basis set: a strategy employed by v1?","volume":"37","author":"Olshausen","year":"1997","journal-title":"Vision Res"},{"key":"2026071101161333300_btag379-B43","first-page":"8024","article-title":"PyTorch: an imperative style, high-performance deep learning library","volume":"32","year":"2019","journal-title":"Adv Neural Inform Process Syst"},{"key":"2026071101161333300_btag379-B44","doi-asserted-by":"crossref","first-page":"2559","DOI":"10.1016\/j.cell.2022.05.013","article-title":"Mapping information-rich genotype-phenotype landscapes with genome-scale perturb-seq","volume":"185","author":"Replogle","year":"2022","journal-title":"Cell"},{"key":"2026071101161333300_btag379-B45","doi-asserted-by":"publisher","DOI":"10.1101\/2023.11.28.56891"},{"key":"2026071101161333300_btag379-B46","doi-asserted-by":"crossref","first-page":"S10","DOI":"10.1038\/nm1066","article-title":"Protein aggregation and neurodegenerative disease","volume":"10","author":"Ross","year":"2004","journal-title":"Nat Med"},{"key":"2026071101161333300_btag379-B47"},{"key":"2026071101161333300_btag379-B48","doi-asserted-by":"crossref","first-page":"1958","DOI":"10.1016\/j.cell.2018.10.024","article-title":"Genome-wide CRISPR screens in primary human T cells reveal key regulators of immune function","volume":"175","author":"Shifrut","year":"2018","journal-title":"Cell"},{"key":"2026071101161333300_btag379-B49","doi-asserted-by":"crossref","first-page":"1721","DOI":"10.1126\/science.290.5497.1721","article-title":"How cells handle cholesterol","volume":"290","author":"Simons","year":"2000","journal-title":"Science"},{"key":"2026071101161333300_btag379-B50","doi-asserted-by":"crossref","first-page":"15545","DOI":"10.1073\/pnas.0506580102","article-title":"Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles","volume":"102","author":"Subramanian","year":"2005","journal-title":"Proc Natl Acad Sci U S A"},{"key":"2026071101161333300_btag379-B51","doi-asserted-by":"crossref","first-page":"D638","DOI":"10.1093\/nar\/gkac1000","article-title":"The STRING database in 2023: protein\u2013protein association networks and functional enrichment analyses for any sequenced genome of interest","volume":"51","author":"Szklarczyk","year":"2023","journal-title":"Nucleic Acids Res"},{"key":"2026071101161333300_btag379-B52","first-page":"eabl4896","article-title":"The Tabula Sapiens: a multiple-organ, single-cell transcriptomic atlas of humans","volume":"376","author":"Tabula Sapiens Consortium","year":"2022","journal-title":"Science (1979)"},{"key":"2026071101161333300_btag379-B53"},{"key":"2026071101161333300_btag379-B54","doi-asserted-by":"crossref","first-page":"616","DOI":"10.1038\/s41586-023-06139-9","article-title":"Transfer learning enables predictions in network biology","volume":"618","author":"Theodoris","year":"2023","journal-title":"Nature"},{"key":"2026071101161333300_btag379-B55","first-page":"5998"},{"key":"2026071101161333300_btag379-B56","first-page":"12388","volume":"33"},{"key":"2026071101161333300_btag379-B57","doi-asserted-by":"crossref","first-page":"1081","DOI":"10.1126\/science.1209038","article-title":"The unfolded protein response: from stress pathway to homeostatic regulation","volume":"334","author":"Walter","year":"2011","journal-title":"Science"},{"key":"2026071101161333300_btag379-B58"},{"key":"2026071101161333300_btag379-B59","doi-asserted-by":"crossref","first-page":"419","DOI":"10.1080\/00401706.1962.10490022","article-title":"Note on a method for calculating corrected sums of squares and products","volume":"4","author":"Welford","year":"1962","journal-title":"Technometrics"},{"key":"2026071101161333300_btag379-B60","doi-asserted-by":"crossref","first-page":"15","DOI":"10.1186\/s13059-017-1382-0","article-title":"SCANPY: large-scale single-cell gene expression data analysis","volume":"19","author":"Wolf","year":"2018","journal-title":"Genome Biol"},{"key":"2026071101161333300_btag379-B61","doi-asserted-by":"crossref","first-page":"852","DOI":"10.1038\/s42256-022-00534-z","article-title":"scBERT as a large-scale pretrained deep language model for cell type annotation of single-cell RNA-seq data","volume":"4","author":"Yang","year":"2022","journal-title":"Nat Mach Intell"}],"container-title":["Bioinformatics"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/academic.oup.com\/bioinformatics\/advance-article-pdf\/doi\/10.1093\/bioinformatics\/btag379\/68530126\/btag379.pdf","content-type":"application\/pdf","content-version":"am","intended-application":"syndication"},{"URL":"https:\/\/academic.oup.com\/bioinformatics\/article-pdf\/42\/7\/btag379\/68530126\/btag379.pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"syndication"},{"URL":"https:\/\/academic.oup.com\/bioinformatics\/article-pdf\/42\/7\/btag379\/68530126\/btag379.pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,7,11]],"date-time":"2026-07-11T05:16:22Z","timestamp":1783746982000},"score":1,"resource":{"primary":{"URL":"https:\/\/academic.oup.com\/bioinformatics\/article\/doi\/10.1093\/bioinformatics\/btag379\/8708321"}},"subtitle":[],"editor":[{"given":"A Ercument","family":"Cicek","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"editor"}]}],"short-title":[],"issued":{"date-parts":[[2026,6,15]]},"references-count":60,"journal-issue":{"issue":"7","published-print":{"date-parts":[[2026,7,2]]}},"URL":"https:\/\/doi.org\/10.1093\/bioinformatics\/btag379","relation":{},"ISSN":["1367-4803","1367-4811"],"issn-type":[{"value":"1367-4803","type":"print"},{"value":"1367-4811","type":"electronic"}],"subject":[],"published-other":{"date-parts":[[2026,7]]},"published":{"date-parts":[[2026,6,15]]},"article-number":"btag379"}}