CoChat
Sign Up
Cited by 119Gene Regulatory Network AnalysisGenomics and Chromatin DynamicsSingle-cell and spatial transcriptomics

Inferring gene regulatory networks from single-cell multiome data using atlas-scale external data

Open access
Yuan, Qiuyue, Duren, Zhana·2025·Nature Biotechnology
DOI 10.1038/s41587-024-02182-7PMC PMC11825371

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Existing methods for gene regulatory network (GRN) inference rely on gene expression data alone or on lower resolution bulk data. Despite the recent integration of chromatin accessibility and RNA sequencing data, learning complex mechanisms from limited independent data points still presents a daunting challenge. Here we present LINGER (Lifelong neural network for gene regulation), a machine-learning method to infer GRNs from single-cell paired gene expression and chromatin accessibility data. LINGER incorporates atlas-scale external bulk data across diverse cellular contexts and prior knowledge of transcription factor motifs as a manifold regularization. LINGER achieves a fourfold to sevenfold relative increase in accuracy over existing methods and reveals a complex regulatory landscape of genome-wide association studies, enabling enhanced interpretation of disease-associated variants and genes. Following the GRN inference from reference single-cell multiome data, LINGER enables the estimation of transcription factor activity solely from bulk or single-cell gene expression data, leveraging the abundance of available gene expression data to identify driver regulators from case-control studies.

AI snapshot

Written from the paper’s full text.

TL;DR

LINGER infers gene regulatory networks from paired single-cell gene expression and chromatin accessibility data by pretraining on atlas-scale bulk ENCODE data and incorporating TF–motif prior knowledge as manifold regularization. It reports a fourfold to sevenfold relative increase in accuracy over existing methods and enables GWAS interpretation and TF activity estimation from expression data alone.

Key findings

  • LINGER improved population-level GRN inference over scNN, BulkNN and PCC across 20 blood ChIP–seq ground-truth datasets, with significantly higher AUC and AUPR ratio; for trans-regulatory potential, other methods averaged AUPR ratios of 1.17–1.29, 0.17–0.29 units above random, whereas LINGER achieved 1.25 units above random, described as a fourfold to sevenfold relative increase.
  • For cell type-specific TF–RE binding in naive CD4 T cells, LINGER reached AUC 0.92 and AUPR ratio 5.17 for ETS1, versus PCC (AUC 0.78; AUPR ratio 2.71) and motif binding affinity (AUC 0.70; AUPR ratio 1.92); for cis-regulatory potential in naive CD4 T cells, LINGER achieved AUC 0.66–0.70 and AUPR ratio 1.81–2.16 across distance groups, while other methods were close to random.
  • In IBD GWAS analysis using risk loci from about 330,000 individuals, classical monocytes showed significantly higher trait regulation scores for top-expressed TFs than random TFs (P = 8.9 × 10−29); among the top 200 STAT1 target genes in classical monocytes, 67 overlapped GWAS genes (P < 0.01), and the SLC24A4 cis-regulatory network was dense near risk SNP rs11626366 (P = 7.4 × 10−3).
  • LINGER estimated TF activity from expression alone: FOXN1 activity was lower in AML patients than healthy donors (P = 0.035) despite no significant expression difference, and high FOXN1 activity was associated with higher survival; upon TCR stimulation, FOXK2 and NR4A1 regulon activity increased at 8 h (P = 0.0057 and P = 0.00081, respectively) without significant expression changes.

Methods

  • LINGER uses a neural network to predict target-gene expression from TF expression and regulatory-element accessibility, pretraining on 201 ENCODE bulk samples across diverse cellular contexts, then refining on single-cell multiome data with elastic weight consolidation and manifold regularization based on TF–RE motif matching; Shapley values infer TF–TG and RE–TG regulatory strengths.
  • Data included 10x Genomics PBMC multiome data (11,909 cells; after filtering, 9,543 cells; metacells used for training), an in silico mixture of H1, BJ, GM12878 and K562 from SNARE-seq, and ground truth from ChIP–seq, eQTLGen/GTEx, promoter-capture Hi-C, single-cell eQTL, KnockTF perturbation data and GWAS summary statistics.
  • Performance was assessed mainly by AUC, AUPR ratio, F1 score and paired t-tests, with comparisons to scNN, BulkNN, PCC, motif binding affinity, GENIE3, PIDC, SCENIC+ and CellOracle.

Limitations

  • The paper states that TF self-regulation is not considered because it is challenging to model without additional data.
  • The paper states that lifelong learning involves a tradeoff between retaining bulk prior knowledge and fitting new single-cell data, and that the flexibility of variation in prior knowledge is not constrained when fitting new data.
  • Evident from the text: benchmarking is largely limited to blood/PBMC and the H1 cell line, with ground truth restricted to available ChIP–seq, eQTL, Hi-C and perturbation datasets; generalization to other tissues or cell types is not established here.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41587-024-02182-7

Cites

96 of 96 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

26 of 48 cited sources checked · 14 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    GRN inference accuracy has remained disappointingly low, marginally exceeding random predictions.

    “Despite extensive efforts, GRN inference accuracy has remained disappointingly low, marginally exceeding random predictions39.”
    • Also stated in: § Main · plain
      “Third, inferred GRN accuracy assessed by experimental data is only marginally better than random prediction39.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [39] Pratapa, A., Jalihal, A. P., Law, J. N., Bharadwaj, A. & Murali, T. M. Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data. Nat. Methods 17, 147–154 (2020). source
  • Premisecitation

    Single-cell RNA sequencing data enables cell type-specific trans-regulation inference through co-expression analysis.

    “Single-cell RNA sequencing (scRNA-seq) data enables cell type-specific trans-regulation inference through co-expression analysis such as PIDC and SCENIC21–30.”
    Quote in the paperSubstantial claim7 of 10 sources checked
    • cites [21] Chan, T. E., Stumpf, M. P. H. & Babtie, A. C. Gene regulatory network inference from single-cell data using multivariate information measures. Cell Syst. 5, 251–267.e3 (2017). source

      The Introduction describes using statistical relationships among single-cell gene-expression profiles to infer putative regulatory relationships, and the Results show PIDC identifying relationships specific to cell subpopulations. But the paper’s experimental applications use single-cell qPCR data, not scRNA-seq, and its inferred relationships do not establish specific trans-regulatory interactions. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [22] Aibar, S. et al. SCENIC: single-cell regulatory network inference and clustering. Nat. Methods 14, 1083–1086 (2017). source
    • cites [23] Matsumoto, H. et al. SCODE: an efficient regulatory network inference algorithm from single-cell RNA-seq during differentiation. Bioinformatics 33, 2314–2321 (2017). source
    • cites [24] Papili Gao, N., Ud-Dean, S. M. M., Gandrillon, O. & Gunawan, R. SINCERITIES: inferring gene regulatory networks from time-stamped single cell transcriptional expression profiles. Bioinformatics 34, 258–266 (2018). source

      The Introduction and Discussion say that single-cell expression data can be used to infer gene regulatory networks while revealing cellular heterogeneity. But the Discussion describes SINCERITIES as inferring networks from time-stamped changes in expression distributions, not co-expression analysis; the paper does not establish the claim specifically for cell type-specific trans-regulation or identify PIDC and SCENIC as examples. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [25] Sanchez-Castillo, M., Blanco, D., Tienda-Luna, I. M., Carrion, M. C. & Huang, Y. A Bayesian framework for the inference of gene regulatory networks from time and pseudo-time series data. Bioinformatics 34, 964–970 (2018). source
    • cites [26] Hu, Y., Peng, T., Gao, L. & Tan, K. CytoTalk: de novo construction of signal transduction networks using single-cell transcriptomic data. Sci. Adv. 7, eabf1356 (2021). source

      The Materials and Methods section (“Construction of intracellular functional gene interaction network”; page unavailable) says CytoTalk builds a coexpression network for each cell type from single-cell data. But the paper uses those networks to infer signal-transduction pathways between cell types; the cited passages do not establish cell type-specific trans-regulatory inference or describe PIDC or SCENIC. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [27] Frankowski, P. C. A. & Vert, J. P. Gene regulation inference from single-cell RNA-seq data with linear differential equations and velocity inference. Bioinformatics 36, 4774–4780 (2020). source

      The Abstract says scRNA-seq offers possibilities for inferring gene regulatory networks, supporting the general inference claim. It describes the paper’s approach as velocity inference and differential-equation modeling, but does not support the more specific claim of cell type-specific inference through co-expression analysis such as PIDC or SCENIC. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [28] Specht, A. T. & Li, J. LEAP: constructing gene co-expression networks for single-cell RNA-sequencing data using pseudotime ordering. Bioinformatics 33, 764–766 (2017). source

      The Methods and Conclusion describe using scRNA-seq co-expression to identify directional associations that may reflect gene regulation. The Results apply LEAP to dendritic cells, but the paper does not establish cell type-specific trans-regulation inference; it characterizes the inferred associations as likely regulatory rather than confirmed regulation. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [29] Moerman, T. et al. GRNBoost2 and Arboreto: efficient and scalable inference of gene regulatory networks. Bioinformatics 35, 2159–2161 (2019). source
    • cites [30] Zhang, S. et al. Inference of cell type-specific gene regulatory networks on cell lineages from single cell omic datasets. Nat. Commun. 14, 3064 (2023).

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Premisecitation

    These TFs have been previously reported to be associated with IBD in the literature.

    “These TFs have been previously reported to be associated with IBD in the literature.”
    Quote in the paperNot a substantial claim2 of 3 sources checked
    • cites [66] Wang, X., Guo, R., Lv, Y. & Fu, R. The regulatory role of Fos related antigen-1 in inflammatory bowel disease. Mol. Med. Rep. 17, 1979–1985 (2018). source

      The Introduction reports an association between Fra-1 expression in intestinal epithelial cells and IBD activity, and the Discussion reports elevated STAT1, STAT6 and CREB signaling in IBD patients. The citing sentence does not identify which transcription factors “these TFs” refers to, so this paper cannot verify the claim for the entire unspecified group. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
    • cites [67] Nowak, J. K. et al. Characterisation of the circulating transcriptomic landscape in inflammatory bowel disease provides evidence for dysregulation of multiple transcription factors including NFE2, SPI1, CEBPB, and IRF2. J. Crohns Colitis 16, 1255–1268 (2022). source
    • cites [68] Broom, O. J., Widjaya, B., Troelsen, J., Olsen, J. & Nielsen, O. H. Mitogen activated protein kinases: A role in inflammatory bowel disease? Clin. Exp. Immunol. 158, 272–280 (2009). source
  • Findingown data

    LINGER outperforms all other methods in predicting direct differentially expressed genes under TF perturbation using eight PBMC datasets from KnockTF (P ≤ 3.72 × 10−4).

    “Extended Data Fig. 4a,b shows that LINGER outperforms all other methods (P ≤ 3.72 × 10−4).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [59] Feng, C. et al. KnockTF: a comprehensive human gene expression profile database with knockdown/knockout of transcription factors. Nucleic Acids Res. 48, D93–D100 (2020). source

      The Introduction identifies KnockTF as a database of human TF knockdown/knockout expression datasets, and Data Source and Processing describes its differentially expressed gene analyses. This supports citing KnockTF as a data source, but our copy does not report eight PBMC datasets or compare LINGER with other methods; it does not support the stated performance result or P-value. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • Rests on an assumption: Perturbation experiment data from KnockTF are valid ground truth for direct DEGs under TF perturbation.
  • Interpretationcitation

    The identified TFs (FOXK2 and NR4A1) have important roles in naive CD4+ T cell response upon TCR stimulation.

    “Other studies have also shown that FOXK2 affects the activation of T lymphocytes77,78 and revealed the essential roles of NR4A1 in regulatory T cell differentiation79,80, suggesting that the identified TFs have important roles in naive CD4+ T cell response upon TCR stimulation.”
    Quote in the paperSubstantial claim2 of 4 sources checked
    • cites [77] Blanchett, S., Boal-Carvalho, I., Layzell, S. & Seddon, B. NF-κB and extrinsic cell death pathways—entwined do-or-die decisions for T cells. Trends Immunol. 42, 76–88 (2021). source
    • cites [78] Oh, H. & Ghosh, S. NF-κB: roles and regulation in different CD4+ T-cell subsets. Immunol. Rev. 252, 41–51 (2013). source

      The Introduction discusses TCR-stimulated T-cell activation, and the section on natural regulatory T-cell development discusses NF-κB and FoxP3. Searches for NR4A1 and FOXK2 returned no passages linking either factor to regulatory T-cell differentiation or naive CD4+ T-cell responses. Our copy therefore does not address the claim about these two factors. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
    • cites [79] Sekiya, T. et al. Essential roles of the transcription factor NR4A1 in regulatory T cell differentiation under the influence of immunosuppressants. J. Immunol. 208, 2122–2130 (2022). source
    • cites [80] Fassett, M. S., Jiang, W., D’Alise, A. M., Mathis, D. & Benoist, C. Nuclear receptor Nr4a1 modulates both regulatory T-cell (Treg) differentiation and clonal deletion. Proc. Natl Acad. Sci. USA 109, 3891–3896 (2012). source
  • Classified as probably not a substantial claim. 24
  • May combine two claims. 30
  • The quote only partly states this claim. 2

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    EWC regularization encourages the posterior to remain close to the prior, retaining knowledge while adapting, preventing excessive changes and ensuring a more stable learning process.

    “EWC regularization encourages the posterior to remain close to the prior, retaining knowledge while adapting, preventing excessive changes and ensuring a more stable learning process47.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [47] Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. Proc. Natl Acad. Sci. USA 114, 3521–3526 (2017). source

  • Classified as probably not a substantial claim.

    Premisecitation

    Lifelong learning has been proven to leverage knowledge learned in previous tasks to learn the new task better.

    “Lifelong learning has been proven to leverage the knowledge learned in previous tasks to learn the new task better45.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [45] Parisi, G. I., Kemker, R., Part, J. L., Kanan, C. & Wermter, S. Continual lifelong learning with neural networks: a review. Neural Netw. 113, 54–71 (2019). source

      Section 3.5 defines positive forward transfer as prior learning improving performance on a later task, and Section 4.3 describes approaches that reuse knowledge across tasks. But Section 4.3 calls transfer learning an open challenge, while Section 3.5 also recognizes negative transfer. The review therefore supports that lifelong-learning approaches can improve later-task learning, not that lifelong learning has generally been proven to do so. Page numbers were not provided in the passages. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Genome-wide measurements of chromatin accessibility locate REs, enabling TF-RE connections by motif matching and connecting REs to nearby TGs.

    “Genome-wide measurements of chromatin accessibility, such as DNase-seq17 and assay for transposase-accessible chromatin sequencing (ATAC-seq)18, locate REs, enabling TF–RE connections by motif matching and connecting REs to their nearby TGs19.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites DNase-seq17

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites ATAC-seq18

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [19] Neph, S. et al. Circuitry and dynamics of human transcription factor regulatory networks. Cell 150, 1274–1286 (2012). source

      The Experimental Procedures (Regulatory network construction; page unavailable) describe matching TF motifs to DNaseI footprints and linking those footprints to genes whose transcription start sites are within 5,000 nucleotides. This supports the proposed connection method, but the study constructed networks for TF genes specifically, not target genes generally; the Discussion also cautions that proximity does not reliably assign distal regulatory elements to their target genes. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    scJoint was developed to transfer labels from scRNA-seq to scATAC-seq data, which may enable improved cell GRN inference.

    “Recently, scJoint38 was developed to transfer labels from scRNA-seq to scATAC-seq data, which may enable improved cell GRN inference.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [38] Lin, Y. et al. scJoint integrates atlas-scale single-cell RNA-seq and ATAC-seq data with transfer learning. Nat. Biotechnol. 40, 703–710 (2022). source

  • May combine two claims.

    Findingown data

    FOXK2 and NR4A1 are activated at 8 h based on regulon activity, consistent with whole proteomics and phosphoproteomics data.

    “FOXK2 and NR4A1 are activated at 8 h based on regulon activity (Fig. 6d), which is consistent with the whole proteomics and phosphoproteomics data (Fig. 6e)76.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [76] Tan, H. et al. Integrative proteomics and phosphoproteomics profiling reveals dynamic signaling networks and bioenergetics pathways underlying T cell activation. Immunity 46, 488–503 (2017). source

      The Results section on TCR-stimulated proteome changes places NR4A1 in a whole-proteome cluster that rises at 2 h and then gradually declines. The cited paper does not establish FOXK2 or NR4A1 activation at 8 h based on regulon activity, nor does the retrieved phosphoproteomics discussion show matching changes for those factors. Its NR4A1 finding therefore supports only part of the claimed consistency. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    GWASs have identified thousands of disease variants, but the active cells and functions involving variant-regulated genes remain largely unknown.

    “GWASs have identified thousands of disease variants, but the active cells and functions involving variant-regulated genes remain largely unknown61.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [61] Jagadeesh, K. A. et al. Identifying disease-critical cell types and cellular processes by integrating single-cell RNA-sequencing and human genetics. Nat. Genet. 54, 1479–1492 (2022). source

  • Classified as probably not a substantial claim.

    Premisecitation

    EWC constrains the parameters of the first layer to stay in a region previously learned from the bulk data.

    “EWC constrains the parameters of the first layer to stay in a region of , which is previously learned from the bulk data45.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [45] Parisi, G. I., Kemker, R., Part, J. L., Kanan, C. & Wermter, S. Continual lifelong learning with neural networks: a review. Neural Netw. 113, 54–71 (2019). source

      Section 3.2 (page unavailable) describes EWC as penalizing changes to task-relevant parameters relative to values learned on a previous task. It does not restrict this constraint to the first layer or describe the previous training data as “bulk data.” (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    These findings align with previous studies linking monocytes to the pathogenesis of IBD.

    “These findings align with previous studies linking monocytes to the pathogenesis of IBD64,65.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [64] Anderson, A. et al. Monocytosis is a biomarker of severity in inflammatory bowel disease: analysis of a 6-year prospective natural history registry. Inflamm. Bowel Dis. 28, 70–78 (2022). source
    • cites [65] Aschenbrenner, D. et al. Deconvolution of monocyte responses in inflammatory bowel disease reveals an IL-1 cytokine network that regulates IL-23 in genetic and acquired IL-10 resistance. Gut 70, 1023–1036 (2021). source

  • The quote only partly states this claim.

    Findingown data

    Among 671 individuals with AML, individuals with high FOXN1 activity level tend to have a higher survival probability (two-sided log-rank test).

    “we calculated the TF activity of the transcriptome profile (bulk RNA-seq data) of 671 individuals with AML73 and performed survival analysis, which indicated that individuals with high FOXN1 activity level tend to have a higher survival probability (Fig. 6c).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [73] Bottomly, D. et al. Integrative analysis of drug response and clinical outcome in acute myeloid leukemia. Cancer Cell 40, 850–864.e9 (2022). source
    • Rests on an assumption: TF activity, focusing on the DNA-binding component of TF proteins, is a more reliable metric than mRNA for identifying driver regulators.
    • Rests on an assumption: Estimating TF activity based on the expression patterns of downstream TGs requires an accurate GRN.
    • The authors state this as an assumption: The GRN structure is consistent for the same cell type across individuals.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The Shapley value measures the contribution of features in a machine-learning model and is widely used in algorithms such as deep learning, graphical models and reinforcement learning.

    “The Shapley value measures the contribution of features in a machine-learning model and is widely used in algorithms such as deep learning, graphical models and reinforcement learning85.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [85] Rozemberczki, B. et al. The Shapley value in machine learning. Preprint at https://doi.org/10.48550/arXiv.2202.05594 (2022). source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Chromatin accessibility alterations occur in age-related macular degeneration.

    “which aligns with the conclusion that chromatin accessibility alterations occur in age-related macular degeneration54.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [54] Wang, J. et al. ATAC-seq analysis reveals a widespread decrease of chromatin accessibility in age-related macular degeneration. Nat. Commun. 9, 1364 (2018). source

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    A previous study found that disease-associated genes from GWAS were enriched in selectively constrained genes, while eQTL genes were depleted in selectively constrained genes.

    “A previous study found that disease-associated genes from GWAS were enriched in selectively constrained genes, while eQTL genes were depleted in selectively constrained genes52.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [52] Mostafavi, H., Spence, J. P., Naqvi, S. & Pritchard, J. K. Systematic differences in discovery of genetic effects on gene expression and complex traits. Nat. Genet. 55, 1866–1875 (2023). source

  • May combine two claims.

    Findingown data

    The estimated success probability of being loss-of-function intolerant is 0.26 for trans-dominant genes and 0.09 for cis-dominant genes.

    “Points show estimated success probability from binomial distribution, at 0.26 and 0.09 for trans-dominant and cis-dominant, respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    TFs contribute more to predicting the expression of trans-dominant genes, while REs contribute more to cis-dominant genes.

    “TFs contribute more to predicting the expression of trans-dominant genes, while REs contribute more to cis-dominant genes.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    LINGER reveals a complex regulatory landscape of GWAS, enabling enhanced interpretation of disease-associated variants and genes.

    “reveals a complex regulatory landscape of genome-wide association studies, enabling enhanced interpretation of disease-associated variants and genes.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Shapley value estimates the contribution of each feature for each gene.
    • Rests on an assumption: The correlation of TF and RE parameters learned in the second layer reflects TF–RE binding strength.

  • May combine two claims.

    Findingown data

    scNN modeled gene expression better than elastic net, with −log10P = 572.09, especially for genes with negative Pearson's correlation coefficient in elastic net predictions (−log10P = 1,060.17).

    “We found that scNN modeled gene expression better than elastic net, with −log10P = 572.09, especially for those substantial proportions of genes that show negative Pearson’s correlation coefficient (PCC) in elastic net predictions (−log10P = 1,060.17; Fig. 2a).”
    • Also stated in: § LINGER improves the cellular population GRN inference · boosted
      “This demonstrates that the three-layer neural network model scNN outperforms the elastic net model in predicting gene expression.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    LINGER performs better than other methods, with significantly higher AUC and AUPR ratio across all ground truth data for trans-regulatory strength.

    “Compared to other methods, LINGER performs better, with a significantly higher AUC (Fig. 2b) and AUPR ratio (Fig. 2c) across all ground truth data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    LINGER enables the estimation of transcription factor activity solely from bulk or single-cell gene expression data, leveraging the abundance of available gene expression data to identify driver regulators from case-control studies.

    “Following the GRN inference from reference single-cell multiome data, LINGER enables the estimation of transcription factor activity solely from bulk or single-cell gene expression data, leveraging the abundance of available gene expression data to identify driver regulators from case-control studies.”
    • Also stated in: § Main · plain
      “Fourth, LINGER enables the estimation of TF activity solely from gene expression data, identifying driver regulators.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Shapley value estimates the contribution of each feature for each gene.
    • Rests on an assumption: The correlation of TF and RE parameters learned in the second layer reflects TF–RE binding strength.

  • The quote only partly states this claim.

    Findingown data

    LINGER improves the cellular population GRN inference.

    “LINGER improves the cellular population GRN inference”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Shapley value estimates the contribution of each feature for each gene.
    • Rests on an assumption: The correlation of TF and RE parameters learned in the second layer reflects TF–RE binding strength.

  • May combine two claims.

    Findingown data

    The sigmoid activation function would not improve gene expression prediction but would improve GRN inference.

    “Note that the sigmoid activation function would not improve the gene expression prediction but would improve the GRN inference (Extended Data Fig. 2a).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In naive CD4 T cells, LINGER achieves AUC ranging from 0.66 to 0.70 and AUPR ratio ranging from 1.81 to 2.16 across all distance groups, while other methods are close to random.

    “In naive CD4 T cells, LINGER achieves AUC ranging from 0.66 to 0.70 (Fig. 3f) and AUPR ratio ranging from 1.81 to 2.16 (Fig. 3g) across all distance groups, while other methods are close to random.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In naive CD4 T cells, LINGER achieves an AUC of 0.92 and an AUPR ratio of 5.17 for ETS1, which is an improvement over PCC (AUC 0.78; AUPR 2.71) and motif binding affinity (AUC 0.70; AUPR 1.92).

    “For example, in naive CD4 T cells, LINGER achieves an AUC of 0.92 and an AUPR ratio of 5.17 for ETS1, which is an improvement over PCC (AUC, 0.78; AUPR ratio, 2.71) and motif binding affinity (AUC, 0.70; AUPR ratio, 1.92) (Fig. 3a,e).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    The use of lifelong learning to leverage external data does not lead to improved gene expression prediction but does improve the GRN inference.

    “However, the use of lifelong learning to leverage external data does not lead to improved gene expression prediction but does improve the GRN inference.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Among the top 200 TGs regulated by STAT1 in classical monocytes, 67 overlap with GWAS genes, and this overlap is statistically significant (P < 0.01) based on a background distribution from a random selection of TFs (one-sided bootstrap hypothesis testing).

    “Among the top 200 TGs regulated by STAT1 in classical monocytes, 67 of them overlap with the GWAS genes, which is statistically significant with a P value of less than 0.01 based on a background distribution from a random selection of TFs (one-sided bootstrap hypothesis testing).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    GATA1 knockout shifted proerythroblasts to a megakaryocytic or erythroid progenitor state, consistent with GATA1's role in inhibiting erythroblast maturation.

    “CellOracle predictions based on the LINGER GRN showed that GATA1 knockout shifted proerythroblasts to a megakaryocytic or erythroid progenitor state (Extended Data Fig. 4e), consistent with the functional role of GATA1 in inhibiting erythroblast maturation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    LINGER consistently performs better than other methods for all 20 TFs in PBMCs, with a significantly higher AUC and AUPR ratio (P ≤ 9.49 × 10−9).

    “Overall, LINGER consistently performs better than other methods for all 20 TFs in PBMCs, with a significantly higher AUC and AUPR ratio (P ≤ 9.49 × 10−9; Fig. 4c,d and Supplementary Table 7).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: ChIP-seq putative targets are valid ground truth for trans-regulatory potential.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Figure 5c lists the top eight candidate TFs in classical monocytes.

    “Figure 5c lists the top eight candidate TFs in classical monocytes.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For all 20 TFs in PBMCs, LINGER consistently exhibits the highest AUC and AUPR ratios, and overall distributions are significantly higher than others in PBMCs (P ≤ 8.72 × 10−5).

    “For all 20 TFs in PBMCs, LINGER consistently exhibits the highest AUC and AUPR ratios, and the overall distributions are significantly higher than others in PBMCs (P ≤ 8.72 × 10−5; Fig. 3b,c and Supplementary Table 6).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The overlap between STAT1 top-ranked TGs and GWAS genes is statistically significant for the top 500, 1,000, 2,000 and 5,000 TGs.

    “The numbers of overlapped TGs are all significant for the top 500, 1,000, 2,000 and 5,000 TGs (Fig. 5d).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The average AUPR ratio across ground truth datasets for other methods was 1.17-1.29, 0.17-0.29 units above random prediction, whereas LINGER achieves 1.25 units above random prediction, indicating a fourfold to sevenfold relative increase.

    “The average AUPR ratio across ground truth datasets for other methods was 1.17–1.29, 0.17–0.29 units above random prediction, whereas LINGER achieves 1.25 units above random prediction, indicating a fourfold to sevenfold relative increase (Fig. 4d).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: ChIP-seq putative targets are valid ground truth for trans-regulatory potential.

  • May combine two claims.

    Findingown data

    Compared to weight 0, weight 0.01 improved performance on 100% of ground truth data based on AUC and 80% based on AUPR ratio.

    “Compared to 0, weight 0.01 improved the performance on 100% (Extended Data Fig. 2c) and 80% (Extended Data Fig. 2d) of ground truth data based on the AUC and AUPR ratio, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Taking eQTL as ground truth, F1 score of LINGER is significantly higher than SCENIC+ and other cell types.

    “Taking eQTL as ground truth, the F1 score of LINGER is significantly higher than SCENIC+ (Fig. 3k) and other cell types (Extended Data Fig. 3k–o).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Methods integrating unpaired scRNA-seq and scATAC-seq link TFs to REs by motif matching and link REs to TGs using covariation of RE-TG or physical base pair distance.

    “Those methods, including IReNA32, SOMatic33, UnpairReg34, CoupledNMF35,36, DC3 (ref. 36) and others37 link TFs to REs by motif matching and link REs to TGs using the covariation of RE–TG or physical base pair distance.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites IReNA32

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites SOMatic33

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites UnpairReg34

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    scATAC-seq can be used to infer trans-regulation.

    “Single-cell sequencing assay for transposase-accessible chromatin (scATAC-seq) can be used to infer trans-regulation, as in DeepTFni31.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [31] Li, H. et al. Inferring transcription factor regulatory networks from single-cell ATAC-seq data based on graph neural networks. Nat. Mach. Intell. 4, 389–400 (2022). source

  • Classified as probably not a substantial claim.

    Premisecitation

    Co-expressions are interpreted as correlations rather than causal regulations.

    “Moreover, co-expressions are interpreted as correlations rather than causal regulations16.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [16] Wang, Y. X. R. & Huang, H. Review on statistical methods for gene network reconstruction using expression data. J. Theor. Biol. 362, 53–61 (2014). source

  • Classified as probably not a substantial claim.

    Premisecitation

    Co-expression-based methods infer TF-TG trans-regulation from gene expression by capturing TF-TG covariation.

    “Co-expression-based methods such as WGCNA14, ARACNe9 and GENIE3 (ref. 15) infer the TF–TG trans-regulation from gene expression by capturing the TF–TG covariation.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites WGCNA14

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites ARACNe9

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites GENIE3 (ref. 15)

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Gene expression prediction ability was assessed using fivefold cross-validation.

    “We assessed the gene expression prediction ability of the two models using fivefold cross-validation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Many methods integrate unpaired scRNA-seq and scATAC-seq data to infer trans-regulation.

    “Many methods integrate unpaired scRNA-seq and scATAC-seq data to infer trans-regulation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    The extent to which the final result deviates from the prior knowledge depends on the loss incurred in fitting the new data.

    “The extent to which the final result deviates from the prior knowledge depends on the loss incurred in fitting the new data.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The significant overlap of both DEGs and GWAS with the top-ranked TGs of STAT1 indicates the robustness and unbiased nature of our method.

    “The lack of significant overlap between DEGs and GWAS genes (P = 0.15, two-sided Fisher’s exact test) but the significant overlap of both DEGs and GWAS with the top-ranked TGs of STAT1 indicates the robustness and unbiased nature of our method.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    LINGER can not only predict gene expression under perturbation but also enable downstream characterizations of differentiation trajectories through integration with complementary analytical frameworks like CellOracle.

    “These results demonstrate that LINGER can not only predict gene expression under perturbation but also enable downstream characterizations of differentiation trajectories through integration with complementary analytical frameworks like CellOracle.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The flexibility of the variation in prior knowledge is not constrained when fitting the new data.

    “The flexibility of the variation in prior knowledge is not constrained when fitting the new data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    LINGER's lifelong learning approach has the advantage of pre-training on bulk collections, allowing users to easily retrain the model for their own studies while capitalizing on publicly available resources without direct access.

    “LINGER’s lifelong learning approach has the advantage of pre-training on bulk collections, allowing users to easily retrain the model for their own studies while capitalizing on publicly available resources without direct access.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationnone given

    The first hidden layer has 64 neurons with rectified linear unit activation that can capture regulatory modules, each of which contains multiple TFs and REs.

    “The first hidden layer has 64 neurons with rectified linear unit activation that can capture regulatory modules, each of which contains multiple TFs and REs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    For a random predictor, the AUPR equals the fraction of positive samples in the dataset.

    “For a random predictor, the AUPR equals the fraction of positive samples in the dataset.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    These regulatory modules are characterized by enriched motifs of the TFs on the corresponding REs.

    “These regulatory modules are characterized by enriched motifs of the TFs on the corresponding REs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The lifelong learning mechanism will encourage the model to retain prior knowledge from the bulk data when adapting to the new single-cell data.

    “The lifelong learning mechanism will encourage the model to retain prior knowledge from the bulk data when adapting to the new single-cell data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    LINGER will learn this tradeoff automatically to obtain a maximized usage of the information from both datasets.

    “LINGER will learn this tradeoff automatically to obtain a maximized usage of the information from both datasets.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    It is a tradeoff between retaining prior knowledge and fitting new data.

    “It is a tradeoff between retaining prior knowledge and fitting new data.”
    Quote in the paperSubstantial claim
49 of 49

  • The paper relies on this without support.

    Most single-cell data points are not independent.

    “Although single-cell data offers a large number of cells, most of them are not independent.”

  • The paper relies on this without support.

    Incorporating prior knowledge such as motif matching into non-linear models is challenging.

    “Second, incorporating prior knowledge such as motif matching into non-linear models is challenging.”

  • The paper relies on this without support.

    Learning a complex mechanism like GRN from limited data points remains a challenge.

    “First, learning such a complex mechanism from limited data points remains a challenge.”

  • The paper relies on this without support.

    Shapley value estimates the contribution of each feature for each gene.

    “we infer the regulatory strength of TF–TG and RE–TG interactions using the Shapley value, which estimates the contribution of each feature for each gene.”

  • The paper relies on this without support.

    The correlation of TF and RE parameters learned in the second layer reflects TF–RE binding strength.

    “The TF–RE binding strength is generated by the correlation of TF and RE parameters learned in the second layer (Fig. 1a).”

  • The paper relies on this without support.

    ChIP-seq putative targets are valid ground truth for trans-regulatory potential.

    “The ground truth data in a–d are putative targets of TFs from ChIP–seq data for the corresponding cell types in PBMCs.”

  • The paper relies on this without support.

    Perturbation experiment data from KnockTF are valid ground truth for direct DEGs under TF perturbation.

    “using perturbation experiment data as ground truth”

  • The paper relies on this without support.

    The underlying regulatory drivers of disease remain elusive despite researchers often identifying DEGs between cases and controls using bulk or single-cell expression data.

    “Researchers often identify DEGs between cases and controls using bulk or single-cell expression data, but the underlying regulatory drivers remain elusive.”

  • The paper relies on this without support.

    TF activity, focusing on the DNA-binding component of TF proteins, is a more reliable metric than mRNA for identifying driver regulators.

    “TF activity, focusing on the DNA-binding component of TF proteins, is a more reliable metric than mRNA for identifying driver regulators.”

  • The paper relies on this without support.

    Estimating TF activity based on the expression patterns of downstream TGs requires an accurate GRN.

    “One feasible approach is to estimate TF activity based on the expression patterns of downstream TGs, which necessitates the availability of an accurate GRN.”

  • The authors state this as an assumption.

    The GRN structure is consistent for the same cell type across individuals.

    “Assuming that the GRN structure is consistent for the same cell type across individuals”

  • The paper relies on this without support.

    Traditionally, GRN inference performance is assessed by gene expression prediction.

    “Traditionally, GRN inference performance is assessed by gene expression prediction.”
12 of 12

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • TCR-stimulation regulator is named FOXK2 in the text but FOXK1 in Fig. 6

    Major

    The text identifies FOXK2 as the regulator activated after stimulation and links that result to Fig. 6d,e, but the caption identifies the regulator in both panels as FOXK1.

  • Fig. 4 reports a P value above its stated maximum

    Minor

    The reported −log10(P) of 7.72 for the Fig. 4c PCC comparison corresponds to P ≈ 1.9 × 10−8, which exceeds the stated upper bound of 9.49 × 10−9 for the Fig. 4c,d comparisons.

2 of 2

Cites

Flagged first, then by citations

4 of the 100 references couldn’t be matched to a known paper.

Research works printed but not linked: 1

  1. Chaudhri, Z. & Liu, B. Lifelong Machine Learning (Springer International Publishing, 2022).

Cited by

Sorted by citations

Related

Authors
Yuan, Q., & Duren, Z.
Affiliations
Clemson University (US)
Corresponding author
Duren, Zhana
ORCID
Venue
Nature Biotechnology
Publisher
Springer Science and Business Media LLC
Published
2024-04-12
Volume
43
Issue
2
Pages
247-257
ISSN
  • 0733-222X (ISSN-L)
  • 1087-0156
  • 1546-1696
DOI
10.1038/s41587-024-02182-7
PMC
PMC11825371
OpenAlex
W4394764901
Semantic Scholar
93e99767340c76e0af7f776ed66f5a1299b84479
Cited by
118
References
100
Citations by year
2026: 48 · 2025: 53 · 2024: 17
FWCI
13.74
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Biotechnology (publishedVersion, cc-by)
  • Research Square (acceptedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion, cc-by)
Funders
U.S. Department of Health & Human Services | NIH | National Institute of General Medical Sciences (P20 GM139769, R35 GM150513)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene Regulatory Network Analysis
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Chromatin Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
Keywords
gene regulatory network inference, chromatin accessibility, gene expression, manifold regularization, transcription factor binding motifs, case-control study, disease-associated variants, genome-wide association study
MeSH terms
Machine Learning, Chromatin, Gene Expression Regulation, Humans, Transcription Factors, Neural Networks, Computer, Computational Biology, Gene Expression Profiling, Gene Regulatory Networks, Genome-Wide Association Study, Single-Cell Analysis

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-14.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-14.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review