CoChat
Sign Up
Cited by 29Genomics and Chromatin DynamicsGene Regulatory Network AnalysisDevelopmental Biology and Gene Regulation

Design principles of cell-state-specific enhancers in hematopoiesis

Open access
Frömel, Robert, Rühle, Julia, Bernal Martinez, Aina, Szu-Tu, Chelsea, Pacheco Pastor, Felix, Martinez-Corral, Rosa, Velten, Lars·2025·Cell
DOI 10.1016/j.cell.2025.04.017PMC PMC12173716

Published version of a preprint

Open this paper in CoChat — free account

Abstract

During cellular differentiation, enhancers transform overlapping gradients of transcription factors (TFs) to highly specific gene expression patterns. However, the vast complexity of regulatory DNA impedes the identification of the underlying cis-regulatory rules. Here, we characterized 64,400 fully synthetic DNA sequences to bottom-up dissect design principles of cell-state-specific enhancers in the context of the differentiation of blood stem cells to seven myeloid lineages. Focusing on binding sites for 38 TFs and their pairwise interactions, we found that identical sites displayed both repressive and activating function as a consequence of cell state, site combinatorics, or simply predicted occupancy of a TF on an enhancer. Surprisingly, combinations of activating sites frequently neutralized one another or gained repressive function. These negative synergies convert quantitative imbalances in TF expression into binary activity patterns. We exploit this principle to automatically create enhancers with specificity to user-defined combinations of hematopoietic progenitor cell states from scratch.

AI snapshot

Written from the paper’s full text.

TL;DR

This paper uses 64,400 fully synthetic DNA sequences and a lentiviral MPRA to dissect how transcription factor binding sites encode cell-state-specific enhancer activity during blood stem cell differentiation. It identifies negative synergies between TF motifs, including activator pairs that become repressors, as a design principle that converts overlapping TF gradients into binary, cell-state-specific expression and enables automated enhancer design.

Key findings

  • Across seven murine hematopoietic progenitor cell states, single-factor motifs were classified as 3/38 inactive, 20/38 activators, and 10/38 repressors; Myc, Gata1/Tal1, Gata2, Creb1, and Nr2c2 showed cell-state-dependent duality, and seven factors showed non-monotonous, occupancy-dependent duality.
  • Pairs of activating motifs frequently neutralized one another or became repressive: Fli1-Spi1 and Gata2-Cebpa individually showed little repression but together elicited strong repression in some cell states; 8/11 duality-associated interactions were significantly affected by spacing and/or orientation, versus 11/34 other motif pairs.
  • The authors trained models on pairwise motif interactions and performed 64 automated design tasks for activation/inactivity across six cell states; the best strategy, model-guided motif placement followed by local search, produced desired activity in 75% of cases and desired inactivity in 64% of cases (F1 for activity: 0.73).
  • In K562 leukemia cells, repressive function was frequently lost: 15/20 HPC activators were also activators in K562, but repressor agreement was poor, and many motif combinations that were neutralizing in primary HPCs drove transcription in K562.

Methods

  • Synthetic enhancer libraries A–H were cloned upstream of a minimal promoter in a lentiviral MPRA vector; primary mouse Lin−Kit+ HSPCs were infected and differentiated for four days, then FACS-sorted into seven progenitor cell states, with additional screens in K562 cells.
  • Libraries A–C tested single TF motifs and pairwise TF motif combinations for 38–42 TFs, varying motif affinity, number, spacing, orientation, and arrangement on random DNA backgrounds; 445,220 HPC and 100,923 K562 measurements passed quality filters.
  • Analyses included random-DNA-based activation/repression thresholds, random forests, shape-constrained additive models, a deep convolutional model, biophysical modeling of polymerase-cycle regulation, AlphaFold3/Chai-1 structural prediction, and model-guided sequence design.

Limitations

  • Stated by the paper: models remain limited to the role of 38 TFs; the approach is limited to correlative analyses across seven cell states; 3D chromatin context effects were removed by averaging across lentiviral integration sites.
  • Stated by the paper: activity measurements were available for only seven cell states, impeding more quantitative analysis of correlations between several TFs’ expression and activity; the assay was more powered to detect activation than repression because of the low baseline activity of random DNA.
  • Evident from the text: the K562 comparison is a leukemia cell line context, and the authors suggest its permissive epigenetic environment may explain the loss of repressive interactions, so those findings may not generalize to primary cells.

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

Integrity signals

Registry

Crossref · 10.1016/j.cell.2025.04.017

Cites

106 of 106 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

57 of 64 cited sources checked · 27 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Altered ratios between TFs can be sufficient to establish specific gene expression programs.

    “Indeed, altered ratios between TFs can be sufficient to establish specific gene expression programs.9,10”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [9] Dahl R. Walsh J.C. Lancki D. Laslo P. Iyer S.R. Singh H. Simon M.C. Regulation of macrophage and neutrophil cell fates by the PU.1:C/EBPα ratio and granulocyte colony-stimulating factor Nat. Immunol. 4 2003 1029 1036 10.1038/ni973 12958595 source
    • cites [10] Hosokawa H. Ungerbäck J. Wang X. Matsumoto M. Nakayama K.I. Cohen S.M. Tanaka T. Rothenberg E.V. Transcription Factor PU.1 Represses and Activates Gene Expression in Early T Cells by Redirecting Partner Transcription Factor Binding Immunity 48 2018 1119 1134.e7 10.1016/j.immuni.2018.04.024 29924977 PMC6063530 source
  • Premisecitation

    Overlapping TF gradients accompany cellular differentiation in systems including exit from pluripotency, organogenesis, and tissue regeneration by adult stem cells.

    “Similar overlapping TF gradients accompany cellular differentiation in systems ranging from the exit from pluripotency5 over organogenesis6 to tissue regeneration by other adult stem cells.7,8”
    Quote in the paperNot a substantial claim4 of 4 sources checked
    • cites [5] Loh K.M. Lim B. A precarious balance: Pluripotency factors as lineage specifiers Cell Stem Cell 8 2011 363 369 10.1016/j.stem.2011.03.013 21474100 source

      The paper’s discussion of pluripotency-factor expression (indexed as Abstract; page unavailable) reports variation in Oct4 and Sox2 levels among embryonic stem cells and cites findings that changes in their levels can elicit differentiation. It proposes that competing factors influence lineage choice, but does not establish that overlapping TF gradients accompany exit from pluripotency. Citation 5 is attached to that part of the citing sentence, not to its separate claims about organogenesis or adult stem cells. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [6] Pijuan-Sala B. Griffiths J.A. Guibentif C. Hiscock T.W. Jawaid W. Calero-Nieto F.J. Mulas C. Ibarra-Soria X. Tyser R.C.V. Ho D.L.L. A single-cell molecular map of mouse gastrulation and early organogenesis Nature 566 2019 490 495 10.1038/s41586-019-0933-9 30787436 PMC6522369 source

      Citation 6 attaches to organogenesis, not to the separately cited claims about exit from pluripotency or adult stem cells. In “Mapping endoderm development” (page unavailable), the paper reports transcription factors with sequential activation profiles during hindgut maturation and specification. This supports changing TF expression during early organogenesis, but the passage does not establish the claimed overlapping TF gradients. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
    • cites [7] Cockburn K. Annusver K. Gonzalez D.G. Ganesan S. May D.P. Mesa K.R. Kawaguchi K. Kasper M. Greco V. Gradual differentiation uncoupled from cell cycle exit generates heterogeneity in the epidermal stem cell layer Nat. Cell Biol. 24 2022 1692 1700 10.1038/s41556-022-01021-8 36357619 PMC9729105 source
    • cites [8] Dulken B.W. Leeman D.S. Boutet S.C. Hebestreit K. Brunet A. Single-Cell Transcriptomic Analysis Defines Heterogeneity and Transcriptional Dynamics in the Adult Neural Stem Cell Lineage Cell Rep. 18 2017 777 790 10.1016/j.celrep.2016.12.060 28099854 PMC5269583 source
  • Premisecitation

    Genome-wide datasets of TF binding and chromatin accessibility are not informative regarding the functional role of each TF in a given sequence context.

    “Genome-wide datasets of TF binding and chromatin accessibility13,22,23 are not informative regarding the functional role of each TF in a given sequence context.”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [13] Wilson N.K. Foster S.D. Wang X. Knezevic K. Schütte J. Kaimakis P. Chilarska P.M. Kinston S. Ouwehand W.H. Dzierzak E. Combinatorial Transcriptional Control In Blood Stem/Progenitor Cells: Genome-wide Analysis of Ten Major Transcriptional Regulators Cell Stem Cell 7 2010 532 544 10.1016/j.stem.2010.07.016 20887958 source
    • cites [22] Subramanian S. Thoms J.A.I. Huang Y. Cornejo-Páramo P. Koch F.C. Jacquelin S. Shen S. Song E. Joshi S. Brownlee C. Genome-wide transcription factor-binding maps reveal cell-specific changes in the regulatory architecture of human HSPCs Blood 142 2023 1448 1462 10.1182/blood.2023021120 37595278 PMC10651876 source
    • cites [23] Corces M.R. Buenrostro J.D. Wu B. Greenside P.G. Chan S.M. Koenig J.L. Snyder M.P. Pritchard J.K. Kundaje A. Greenleaf W.J. Lineage-specific and single-cell chromatin accessibility charts human hematopoiesis and leukemia evolution Nat. Genet. 48 2016 1193 1203 10.1038/ng.3646 27526324 PMC5042844 source

      The Results section ('Regulatory networks of normal hematopoiesis') says a motif sequence alone cannot identify the precise TF causing accessibility, especially when TFs share similar motifs. That supports a limitation of accessibility data, but not the broader claim that genome-wide TF-binding and accessibility datasets are uninformative about each TF’s functional role in a given sequence context. The paper also uses accessibility data, integrated with RNA-seq, to identify putative regulators. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
  • Interpretationreasoning

    Specific protein-protein interactions between these TFs or their co-factors may play a role in the switch from activators to repressors.

    “While the capability of these models to predict structures from intrinsically disordered regions is limited,50 together with our functional data on spacing dependency, these results suggest that specific protein-protein interactions between these TFs or their co-factors may play a role in the switch from activators to repressors.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [50] Abramson J. Adler J. Dunger J. Evans R. Green T. Pritzel A. Ronneberger O. Willmore L. Ballard A.J. Bambrick J. Accurate structure prediction of biomolecular interactions with AlphaFold 3 Nature 630 2024 493 500 10.1038/s41586-024-07487-w 38718835 PMC11168924 source
  • Interpretationcitation

    Structural modeling of Fli1-Spi1 and Cebpa-Gata2 interactions suggested that steric hindrance cannot explain these effects.

    “Structural modeling50,51 of the Fli1-Spi1 and Cebpa-Gata2 interactions suggested that steric hinderance cannot explain these effects (Figures 4F and S3H–S3K; see STAR Methods, section structural prediction of TF-TF-DNA complexes).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [50] Abramson J. Adler J. Dunger J. Evans R. Green T. Pritzel A. Ronneberger O. Willmore L. Ballard A.J. Bambrick J. Accurate structure prediction of biomolecular interactions with AlphaFold 3 Nature 630 2024 493 500 10.1038/s41586-024-07487-w 38718835 PMC11168924 source
    • cites [51] Chai Discovery Boitreaud J. Dent J. McPartlon M. Meier J. Reis V. Rogozhnikov A. Wu K. Chai-1: Decoding the molecular interactions of life Preprint at bioRxiv 2024 10.1101/2024.10.10.615955 source
  • May combine two claims. 59
  • Classified as probably not a substantial claim. 29
  • The quote only partly states this claim. 2

  • May combine two claims.

    Findingcitation

    In K562 cells, activators usually dominate, and repressive interactions are scarce.

    “In a human leukemia cell line commonly used for MPRA studies42,43 (K562), by contrast, activators usually dominate, and repressive interactions are scarce.”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [42] Agarwal V. Inoue F. Schubach M. Penzar D. Martin B.K. Dash P.M. Keukeleire P. Zhang Z. Sohota A. Zhao J. Massively parallel characterization of transcriptional regulatory elements Nature 639 2025 411 420 10.1038/s41586-024-08430-9 39814889 PMC11903340 source
    • cites [43] Kheradpour P. Ernst J. Melnikov A. Rogov P. Wang L. Zhang X. Alston J. Mikkelsen T.S. Kellis M. Systematic dissection of regulatory motifs in 2000 predicted human enhancers using a massively parallel reporter assay Genome Res. 23 2013 800 811 10.1101/gr.144899.112 23512712 PMC3638136 source

  • May combine two claims.

    Findingown data

    In some cases, activity of enhancers with Gfi1 sites was correlated with expression of Gfi1b, with weak evidence for stronger correlations with other factors from the same motif family.

    “In some cases, we found statistically weak evidence for stronger correlations with other factors from the same motif family,45 e.g., the activity of enhancers with Gfi1 sites was correlated with the expression of Gfi1b (Figure S3C).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [45] Vierstra J. Lazar J. Sandstrom R. Halow J. Lee K. Bates D. Diegel M. Dunn D. Neri F. Haugen E. Global reference mapping of human transcription factor footprints Nature 583 2020 729 736 10.1038/s41586-020-2528-x 32728250 PMC7410829 source

      The ‘Assigning TFs to footprints’ and ‘Cell-selective TF occupancy landscapes’ sections (page not available) say that closely related transcription factors can share recognition sequences, supporting the same-motif-family premise. They do not report enhancer-activity correlations with factor expression or evidence that those correlations are stronger for other family members. The Gfi1/Gfi1b example is presented as the citing authors’ own finding, with a citation to their Figure S3C. (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Runx1, Meis1, Gata1/Tal1, Gata2, and Lyl1 are part of a heptad of TFs with frequent co-binding to stem cell enhancers, and Gata1/Tal1, Spi1, Fli1, and Cebpa are master regulators of lineage differentiation.

    “Runx1, Meis1, Gata1/Tal1, Gata2, and Lyl1 are part of a heptad of TFs with frequent co-binding to stem cell enhancers,13 and Gata1/Tal1, Spi1, Fli1, and Cebpa are master regulators of lineage differentiation.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [13] Wilson N.K. Foster S.D. Wang X. Knezevic K. Schütte J. Kaimakis P. Chilarska P.M. Kinston S. Ouwehand W.H. Dzierzak E. Combinatorial Transcriptional Control In Blood Stem/Progenitor Cells: Genome-wide Analysis of Ten Major Transcriptional Regulators Cell Stem Cell 7 2010 532 544 10.1016/j.stem.2010.07.016 20887958 source

      The Results identify a seven-factor group that frequently binds together in blood stem/progenitor cells: SCL/TAL1, LYL1, LMO2, GATA2, RUNX1, ERG, and FLI-1. MEIS1 and GATA1 are not members of that heptad. The citation is attached to the heptad claim, not the subsequent lineage-differentiation claim. (LLM, conf=0.91)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation

    Under common biophysical assumptions, increases in motif number and affinity lead to higher TF occupancy.

    “Under common biophysical assumptions, increases in these two quantities lead to higher TF occupancy.46,47”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [46] Martinez-Corral R. Friedrich D. Frömel R. Velten L. Gunawardena J. DePace A.H. Emergence of activation or repression in transcriptional control under a fixed molecular context Preprint at bioRxiv 2024 10.1101/2024.05.29.596388 PMC12501156 40982681 source
    • cites [47] Barr K.A. Reinitz J. A sequence level model of an intact locus predicts the location and function of nonadditive enhancers PLoS One 12 2017 e0180861 10.1371/journal.pone.0180861 PMC5513433 28715438 source

      The Results section (‘Sequence level model without enhancer competition and eve expression’; page unavailable) says the model calculates equilibrium TF occupancy from DNA sequence and binding preferences. Materials and methods (‘Parameter estimation’; page unavailable) links binding scores to relative affinity. But the passages on binding-site number discuss the computational cost of calculating occupancy, not whether adding sites raises it; our copy does not state the claimed increase in occupancy from increasing either quantity. (LLM, conf=0.77)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    Factor pairs were only correlated between closely related lineages and separated lineages hierarchically according to known transcriptional and developmental relationships.

    “By contrast, factor pairs were only correlated between closely related lineages and separated lineages hierarchically according to the known transcriptional and developmental relationships between them3 (Figure 2C, right).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [3] Tusi B.K. Wolock S.L. Weinreb C. Hwang Y. Hidalgo D. Zilionis R. Waisman A. Huh J.R. Klein A.M. Socolovsky M. Population snapshots predict early haematopoietic and erythroid hierarchies Nature 555 2018 54 60 10.1038/nature25741 29466336 PMC5899604 source

      The Results section (‘The transcriptional state continuum of HPCs is hierarchical, but not a strict tree’; page not provided) describes correlated pairs of terminal fates used to construct a haematopoietic hierarchy, including specific related fate groups. It does not report that factor pairs were correlated only between closely related lineages; the correlated pairs used to build the hierarchy were fates, not factors. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

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

    Premisecitation

    The natural genome contains only few instances of many TF combinations and binding site arrangements and is statistically underpowered to learn the rules by which it is decoded.

    “it has been argued that the natural genome contains only few instances of many TF combinations and binding site arrangements, and it is in a statistical sense underpowered to “learn the rules by which it is decoded.”28,29”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [28] Lambert S.A. Jolma A. Campitelli L.F. Das P.K. Yin Y. Albu M. Chen X. Taipale J. Hughes T.R. Weirauch M.T. The Human Transcription Factors Cell 172 2018 650 665 10.1016/j.cell.2018.01.029 29425488 PMC12908702 source
    • cites [29] De Boer C.G. Taipale J. Hold out the genome: a roadmap to solving the cis-regulatory code Nature 625 2024 41 50 10.1038/s41586-023-06661-w 38093018 source

  • May combine two claims.

    Premisecitation

    Key hematopoietic TFs are expressed in smooth, overlapping gradients throughout stem and progenitor populations, whereas their target genes display cell-state-specific expression patterns.

    “single-cell transcriptomics1,2,3 and imaging-based4 studies have demonstrated that key TFs are expressed in smooth, overlapping gradients throughout the stem and progenitor populations (Figure 1A, top row); by contrast, the target genes regulated by these factors display cell-state-specific expression patterns (Figure 1A, bottom row).”
    Quote in the paperSubstantial claim3 of 4 sources checked
    • cites [1] Paul F. Arkin Y. Giladi A. Jaitin D.A. Kenigsberg E. Keren-Shaul H. Winter D. Lara-Astiaso D. Gury M. Weiner A. Transcriptional heterogeneity and lineage commitment in myeloid progenitors Cell 163 2015 1663 1677 10.1016/j.cell.2015.11.013 26627738 source
    • cites [2] Velten L. Haas S.F. Raffel S. Blaszkiewicz S. Islam S. Hennig B.P. Hirche C. Lutz C. Buss E.C. Nowak D. Human haematopoietic stem cell lineage commitment is a continuous process Nat. Cell Biol. 19 2017 271 281 10.1038/ncb3493 28319093 PMC5496982 source

      The Results sections “Early haematopoiesis is a continuous process” and “Molecular processes underlying HSC commitment” describe gradual differences among primitive cells and non-exclusive expression of early lineage-priming modules. They do not establish smooth, overlapping gradients of key transcription factors throughout both stem and progenitor populations, or the claimed contrast between those factors and their target genes; the paper also describes lineage-specific transcription-factor modules in later progenitors. Page numbers were not available in our copy. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [3] Tusi B.K. Wolock S.L. Weinreb C. Hwang Y. Hidalgo D. Zilionis R. Waisman A. Huh J.R. Klein A.M. Socolovsky M. Population snapshots predict early haematopoietic and erythroid hierarchies Nature 555 2018 54 60 10.1038/nature25741 29466336 PMC5899604 source

      The Results section (‘The erythroid differentiation trajectory’; page unavailable) reports that Gata1 is expressed early while most of its canonical targets are induced only at the transition to ETD, supporting a difference between TF and target-gene expression timing. It does not establish that key TFs generally form smooth, overlapping gradients throughout stem and progenitor populations, or that their targets generally have cell-state-specific patterns. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [4] Wheat J.C. Sella Y. Willcockson M. Skoultchi A.I. Bergman A. Singer R.H. Steidl U. Single-molecule imaging of transcription dynamics in somatic stem cells Nature 583 2020 431 436 10.1038/s41586-020-2432-4 32581360 PMC8577313 source

      The “PU.1 and Gata TF co-expression in HSPC” section reports substantial co-expression across the progenitor populations studied, and “Single molecule imaging in primary HSPC” reports different transcription of myeloid and erythroid genes in PU.1-high cells. These passages support overlapping TF expression and state-associated differences in selected genes, but do not establish smooth gradients throughout stem and progenitor populations or cell-state-specific expression patterns for target genes generally. Page numbers are unavailable in our copy. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    The study categorized hematopoietic transcription factors of Library A with generalized additive models under shape constraints.

    “We further categorized the hematopoietic TFs of Library A with generalized additive models under shape constraints92 (R package scam).”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [92] Pya N. Wood S.N. Shape constrained additive models Stat. Comput. 25 2015 543 559 10.1007/s11222-013-9448-7 source

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The design rules identified here are transparent insights into pairwise interactions between TF motifs.

    “Unlike recent studies that used genomic data and deep learning for designing cell-type-specific enhancers in the fruit fly61,62 or in cancer cell lines,63 the design rules identified here are transparent insights into pairwise interactions between TF motifs.”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [61] De Almeida B.P. Schaub C. Pagani M. Secchia S. Furlong E.E.M. Stark A. Targeted design of synthetic enhancers for selected tissues in the Drosophila embryo Nature 626 2024 207 211 10.1038/s41586-023-06905-9 38086418 PMC10830412 source
    • cites [62] Taskiran I.I. Spanier K.I. Dickmänken H. Kempynck N. Pančíková A. Ekşi E.C. Hulselmans G. Ismail J.N. Theunis K. Vandepoel R. Cell-type-directed design of synthetic enhancers Nature 626 2024 212 220 10.1038/s41586-023-06936-2 38086419 PMC10830415 source
    • cites [63] Gosai S.J. Castro R.I. Fuentes N. Butts J.C. Mouri K. Alasoadura M. Kales S. Nguyen T.T.L. Noche R.R. Rao A.S. Machine-guided design of cell-type-targeting cis-regulatory elements Nature 634 2024 1211 1220 10.1038/s41586-024-08070-z 39443793 PMC11525185 source

  • May combine two claims.

    Interpretationreasoning

    Such epigenetic alterations can potentially explain the shift toward activation observed for many TF combinations in K562 cells and the co-expression of stem and progenitor genes frequently observed in leukemia.

    “Such epigenetic alterations can potentially explain the shift toward activation observed for many TF combinations in K562 cells and the co-expression of stem and progenitor genes11 frequently observed in leukemia.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [11] Beneyto-Calabuig S. Merbach A.K. Kniffka J.-A. Antes M. Szu-Tu C. Rohde C. Waclawiczek A. Stelmach P. Gräßle S. Pervan P. Clonally resolved single-cell multi-omics identifies routes of cellular differentiation in acute myeloid leukemia Cell Stem Cell 30 2023 706 721.e8 10.1016/j.stem.2023.04.001 37098346 source

      In “LSCs retain erythroid capacity” (page unavailable), the paper reports that leukemic C6 cells expressed HSC-associated genes in some patients and genes associated with early erythromyeloid priming. This supports the cited point that stem- and progenitor-associated expression can coexist in AML cells, but does not establish that such co-expression is frequent across leukemia generally. The citation is attached to that co-expression point, not to the proposed epigenetic explanation or the K562 result. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    All primary hematopoietic cells activate Trp53 upon lentiviral infection without affecting the differentiation of cells in the ex vivo system used here.

    “all primary hematopoietic cells activate Trp53 upon lentiviral infection87 without affecting the differentiation of cells in the ex vivo system used here.37,39”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [87] Piras F. Riba M. Petrillo C. Lazarevic D. Cuccovillo I. Bartolaccini S. Stupka E. Gentner B. Cittaro D. Naldini L. Lentiviral vectors escape innate sensing but trigger p53 in human hematopoietic stem and progenitor cells EMBO Mol. Med. 9 2017 1198 1211 10.15252/emmm.201707922 28667090 PMC5582409 source

      The Results section (‘Lentiviral reverse-transcribed DNA triggers p53 signaling independently of integration in human HSPC’; page unavailable) reports p53 activation after lentiviral exposure in human hematopoietic stem and progenitor cells, not all primary hematopoietic cells. It also reports that primary CD4+ T cells and murine HSPC did not up-regulate p21, a marker used for p53 activation. The differentiation clause is attached to other citations and is not assessed here. (LLM, conf=0.90)

      Experimental — accuracy not yet measured
    • cites [37] Weinreb C. Rodriguez-Fraticelli A.E. Camargo F.D. Klein A.M. Lineage tracing on transcriptional landscapes links state to fate during differentiation Science 367 2020 eaaw3381 10.1126/science.aaw3381 PMC7608074 31974159 source

      The Results section, “A simultaneous assay of clonal states and fates” (page unavailable), describes lentivirally barcoded hematopoietic progenitors differentiating into multiple lineages in vitro. That supports differentiation after transduction, but the paper does not establish that infection had no effect on differentiation. The Trp53 assertion is attached to a different citation in the citing sentence. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
    • cites [39] Jindal K. Adil M.T. Yamaguchi N. Yang X. Wang H.C. Kamimoto K. Rivera-Gonzalez G.C. Morris S.A. Single-cell lineage capture across genomic modalities with CellTag-multi reveals fate-specific gene regulatory changes Nat. Biotechnol. 42 2024 946 959 10.1038/s41587-023-01931-4 37749269 PMC11180607 source

      In “Benchmarking CellTag-multi using in vitro hematopoiesis” (page unavailable), the paper reports that CellTag-labeled mouse hematopoietic cells produced all major lineages in culture. That supports differentiation occurring after lentiviral tagging, but it does not establish that infection left differentiation unaffected relative to uninfected cells. The Trp53 assertion is attached to a different citation in the citing sentence and is not judged here. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation

    Non-monotonous effects of a transcription factor with respect to binding site affinity can arise when gene regulation is limited at multiple steps and the transcription factor acts incoherently on those steps, promoting transcription at some steps and hindering it at others.

    “In that work, we show that non-monotonous effects of a TF with respect to binding site affinity can arise when gene regulation is limited at multiple steps, and a TF acts incoherently on those steps, such that it acts on one or more steps to promote transcription, and on other steps to hinder it.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [46] Martinez-Corral R. Friedrich D. Frömel R. Velten L. Gunawardena J. DePace A.H. Emergence of activation or repression in transcriptional control under a fixed molecular context Preprint at bioRxiv 2024 10.1101/2024.05.29.596388 PMC12501156 40982681 source

  • May combine two claims.

    Findingown data

    Combinatorial Fli1-Spi1 enhancers were repressive if sites for both factors were present at roughly equal ratio or at an excess of Fli1 sites.

    “We observed that combinatorial Fli1-Spi1 enhancers were repressive if sites for both factors were present at roughly equal ratio or at an excess of Fli1 sites (Figure 4H).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The gene expression landscape measured by scRNA-seq reflects functional TF activity levels.

  • May combine two claims.

    Findingown data

    For Fli1-Spi1, both factors were expressed over the entire landscape and activated enhancers containing either individual binding site in all cell states.

    “In the case of the Fli1-Spi1, both factors were expressed over the entire landscape and activated enhancers containing either individual binding site in all cell states (Figure S5D, left and right columns).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The gene expression landscape measured by scRNA-seq reflects functional TF activity levels.

  • May combine two claims.

    Findingown data

    The random background DNA is not biochemically inert but does affect enhancer activity.

    “This random background DNA is not biochemically inert but does affect enhancer activity, as we analyze in detail below.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Across the seven cell states, 3/38 factor motifs were inactive, 20/38 were activators, and 10/38 were repressors.

    “Across the seven cell states, 3/38 factor motifs were inactive, 20/38 were activators, and 10/38 were repressors (Figure 2B, bottom row).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Several cases of combinatorial duality were observed, where a combination of activators turns into a repressor, as well as neutralizing interactions between TF motifs.

    “We observe several cases of combinatorial duality, where a combination of activators turns into a repressor, as well as neutralizing interactions between TF motifs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The deep learning model mostly learned rules on the placed motifs but also extracted information from the DNA background.

    “This model mostly learned rules on the placed motifs but also extracted information from the DNA background (Figures S4D and S4E).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Across the seven cell states, activity levels were positively correlated with mRNA expression of the corresponding TF in the case of 14/20 activators and negatively correlated in the case of 6/10 repressors.

    “Across the seven cell states, activity levels were positively correlated with mRNA expression of the corresponding TF in the case of 14/20 activators and negatively correlated in the case of 6/10 repressors (Figures 3A and 3D).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Cell state and spacing modulated the behavior of combinatorial Fli1-Spi1 enhancers.

    “Cell state (and spacing) modulated this behavior (Figure S5B).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The gene expression landscape measured by scRNA-seq reflects functional TF activity levels.

  • May combine two claims.

    Findingown data

    Sequences containing motifs for any given single factor were often generally active in most cell states or not active in any cell state.

    “We thereby found that sequences containing motifs for any given single factor were often generally active in most cell states (Figure 2B, top row) or not active in any cell state.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Across libraries, 11%-22% of sequences were classified as active in any cell state and 4%-6% as repressed.

    “In total, for the different libraries, 11%–22% of the sequences were thereby classified as active in any cell state, and 4%–6% were classified as repressed (Figure 2A).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For 16/20 activators, more (and/or stronger) binding sites resulted in a monotonous increase in activity, whereas for 6/10 repressors, more (and/or stronger) binding sites resulted in a monotonous decrease in activity.

    “We found that for 16/20 activators, more (and/or stronger) binding sites resulted in a monotonous increase in activity, whereas for 6/10 repressors, more (and/or stronger) binding sites resulted in a monotonous decrease in activity (Figures 3B, 3C, and 3E).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The ability of single-factor motifs to activate transcription was highly correlated between all lineages and only roughly separated lineages into an erythromyeloid and a myeloid branch.

    “Quantitatively, the ability of single-factor motifs to activate transcription was highly correlated between all lineages and only roughly separated lineages into an erythromyeloid and a myeloid branch (Figure 2C, left).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Specific combinations of lineage factor binding sites are inactive or even act as repressors.

    “but specific combinations of these sites are inactive or even act as repressors”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For Fli1-Spi1 and Cebpa-Gata2, repression was achieved by strong sites that were close to each other, whereas more distantly located sites were permissive of transcription.

    “In the case of Fli1-Spi1 and Cebpa-Gata2, repression was achieved by strong sites that were close to each other, whereas more distantly located sites were permissive of transcription (Figures 4E and S3E).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    8 of 11 interactions that gave rise to duality were significantly impacted by spacing and/or orientation of the sites, whereas this was the case for only 11 of the 34 remaining motif pairs.

    “we found that 8 of 11 interactions that gave rise to duality were significantly impacted by spacing and/or orientation of the sites (Figure 4D), whereas this was the case for only 11 of the 34 remaining motif pairs.”
    Quote in the paperSubstantial claim

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

    Findingown data

    Steric hindrance was not observed in the Cebpa-Gata1/2 AF3 models.

    “Steric hindrance was not observed.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The split Fli1-Spi1 model resulted in a pTM score of 0.49.

    “The split model resulted in a pTM score of 0.49 (Figure S3I).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In HPCs, Runx1 and Meis1 motifs were repressive on their own but permissive of activity in a larger number of combinations, compared with other repressors.

    “In HPCs, Runx1 and Meis1 motifs were repressive on their own but permissive of activity in a larger number of combinations, compared with other repressors (Figure S7I).”
    Quote in the paperSubstantial claim

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

    Findingown data

    For Cebpa and Gata1/2, structural prediction outside of the DNA binding domain was poor.

    “In the case of Cebpa and Gata1/2, structural prediction outside of the DNA binding domain was poor.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Findingown data

    The AF3 structural prediction had a pTM score between 0.2-0.24 for all tested Cebpa-Gata1/2 models.

    “The AF3 structural prediction had a pTM score between 0.2-0.24 for all tested models.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    Only the DNA-binding domains of Cebpa and Gata1/2 were modeled properly and were binding to each TFBS respectively.

    “Only the DNA-binding domains of the two proteins were modeled properly and were binding to each TFBS respectively (Figures S3J and S3K).”
    Quote in the paperNot a substantial claim

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

    Findingown data

    The Fli1-Spi1 model had a pTM score of 0.29 and the DNA binding domains had a high pLDDT score.

    “Our model had a pTM score of 0.29 and the DNA binding domains had a high pLDDT score (Figure S3H).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Residual immature progenitors, megakaryocyte-erythroid progenitors, eosinophil, basophil, neutrophil precursors, and early and late monocyte precursors were effectively separated by the sorting scheme.

    “Residual immature progenitors, megakaryocyte-erythroid progenitors, eosinophil, basophil, neutrophil precursors, as well as early and late monocyte precursors were effectively separated.”
    Quote in the paperSubstantial claim

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

    Findingown data

    The Fli1-Spi1 models had higher confidence compared to the Cebpa-Gata2 models.

    “The Fli1-Spi1 models had higher confidence compared to the Cebpa-Gata2 models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The best design strategy was model-guided placement of motifs/motif pairs, followed by a local search, which outperformed the purely deep learning-based procedure.

    “The best design strategy was model-guided placement of motifs/motif pairs, followed by a local search, which outperformed the purely deep learning-based procedure.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The deep model trained on all data except Libraries G + H predominantly learns rules on pairwise motif interactions.

  • May combine two claims.

    Findingown data

    The identical set of 270 Trp53 binding site sequences included in all libraries was characterized by a broad range of activities and a high correlation across different cell types (Pearson R>0.95) and experiments (Pearson R>0.86).

    “These sequences were characterized by a broad range of activities, and a high correlation across different cell types (Pearson R>0.95) and experiments (Pearson R>0.86).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The resulting sequences generated activity relative to random DNA controls, where desired, in 75% of cases and were inactive, where desired, in 64% of cases (F1 score for activity: 0.73).

    “The resulting sequences generated activity relative to random DNA controls, where desired, in 75% of cases and were inactive, where desired, in 64% of cases (F1 score for activity: 0.73, Figures 6C–6E, S6I, and S6J).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The deep model trained on all data except Libraries G + H predominantly learns rules on pairwise motif interactions.

  • May combine two claims.

    Findingown data

    At the level of activators, there was good agreement between primary HPCs and K562: 15/20 activators identified in HPCs were also activators in K562 cells, and the fraction of sequences activated by each factor in the two cell types was quantitatively similar (Pearson R: 0.82).

    “At the level of activators, there was good agreement between primary HPCs and K562: 15/20 activators identified in HPCs were also activators in K562 cells, and the fraction of sequences activated by each factor in the two cell types was quantitatively similar (Figures S7A, Pearson R: 0.82).”
    Quote in the paperSubstantial claim

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

    Findingown data

    Both modelers yielded similar structures, but AF3 allowed a larger input of tokens and model performance, based on pTM score, was slightly better.

    “Both modelers yielded similar structures, but since AF3 allowed a larger input of tokens and model performance, based on pTM score, was slightly better, we continued with these models.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingown data

    Cebpa-Gata2 and Fli1-Spi1 were the factor pairs with the strongest repressive interactions.

    “More precisely, we modeled the interaction of Cebpa and Gata2, as well as the interaction of Fli1 and Spi1, as these were the factor pairs with the strongest repressive interactions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Runx1 and Meis1 motifs activated transcription in combination with weak activators, motifs that on their own were inactive, or even repressors, indicating that they can turn from repressors into co-activators depending on co-occurring motifs.

    “Interestingly, these motifs activated transcription in combination with weak activators (Meis1-Gata2 and Runx1-Myb), motifs that on their own were inactive (Meis1-Tcf3 and Meis1-Tfap2a), or even repressors (Runx1-Tcf12 combination), indicating that they can turn from repressors into co-activators depending on co-occurring motifs (Figures 7I and S7J).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In HPCs, Cebpa was antagonized by several motifs, including strong activators such as Gata1/Tal1, Nfyc, and Elk1, and it was entirely inactive in combination with Mecom.

    “In HPCs, Cebpa was antagonized by several motifs, including strong activators such as Gata1/Tal1, Nfyc, and Elk1, and it was entirely inactive in combination with Mecom (Figures 7E and 7G).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Correlation of predictions and measurements across cell states for each sequence separately was positive and higher for sequences with a high variance across cell states.

    “We also computed the correlation of predictions and measurements across cell states for each sequence separately and observed positive correlations that were higher for sequences with a high variance across cell states (Figure S6E).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The deep model trained on all data except Libraries G + H predominantly learns rules on pairwise motif interactions.

  • May combine two claims.

    Findingown data

    Combinations of activators and repressors were usually inactive in HPCs, but they often displayed activity in K562, also when controlling for different levels of measurement noise.

    “Specifically, combinations of activators and repressors were usually inactive in HPCs, but they often displayed activity in K562 (Figure 7D), also when controlling for different levels of measurement noise between primary cells and K562 (Figure S7F).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    After scaling adjustment, random DNA controls measured in both experiments displayed a similar mean/variance relationship and a similar dynamic range.

    “After this adjustment, random DNA controls measured in both experiments displayed a similar mean/variance relationship and a similar dynamic range.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Only Trp53 was able to activate transcription in all (except one) combinations in HPCs, whereas Elk1, Fli1, Nfyc, Klf1/Klf4, and Sp1 were always activators in K562.

    “Only Trp53 was able to activate transcription in all (except one) combinations in HPCs, whereas Elk1, Fli1, Nfyc, Klf1/Klf4, and Sp1 were always activators in K562.”
    Quote in the paperSubstantial claim

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

    Findingreasoning

    Short motifs such as GATA, and low affinity motifs, did pass the spacer filters.

    “However, short motifs such as GATA, and low affinity motifs, did pass our filters, see also Figures S4F–S4H.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Activity and specificity of these sequences were driven by the placed motifs, with contributions of putative motifs introduced in the background DNA during the optimization procedure.

    “Activity and specificity of these sequences were driven by the placed motifs, with contributions of putative motifs introduced in the background DNA during the optimization procedure (Figure 6F).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The deep model trained on all data except Libraries G + H predominantly learns rules on pairwise motif interactions.

  • May combine two claims.

    Findingown data

    The model's performance was to some extent limited by the technical noise of the measurements; Library E measurements were essentially noise-free (R2 between replicates: 0.97) and the model explained 71% of the variance.

    “This performance was to some extent limited by the technical noise of the measurements: when we measured a small-scale Library E of 30 enhancers with similar designs as Library A + B, we obtained essentially noise-free measurements (R2 between replicates: 0.97), and the model explained 71% of the variance (Figures S6C and S6D), underscoring the quality of our model.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The deep model trained on all data except Libraries G + H predominantly learns rules on pairwise motif interactions.

  • May combine two claims.

    Findingown data

    On the natural genome Library G, performance strongly depended on the number of motifs for the focus 42 TFs present and reached an R2 of >0.3 for enhancers with >8 motifs for these TFs.

    “On the natural genome Library G, performance strongly depended on the number of motifs for the focus 42 TFs present and reached an R2 of >0.3 for enhancers with >8 motifs for these TFs (Figure 5G).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The deep model trained on all data except Libraries G + H predominantly learns rules on pairwise motif interactions.

  • May combine two claims.

    Findingown data

    At the level of repressors, there was little agreement between HPCs and K562; several TFs lost repressive ability in K562 cells.

    “At the level of repressors, however, there was little agreement between HPCs and K562; several TFs lost repressive ability in K562 cells (Figure S7B).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    Monocyte precursors generally behaved like monocytes in the cell-state activity measurements.

    “all cell states except monocyte precursors, which generally behaved like monocytes”
    Quote in the paperSubstantial claim

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

    Findingown data

    Cloning was very even, leading to a 5-10 fold difference in the representation of the different GREs within the library, with one exception where the difference was larger.

    “Cloning was very even, leading to a 5-10 fold difference in the representation of the different GREs within the library, with one exception, where the difference was larger (Figure S2B).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationown data

    The model can explain Myc behavior, assuming that Myc concentration is the only difference among cell states.

    “Our model can explain this behavior, assuming that Myc concentration is the only difference among cell states (Figure S3A).”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Myc concentration is the only difference among cell states.

  • May combine two claims.

    Study design and conductown data

    The choice of core promoter and the measurement time point post-lentiviral integration did not exert a major effect on enhancer activity in K562 cells.

    “We performed several additional screens in K562 cells to demonstrate that the choice of core promoter (Figure S2G) or the choice of the measurement time point post-lentiviral integration (Figure S2H) did not exert a major effect.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Enhancers composed of binding sites for single factors mostly relay TF occupancy to a monotonous activity response, but they also describe a first type of activator-repressor duality, cell-state-dependent duality.

    “They illustrate that enhancers composed of binding sites for single factors mostly relay TF occupancy to a monotonous activity response, but they also describe a first type of activator-repressor duality, cell-state-dependent duality.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Enhancers with binding site pairs can establish binary activity patterns from overlapping TF expression gradients, whereas enhancers composed of single motifs can in most cases only establish gradual differences.

    “Altogether, enhancers with binding site pairs can establish binary activity patterns from overlapping TF expression gradients, whereas enhancers composed of single motifs can in most cases only establish gradual differences.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductreasoning

    The libraries were not biased toward a specific DNA context or motif instance because each candidate enhancer had a different random DNA background and all motif instances were independently sampled.

    “Each candidate enhancer had a different random DNA background, and all motif instances were independently sampled. Thereby, we ensured that our libraries were not biased toward a specific DNA context or motif instance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    These combinations create enhancers that are sensitive to ratios in TF expression and specific to cell states where either factor dominates.

    “creating enhancers that are sensitive to ratios in TF expression and specific to cell states where either factor dominates.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductown data

    Correlations between replicate infections were 0.95-0.99 in K562s and 0.74-0.87 in HPCs.

    “Correlations between replicate infections were 0.95–0.99 in K562s and 0.74–0.87 in HPCs (Figures S2E–S2J).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Occupancy-dependent duality can lead to activator-repressor duality in an unaltered cellular environment and can create enhancers that are only active at an optimal, not necessarily maximal or minimal, TF occupancy.

    “This behavior, which we term “occupancy-dependent duality,” can lead to activator-repressor duality in an unaltered cellular environment and can create enhancers that are only active at an optimal, not necessarily maximal or minimal, TF occupancy.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    Activity measurements were only available for seven cell states, impeding a more quantitative analysis of correlations between gene expression of several factors and activity.

    “A limitation here is that activity measurements were only available for seven cell states, impeding a more quantitative analysis of correlations between gene expression of several factors and activity.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims. The quote only partly states this claim.

    Study design and conductown data

    The study used a genetic algorithm starting from 100 different initial conditions and chose the best solution to fit the biophysical model.

    “In addition, we run the algorithm starting from 100 different initial conditions and chose the best solution.”
    Quote in the paperNot a substantial claim

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

    Premisenone given

    Cryptic TF motifs in a fixed background sequence might have synergistic or antagonistic effects on only a subset of TFs, ultimately leading to biases.

    “The reason for this is that cryptic TF motifs in a fixed background sequence might have synergistic or antagonistic effects to only a subset of TFs, ultimately leading to biases (see also Figure S4).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Synthetic enhancers characterized here have a well-defined motif composition and can establish specificity from highly overlapping TF expression patterns.

    “Synthetic enhancers characterized here have a well-defined motif composition and can establish specificity from highly overlapping TF expression patterns.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    The final parameter set was picked manually to not sacrifice replicate reproducibility.

    “The final parameter set was picked manually to not sacrifice replicate reproducibility (Figure S2E).”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: At the complexity of barcodes used here, a single barcode is likely to reflect a single integration site.

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

    Study design and conductnone given

    Effects of 3D chromatin context were removed by averaging across lentiviral integration sites.

    “Effects of 3D chromatin context were removed by averaging across lentiviral integration sites.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Cell-state-dependent duality can emerge from occupancy-dependent duality, differences in co-factor environment, splicing, or posttranslational modifications in different cell states.

    “Such cell-state-dependent duality can potentially emerge from occupancy-dependent duality, e.g., when only some cell states express a TF dose compatible with activation. However, cell-state-dependent duality can also emerge from differences in co-factor environment, splicing, or posttranslational modifications in different cell states.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation

    The performance of our models on sequence design and genome interpretation tasks highlights the importance of pairwise TF interactions in achieving regulatory specificity and illustrates the benefits of working with non-natural sequences for learning the rules of gene regulation.

    “The performance of our models on sequence design and genome interpretation tasks highlights the importance of pairwise TF interactions in achieving regulatory specificity and illustrates the benefits of working with non-natural sequences for learning the rules of gene regulation.29”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [29] De Boer C.G. Taipale J. Hold out the genome: a roadmap to solving the cis-regulatory code Nature 625 2024 41 50 10.1038/s41586-023-06661-w 38093018 source

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

    Study design and conductnone given

    Before MPRA screens, all lentivirus preparations were tested in K562 cells and HPCs to assure lentiviral efficiency, cell viability and activity of the GFP reporter.

    “Before MPRA screens, all lentivirus preparations were tested in K562 cells and HPCs to assure lentiviral efficiency, cell viability and activity of the GFP reporter.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    In primary HPCs, TFs frequently neutralize each other's effect, and only specific combinations of factors are permissive of expression; by contrast, in K562 cells many combinations of factors were able to drive transcription.

    “Together, these analyses suggest that in primary HPCs, TFs frequently neutralize each other’s effect, and only specific combinations of factors are permissive of expression. By contrast, in K562 cells many combinations of factors were able to drive transcription.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The DNA libraries were cloned such that each GRE was associated with 100s-1000s of unique barcodes in the 5′UTR.

    “We cloned the DNA libraries such that each GRE was associated with 100s-1000s of unique barcodes in the 5′UTR of the (Figure S2A).”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    The study split the entire training data into ten train-validation-test splits, trained an ensemble of ten models on each train set, and averaged their prediction on the test sets.

    “For model training, we split the entire training data (Library A, B, C, F, as well as control sequences included into the various screens) into ten train-validation-test splits. On each train set, we trained an ensemble of ten models and averaged their prediction on the test sets.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The study discretized measurements into active and inactive by using random DNA controls.

    “For data analysis, we discretized measurements into “active” and “inactive” by using random DNA controls, as described in the section identification of active and repressed sequences.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The study constructed ten-fold cross-validated random forests to characterize the regulatory grammar for Library A and B.

    “To characterize the regulatory grammar for Library A and B we constructed ten-fold cross-validated random forests.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    These interactions depend on local motif "grammar" and possibly involve direct or indirect protein interactions between TFs.

    “These interactions depend on local motif “grammar” and possibly involve direct or indirect protein interactions between TFs.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Motif interactions are frequently neutralizing in primary cells but not in a leukemia cell line.

    “These global analyses suggest that motif interactions are frequently neutralizing in primary cells but not in a leukemia cell line.”
    Quote in the paperSubstantial claim
79 of 79

  • The paper relies on this without support.

    There is a shortage of large-scale, functional datasets characterizing systematically designed enhancers during cellular differentiation.

    “There is, however, a shortage of large-scale, functional datasets characterizing systematically designed enhancers during cellular differentiation.”

  • The authors state this as an assumption.

    Myc concentration is the only difference among cell states.

    “assuming that Myc concentration is the only difference among cell states”

  • The paper relies on this without support.

    The gene expression landscape measured by scRNA-seq reflects functional TF activity levels.

    “To relate this functional behavior to TF expression levels, we visually superimposed cell-state-specific patterns of activation and repression (Figures S5D–S5F, arrows) on a gene expression landscape measured by scRNA-seq (Figures S5D–S5F, UMAP).”

  • The paper relies on this without support.

    The deep model trained on all data except Libraries G + H predominantly learns rules on pairwise motif interactions.

    “By nature of the training data, this model is predominantly trained to learn rules on pairwise motif interactions.”

  • The authors state this as an assumption.

    At the complexity of barcodes used here, a single barcode is likely to reflect a single integration site.

    “reflecting the assumption that at the complexity of barcodes used here, a single barcode is likely to reflect a single integration site”

  • The authors state this as an assumption.

    Binding among the sites is independent.

    “The model assumes independent binding among the sites, although binding cooperativity could be included at the expense of more free parameters.”

  • The authors state this as an assumption.

    The model assumes equilibrium binding.

    “Then, we consider equilibrium binding, as described.88”

  • The authors state this as an assumption.

    The model assumes a 3-state polymerase cycle where the transition between states 1 and 2 is reversible and the other transitions are irreversible.

    “We then consider a 3-state polymerase cycle where the transition between states 1 and 2 is reversible, and the other irreversible, as in,89 so that in the absence of regulation the cycle rates are k1, km1, k2, k3 and this yields a steady-state transcription rate:”

  • The authors state this as an assumption.

    Each rate can be non-linearly modulated by the transcription factor according to the given equation.

    “We then assume each rate can be non-linearly modulated by the TF according to:”

  • The authors state this as an assumption.

    mRNA degradation is linear.

    “We finally assume linear mRNA degradation and calculate the natural logarithm of the fold change in gene expression as ln(mTF/m0)”

  • The authors state this as an assumption.

    The effects of space on Elk1 behavior can be modeled phenomenologically through the parameter x, so that all affinities are equally affected by a given spacing.

    “For the fits of the space-dependent Elk1 behavior, we modelled the effects of space phenomenologically, through x, so that all affinities would be equally affected by a given spacing. We note that this is just a proxy for space-dependent cooperativities, but avoids the parameter explosion that would involve explicitly accounting for those.88”
11 of 11

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.

  • Library H has two reported sequence counts

    Minor

    The Results report 3,946 sequences in Library H, while the library-design methods report 3,964.

1 of 1

Cites

Flagged first, then by citations

Cited by

Still being checked.

Related

Authors
Frömel, R., Rühle, J., Bernal Martinez, A., Szu-Tu, C., Pacheco Pastor, F., Martinez-Corral, R., & Velten, L.
Affiliations
  • Universitat Pompeu Fabra (ES)
  • Centre for Genomic Regulation (ES)
Corresponding author
Velten, Lars
ORCID
Venue
Cell
Publisher
Elsevier BV
Published
2025-06
Volume
188
Issue
12
Pages
3202-3218.e21
ISSN
0092-8674 (ISSN-L)
DOI
10.1016/j.cell.2025.04.017
PMC
PMC12173716
OpenAlex
W4410206253
Semantic Scholar
fd755db04cb56af793e94bf133da8577dcbab8cb
Cited by
29
References
106
Citations by year
2026: 16 · 2025: 13
FWCI
8.61
Type
journal article
Language
en
Open access
Licence
Locations
  • Cell (publishedVersion, cc-by-nc)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
  • Repositori digital de la UPF (Universitat Pompeu Fabra) (publishedVersion, cc-by-nc)
Funders
  • National Plan for Scientific and Technical Research and Innovation (RYC2021-033860-I, CEX2020-001049-S, PID2019-108082GA-I00, PID2022-142210NA-I00)
  • European Regional Development Fund (MCIN/AEI/10.13039/501100011033)
  • Generalitat de Catalunya
  • European Commission
  • Ministerio de Ciencia e Innovación
  • Ministerio de Ciencia, Innovación y Universidades (MCIU/AEI/10.13039/501100011033)
  • European Social Fund Plus
  • European Molecular Biology Laboratory
  • European Research Council
  • Government of Catalonia Agency for Administration of University and Research Grants (2023 FI_I 00776)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Chromatin Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene Regulatory Network Analysis
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Developmental Biology and Gene Regulation
Keywords
hematopoiesis, transcription factor binding sites, transcription factor interaction, myeloid differentiation, hematopoietic progenitor cells, synthetic DNA sequences, transcription factor occupancy
MeSH terms
Animals, Binding Sites, Cell Differentiation, Enhancer Elements, Genetic, Hematopoiesis, Hematopoietic Stem Cells, Humans, Transcription Factors, Mice

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 2026-10-27.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-27.

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