Enhancer grammar of liver cell types and hepatocyte zonation states

Classified as probably not a substantial claim.

Previous single-cell and spatial transcriptomics studies have shown that hepatocyte transcriptome varies along the porto-central axis.

“Previous single-cell and spatial transcriptomics studies have shown that not only hepatocyte function, but also transcriptome, varies along this axis8.”

• cites [8] sourceSupported

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

In Drosophila, approximately 70-90% of accessible regions are active enhancers when their accessibility is cell type specific.

“In Drosophila, different studies have reported that ~70-90% of accessible regions are active enhancers, when their accessibility is cell type specific10–12”

• cites [10] sourcePartly supported

  The results passage labeled Abstract reports that 29 of 31 tested clade-specific, promoter-distal accessible regions (94%) functioned as developmental enhancers in Drosophila embryos. This supports the association between cell-type-specific accessibility and enhancer activity, but the reported rate is 94%, not approximately 70–90%, and it comes from a selected set of 31 regions rather than all accessible regions. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [11] Bravo González-Blas, C., et al. Identification of genomic enhancers through spatial integration of single-cell transcriptomics and epigenomics. Mol. Syst. Biol. 16, (2020). sourcePartly supported

  The Results (page unavailable) report that accessibility and activity correlate in 77% of tested enhancer–reporter regions, and that specifically accessible enhancers tend to show matching activity. The Introduction (page unavailable) also reports that about 74% of 31 predicted cell-type-specific enhancers showed expected activity. These figures concern selected or predicted enhancers, not the proportion of all cell-type-specifically accessible regions that are active enhancers. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [12] 2022, Decoding gene regulation in the fly brain, Nature sourcePartly supported

  The main text (p. 4) says that differentially accessible regions (DARs) often function as enhancers, but reports 588 active enhancers among 1,017 Kenyon-cell DARs overlapping tested reporter regions—about 58%. This supports the relationship between cell-type-specific accessibility and enhancer activity, but not the claimed 70–90% rate. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

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

Chromatin accessibility is necessary but not sufficient for enhancer activity.

“chromatin accessibility is necessary but not sufficient for activity9.”

• cites [9] sourceSupported

Classified as probably not a substantial claim.

A potentially large fraction of accessible regions captured by scATAC-seq may not represent functional enhancers.

“Thus, a potentially large fraction of accessible regions, as captured by scATAC-seq, may not represent functional enhancers.”

• cites [9] sourcePartly supported

  In Results, “Grainyhead regulates enhancer accessibility but not activity” (page unavailable), 15 of 21 tested Grh-target regions showed reporter activity in the eye-antennal disc, and single-cell ATAC-seq accessibility did not establish activity. This supports the caution that accessibility alone does not establish enhancer function, but the paper does not estimate a potentially large nonfunctional fraction across scATAC-seq accessible regions; regions inactive in that disc could function elsewhere. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

Massively Parallel Reporter Assays have been established to directly test and identify active enhancers.

“new experimental approaches, such as Massively Parallel Reporter Assays (MPRAs), have been established to directly test and identify active enhancers14–18.”

• cites [14] sourceSupported
• cites [15] sourceSupported
• cites [16] sourceSupported
• cites [17] 2020, A systematic evaluation of the design and context dependencies of massively parallel reporter assays, Nat. Methods sourceSupported
• cites [18] sourceSupported

May combine two claims.

Animals fed ad libitum mapped to early zeitgeber timepoints, while samples for which food was removed the night before mapped to late zeitgeber timepoints.

“we confirmed that the animals that were fed ad libitum mapped to early zeitgeber timepoints, while samples for which food was removed the night before mapped to late zeitgeber timepoints”

• cites [34] 2021, Space-time logic of liver gene expression at sub-lobular scale, Nat. Metab sourceCould not verify

  The Results section reports ad libitum-fed mice sampled at ZT0, ZT6, ZT12 and ZT18; it does not report that ad libitum-fed samples mapped specifically to early timepoints. The Methods and Results passages found for overnight food removal and sample mapping do not describe food being removed the night before or those samples mapping to late timepoints. (LLM, conf=0.77)

  Experimental — accuracy not yet measured

May combine two claims.

Tcf7l1 and Tcf7l2 may bind to the same motif, with Tcf7l2 acting pericentrally for activation and Tcf7l1 periportally for repression.

“this may suggest that Tcf7l1 and Tcf7l2 bind to the same motif: Tcf7l2 pericentrally for activation, and Tcf7l1 periportally for repression36.”

• cites [36] sourcePartly supported

  The Introduction and Results report that Tcf7l1 acts as a transcriptional repressor. The paper does not establish that Tcf7l1 and Tcf7l2 bind the same motif or assign them periportal and pericentral roles, respectively (searched: shared DNA-binding motif, pericentral, periportal). (LLM, conf=0.82)

  Experimental — accuracy not yet measured

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

Methods for inferring enhancer-driven GRNs that integrate enhancer sequence, chromatin accessibility and gene expression found that accuracy to predict active enhancers increases by approximately 10% compared to using accessibility alone.

“methods for inferring enhancer-driven GRNs (eGRNs) that integrate enhancer sequence, chromatin accessibility and gene expression22–24 found that the accuracy to predict active enhancers increases by ~10% compared to using accessibility alone25.”

• cites [22] sourceSupported
• cites [23] sourcePartly supported

  The Discussion and ‘Regulatory network inference with Pando’ describe a GRN method combining sequence-derived regulatory information, chromatin accessibility and gene expression. That section reports that 94% of H3K27ac-overlapping accessible peaks were among its candidate regions, but it does not report an approximately 10% improvement in active-enhancer prediction accuracy over accessibility alone. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [24] Bravo González-Blas, C. B. , et al. SCENIC+: single-cell multiomic inference of enhancers and gene regulatory networks. 2022.08.19.504505 Preprint at https://doi.org/10.1101/2022.08.19.504505 (2022). sourceSupported
• cites [25] sourcePartly supported

  Extended Data Fig. 10 (p. 27) identifies accessibility and correlation with nearby gene expression as key features for predicting GFP activity. The paper also describes using enhancer sequence in its eGRN analysis, but the passages in our copy do not establish an approximately 10% accuracy gain over accessibility alone. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

May combine two claims.

This approach identified PU.1, Runx, and Irf as candidate TFs for Kupffer cell enhancers; Pax5 and Ebf1 for B cells; Tbx21 for T cells; Meis1, Maf and Gata4 for LSECs; Mef2c and Lhx2 for HSC; and Tead4, Sox9 and Hnf1b for BECs.

“this approach identified PU.1, Runx, and Irf as candidate TFs for Kupffer cell enhancers37; Pax5 and Ebf1 for B cells38; Tbx21 for T cells39; Meis1, Maf and Gata4 for LSECs40; Mef2c and Lhx2 for HSC41; and Tead4, Sox9 and Hnf1b for BECs42”

• cites [37] 2021, Epigenetic Regulation of Kupffer Cell Function in Health and Disease, Front. Immunol sourcePartly supported

  For the clause citing this paper, the Homeostatic Transcriptional Regulation of KC Function section says PU.1 motifs are enriched in Kupffer-cell-specific enhancers. The copy discusses IRFs as Kupffer-cell transcription factors but does not identify IRF enhancer motifs or Runx as a candidate for those enhancers. The other cell types in the sentence have separate citations and are not assessed here. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [38] 2021, EBF1 drives hallmark B cell gene expression by enabling the interaction of PAX5 with the MLL H3K4 methyltransferase complex, Sci. Rep sourceSupported
• cites [39] sourceSupported
• cites [40] 2020, Unraveling the transcriptional determinants of liver sinusoidal endothelial cell specialization, Am. J. Physiol. Gastrointest. Liver Physiol sourcePartly supported

  The Results (p. 4) identify C-MAF, GATA4, and MEIS2 as regulators of LSEC marker expression, supporting the LSEC attribution for Maf and Gata4. The cited paper identifies MEIS2, not Meis1, and tests marker expression after TF overexpression rather than identifying candidates for LSEC enhancers. (LLM, conf=0.91)

  Experimental — accuracy not yet measured
• cites [41] sourceNot checked — no full text of the source
• cites [42] sourceNot checked — no full text of the source

May combine two claims.

Ntn4+ LSECs and Ngfr+ HSC are located periportally, while Kit+ LSECs and Spon2+ HSC are located pericentrally.

“This integrated spatial and single-cell analysis also confirmed that Ntn4+ LSECs and Ngfr+ HSC were located periportally; while Kit+ LSECs and Spon2+ HSC were located pericentrally, as previously reported31,32”

• cites [31] 2021, Single-Cell Transcriptomics Reveals Zone-Specific Alterations of Liver Sinusoidal Endothelial Cells in Cirrhosis, Cell. Mol. Gastroenterol. Hepatol sourcePartly supported

  The Results section, “Spatial Lobular Locations of Heterogeneous Liver EC Populations Were Determined” (page not provided), identifies Ntn4 as a periportal landmark and Kit as a marker of pericentral LSECs in mice. It does not establish the claimed periportal location of Ngfr+ HSCs or pericentral location of Spon2+ HSCs. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites [32] sourcePartly supported

  The Results section, “Uncovering HSC Zonation across the Healthy Liver Lobule” (Figure 2; page unavailable), places NGFR-high HSCs near the portal vein. It identifies Adamtsl2-high HSCs near the central vein, but does not establish the claimed locations of Spon2+ HSCs or Ntn4+ and Kit+ LSECs. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Regulons that depend on nutritional status include Agmat and Mlxipl; on hormone levels include Nr1i2, Nr1i3 and Rxra; and on circadian rhythm include Clock.

“This allowed us to identify regulons that depend on nutritional status, such as Agmat52 and Mlxipl53; on hormone levels, such as Nr1i2, Nr1i3 and Rxra54; and on circadian rhythm, such as Clock55.”

• cites [52] sourceNot checked — no full text of the source
• cites [53] sourceSupported
• cites [54] sourcePartly supported

  Only the hormone-level examples are attributed to this source by citation 54. The paper’s nuclear-receptor discussion (section labeled Abstract; page unavailable) describes hormone-responsive transcription and identifies RXR as a receptor that partners with the thyroid hormone receptor. It does not identify Nr1i2 or Nr1i3 or establish that regulons for all three depend on hormone levels. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

May combine two claims.

SCENIC+ eGRNs were validated using publicly available Hi-C data and TF ChIP-seq, and transcriptome prediction using SCENIC+ TFs as features is more accurate compared to random.

“we validated the SCENIC+ eGRNs using publicly available Hi-C data50 and TF ChIP-seq51, and found that transcriptome prediction using SCENIC+ TFs as features is more accurate compared to random (see Methods, p-value < 2.2 x 10-16”

• cites [50] sourceSupported
• cites [51] Ballester, B. et al. Multi-species, multi-transcription factor binding highlights conserved control of tissue-specific biological pathways. eLife 3, e02626 (2014). sourceSupported

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

Convolutional neural networks can model, predict, and interpret enhancer accessibility or activity from sequence.

“we and others have recently shown the power of convolutional neural networks (CNN) to model, predict, and interpret an enhancer accessibility or activity from its sequence12,19–21.”

• cites [12] 2022, Decoding gene regulation in the fly brain, Nature sourceSupported
• cites [19] sourceSupported
• cites [20] sourceSupported
• cites [21] sourceSupported

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

In mammals, the ratio of accessible regions that are active enhancers drops to approximately 30%.

“but in mammals this ratio drops to ~30%13.”

• cites [13] sourcePartly supported

  The Discussion (page unavailable) reports that approximately 26% of predicted enhancer sequences were active in the assay, close to the cited ~30%. But the denominator was ENCODE-predicted enhancers tested in a human K562 cell-line assay, not all accessible regions in mammals; the paper also calls its estimate a lower bound. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

May combine two claims.

BEC clusters are located in bile ducts; pericentral and periportal VECs surround corresponding vessels with fibroblasts; and Kupffer cells are preferentially located in periportal and mid zones without a strong zonation pattern.

“In addition, BEC clusters could be clearly located in the bile ducts; pericentral and periportal VECs surround the corresponding vessels, together with fibroblasts; and Kupffer cells are preferentially located in periportal and mid zones without a strong zonation pattern, in agreement with recent studies33.”

• cites [33] 2022, Spatial proteogenomics reveals distinct and evolutionarily conserved hepatic macrophage niches, Cell sourcePartly supported

  The spatial-niches section (pp. 7–8) identifies central- and portal-vein endothelial cells and fibroblasts at both vessels; the spatial-orientation section (p. 5) places cholangiocytes in portal zones and Kupffer cells preferentially in periportal and mid zones. But p. 5 says the lack of a strong zonation pattern concerns Kupffer-cell gene expression, while their location is zonated. The claim does not preserve that distinction. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

Public human liver scATAC-seq data reveal that predicted Tbx3 and Tcf7l1 repressive sites are conserved between species, with similar accessibility patterns along the porto-central axis.

“As a third validation experiment, we analyzed a public human liver scATAC-seq data set62, revealing that the predicted Tbx3 and Tcf7l1 repressive sites are conserved between species, with similar accessibility patterns along the porto-central axis (Fig 5j).”

• cites [62] sourceSupported

May combine two claims.

Human HepG2 cells express Tbx3 and may therefore serve as a model to test the effect of mutating Tbx3 sites in hepatocyte enhancer sequences.

“Human HepG2 cells express Tbx3 and may therefore serve as a model to test the effect of mutating Tbx3 sites in hepatocyte enhancer sequences63.”

• cites [63] sourcePartly supported

  The Results section, “Expression of Tbx3 in HCC Tissues and Cell Lines” (page unavailable), reports Tbx3 protein expression in HepG2 cells. The paper does not address mutating Tbx3 sites in hepatocyte enhancer sequences or evaluate HepG2 for that purpose; the proposed use is an inference beyond what this source establishes. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

Predicted binding hits were validated using ChIP-seq data for Hnf4a, Onecut1, Cebpa and Foxa1.

“To validate the predicted binding hits, we used the ChIP-seq data for Hnf4a, Onecut1, Cebpa and Foxa151, as described for the SCENIC+ regulons.”

• cites [51] Ballester, B. et al. Multi-species, multi-transcription factor binding highlights conserved control of tissue-specific biological pathways. eLife 3, e02626 (2014). sourceSupported

May combine two claims.

Our data suggest a model in which the Wnt ligand is produced by endothelial cells in the central vein, creating a decreasing gradient of Wnt from the central to the portal vein.

“Our data suggests a model (Fig 6) in which the Wnt ligand is produced by endothelial cells in the central vein70, creating a decreasing gradient of Wnt from the central to the portal vein.”

• cites [70] 2018, Endothelial Wnts regulate β-catenin signaling in murine liver zonation and regeneration: A sequel to the Wnt–Wnt situation, Hepatol. Commun sourceSupported

Classified as probably not a substantial claim.

Recent work from the Morris lab has shown that inferred eGRNs can be used to predict the effect of a TF perturbation.

“Recent work from the Morris lab72 has shown that inferred eGRNs can be used to predict the effect of a TF perturbation.”

• cites [72] sourceSupported

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

Tcf7l1 and Tbx3 are well-known repressors in development, but had not been directly linked to hepatocyte zonation.

“Tcf7l1 and Tbx3 are indeed well-known repressors in development68,69, but had not been directly linked to hepatocyte zonation.”

• cites [68] Athanasouli, P. et al. The Wnt/TCF7L1 transcriptional repressor axis drives primitive endoderm formation by antagonizing naive and formative pluripotency. 2022.05.18.492419 Preprint at https://doi.org/10.1101/2022.05.18.492419 (2022). sourceSupported
• cites [69] 2020, The roles and regulation of TBX3 in development and disease, Gene sourceSupported

Classified as probably not a substantial claim.

To correct for batch effects, Harmony was used on the selected principal components.

“To correct for batch effects, we used Harmony on the selected principal components (34), using Leiden clustering with resolution 0.6, resulting in 15 clusters.”

• cites [34] 2021, Space-time logic of liver gene expression at sub-lobular scale, Nat. Metab sourceCould not verify

  The Methods section on t-SNE clustering describes a low-dimensional representation, but does not mention Harmony or batch-effect correction. Searches for Harmony, batch correction, and selected principal components found no passage addressing the attributed method; page numbers are unavailable in our copy. (LLM, conf=0.73)

  Experimental — accuracy not yet measured

May combine two claims.

TF footprinting suggests that direct repressor binding may occur within nucleosome-occupied regions, while activator binding is strongly associated with nucleosome depletion.

“As a potential mechanism, how repressor binding could result in the absence of an ATAC peak, TF footprinting indeed suggests that direct repressor binding may occur within nucleosome-occupied regions, while activator binding is strongly associated with nucleosome depletion64.”

• cites [64] sourcePartly supported

  In “Determination of the molecular function of TFs” (Fig. 4a; page not available), REST’s footprint shows locally increased accessibility with compact surrounding chromatin, whereas STAT2’s shows increased accessibility across the surrounding region. This supports the contrast between repressor and activator footprints, but the paper does not establish binding within nucleosome-occupied regions or directly measure nucleosome depletion. The figure caption also says the analyzed binding sites overlapped open chromatin. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

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

CNN-based enhancer modeling has recently gained traction due to its capacity for interpretation of enhancer grammar.

“CNN-based enhancer modeling has recently gained traction, due to their capacity for the interpretation of enhancer grammar 12,19,20,74”

• cites [12] 2022, Decoding gene regulation in the fly brain, Nature sourcePartly supported

  In “Combinatorial TF expression is reflected by enhancer architecture” (page not indexed), the authors describe training a convolutional neural network on enhancer-region sequences. Figure 3 (p. 11) shows that interpreting nucleotide contributions reveals transcription-factor motifs, supporting the claimed capacity to interpret enhancer architecture. The paper does not establish that CNN-based modeling has recently gained traction or that this capacity caused its uptake. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [19] sourceSupported
• cites [20] sourceSupported
• cites [74] sourceSupported

Classified as probably not a substantial claim.

Batch effects between samples were corrected using harmonypy on the scaled topic distributions.

“Batch effects between samples were corrected using harmonypy89 (v0.0.6) on the scaled topic distributions, and Leiden clustering with resolution 0.6 resulted in 11 clusters, corresponding to 14 cell types based on previous labelling.”

• cites [89] 2019, Fast, sensitive and accurate integration of single-cell data with Harmony, Nat. Methods sourcePartly supported

  The Online Methods > Harmony section describes an algorithm that corrects batch effects in cell embeddings, supporting the citation for the batch-correction method. Our copy identifies an R implementation, however, not harmonypy or version 0.0.6; it also does not describe correction of scaled topic distributions. No page number is available in our copy. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

To validate these regulons, publicly available Hi-C and ChIP-seq data were used.

“To validate these regulons, we used publicly available Hi-C and ChIP-seq data50,51.”

• cites [50] sourceSupported
• cites [51] Ballester, B. et al. Multi-species, multi-transcription factor binding highlights conserved control of tissue-specific biological pathways. eLife 3, e02626 (2014). sourceSupported

Classified as probably not a substantial claim.

Transfer learning from sequence models trained with large data sets has been recently shown as a robust alternative.

“While training on small data sets may lead to overfitting, transfer learning from sequence models trained with large data sets has been recently shown as a robust alternative75,”

• cites [75] Salvatore, M. , Horlacher, M. , Winther, O. & Andersson, R. Transfer learning reveals sequence determinants of regulatory element accessibility. 2022.08.05.502903 Preprint at https://doi.org/10.1101/2022.08.05.502903 (2022). sourceSupported

May combine two claims.

The zonation-CNN added Tbx3 and Tcf7l1 motifs as crucial sequence features to hepatocyte enhancers, while the activity-CNN added ETS and AP-1 sites underlying higher enhancer activity, in agreement with previous MPRA studies in the liver.

“The zonation-CNN added Tbx3 and Tcf7l1 motifs as crucial sequence features to the hepatocyte enhancers; while the activity-CNN added ETS and AP-1 sites underlying higher enhancer activity, in agreement with previous MPRA studies in the liver17.”

• cites [17] 2020, A systematic evaluation of the design and context dependencies of massively parallel reporter assays, Nat. Methods sourcePartly supported

  In the Results sections on MPRA design and sequence context (page not available), the paper reports that ETS-variant factors predict episomal assay activity and that predicted binding of Jun and FOS, components of AP-1, helps explain activity differences between enhancer lengths in HepG2 experiments. It does not report zonation-CNN or activity-CNN adding motifs, or identify Tbx3 and Tcf7l1 as crucial hepatocyte-enhancer features. The reported associations therefore support only part of the claimed agreement with prior MPRA work. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

SCENIC+ identifies Tbx3 and Tcf7l1 as transcriptional repressors in pericentral and periportal hepatocytes, targeting 193 and 520 regions and 77 and 119 genes, respectively.

“SCENIC+ identifies Tbx3 and Tcf7l1 as transcriptional repressors in pericentral and periportal hepatocytes, targeting 193 and 520 regions and 77 and 119 genes, respectively”

May combine two claims.

In all animals, there is a unidirectional gradient within hepatocytes, corroborated by gene expression and region accessibility.

“In all animals, we found a unidirectional gradient within hepatocytes, corroborated both by gene expression and region accessibility”

May combine two claims.

Tcf7l1 is expressed periportally, and its candidate target regions are accessible pericentrally, while Tcf7l2 is expressed in all hepatocytes.

“Tcf7l1 is expressed periportally, and its candidate target regions are accessible pericentrally, while Tcf7l2 is expressed in all hepatocytes”

May combine two claims.

220 and 275 genes and 281 and 475 regions vary along the porto-central axis in HSC and LSECs, respectively.

“We also identified 220 and 275 genes and 281 and 475 regions that vary along the porto-central axis in HSC and LSECs, respectively (adjusted p-value < 0.01”

May combine two claims.

There is a mid-lobular area where Cyp2e1 and Cps1 are co-expressed, reflected by overlap of topic 60 and topic 66.

“We additionally identified a mid-lobular area where Cyp2e1 and Cps1 are co-expressed, also reflected by the overlap of topic 60 (pericentrally intermediate) and topic 66 (periportal) cell contributions.”

May combine two claims.

SCENIC+ analysis revealed 180 regulons, including Spi1, Jdp2, Runx1 in Kupffer cells; Ebf1 and Pax5 in B cells; Gata3 and Lef1 in T cells; Gata4, Meis1 and Maf in LSECs; Lhx2 and Tead1 in HSC; Wt1 in fibroblasts; and Tead4, Dox4 and Hnf1b in BECs.

“This analysis revealed 180 regulons, including Spi1, Jdp2, Runx1 in Kupffer cells; Ebf1 and Pax5 in B cells; Gata3 and Lef1 in T cells; Gata4, Meis1 and Maf in LSECs; Lhx2 and Tead1 in HSC; Wt1 in fibroblasts; and Tead4, Dox4 and Hnf1b in BECs”

May combine two claims.

Chromatin regions where accessibility is negatively associated with Tbx3 expression are located nearby genes that are anti-expressed with Tbx3, and the same holds for Tcf7l1.

“the chromatin regions where accessibility is negatively associated with Tbx3 expression are located nearby genes that are anti-expressed with Tbx3 (same for Tcf7l1).”

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

Two complementary topics, Topic 17 and Topic 75, represent different phases of the circadian rhythm, with enrichment for positive and negative regulation.

“two complementary topics (Topic 17 and Topic 75) represent different phases of the circadian rhythm, with enrichment for positive and negative regulation (adjusted p-value = 10-19 and 10-7, respectively)”

May combine two claims.

Motif enrichment analysis identified Hnf1a, Ppara, Nfi, Hnf4a, Cebpa, Foxa1 and Onecut1 motifs in regions accessible across all hepatocytes, and Tbx3 and Tcf7l1/2 motifs in regions accessible periportally and pericentrally, respectively.

“This resulted in the identification of Hnf1a, Ppara, Nfi, Hnf4a, Cebpa, Foxa1 and Onecut1 motifs in regions accessible across all hepatocytes, and Tbx3 and Tcf7l1/2 motifs in regions accessible periportally and pericentrally, respectively”

May combine two claims.

On a Cdh1 enhancer, DeepLiver finds Foxa1, Hnf4a and Hnf1a sites as drivers of enhancer accessibility and activity, while Tbx3 sites are predicted to make the enhancer periportal.

“For instance, on a Cdh1 enhancer, DeepLiver finds Foxa1, Hnf4a and Hnf1a sites as drivers of enhancer accessibility and activity, while Tbx3 sites (one dimer motif, and two monomers) are predicted to make the enhancer periportal.”

May combine two claims.

SCENIC+ predicted target regions of Hnf1a, Hnf4a, Foxa1, Cebpa, Onecut1, Nfib and Tcf7l1 are all significantly more active compared to shuffled regions, with 45%, 39%, 43%, 39%, 35%, 32% and 26% active, respectively.

“The SCENIC+ predicted target regions of Hnf1a, Hnf4a, Foxa1, Cebpa, Onecut1, Nfib and Tcf7l1 are all significantly more active compared to shuffled regions, with 45%, 39%, 43%, 39%, 35%, 32% and 26% of their predicted target regions active, respectively”

May combine two claims.

Hnf1a target regions are predicted to be more active (median 0.52), while Nfib target regions have low activity scores (median 0.16).

“In agreement with results above, Hnf1a target regions are predicted to be more active (median 0.52), while Nfib target regions have low activity scores (median 0.16, Fig S18b).”

May combine two claims.

Tcf7l1 expression is anticorrelated with region accessibility and gene expression (-0.32 and -0.40).

“Tcf7l1 expression is anticorrelated with region accessibility and gene expression (-0.32 and -0.40).”

May combine two claims.

Pericentral enhancers show higher activity in the pericentral fraction, and vice-versa.

“As expected, pericentral enhancers show higher activity in the pericentral fraction, and vice-versa (Fig 5h, Fig S22b,c).”

May combine two claims.

Core regulons shared across all mice include Onecut1, Cebpa, Hnf1a, Foxa1 and Nfib; Tbx3 and Tcf7l1 are core pericentral and periportal repressive regulons, respectively.

“Among the general 'core' (shared across all mice) regulons we identified Onecut1, Cebpa, Hnf1a, Foxa1 and Nfib; while Tbx3 and Tcf7l1 are core pericentral and periportal (repressive) regulons, respectively”

May combine two claims.

Variants of the predicted binding sites of the zonation TFs Tbx3 and Tcf7l1 also show changes, but these are more difficult to assess from bulk experiments where periportal and pericentral hepatocytes are pooled.

“Variants of the predicted binding sites of the zonation TFs Tbx3 and Tcf7l1 also show changes, but these are more difficult to assess from these bulk experiments where periportal and pericentral hepatocytes are pooled (Fig 5e, S19f-g, S20, S21).”

May combine two claims.

A Nfib site contributes to accessibility but not activity, and a Tcf7l1/2 site is uniquely found in the zonation model, contributing to make the Cyp2e1 enhancer pericentral.

“On the other hand, a Nfib site contributes to accessibility (but not activity), and a Tcf7l1/2 site is uniquely found in the zonation model, contributing to make the enhancer pericentral.”

May combine two claims.

On a pericentral Cyp2e1 enhancer, DeepLiver identifies Hnf1a, Foxa1 and Onecut1 sites that contribute to enhancer accessibility and activity, and an Ets site that contributes to activity but not accessibility.

“On a pericentral Cyp2e1 enhancer, DeepLiver identifies Hnf1a, Foxa1 and Onecut1 sites that contribute to enhancer accessibility and activity, and an Ets site that contributes to activity but not accessibility, as observed in other enhancers too (Supplementary Note 1).”

May combine two claims.

DeepLiver predicts F9 and SORT1 mutagenesis effects with a Spearman correlation of 0.64 and 0.36, respectively, while other methods range from -0.04 to 0.52 and -0.28 to 0.29.

“For example, DeepLiver predicts F9 and SORT1 mutagenesis effects with a spearman correlation of 0.64 and 0.36, respectively, while for other methods ranges from -0.04 to 0.52 and -0.28 to 0.29 (Fig S17).”

May combine two claims.

Both accessibility of the Cdh1 enhancer and Cdh1 gene expression are anticorrelated with Tbx3 expression (-0.44 and -0.17, respectively).

“Importantly, both accessibility of this enhancer, and Cdh1 gene expression, are anticorrelated with Tbx3 expression (-0.44 and -0.17, respectively).”

May combine two claims.

Tbx3 LOF and Tcf7l1 GOF resulted in an activity increase in the pericentral fraction, and Tcf7l1 LOF in the periportal population.

“Tbx3 LOF and Tcf7l1 GOF resulted in an activity increase in the pericentral fraction and Tcf7l1 LOF in the periportal population (Fig 5h,i, Fig S22d, S23).”

May combine two claims.

DeepLiver classifies Tbx3 target regions as periportal, Tcf7l1 and Nfib as pericentral/general, and Hnf4a, Foxa1, Hnf1a, Onecut1 and Cebpa regions as general.

“Importantly, DeepLiver classifies Tbx3 target regions as periportal, Tcf7l1 and Nfib as pericentral/general, and Hnf4a, Foxa1, Hnf1a, Onecut1 and Cebpa regions as general (Fig S18a).”

May combine two claims.

Hnf1a and Hnf4a GOF variants resulted in increased activity in both fractions, with milder effects for Cebpa and Foxa1 variants.

“Hnf1a and Hnf4a GOF variants resulted in increased activity in both fractions, with milder effects for Cebpa and Foxa1 variants.”

May combine two claims.

Among mouse regions active in vivo in mouse liver, 64% are distal enhancers and 27% are promoters.

“Among the mouse regions that are active in vivo in the mouse liver, 64% are distal enhancers and 27% are promoters.”

May combine two claims.

Between 1,235 and 6,991 target regions were identified for Tbx3, Tcf7l1, Foxa1, Hnf1a, Hnf4a, Nfib, Onecut1 and Cebpa, with 17%-70% overlap with SCENIC+ predicted target regions.

“We identified between 1,235 and 6,991 target regions for Tbx3, Tcf7l1, Foxa1, Hnf1a, Hnf4a, Nfib, Onecut1 and Cebpa, with good overlap with SCENIC+ predicted target regions (17%-70%, Fig S18a-c).”

May combine two claims.

Sequences with more than two Hnf1a motif instances were predicted as highly active by DeepLiver, in agreement with experimental results.

“Sequences with more than two Hnf1a motif instances were predicted as highly active by DeepLiver, in agreement with experimental results (Fig S16h-j).”

May combine two claims.

Simulating Tbx3 or Tcf7l1 knockdown and overexpression in hepatocytes suggests that Tbx3 overexpression and Tcf7l1 knockdown can switch periportal hepatocytes to a pericentral state, while Tbx3 knockdown or Tcf7l1 overexpression can switch pericentral hepatocytes to a periportal state.

“Simulating Tbx3 or Tcf7l1 knockdown and overexpression in hepatocytes (see Methods), suggests that Tbx3 overexpression and Tcf7l1 knockdown can switch periportal hepatocytes to a pericentral state, while Tbx3 knockdown or Tcf7l1 overexpression can switch pericentral hepatocytes to a periportal state (Fig 5a-c).”

May combine two claims.

Among candidate enhancers near the pericentral gene Axin2 there are six Tcf7l1 target regions; among candidate enhancers near the periportal gene Cdh1 there are two Tbx3 target regions; in both cases these regions are bound by additional core general TFs.

“among the candidate enhancers near the pericentral hepatocyte gene Axin2 we find six Tcf7l1 target regions. Among the candidate enhancers near the periportal gene Cdh1 we identify two Tbx3 target regions. In both cases, these regions are bound by additional core general TFs”

May combine two claims.

In the Cdh1 enhancer region, Hnf4a and Foxa1 ChIP-seq signal is present, but no Cebpa nor Onecut1 ChIP-seq signal is present.

“In agreement, we find Hnf4a and Foxa1 ChIP-seq signal in this region, but no Cebpa nor Onecut1 ChIP-seq signal.”

May combine two claims.

Tbx3 target regions are more active in vivo, with 20% and 5% of the regions active in vivo and HepG2, respectively.

“Tbx3 target regions are more active in vivo (20% and 5% of the regions are active in vivo and HepG2, respectively).”

May combine two claims.

Simulated knockdown of Tbx3 or Tcf7l1 results in upregulation of their predicted target genes, while their overexpression results in target gene downregulation.

“Simulated knockdown of these TFs results in upregulation of their predicted target genes, while their overexpression results in target gene downregulation.”

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

Topic modelling of scATAC-seq with cisTopic resulted in a model with 19 topics.

“This resulted in a model with 19 topics.”

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

The libraries cloned in the pSA293-CHEQseq-3'BC plasmids yielded 723,805 and 71,658 unique enhancer-barcode assignments, with 85.5% and 97.8% of barcodes assigned to a unique enhancer, respectively.

“This resulted in libraries with 723,805 and 71,658 unique enhancer-barcode assignments, with 85.5% and 97.8% barcodes assigned to a unique enhancer, respectively.”

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

The 5' 12K library yielded 8,835,050 enhancer-barcode assignments, with 78.3% of barcodes assigned to a unique enhancer.

“This resulted in 8,835,050 enhancer-barcode assignments for the 5' 12K library, with 78.3% barcodes assigned to a unique enhancer.”

May combine two claims.

220, 275 and 2697 genes and 281, 475 and 8.805 regions vary along the porto-central axis in HSC, LSEC and hepatocytes, respectively.

“We identified 220, 275 and 2697 genes and 281, 475 and 8.805 regions that vary along the porto-central axis in HSC, LSEC and hepatocytes, respectively.”

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

Leiden clustering with resolution 0.6 resulted in 15 clusters.

“using Leiden clustering with resolution 0.6, resulting in 15 clusters.”

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

Using 14 PCs, Leiden clustering resulted in 19 clusters that corresponded to 11 cell types, annotated based on marker gene expression.

“Using 14 PCs, we performed Leiden clustering, resulting in 19 clusters that corresponded to 11 cell types, that were annotated based on marker gene expression.”

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

The samples were merged, obtaining 29,798 high-quality cells.

“The samples were merged, obtaining 29,798 high-quality cells, and reanalyzed with VSN-Pipelines.”

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

pycisTopic peak calling and consensus peak merging resulted in 486,888 regions.

“This resulted in 486,888 regions.”

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

Topic modelling with Mallet selected a model with 82 topics.

“we selected a model with 82 topics.”

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

Leiden clustering with resolution 0.6 after harmonypy correction resulted in 11 clusters, corresponding to 14 cell types based on previous labelling.

“and Leiden clustering with resolution 0.6 resulted in 11 clusters, corresponding to 14 cell types based on previous labelling.”

May combine two claims.

Using a similar strategy within SCENIC+ revealed that Tbx3 knock-down drives cells to the periportal state, while Tcf7l1 knock-down drives cells to the pericentral state.

“We employed a similar strategy within SCENIC+, which revealed that Tbx3 knock-down indeed drives cells to the periportal state, while Tcf7l1 knock-down drives cells to the pericentral state.”

Classified as probably not a substantial claim.

Inference of consensus peaks from the human liver scATAC-seq data yielded a dataset of 121,593 regions and 6,366 cells.

“we first inferred consensus peaks as previously described, resulting in a data set with 121,593 regions and 6,366 cells.”

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

Hi-C data processing resulted in 890,488 region-gene links.

“resulting in 890,488 region-gene links.”

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

scATAC-seq quality filtering resulted in 22,600 high-quality cells.

“resulting in 22,600 high-quality cells.”

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

The selected pycisTopic topic model for the mouse liver epigenome comprises 82 topics.

“The selected topic model (82 topics) is indicated with a vertical grey line.”

May combine two claims.

Ntgr is a periportal HSC marker and Adamtsl2 is a pericentral HSC marker in the mouse liver.

“Liver section image showing smFISH profiles for Ntgr (HSC PP marker) and Adamtsl2 (PC HSC marker).”

May combine two claims.

Ntn4 is a periportal LSEC marker and Kit is a pericentral LSEC marker in the mouse liver.

“Liver section image showing smFISH profiles for Ntn4 (PP LSEC marker) and Kit (PC LSEC marker).”

May combine two claims.

Among high-confidence enhancer classes in the mouse liver and HepG2, 4,285 are not active, 806 are active only in vivo, 921 are active only in HepG2, and 1,186 are active in both HepG2 and in vivo.

“None: Not active (4,285), In vivo: Active only in vivo (806), HepG2: Active only in HepG2 (921), Both: Active in HepG2 and in vivo (1,186).”

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

Hepatocyte function varies depending on position along the porto-central axis, as hepatocytes are exposed to different microenvironments, a phenomenon known as zonation.

“Hepatocyte function varies depending on the position along this porto-central axis, as they are exposed to different microenvironments, a phenomenon known as zonation7.”

• cites [7] sourceNot checked — no full text of the source

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

Periportal hepatocytes are more involved in gluconeogenesis and beta-oxidation, while pericentral hepatocytes are more involved in glycolysis, lipogenesis, and detoxification.

“For example, periportal hepatocytes are more involved in gluconeogenesis and beta-oxidation, while pericentral hepatocytes are more involved in glycolysis, lipogenesis, and detoxification7.”

• cites [7] sourceNot checked — no full text of the source

May combine two claims.

The hepatocyte zonation eGRN is a subset or a layer of the general hepatocyte eGRN.

“This suggests that the hepatocyte zonation eGRN is a subset (or a layer) of the general hepatocyte eGRN.”

May combine two claims.

These observations suggest that Tbx3 and Tcf7l1 may repress these regions.

“These observations suggest that Tbx3 and Tcf7l1 may repress these regions.”

May combine two claims.

The DeepLiver motif results agree with motifs identified by random forest feature selection from MPRA data and with hepatocyte and zonation regulators identified by SCENIC+.

“These results thus agree with the motifs identified by random forest feature selection from the MPRA data (Fig S15), and the hepatocyte (and zonation) regulators identified by SCENIC+ (Fig 2).”

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

Distal enhancer activation is more difficult to recapitulate in HepG2 cells, or distal enhancers are less conserved than promoters between mouse and human.

“suggesting either that distal enhancer activation is more difficult to recapitulate in HepG2 cells, or that distal enhancers are less conserved than promoters between mouse and human”

May combine two claims.

The grammar of hepatocyte enhancers that encodes their zonation pattern includes Tbx3 and Tcf7l1/2 binding sites, while Hnf1a and Hnf4a are the most relevant binding sites regarding activity.

“In summary, our results suggest that the grammar of hepatocyte enhancers that encodes their zonation pattern includes Tbx3 and Tcf7l1/2 binding sites, while Hnf1a and Hnf4a are the most relevant binding sites regarding activity.”

May combine two claims.

DeepLiver assigns given DNA sequences to cell types represented by topics in the liver and predicts activity and zonation patterns of hepatocyte enhancers.

“Thus, DeepLiver assigns given DNA sequences to cell types (represented by topics) in the liver and predicts activity and zonation patterns of hepatocyte enhancers (Fig 4a).”

May combine two claims.

Around 40% of accessible regions in hepatocytes are active, and the enhancer sequence is predictive of both enhancer accessibility and enhancer activity.

“In summary, around 40% of accessible regions in hepatocytes are active, and the enhancer sequence is not only predictive of enhancer accessibility, but also of enhancer activity.”

May combine two claims.

DeepLiver decodes enhancer accessibility, activity, and zonation at base pair resolution, and can predict variants that modulate enhancer activity and zonation in hepatocytes.

“In summary, DeepLiver decodes enhancer accessibility, activity, and zonation at base pair resolution, and can predict variants that modulate enhancer activity and zonation in hepatocytes.”

May combine two claims.

SCENIC+ identified Hnf4a, Hnf1a, Cebpa, Onecut1, Foxa1 and Nfib as core general hepatocyte TFs and Tbx3 and Tcf7l1 as key repressors of the zonation programs.

“In summary, SCENIC+ identified Hnf4a, Hnf1a, Cebpa, Onecut1, Foxa1 and Nfib as core general hepatocyte TFs and the repressors Tbx3 and Tcf7l1 as key repressors of the zonation programs”

May combine two claims.

The DeepLiver model provides meaningful interpretations of hepatocyte enhancers and predicts that these enhancers consist of a core hepatocyte code mixed with zonated repressor TFs Tbx3 and Tcf7l1, which bias enhancer activity to pericentral or periportal zones respectively.

“The DeepLiver model provides meaningful interpretations of hepatocyte enhancers and predicts that these enhancers consist of a core hepatocyte code, mixed with zonated repressor TFs, Tbx3 and Tcf7l1, which bias enhancer activity to either pericentral or periportal zones respectively.”

May combine two claims.

Consequently, knockdown or overexpression of one of Tbx3 or Tcf7l1 provokes upregulation or downregulation of the other, respectively, and downregulation and upregulation of the other's target genes.

“Consequently, the knockdown or overexpression of one of the TFs provokes the upregulation or downregulation of the other, respectively (and downregulation and upregulation of the target genes of the other as well; Fig 5c).”

May combine two claims.

In periportal hepatocytes with low Wnt concentration, b-catenin is degraded and Tcf7l1 activated; this represses the pericentral program including Tbx3 and activates the periportal program.

“In periportal hepatocytes, with low Wnt concentration, b-catenin is degraded and Tcf7l1 activated. This represses the pericentral program, including Tbx3, and activates the periportal program.”

Classified as probably not a substantial claim.

Dimensionality reduction using only the 100 selected genes ensured that all cell types could be distinguished with this gene panel.

“In addition, we performed dimensionality reduction only using these 100 genes to ensure that all cell types could be distinguished with this gene panel.”

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

Hepatocyte clusters with low gene expression and high percentage of mitochondrial reads were removed, as well as doublets called with Scrublet.

“Hepatocyte clusters with low gene expression and high percentage of mitochondrial reads were removed, as well as doublets called with Scrublet (v0.2.3)87.”

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

The luciferase reporter assays were performed with three biological replicates per condition.

“Three biological replicates were done per condition.”

Classified as probably not a substantial claim.

To rank regulons based on how affected they are by zonation and/or sample, ANOVA was performed over the AUC values along the pseudotime per sample, adjusting p-values with Bonferroni.

“To rank regulons (or signatures) based on how affected they are by zonation and/or sample, we performed ANOVA over the AUC values along the pseudotime per sample (adjusting the p-values with Bonferroni).”

May combine two claims.

The pericentral topics 58 and 60 consist of regions with Tcf7l1/2 motifs while the periportal topic 66 contains Tbx3 binding sites, suggesting that Tcf7l1 and Tbx3 are directly involved in liver metabolism.

“The pericentral topics (topic 58 and topic 60) consist of regions with Tcf7l1/2 motifs while the regions in the periportal topic (topic 66) contain Tbx3 binding sites, suggesting that Tcf7l1 and Tbx3 are directly involved in liver metabolism.”

Classified as probably not a substantial claim.

After training, model performance was assessed on the non-augmented test set by scoring the test set regions with the models.

“After training, we assessed the performance of the models on the non-augmented test set by scoring the test set regions with the models.”

Classified as probably not a substantial claim.

Model performance was assessed by 9-fold cross validation.

“To assess the performance of the models, we performed 9-fold cross validation.”

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

Cells with at least 350 genes expressed and a percentage of mitochondrial reads below 10% were kept.

“Briefly, cells with at least 350 genes expressed and a percentage of mitochondrial reads below 10% were kept.”

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

scATAC-seq quality filtering kept cells with at least 1,000 fragments, FRiP > 0.4 and TSS enrichment > 7.

“We further filtered the data set based on the scATAC-seq quality as well, keeping cells with at least 1,000 fragments, FRiP > 0.4 and TSS enrichment > 7, resulting in 22,600 high-quality cells.”

Classified as probably not a substantial claim.

High quality cells were selected using minimum thresholds on FRIP, number of fragments and TSS enrichment.

“The blue dotted lines indicate the minimum threshold in FRIP, number of fragments and TSS enrichment to select high quality cells.”

May combine two claims.

High Wnt concentration leads to b-catenin stabilization, Tcf7l1 degradation, and Tcf7l2 activation, resulting in activation of the pericentral gene program that includes Tbx3.

“High Wnt concentration leads to b-catenin stabilization71, Tcf7l1 degradation, and Tcf7l2 activation, resulting in the activation of the pericentral gene program, that includes Tbx3.”

• cites [71] sourceNot checked — no full text of the source

Classified as probably not a substantial claim.

Genes were filtered for p-adj < 0.01.

“After filtering for genes fitted with p-adj < 0.01.”

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

The ratio of signals matching to codes used in the experiment and signals matching to codes not used in the experiment was used to estimate specificity (false positives).

“The ratio of signals matching to codes used in the experiment and signals matching to codes not used in the experiment were used as estimation for specificity (false positives).”

Classified as probably not a substantial claim.

Transcriptome and epigenome UMAPs (29,798 and 22,600 cells, respectively) were batch corrected with harmony.

“Transcriptome and epigenome based UMAPs (29,798 and 22,600 cells, respectively) before and after batch correction with harmony.”

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

The number of counts per enhancer correlates across samples and replicates.

“Correlation between number of counts per enhancer across samples and replicates.”

• Also stated in: § Corresponding author: stein.aerts@kuleuven.be · plain

  “Correlation between number of counts per enhancer across samples and replicates.”