Obesity is a major risk factor for type-II diabetes (T2D), yet many people living with obesity never develop it, and what makes some individuals susceptible is still unclear. To examine this at the level of adipose tissue, we collected paired omental and subcutaneous fat, together with blood, from 58 people with obesity undergoing bariatric surgery. Participants either had (Ob+T2D) or did not have (Ob) a clinical diagnosis of T2D. Subsets of the cohort were analysed with several methods: 1) single-cell RNA sequencing of adipose stromal cells (n = 10), 2) flow cytometry of adipose stromal cells and blood mononuclear cells, at baseline and after stimulation (n = 42), 3) stimulation of omental stromal cells with Toll-like receptor agonists (n = 20), 4) histology of adipose sections (n = 17), and 5) multiplex chemokine measurement in serum (n = 35). In T2D, the omental stromal compartment was remodelled, most prominently in mesothelial and endothelial cells. Mesothelial and IGFBP2+ cells took on a fibrotic, epithelial–mesenchymal transition-like state. Omental endothelial cells carried signatures of lipid overload, inflammation and antigen presentation. Innate T cells and macrophages differed in phenotype between the two fat depots, and innate T cells showed inflammatory profiles that diverged between omental and subcutaneous tissue. Together, the data point to depot-specific changes in adipose tissue in T2D, and to compensatory cell networks that could be targeted therapeutically. This deposit contains: 1) adipose histology images and adipocyte size measurements, 2) gated flow cytometry statistics 3) serum chemokine data, and 4) de-identified patient clinical metadata. Single-cell sequencing data are available from GEO (GSE278526).

Full experimental detail is available in the associated article. A summary relevant to users of this dataset follows.

Participants, ethics and group assignment

Participants were 58 adults with obesity having elective bariatric surgery at St. Vincent’s University Hospital and St. Michael’s Hospital, Dublin, Ireland. They were recruited consecutively from July 2020 to June 2021, and each provided written informed consent. Each donated venous blood plus paired abdominal subcutaneous and omental adipose tissue, with omental tissue taken from the left lower edge of the greater omentum. The study was approved under PHS IRB protocol 2019P001128 and followed the Declaration of Helsinki (2013). Specimens and clinical records were separated from identifying information and given an anonymised code (P##).

Participants were assigned to one of two groups: obesity without T2D (Ob) or obesity with T2D (Ob+T2D). Assignment to Ob+T2D required both a clinical T2D diagnosis and an HbA1c of at least 39 mmol/mol (5.7 %, the pre-diabetes threshold) on routine pre-operative bloods (1). For exploratory analyses in the flow cytometry cohort, participants were also classified by whether they had hypertension and dyslipidaemia.

Not every participant contributed to every analysis. Allocation depended on how much tissue and blood was available on the day of surgery:

• flow cytometry of adipose stromal cells and PBMCs, unstimulated and stimulated: n = 42
• single-cell RNA sequencing (scRNA-seq) of adipose immune and non-immune stromal cells: n = 10. Omental tissue from these same participants was also used for TLR stimulation with LPS and Poly(I:C)
• TLR stimulation with LPS, Poly(I:C), R848 and CpG: omental tissue from 10 further participants

• histology: n = 17, using tissue left over after the other analyses

• serum chemokine/cytokine multiplex: n = 35, all available serum samples

Clinical characteristics for each analysis are tabulated in the associated article (Tables S1–S8 in associated article).

Sample collection and transport

Blood was drawn at the start of surgery and adipose tissue was taken during the procedure. Tissue went straight into plain Hanks’ Balanced Salt Solution (HBSS; Sigma-Aldrich, Darmstadt, DE) pre-chilled to 4 °C and reached the laboratory within 20–30 minutes. Venous blood from an antecubital vein was collected into BD Vacutainer tubes (BD, Berkshire, UK): serum-separation tubes for serum, and EDTA tubes for plasma and leukocytes.

Blood processing

Serum tubes were left to clot for about 30 minutes at room temperature (RT), then spun at 3,000 × g for 10 minutes at 4 °C. Serum was removed, held at −20 °C for about 1 hour and then moved to −80 °C until use.

PBMCs were isolated from the EDTA blood over a Ficoll density gradient (Sigma-Aldrich), and plasma was collected from the same separation. Part of the fresh PBMC fraction was set aside for flow cytometry. The rest was cryopreserved in RPMI 1640 (Gibco, Thermo Fisher Scientific, Leicestershire, UK) with 40 % foetal bovine serum (FBS) and 10 % DMSO (both Sigma-Aldrich), cooled at a controlled rate to −80 °C in Mr. Frosty containers (Thermo Fisher Scientific). Plasma was stored the same way as serum.

Adipose tissue processing and digestion

Tissue (1–12 g) was taken out of HBSS and rinsed on 100 µm gauze with PBS (Thermo Fisher Scientific) warmed to 37 °C to remove residual blood. It was then weighed and divided three ways:

• a 0.5–5 g piece for immediate enzymatic digestion
• a small piece into 10 % neutral buffered formalin at RT for histology
• the remainder frozen in CryoStor CS10 (STEMCELL Technologies, Cambridge, UK), cooled at a controlled rate to −80 °C in Mr. Frosty containers

For digestion, tissue was finely minced and placed in 6 mL of adipose buffer pre-warmed to 37 °C. Adipose buffer was pH 7.4 and contained:

• 140 mM NaCl, 4.7 mM KCl, 1.25 mM MgSO₄·7H₂O, 2.5 mM CaCl₂·2H₂O and 2.5 mM NaH₂PO₄·2H₂O
• 10 mM HEPES
• 5.5 mM D-glucose
• 2 % BSA fraction V

All buffer components were from Sigma-Aldrich. Type-1 collagenase (250 U/mL; Worthington Biochemical, NJ, US) was added. Samples were shaken at 180–200 rpm and 37 °C for 45–60 minutes, and digestion was stopped by adding EDTA to 2 mM.

The digest was passed through 400 µm gauze and spun at 100 × g for 1 minute at RT. The remaining solution was spun at 300 × g for 7 minutes at RT. Red cells were then lysed for 10 minutes at RT in 9 volumes of lysis buffer (155.1 mM NH₄Cl, 9.9 mM KHCO₃, 0.1 mM EDTA; Sigma-Aldrich). After lysis, cells were diluted in adipose buffer and spun again (300 × g, 7 minutes, RT), and the supernatant was discarded. The cells were then resuspended in adipose buffer, filtered through 100 µm gauze and spun a final time under the same conditions. The final cell suspension was taken up in FACS buffer (PBS with 2 mM EDTA and 2 % FBS).

Stimulation of PBMCs and adipose leukocytes

PMA/ionomycin (Panel 4). PBMCs and adipose leukocytes were resuspended in complete RPMI and cultured at 37 °C, 5 % CO₂ and 95 % humidity. Complete RPMI was RPMI with 10 % FBS, 100 U/mL penicillin and 0.1 mg/mL streptomycin (Thermo Fisher Scientific), plus 2 mM L-glutamine. Stimulated wells received Cell Activation Cocktail without brefeldin A, which contains PMA and ionomycin (BioLegend). Brefeldin A (BioLegend; 5 µg/mL final) was added to both stimulated and unstimulated wells, and cells were cultured for 4 hours. They were then spun at 1,500 × g for 5 minutes, washed twice in FACS buffer and stained. Unstimulated wells were used to set cytokine gates.

TLR agonists (Panel 5). Cryopreserved omental tissue from the 10 scRNA-seq participants was thawed and digested as above, and the cells were cultured in complete RPMI under the same conditions. Stimulated wells received LPS (100 ng/mL; Enzo Life Sciences, NY, US) and Poly(I:C) (50 µg/mL; InvivoGen, CA, US). Omental cells from another 10 participants (5 per group) were stimulated in the same way with a broader TLR agonist cocktail:

• Poly(I:C), 50 µg/mL
• LPS, 100 ng/mL
• resiquimod (R848), 10 µg/mL
• CpG oligonucleotides, 1 µM

Except for LPS, the agonists came from InvivoGen. All wells contained brefeldin A (5 µg/mL), and cells were cultured for 6 hours with or without stimulation. They were then washed and stained as for Panel 4, and unstimulated wells were again used for gating.

Flow cytometry

Cells were stained with Zombie Aqua or Zombie Yellow fixable viability dye (BioLegend; 1:200 in FBS-free PBS) for 20 minutes at RT in the dark. Each sample was about 250,000 PBMCs or up to 500,000 adipose stromal cells. After a wash in FACS buffer (1,500 × g, 5 minutes), surface markers were stained for 20 minutes at RT in the dark. Staining was done in 100 µL of FACS buffer containing 5 % Human TruStain FcX Fc-receptor block (BioLegend) and antibodies at 1:50. Cells were then fixed in BioLegend fixation buffer and permeabilised according to the manufacturer’s instructions. Nuclear transcription factors used the FOXP3 Fix/Perm buffer set (BioLegend), and intracellular cytokines used the Intracellular Staining buffer (BioLegend). As a last step before acquisition, LipidTOX Green (Invitrogen) was added at 1:100 for 30 minutes at RT in the dark.

Fresh adipose leukocytes were stained with three panels: lymphocytes (Panel 1), innate lymphocytes (Panel 2), and myeloid cells and granulocytes (Panel 3). Panels 4 and 5 were used for the PMA/ionomycin and TLR stimulations respectively. PBMCs were stained with Panels 1, 2 and 4. Gating trees are shown in the associated article (Figs. S3–S6).

Precision Count Beads (BioLegend) were added to every sample to calculate absolute cell numbers. Data were acquired on BD FACSAria III and LSRFortessa cytometers. Compensation used UltraComp eBeads (Thermo Fisher Scientific), with one exception: in the TLR experiments, LipidTOX Green and Zombie Aqua were compensated using adipose tissue cells. Gating and export were done in FlowJo v10.8.1 (FlowJo LLC, OR, US).

Antibodies are listed below by fluorochrome as target (clone; supplier). BD = BD Biosciences; BL = BioLegend.

• BUV395: CD16 (3G8; BD)
• FITC: CD36 (5-271; BL); CD235a (GA-R2; BD); Vα7.2 (3C10; BL); CD8a (RPA-T8; BL)
• PerCP: CD45 (2D1; BD); CD34 (8G12; BD)
• PerCP-Cy5.5: IL-4 (MP4-25D2; BL)
• APC: CD15 (HI98; BD); HLA-DR (L243; BL); CD4 (A161A1; BL); Vδ3 (P11.5B; Beckman Coulter, CA, USA); CD1d tetramer (PBS-57; NIH Tetramer Core Facility, GA, US)
• Alexa Fluor 700: TNF (MAb11; BL); CD3 (UCHT1; BD); CD11b (ICRF44; BD)
• APC-H7: CD14 (MφP9; BD)
• APC-Cy7: CD31 (WM59; BL); CD4 (OKT4; BL); IL-17A (BL168; BL)
• BV421: TNF (MAb11; BL); HLA-DR (G46-6; BD); CD66b (G10F5; BD); CD1d tetramer (PBS-57; NIH Tetramer Core Facility); TCR-γδ (B1; BL)
• BV510: CD3 (OKT3; BL)
• BV605: CD14 (M5E2; BL); CD3 (SK7; BD); Siglec-8 (837535; BD); CD8 (SK1; BL); CD45 (HI30; BL)
• BV650: IFNγ (4S.B3; BD); CD8 (SK1; BL); CD161 (DX12; BD); CD206 (19.2; BD); RORγt (Q21-559; BD)
• BV711: CD3 (OKT3; BL); Vα7.2 (OF-5A12; BD); Vδ2 (B6; BL); CD11b (ICRF44; BD); CD19 (HIB19; BD); CD63 (H5C6; BL)
• BV785/BV786: CD45RA (HI100; BL); CD45 (HI30; BD, BV786); CD45 (HI30; BL, BV786); HLA-DR (G46-6; BD); CD4 (OKT4; BL); CD8 (SK1; BL)
• PE: CD161 (HP-3G10; BL); IL-10 (JES3-19F1; BL); IL-4 (MP4-25D2; BL); CD27 (L128; BD); MSLN (420411; R&D Systems, MN, USA); TREM2 (237920; R&D Systems)
• PE-CF594: PLZF (R17-809; BD); IL-6 (MQ2-13A5; BD); CD9 (M-L13; BD); IgD (IA6-2; BD); IFNγ (B27; BD); IL-17F (O33-782; BD)
• PE-Cy5: CD56 (B159; BD)
• PE-Cy7: IL-17A (BL168; BL); Vδ1 (TS8.2; Invitrogen); CD16 (3G8; BD); TCR-γδ (B1; BL); IFNγ (B27; BL); CD44 (G44-26; BD)

Adipose histology

Tissue was fixed in formalin for 24 hours at RT and taken through 2-hour ethanol steps of increasing concentration, from 30 % to 70 % in distilled water. It was then held overnight in 70 % ethanol at 4 °C. Samples were placed in cassettes (VWR International, Dublin, IE), dehydrated and embedded in paraffin. Sections of 7–10 µm were cut on a microtome (Leica Biosystems, IL, US), mounted on glass slides (VWR) and baked overnight at 60 °C. H&E staining was done on a Leica Autostainer XL ST5010, and slides were coverslipped on a Leica CV5030. Slides were imaged at 10× on an Olympus BX51 upright microscope. Adipocyte size was measured in batch with the Adiposoft plugin (2) for ImageJ (3). Crown-like structures were counted by hand and expressed per cm² of imaged tissue.

Serum chemokines

Serum chemokines were measured with the LEGENDplex human pro-inflammatory chemokine 13-plex bead assay (BioLegend), following the kit instructions. Data were acquired on a BD LSRFortessa and analysed with BioLegend’s online LEGENDplex analysis software. Two analytes were not reported in the associated article because too few samples gave usable data: MIG (monokine induced by gamma interferon) and MIP-3α (macrophage inflammatory protein-3α). Measurements whose technical replicates differed by a coefficient of variation above 30 % were excluded.

Single-cell RNA sequencing

Paired omental and subcutaneous tissue from 10 participants (5 per group) was digested as above. Stromal cells were sorted into CD45+ and CD45− fractions on a BD FACSAria III using the following stains (gating in Fig. S1 of the associated article):

• CD45–BV785 (BL)
• CD235a–FITC (BL)
• Zombie Aqua (BL)
• propidium iodide (Thermo Fisher Scientific)

Up to 5,000 cells per sample, half CD45+ and half CD45−, went forward. Cells from the five participants in each group were pooled and loaded on a single lane of a Chromium chip (10x Genomics, CA, US) with the Single Cell 3′ kit. cDNA and libraries were made according to the manufacturer’s protocol. Sequencing was done on an Illumina NextSeq 500 (Illumina, CA, US) at an average of 25,726 reads per cell. Per-cell transcript counts were generated with Cell Ranger v3.1, using the STAR aligner against the GRCh38-3.0.0 reference.

Demultiplexing. Cells were assigned to participants by genotype using souporcell (4), with no reference genotypes and the number of clusters set to five, one per participant in each pool. As validation, souporcell was first run separately on the omental and subcutaneous pools, and the SNP calls of the resulting clusters were compared by intersection over union. Each omental cluster matched exactly one subcutaneous cluster, as expected for paired samples. Sex-specific gene expression gave three female and two male clusters per pool, matching the three women and two men in each group (Fig. S2 of the associated article). For the final assignment, omental and subcutaneous data were merged within each group to increase coverage, and souporcell was rerun with five clusters (Ob and Ob+T2D kept separate). Cells not assigned to a single participant were discarded.

Quality control and clustering. Cells were kept if they expressed at least 100 genes and had less than 20 % mitochondrial reads, and genes were kept if detected in at least three cells. This left 25,830 omental and 9,127 subcutaneous cells. Downstream analysis used Scanpy (5) and Seurat (6) in R (7). Counts were normalised with analytic Pearson residuals (8), and the top 1,000 highly variable genes were used for PCA. A neighbourhood graph built from the PCA coordinates was used for UMAP embedding and Leiden clustering (9). Cell types were annotated by hand from marker gene expression, guided by published adipose tissue atlases (10–12). Participant and cell-type annotations are provided on the GEO record.

Downstream analyses.

• Differential abundance: per-participant cell-type proportions were tested with scCODA (13) within the pertpy framework (14).
• Cell–cell communication: predicted with MultiNicheNet (15), using the NicheNet v2 ligand–target and ligand–receptor networks (16, 17). To find shared outgoing signalling patterns, MultiNicheNet prioritisation scores summarised at group level (Ob+T2D vs Ob) were factorised separately for each depot by non-negative matrix factorisation (NMF R package (18)). Each sender cell type × ligand score matrix was split into four patterns, each a set of prioritised ligands shared across cell types, with scores giving each cell type’s contribution to each pattern.
• Multicellular programmes: inferred with DIALOGUE (19), which gave three programmes in omental tissue. The third captured genes involved in immune–stromal interactions.
• Trajectories: groups were integrated with Harmony (20). Diffusion pseudotime (21) in Scanpy was used for two proposed transitions: mesothelial cells → IGFBP2+ cells → fibro-adipogenic progenitors (FAPs), and FAPs → pre-adipocytes. Slingshot (22) trajectories were tested with tradeSeq (23) and visualised with ForceAtlas2 (24) at the cell level and PAGA (25) at the cluster level.
• Differential expression and enrichment: Ob+T2D was compared with Ob within each cell type using MAST (26) on single cells, with participant sex and cell coverage as covariates. Gene set enrichment was run with fgsea (27) against KEGG (28) and Gene Ontology (29, 30), ranking genes by sign(logFC) × −log10(FDR).
• Visualisation: ComplexHeatmap (31), seaborn (32) and ggplot2 (33).

Code and package versions are available at https://github.com/TRIMW/Defining-Metabolically-Healthy-Obesity-in-Humans. Raw data are at GEO (GSE278526). Before deposit, genetic variation was removed from the FASTQ files using BAMboozle v0.5.0 (34). Claude Code (Anthropic) was used to help customise the R and Python scripts for the MultiNicheNet, DIALOGUE, pseudotime and MOFA2 analyses.

Statistical analysis

Group comparisons. Normality was checked with the Shapiro–Wilk test, and p < 0.05 was taken as non-normal. Ob and Ob+T2D were compared with Welch’s or independent-samples t-tests for normal data. Non-normal data were compared with Mann–Whitney U tests with tie and continuity correction. Adipocyte size distributions were compared by two-way mixed-model ANOVA with Bonferroni post hoc correction. Correlations used Spearman’s ρ or Pearson’s r.

Covariates and comorbidities.

• Covariate effects were assessed by comparing ANOVA with ANCOVA for normal data, or Kruskal–Wallis with Quade tests for non-normal data.
• The cumulative burden of clinically diagnosed comorbidities was tested by simple linear regression, or by the Jonckheere–Terpstra test for non-normal data.
• The effects of combinations of T2D, hypertension and dyslipidaemia on flow cytometry readouts were tested by two-way ANOVA with type III sums of squares for normal data, or by the Scheirer–Ray–Hare test for non-normal data. Comparisons against the group negative for both conditions used Dunnett’s test, and comparisons across all groups used Benjamini–Hochberg FDR correction. Each pair of conditions was analysed by crossing positive/negative status for one (e.g. T2D −/+) with positive/negative status for the other (e.g., hypertension −/+).
• Sex differences were tested with the same two-way approach, followed by Tukey’s HSD.

Effect sizes and reporting. Effect sizes are Hedges’ g for n ≤ 20 and Cohen’s d for n > 20. Descriptive statistics are mean ± SD. Analyses were run in R, Python, SPSS (IBM, NY, US) and Prism (GraphPad Software, CA, US).

MOFA2. Proportional flow cytometry variables were analysed with Multi-Omics Factor Analysis v2 (MOFA2) (35). Variables were arranged into six views, one per sample type (blood, omental adipose, subcutaneous adipose) and condition (basal or PMA/ionomycin-stimulated), with participants shared across views. Basal data came from the fresh unstimulated panels (Panels 1–3). The unstimulated wells of the stimulation panels were not used: they received only 5–10 % of the cells and served solely as gating controls, so no basal cytokine data were generated. Variables were filtered by variance and inverse-normal transformed before training. Training requested 10 factors in slow convergence mode. Factors that explained less than 1 % of variance across all views were dropped, leaving 8. Per-participant factor scores were correlated with clinical and biochemical variables by Pearson correlation, with FDR correction across all factor–variable tests. The flow cytometry variables behind each factor were identified from its feature weights.

References

1. American Diabetes Association Professional Practice Committee, 2. Classification and Diagnosis of Diabetes: Standards of Medical Care in Diabetes—2022. Diabetes Care 45, S17–S38 (2022). doi: 10.2337/dc22-S002
2. M. Galarraga et al., Adiposoft: automated software for the analysis of white adipose tissue cellularity in histological sections. J Lipid Res 53, 2791–2796 (2012). doi: 10.1194/jlr.D023788
3. C. A. Schneider, W. S. Rasband, K. W. Eliceiri, NIH Image to ImageJ: 25 years of image analysis. Nat Methods 9, 671–675 (2012). doi: 10.1038/nmeth.2089
4. H. Heaton et al., Souporcell: robust clustering of single-cell RNA-seq data by genotype without reference genotypes. Nat Methods 17, 615–620 (2020). doi: 10.1038/s41592-020-0820-1
5. F. A. Wolf, P. Angerer, F. J. Theis, SCANPY: large-scale single-cell gene expression data analysis. Genome Biol 19, 15 (2018). doi: 10.1186/s13059-017-1382-0
6. Y. Hao et al., Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nat Biotechnol (2023). doi: 10.1038/s41587-023-01767-y
7. RStudio Team, RStudio (RStudio, PBC, Boston, MA, 2020).
8. J. Lause, P. Berens, D. Kobak, Analytic Pearson residuals for normalization of single-cell RNA-seq UMI data. Genome Biol 22, 258 (2021). doi: 10.1186/s13059-021-02451-7
9. V. A. Traag, L. Waltman, N. J. van Eck, From Louvain to Leiden: guaranteeing well-connected communities. Sci Rep 9, 5233 (2019). doi: 10.1038/s41598-019-41695-z
10. A. D. Hildreth et al., Single-cell sequencing of human white adipose tissue identifies new cell states in health and obesity. Nat Immunol 22, 639–653 (2021). doi: 10.1038/s41590-021-00922-4
11. R. Ferrero et al., A human omentum-specific mesothelial-like stromal population inhibits adipogenesis through IGFBP2 secretion. Cell Metab (2024). doi: 10.1016/j.cmet.2024.04.017
12. M. P. Emont et al., A single-cell atlas of human and mouse white adipose tissue. Nature 603, 926–933 (2022). doi: 10.1038/s41586-022-04518-2
13. M. Büttner et al., scCODA is a Bayesian model for compositional single-cell data analysis. Nat Commun 12, 6876 (2021). doi: 10.1038/s41467-021-27150-6
14. L. Heumos et al., Pertpy: an end-to-end framework for perturbation analysis. Nat Methods 23, 350–359 (2026). doi: 10.1038/s41592-025-02909-7
15. R. Browaeys et al., MultiNicheNet: a flexible framework for differential cell-cell communication analysis from multi-sample multi-condition single-cell transcriptomics data. bioRxiv 2023.06.13.544751 (2023). doi: 10.1101/2023.06.13.544751
16. R. Browaeys, NicheNet-v2: final networks and ligand-target matrices [Data set] (2022). doi: 10.5281/zenodo.7074291
17. R. Browaeys, NicheNet-v2 ligand-receptor networks: integrated information from Omnipath [Data set] (2023). doi: 10.5281/zenodo.10229222
18. R. Gaujoux, C. Seoighe, A flexible R package for nonnegative matrix factorization. BMC Bioinformatics 11, 367 (2010). doi: 10.1186/1471-2105-11-367
19. L. Jerby-Arnon, A. Regev, DIALOGUE maps multicellular programs in tissue from single-cell or spatial transcriptomics data. Nat Biotechnol 40, 1467–1477 (2022). doi: 10.1038/s41587-022-01288-0
20. I. Korsunsky et al., Fast, sensitive and accurate integration of single-cell data with Harmony. Nat Methods 16, 1289–1296 (2019). doi: 10.1038/s41592-019-0619-0
21. L. Haghverdi et al., Diffusion pseudotime robustly reconstructs lineage branching. Nat Methods 13, 845–848 (2016). doi: 10.1038/nmeth.3971
22. K. Street et al., Slingshot: cell lineage and pseudotime inference for single-cell transcriptomics. BMC Genomics 19, 477 (2018). doi: 10.1186/s12864-018-4772-0
23. K. Van den Berge et al., Trajectory-based differential expression analysis for single-cell sequencing data. Nat Commun 11, 1201 (2020). doi: 10.1038/s41467-020-14766-3
24. M. Jacomy et al., ForceAtlas2, a continuous graph layout algorithm for handy network visualization designed for the Gephi software. PLoS One 9, e98679 (2014). doi: 10.1371/journal.pone.0098679
25. F. A. Wolf et al., PAGA: graph abstraction reconciles clustering with trajectory inference through a topology preserving map of single cells. Genome Biol 20, 59 (2019). doi: 10.1186/s13059-019-1663-x
26. G. Finak et al., MAST: a flexible statistical framework for assessing transcriptional changes and characterizing heterogeneity in single-cell RNA sequencing data. Genome Biol 16, 278 (2015). doi: 10.1186/s13059-015-0844-5
27. G. Korotkevich et al., Fast gene set enrichment analysis. bioRxiv 060012 (2021). doi: 10.1101/060012
28. M. Kanehisa, S. Goto, KEGG: Kyoto Encyclopedia of Genes and Genomes. Nucleic Acids Res 28, 27–30 (2000). doi: 10.1093/nar/28.1.27
29. M. Ashburner et al., Gene Ontology: tool for the unification of biology. Nat Genet 25, 25–29 (2000). doi: 10.1038/75556
30. The Gene Ontology Consortium, The Gene Ontology knowledgebase in 2026. Nucleic Acids Res 54, D1779–D1792 (2026). doi: 10.1093/nar/gkaf1292
31. Z. Gu, R. Eils, M. Schlesner, Complex heatmaps reveal patterns and correlations in multidimensional genomic data. Bioinformatics 32, 2847–2849 (2016). doi: 10.1093/bioinformatics/btw313
32. M. Waskom, seaborn: statistical data visualization. J Open Source Softw 6, 3021 (2021). doi: 10.21105/joss.03021
33. H. Wickham, ggplot2: Elegant Graphics for Data Analysis (Springer-Verlag, New York, 2016).
34. C. Ziegenhain, R. Sandberg, BAMboozle removes genetic variation from human sequence data for open data sharing. Nat Commun 12, 6216 (2021). doi: 10.1038/s41467-021-26152-8
35. R. Argelaguet et al., Multi-Omics Factor Analysis—a framework for unsupervised integration of multi-omics data sets. Mol Syst Biol 14, e8124 (2018). doi: 10.15252/msb.20178124