Human and mouse scRNA-seq meta-analysis
To systematically analyse scRNA-seq cancer studies in the public domain, we evaluated all datasets published from January 2014 (when scRNA-seq was first described69) to July 2025. We only included studies using droplet-based methods, which have emerged as the dominant scRNA-seq approach, to minimize potential technique-specific artefacts when analysing integrated data70. Separate queries were evaluated independently for human and mouse datasets in the two largest public scRNA-seq repositories: the GEO71, managed by the National Institutes of Health (NIH), and ArrayExpress72, run by the European Bioinformatics Institute (EBI). Datasets that were specific to either blood or central nervous system cancers, or that failed to meet specific data architecture requirements (see Fig. 1a), were excluded from our analysis. We also excluded studies that removed stromal cells to ensure that these comparatively understudied cell populations were appropriately represented in the final atlas.
Our final comprehensive human TME atlas comprises 2,855 total samples from 190 scRNA-seq datasets. Our mouse atlas contains 526 total samples from 68 scRNA-seq datasets. Both human and mouse integrated TME atlases were generated (separately) using the following steps:
-
(1)
Comprehensive curation of all publicly available cancer scRNA-seq datasets (Extended Data Fig. 1a).
-
(2)
Processing, filtering and quality control for individual scRNA-seq datasets (Extended Data Fig. 1b).
-
(3)
Disk/memory read optimization using the R package BPCells73,74 (Extended Data Fig. 1c,d).
-
(4)
Sketch-based sampling and integration across datasets using three independent approaches to reduce batch effects: Harmony, FastMNN or scVI75,76,77 (Extended Data Figs. 1e and 2a–f and Supplementary Fig. 1a–d).
-
(5)
Supervised annotation and filtering using a combination of canonical marker genes, estimated copy number variation (inferCNV, v1.14.2)78, SingleR79 and scATOMIC27 into seven broad cell types: epithelial, fibroblast, endothelial, mural, nerve, lymphoid and myeloid (Extended Data Fig. 1f and Supplementary Fig. 2).
-
(6)
Initial pooling of all cells that had been assigned to a given cell type by any one of the three approaches, permitting overlapping annotations (that is, the same cell being initially assigned to more than one cell type; Extended Data Fig. 1g).
-
(7)
Iterative removal of ‘misassigned’ cells through re-normalization, re-finding of variable features, re-scaling of data, re-performing of PCA, re-integration using Harmony (see the section ‘Selection of Harmony for downstream integration’ below for justification of this selection), re-clustering and re-annotation to generate the final ‘pure’ atlases for fibroblast, endothelial, mural, nerve, lymphoid and myeloid cells (Extended Data Fig. 1h), followed by final annotation of each population (Fig. 2d–i, Extended Data Fig. 1i,j and Supplementary Tables 6 and 7).
-
(8)
Generation of the final pan-cancer, all-cells TME human and mouse atlases.
-
(9)
Cross-species integration of human and mouse TME atlases.
These steps are described in detail in the following sections.
Comprehensive curation of all publicly available cancer scRNA-seq datasets
Publicly available human and mouse cancer scRNA-seq datasets were identified by searching two functional genomics databases: GEO and ArrayExpress (Extended Data Fig. 1a). The following search terms were used:
Human GEO: (‘cancer’ OR ‘tumor’ OR ‘tumour’ OR ‘*carcinoma*’) AND ((‘single cell*’ OR ‘single-cell*’ OR ‘scRNA*’ OR ‘sc-RNA*’) OR (‘single nuc*’ OR ‘single-nuc*’ OR ‘snRNA*’ OR ‘sn-RNA*’)) AND 2014/01:2025/07[Publication Date] AND ‘Homo sapiens’[organism] AND ‘expression profiling by high throughput sequencing’[DataSet Type].
Mouse GEO: (‘cancer’ OR ‘tumor’ OR ‘tumour’ OR ‘*carcinoma*’) AND ((‘single cell*’ OR ‘single-cell*’ OR ‘scRNA*’ OR ‘sc-RNA*’) OR (‘single nuc*’ OR ‘single-nuc*’ OR ‘snRNA*’ OR ‘sn-RNA*’)) AND 2014/01:2025/07[Publication Date] AND ‘Mus musculus’[organism] AND ‘expression profiling by high throughput sequencing’[DataSet Type].
ArrayExpress: (‘cancer’ OR ‘tumor’ OR ‘tumour’ OR ‘*carcinoma*’) AND ((‘single cell*’ OR ‘single-cell*’ OR ‘scRNA*’ OR ‘sc-RNA*’) OR (‘single nuc*’ OR ‘single-nuc*’ OR ‘snRNA*’ OR ‘sn-RNA*’)).
A total of 2,168 human and 1,807 mouse datasets were manually screened to identify datasets that met the study inclusion criteria: (1) datasets must contain at least three scRNA-seq samples from primary tumour or non-malignant tissue samples; (2) the scRNA-seq platform must be droplet based (for example, 10X Genomics, Drop-seq, among others); (3) the cancer type must be a non-central nervous system, solid-organ tumour; and (4) the scRNA-seq sample must include at least some stromal cells. The mouse datasets were further screened to meet two additional inclusion criteria: (1) the tumour model must be either a genetically engineered mouse model or an orthotopic cancer cell line mouse model; and (2) the mouse bearing the tumour must not have received treatment. A total of 210 human scRNA-seq datasets and 79 mouse scRNA-seq datasets met these inclusion criteria (Extended Data Fig. 1a).
Data for each dataset were downloaded from their public repositories and converted into base Seurat objects using the functions read10X, read10X_h5, ReadMtx + CreateSeuratObject or read.table + CreateSeuratObject in R (v4.2.2). Each newly created Seurat object (corresponding to a single scRNA-seq sample) was then populated with the relevant sample-level meta-data. This was accomplished by first constructing a single large.csv file containing: the sample name, study name, organ, cancer, site (that is, tumour, tumour-adjacent or non-cancer) and FACS-sorting strategy (if any). For human samples, additional clinical meta-data — including patient age, sex, tumour clinical stage and treatment response — were recorded if available. For all mouse samples, the precise tumour model (for example, MMTV-PyMT) and tumour model type (for example, genetically engineered mouse model) were also recorded. These meta-data were manually collected from the corresponding database entries and articles across all 2,855 human samples and all 526 mouse samples.
Processing and filtering of individual scRNA-seq datasets
Once populated with the above meta-data, samples were then analysed individually for quality control using the Seurat single-cell analysis pipeline (v4.9.9.9050)73. To ensure comparability across datasets, ENSEMBL transcript IDs and gene synonyms were converted to human or mouse gene symbols using biomaRt (v2.56), and only genes found in the human or mouse reference genome GRCh38.p14 or GRCm39 were retained. Afterwards, the raw count matrices for each dataset were confirmed to contain only integer values. Low-quality cells with nFeature_RNA ≤ 200 or nFeature_RNA ≥ 20% mitochondrial reads were then filtered out80 (Extended Data Fig. 1b). To remove potential doublets, data from individual samples were separately exported to Python using the writeH5AD function from the zellkonverter R package, followed by application of the Scrublet function from the Scrublet package81 and exported back to R. Cells ascribed as ‘doublets’ by scrublet using the default parameters were then removed. Finally, all samples from each dataset were merged and ‘flattened’ (using the JoinLayers function) into single Seurat objects for downstream analysis.
Disk read/write optimization and sketch-based sampling
Given the large number of cells (more than 10 million), we used the BPCells package in R to abstract our sparse data matrices using bit-packing compression to minimize disk read/write requirements74 (Extended Data Fig. 1c). All subsequent operations were performed using objects in ‘BPCells format’, with conversion back into memory as dgCMatrix objects only when explicitly required (for example, R functions such as FindAllMarkers, as well as conversion to anndata objects for Python analyses). Conversion into BPCells objects was performed at the per-study level.
Sketch-based sampling and integration across datasets
The human and mouse datasets were separately integrated following the same steps: using Seurat (v4.9.9.9050) on study-level objects, the raw RNA counts were log normalized, and the top 2,000 variable features (highly variable genes) were identified. To improve the computational efficiency of analysing datasets spanning millions of cells, we selected a representative sketch of 1,000 cells per dataset based on the leverage scores calculated for each cell73 (Extended Data Fig. 1d). Using the sketched datasets, variable features were re-identified, expression levels were scaled and PCA was performed (npcs = 30). Next, integration of the sketched data was performed. Given that there is an element of stochasticity to any one method for integration and downstream analyses, which may result in the misclassification of a given individual cell35,79, we initially applied three distinct integration approaches in Seurat: Harmony, scVI and FastMNN (Extended Data Fig. 1e). The batch key for this integration was ‘study’, and no covariate regression was performed for any of the three methods. After each integration approach, UMAP dimensionality reduction was performed (dims = 1:30, min.dist = 0.001), neighbours were found (dims = 1:30) and initial clusters were identified (resolution varied based on dataset). The integration results and clusters from the sketched dataset were then projected onto the full dataset (Supplementary Fig. 2). All steps were performed using Seurat functions (NormalizeData, FindVariableFeatures, SketchData, ScaleData, RunPCA, IntegrateLayers, RunUMAP, FindNeighbours and FindClusters) with default parameters, except as mentioned above.
Supervised annotation
Distinct cell types in each of the Harmony, scVI and FastMNN atlases were identified and grouped into seven broad categories using canonical marker genes in conjunction with SingleR79, scATOMIC27, EnrichR82 and inferCNV78 (Supplementary Fig. 2). The predominant markers for each population were: epithelial cells (EPCAM), fibroblasts (PDGFRA), endothelial cells (PECAM1 (also known as CD31)), mural cells (RGS5), lymphoid cells (CD19 and CD3E), myeloid cells (CD14) and nerve cells (PLP1; Extended Data Fig. 1f).
Initial pooling
To separately generate pan-cancer atlases for human and mouse CAFs, cells identified as fibroblasts by any of the Harmony, scVI or FastMNN algorithms were extracted in silico for further analysis (Extended Data Fig. 1g).
Iterative filtering and final annotation
Using Seurat (v4.9.9.9050) and BPCells74, each cell type (for example, endothelial cells) was split by dataset, the raw RNA counts were log normalized, the top 2,000 variable features (highly variable genes) were identified, expression levels were scaled and PCA (npcs = 30) was performed. Next, integration of the data was performed using Harmony alone (see justification provided in the section ‘Selection of Harmony for downstream integration’). UMAP dimensionality reduction was performed (dims = 1:30, min.dist = 0.001), neighbours were found (dims = 1:30) and clusters were identified. Distinct cell types were annotated using the top 100 marker genes for each cluster, and contaminating non-endothelial clusters were removed. This computational pipeline was iterated four total times to generate the final human pan-cancer endothelial cell atlas and four total times to generate the final mouse pan-cancer endothelial cell atlas (Extended Data Fig. 1h,i). The specific details for the, as an example, human endothelial cell in silico ‘isolation’ are contained within Supplementary Fig. 3. Each subpopulation in the final human and mouse endothelial cell atlases were annotated in a semi-supervised manner using SingleR79, scATOMIC27 and canonical endothelial markers (PECAM1, VWF and CDH5).
The human and mouse pan-cancer fibroblast, mural, nerve, lymphoid and myeloid cell atlases were similarly generated by including cells identified to be the given cell type by any of the three algorithms. The computational pipeline was iterated six, five, four, five and three times to generate the human endothelial, mural, nerve, lymphoid and myeloid cell atlases, respectively. Similarly, the computational pipeline was iterated three, three, two, four and four times to generate the mouse endothelial, mural, nerve, lymphoid and myeloid cell atlases, respectively. The subpopulations in the cell-type atlases were annotated as above (Extended Data Fig. 1j). Downstream analyses were not performed on epithelial cells due to the challenges in distinguishing epithelial from tumour cells in cancers of epithelial origin. It is worth noting that our approach does not explicitly account for the potential presence of ambient RNA within the droplets used to extract transcriptional profiles83. Unfortunately, owing to the format of most publicly available scRNA-seq datasets (lacking ‘unfiltered’ count matrices), robust methods for ambient RNA correction such as CellBender84 or SoupX85 could not be applied.
Generation of the final pan-cancer, all-cells TME atlases
The human cell subpopulation annotations and mouse subpopulation annotations from the six separate cell-type atlases were subsequently mapped back onto the respective all-cells atlases. Cells labelled as epithelial cells in the initial broad categorization in (5) were annotated as epithelial cells. Cells that were filtered out through the iterative cleaning steps were removed. In total, the final human TME atlas had 45 cell populations, and the final mouse TME atlas had 44 cell populations (Fig. 2a and Extended Data Fig. 9a).
Cross-species integration
To evaluate the conservation of cell subpopulations within the TME between humans and mice, the human and mouse cell-type atlases were integrated for each of six primary TME cell types in Seurat (v4.9.9.9050). To perform this integration, orthologue transformation was performed using the convert_mouse_to_human_symbols function in the NicheNet R package (v2.0.4)86 to convert the mouse gene names in the mouse atlases into human orthologue gene names. The six human cell-type atlases were then merged separately with the corresponding mouse cell-type atlas. For each of the combined human–mouse atlases, Seurat (v4.9.9.9050) was used to split the object by dataset, normalize data, find variable features, scale expression levels, perform PCA (npcs = 30), integrate the data using Harmony and perform UMAP dimensionality reduction (dims = 1:30, min.dist = 0.001). To assess human–mouse conservation of cell types, the UMAP results were plotted such that each cell was coloured by the cell-type annotation from the respective human or mouse cell-type atlas (Fig. 5a–f). To assess the quality of this cross-species analysis, we used two approaches. First, we developed a correlation matrix based on co-incidence among underlying clusters (Fig. 5a–f, right); second, we applied anchor-based label transfer to map mouse annotations onto human cells. The latter was performed by separately applying SCTransform in Seurat to both mouse and human objects for each of the principal cell types used in Fig. 5a–f. Then, FindTransferAnchors was applied with dimensions of 1:30, reduction = ‘pca’, followed by TransferData with dimensions of 1:30. No changes to other default parameters were made. Heatmaps of overlap were generated using ggplot2 (Supplementary Fig. 24a–f).
Selection of Harmony for downstream integration
Harmony was selected as the approach for all subsequent downstream integration (beyond the initial first-pass using all three approaches) based on its performance in biological conservation and batch metrics using scIB35 (Supplementary Table 5). For this analysis, Seurat objects integrated by each of the three approaches (Harmony, scVI and FastMNN) were converted into SingleCellExperiment objects using the SingleCellExperiment library. The resulting objects were imported into Python using the anndata2ri package. Individual metrics were then calculated using the scib.metrics function set within scIB.
Downstream analyses using the human pan-cancer TME atlas
The human TME atlas was used to perform six downstream analyses: (1) cell subpopulation enrichment analysis, (2) SCENIC analysis, (3) scMetabolism analysis, (4) TCGA survival analysis, (5) CellChat analysis, and (6) NicheNet analysis. The methods for each of these six analyses are detailed below.
Cell subpopulation enrichment analysis
The relative enrichment of each cell subpopulation in the human scRNA-seq TME atlas was directly calculated for three clinically relevant meta-data categories: between samples from non-malignant tissue versus tumour; between early-stage (I and II) and late-stage (III and IV) tumours; and between treatment responsive and non-responsive tumours. The relative enrichment of each cell subpopulation in the mouse scRNA-seq TME atlas was directly calculated between non-malignant versus tumour tissue. These meta-data were annotated in a consistent manner for each scRNA-seq sample if available in the original article. To calculate relative enrichment for each cell subpopulation, the percentage of cells in each cell subpopulation (for example, ISG15+ iCAFs) out of all cells in the broader cell type (for example, fibroblast) was calculated for each sample. To avoid unstable proportion estimates from sparsely represented samples, any sample contributing 30 or fewer cells to a given parent cell type was excluded from the analysis of the subpopulations of that cell type. This criterion was applied independently for each parent lineage, so the number of samples contributing to a given comparison depends on that lineage and was substantially smaller than the total number of samples in the cohort; within a lineage, the same set of samples contributes to every constituent subpopulation. The exact n for every comparison is reported in the corresponding figure legends and in Supplementary Tables 8 and 13.
Statistical significance for cell subpopulation enrichment was determined using a two-sided weighted Welch t-test for two independent samples (wtd.t.test, weights R package (v1.0.4), with samedata = FALSE), in which the unit of replication was the individual scRNA-seq sample and the tested quantity was the per-sample subpopulation proportion. To account for highly variable single-cell capture rates, sample weights were assigned as the square root of the number of cells per sample. Degrees of freedom were obtained by the Welch–Satterthwaite approximation and are therefore non-integer. Separately, the weighted standard errors of the group means that define the plotted error bars were computed using the Kish effective sample size, neff = (Σw)2/Σw2, which downweights the contribution of samples with disproportionately large cell numbers. Because the cell subpopulations are nested within parent cell types, Benjamini–Hochberg adjustment was applied separately within each parent lineage, so that each family of tests corresponds to the set of subpopulations that partition a single parent cell type; Benjamini–Hochberg adjusted P values computed across all subpopulations simultaneously are additionally reported for completeness. Full statistical outputs for every comparison, including the exact n per group, weighted group means and standard errors, fold enrichment with its 95% confidence interval (obtained by the delta method on the log scale), the t-statistic, the degrees of freedom, and exact raw and adjusted P values, are provided in Supplementary Tables 8 and 13.
SCENIC analysis
SCENIC87 was used to calculate the area under the recovery curve (AUC) for the enrichment of regulons in human scRNA-seq. Seurat objects were generated with AUC scores and clustered by running the standard Seurat pipeline. Differentially expressed regulons were identified by running the FindAllMarkers function and plotted using the standard DimPlot function (Extended Data Fig. 5a–d).
scMetabolism analysis
To evaluate whether specific cell populations may be correlated with differences in metabolic programs, we used the scMetabolism package in R88. Seurat objects for all pooled stromal cell populations (Extended Data Fig. 5e) or individually separate Seurat objects (Extended Data Fig. 5f–h) were evaluated using the sc.metabolism.Seurat function with parameters: method = ‘AUCell’, imputation = F, ncores = 2 and metabolism.type = ‘KEGG’. Using default input pathways of ‘glycolysis/gluconeogenesis’, ‘oxidative phosphorylation’ and ‘citrate cycle (TCA cycle)’, metabolic plots in (Extended Data Fig. 5e–h) were generated using the DotPlot.metabolism function for each population grouping.
TCGA survival analysis
Bulk RNA-seq data from TCGA2 were used to orthogonally calculate the relative enrichment of each cell subpopulation in tumour-adjacent versus tumour samples and to assess the effect of each cell subpopulation on patient survival. TCGA dataset included 9,880 primary tumour and 740 tumour-adjacent samples across 31 solid tumour types, and the dataset was downloaded using TCGAbiolinks (v3.18)2. To deconvolve TCGA bulk RNA-seq data, we first generated a downsampled scRNA-seq reference that contained 1,000 randomly chosen cells for each of the cell subpopulations from the human TME atlas. Using this downsampled scRNA-seq reference, bulk deconvolution using CIBERSORTx with S-mode batch correction was performed for each of the 31 TCGA solid-organ cancer types. The output of bulk deconvolution is an estimate of the percentage of each cell subpopulation in each sample. For each cell subpopulation, the mean and standard error of the cell subpopulation proportions in tumour and tumour-adjacent samples were calculated. The relative enrichment in tumour-adjacent versus tumour samples was computed as the ratio between the two groups (Supplementary Fig. 6a).
The CIBERSORTx cell subpopulation proportion estimates from the tumour samples were additionally used to assess the effect of each cell subpopulation on patient survival. First, for each cell subpopulation, patient tumour samples for each cancer type were stratified into two groups based on the relative proportion of the cell subpopulation. For instance, patients whose tumour samples had a proportion greater than the mean estimated proportion for that subpopulation were considered high-abundance groups, and conversely, patients whose tumour samples had a proportion less than the mean estimated proportion were considered low-abundance groups. Cox regression analysis implemented with the survcomp R package was then used to estimate hazard ratios between high and low expressors for each cancer type by using survival time clinical meta-data from TCGA. A pan-cancer hazard ratio for each cell subpopulation was then estimated by calculating the inverse-variance weighted average of the hazard ratios from each of the 31 cancer types (Supplementary Fig. 6b).
To correct for multiple hypothesis testing, Benjamini–Hochberg adjusted P values were calculated for all comparisons. Full statistical outputs, including fold enrichment or hazard ratio, weighted standard errors and adjusted P values for all clinical group comparisons, are provided in Supplementary Table 9.
CellChat analysis
CellChat (v1.5.0)89 was used to analyse putative cell–cell communication networks among cell subpopulations in our human TME atlas. To perform CellChat, the associated TME Seurat object was first downsampled to only include 100,000 randomly chosen cells, with at least 100 cells per cell subpopulation. The raw counts corresponding to these downsampled data (and their associated meta-data) were extracted from the initial Seurat object and used to populate a new CellChat object. Labels for the non-nerve TME subpopulations were used as groupings. This object was processed through the standard CellChat pipeline, followed by network centrality analysis for pathways of interest (Fig. 4a, Extended Data Fig. 7a,c and Supplementary Fig. 15a).
NicheNet analysis
NicheNet (v2.0.4)86 was used to provide orthogonal support for cell–cell communication networks predicted by CellChat. Seurat objects were downsampled to 100,000 randomly chosen cells (at least 100 per subpopulation), and then fed into NicheNet using sender and receiver populations (as described below) to assess interactions. Regulatory potential heatmaps were generated to visualize these interactions, systematically displaying prioritized sender ligands on the y axis and their regulatory potential against predicted target genes on the x axis. Specifically, this framework was applied to evaluate signalling from: (1) T CD8+GZMB+ cells to ISG15+ iCAFs (IFNγ signalling, Fig. 4b); (2) dendritic cells to ISG15+ iCAFs (MIF signalling; Supplementary Fig. 15b–d); (3) macrophage–monocyte to ESM1+ tip cells (VEGF signalling; Extended Data Fig. 7b); and (4) mural cells to ESM1+ tip cells (Extended Data Fig. 7d).
Histology and immunofluorescence staining
Formalin-fixed, paraffin-embedded (FFPE) tumour and tumour-adjacent tissue blocks were acquired from the Stanford Department of Pathology for seven PDAC, three PTC and three CRC specimens. Immunofluorescence staining was performed on 8-µm histological sections. For immunofluorescence staining, slides were immersed for 10 min in pre-warmed sodium citrate buffer (10 mM, pH 6.0) for antigen retrieval, and slides were blocked for 1 h in Power Block (BioGenex). For PDGFRα–HLA-DR co-localization, slides were incubated overnight at 4 °C with the following primary antibodies: anti-human PDGFRα rabbit monoclonal antibody (3174, CST; clone D1E1E) and anti-human HLA-DR mouse monoclonal antibody (327002, BioLegend; clone LN3). Slides were then incubated for 1 h with the following secondary antibodies: Alexa Fluor 647-conjugated donkey anti-rabbit polyclonal antibody (A-31573, Invitrogen) and Alexa Fluor 555-conjugated donkey anti-mouse polyclonal antibody (A-31570, Invitrogen). Finally, slides were mounted using Fluoromount-G mounting medium with DAPI (00-4959-52, Invitrogen). Fluorescent images at ×20 magnification were acquired with a Leica SP8 inverted confocal microscope.
CODEX and matrix architecture analysis
CODEX/Phenocycler, a highly multiplexed immunofluorescence technique, was previously used to quantify the spatial expression of 30 proteins for 3 PDAC FFPE histological sections by our laboratory90. The CODEX panel assessed expression of several fibroblast markers (αSMA and COL1), an apCAF marker (HLA-DR), immune cell markers (CD11c, CD20, CD68, CD8 and CD4) and epithelial/tumour cell markers (E-cadherin and pan-cytokeratin (CK)). Separately, a fibre quantification algorithm developed by our laboratory was previously used to quantify the matrix architecture from 85 FFPE PDAC specimens91. In brief, 8-µm haematoxylin and eosin (H&E)-stained histological sections were imaged at ×40, and 354,838 200 × 200 µm image tiles were programmatically generated. Afterwards, the histological architecture was quantified for each tile by examining 147 fibre features, and UMAP dimensionality reduction and trajectory analysis using Monocle3 were performed to assign pseudotime values to individual H&E tiles, with high average matrix pseudotime for a given patient correlating with worse patient survival. To examine the matrix architecture in the CODEX sections, 8-µm histological sections consecutive to the CODEX sections were previously stained with H&E, and 200 × 200 µm image tiles were previously analysed using the fibre quantification algorithm. The H&E tiles were spatially aligned with the CODEX sections using GIMP to generate a spatial map of matrix pseudotime for each of the three CODEX sections. In this present study, the CODEX Multiplex Analysis Viewer was used to visualize the multiplex immunofluorescence staining for 11 protein markers in regions of high and low matrix pseudotime in three PDAC CODEX sections. Spindle-like HLA-DR+αSMA+COL1+CD68−CD11c−CD20−CD8−CD4−E-cadherin−pan-CK−CD31− cells were identified by inspection.
Analysis of publicly available Xenium data
To provide further validation for our immunofluorescence and CODEX observations above, we used a publicly Xenium dataset from the 10X Genomics website: Pancreatic Cancer with Xenium Human Multi-Tissue and Cancer Panel (https://www.10xgenomics.com/datasets/pancreatic-cancer-with-xenium-human-multi-tissue-and-cancer-panel-1-standard; dataset listing: https://www.10xgenomics.com/datasets?query=Xenium). The scRNA-seq reference was constructed by selecting cell populations from our initial global populations with our fibroblast subtyping. For deconvolution, the NiCo 1.1.0 pipeline63 was implemented as recommended by filtering and normalizing the reference dataset and Xenium ST datasets, finding commonly shared genes, performing SCTransform from Scanpy, and Leiden clustering the Xenium data at a resolution of 0.7. Then, the find_anchor_cells_between_ref_and_query and nico_based_annotation function was run to find anchors between the reference and Xenium ST datasets with a spatial dispersion cut-off of 0.15. Colocalization of genes (transcripts) and populations (labelled cells) were then used to generate the accompanying figures.
Stimulation of human primary PDAC CAFs and bulk RNA-seq
Primary CAFs were isolated using the Miltenyi Biotec Tumor Dissociation Kit (130-095-929) from one fresh pancreatic PDAC surgical specimen under an institutional review board (IRB)-approved protocol, with informed consent obtained preoperatively under Stanford’s IRB-65893 (ref. 92). The donor was one male patient. No established or immortalized cell lines were used in this study; fibroblast identity follows from the primary outgrowth isolation protocol of ref. 92 and was not separately confirmed by marker analysis. The cultures were not tested for mycoplasma contamination. Isolated CAFs were cultured for four passages in DMEM/F-12 with GlutaMAX (10565018, Gibco), 10% FBS (16000044, Gibco), 1% antibiotic–antimycotic (B22110, R&D) and 0.2% primocin (ant-pm-05, InvivoGen). Passage 4 cells (105 cells per replicate; n = 3 replicates per condition) were allowed to adhere for 12 h, after which cells were treated for 24 h with control media, recombinant human IFNγ (20 ng ml−1; 285-IF, R&D) or LPS (100 ng ml−1; tlrl-eblps, InvivoGen). Cells from a single passage-4 pool were split into nine plates, which were not further passaged, and each plate was then treated and sequenced independently (three plates per condition). RNA was then extracted using TRIzol (15596026, Invitrogen). Nine RNA-seq libraries were prepared with the NEBNext Ultra II Directional RNA Library Prep Kit for Illumina (E7760, NEB) following the manufacturer’s protocol for the NEBNext rRNA Depletion Kit v2 (human/mouse/rat) (E7400L, NEB).
For each library, 50 ng of each replicate was used as input. During the adaptor ligation step, Singular Genomics Universal Library Prep Adapter (#700110) was used to adapt libraries to be sequenced on the Singular Genomics G4 sequencer. PCR enrichment of adaptor-ligated DNA was performed for 12 cycles. No index reads were performed (read 1, 75 bp; read 2, 75 bp; no index cycles), as each library was sequenced in a dedicated lane. All RNA-seq libraries were sequenced on a G4 sequencer instrument for paired-end 2 × 75 bp sequencing, across three sequencing runs performed on the same day, with a single library sequenced per lane (nine lanes in total).
Bulk RNA-seq data were extracted from the sequencer as per-lane fastq files, one library per lane and transferred to the Stanford SCG server. Individual fastq files were aligned to the human GRCh37-hg19 reference genome using TopHat (v2.0.13), resulting in nine counts matrices (n = 3 biological replicates for each of the three experimental groups). These data were read into R using the base read.delimn function, were merged into a single matrix using cbind, and then were annotated according to sample name, replicate and condition. To compute the top differentially expressed genes between each pair of conditions, counts per million (CPM) normalization was first performed on the counts matrices using the cpm function. A gene threshold of at least 2 samples with at least 0.5 CPM was used, resulting in 26,363 genes being preserved for downstream analyses. PCA was performed using prcomp, and the log-transformed PCA data were plotted using autoplot. Differentially expressed genes between groups were determined using DESeq2 (v1.38.4), with its functions calcNormFactors, esimateCommonDisp, estimageGLMTrededDisp, estimatedTagwiseDisp, DESeq and lfcShrink. Volcano plots of the resulting resLFC objects were then generated using EnhancedVolcano (v1.13.2). The biological coefficient of variation was calculated using plotBCV (Supplementary Fig. 16a). EnrichR analyses were performed on the top 200 genes upregulated following treatment with either IFNγ or LPS (Supplementary Fig. 16b,c).
To generate a signature for ISG15+ and CXCL8+ iCAFs, the top 50 genes defining the ISG15+ iCAF and CXCL8+ iCAF subpopulations were first identified using the FindAllMarkers function. To calculate a signature score for each iCAF subpopulation, the expression values of the 50 marker genes in each sample was examined. First, to equally weight the 50 marker genes, row-wise normalization of the counts matrix was performed (for example, Vi,j = Vi,j/Σ(Vi)). Then, to calculate the signature score for each iCAF subpopulation, the normalized expression values for the 50 marker genes were summed for each sample. Standard deviation and standard error were then computed based on the dispersal of cluster signature scores for biological replicates in each condition.
Comparison of our human TME annotations to other large atlases
We examined four large, recent TME atlases published by Nofech-Mozes et al.27, Nieto et al.28, Kang et al.20 and Han et al.21. For each atlas, we downloaded the corresponding full atlas object from its respective repository and converted to SingleCellExperiment objects. Large datasets were downsampled to 50,000 cells. The scTME toolkit was then applied to blindly annotate each cell separately within the query dataset according to cell populations from our TME atlas. The resulting predictive scTME annotations were then compared against the annotations provided for that specific dataset, and heatmaps were generated based on co-incidence.
scTME
Using this large atlas, we developed a new R package, scTME, that utilizes population-level data to provide probability estimates for the annotation of cells within a given cluster in a candidate dataset. This package utilizes many of the classification functions within the SingleR toolkit79, which is a prerequisite for scTME. The program takes a single required input:
-
(1)
A query dataset (obj), supplied as either a Seurat object or a SingleCellExperiment object of genes (g) × cells (c) dimensions. Data that have not been log normalized are normalized automatically.
All remaining inputs are optional.
-
(2)
An optional reference dataset (ref). When omitted, the integrated human TME atlas distributed with the package is used.
-
(3)
An optional vector of cluster labels (clusters), of (1 × c) dimensions, or the name of a metadata column containing them, corresponding to the cell groupings over which the object will be evaluated. When supplied, annotation was performed at the level of these groupings and the assigned label was propagated to each constituent cell; when omitted, each cell was annotated individually.
-
(4)
An optional factor (level) that was set to 2 by default, corresponding to the 45 subpopulations defined in Fig. 2a,d–i. It can alternatively be switched to one, in which case the output will correspond to the seven broad cell types defined in Fig. 2a.
-
(5)
An optional parameter, label.col, naming the column of reference metadata from which labels were drawn, for use when a custom reference is supplied.
-
(6)
An optional threshold, max.cells (default 50,000), above which the data will be randomly subsampled to a final object with the number of cells equal to max.cells
-
(7)
An optional parameter, return.details (default = FALSE), which will output the scoring matrices in addition to the final cell annotations.
The output of the program is:
-
(1)
A (1 × c dimensional) label vector corresponding to the annotations assigned to each cell, based on their original input grouping and the output annotated to each grouping.
-
(2)
A scoring matrix corresponding to the probabilities assessed to each group-level assignment. This is only returned if return.details = TRUE. If so, these two outputs are returned as a list.
Details of the code itself can be found within the GitHub documentation (https://github.com/michaeljanuszyk/scTME) and it can be installed using the devtools library in R with: install_github(‘michaeljanuszyk/scTME’).
Illustration
Original illustrations were drawn using an Apple iPad Pro and Apple Pencil 2.0 in Procreate (v5.2.7) software. Post-editing adjustments of parameters such as brightness, contrast and saturation were applied in Adobe Photoshop (v23.4.2). Finally, text and object overlays were added to the figures using Biorender (https://biorender.com; see relevant figure legends for licenses).
Statistics and reproducibility
No statistical method was used to predetermine sample size. For the meta-analysis, sample size was determined by the number of published scRNA-seq datasets meeting the inclusion criteria defined in Fig. 1a. No data were excluded from the analyses other than by the pre-specified dataset-level and cell-level quality control criteria described above. No randomization was used. For the meta-analysis, all published datasets meeting the inclusion criteria were included, so no allocation step arose. For the in vitro experiments, the replicate cultures were assigned to the control, IFNγ and LPS conditions without randomization and handled in parallel. Unless stated otherwise in the corresponding figure legend, all statistical tests were two-sided and all P values were adjusted for multiple comparisons using the Benjamini–Hochberg procedure. Exact adjusted P values are reported in Supplementary Tables 8–10 and 13.
Cell subpopulation enrichment analyses (Fig. 3a–c and Supplementary Fig. 21) were performed using a two-sided weighted Welch t-test for two independent samples, in which the unit of replication was the individual scRNA-seq sample and sample weights were equal to the square root of the number of cells per sample, with Welch–Satterthwaite degrees of freedom. Within each parent lineage, these analyses were restricted to samples contributing more than 30 cells to that lineage. Sample numbers therefore differ between lineages and are stated per lineage in the legend of Fig. 3 and for every individual comparison in Supplementary Table 8.
Across all lineages, the number of contribution samples are 859–1,777 (Fig. 3a), 311–658 (Fig. 3b) and 48–108 (Fig. 3c) samples per parent cell lineage. Exact values per lineage are given in Supplementary Table 8. No statistic reported in this study is derived from fewer than three biologically independent samples per group. Within a given lineage, the same set of samples contributes to every constituent subpopulation. Benjamini–Hochberg adjustment was applied separately within each parent lineage. No statistics were derived from any comparison with fewer than three biologically independent samples per group. TCGA deconvolution analyses (Supplementary Fig. 6) used n = 9,880 primary tumour and 740 normal samples across 31 solid tumour types, with hazard ratios estimated by Cox regression and pooled across cancer types by inverse-variance weighting.
Immunofluorescence co-staining for PDGFRα and HLA-DR (Fig. 3d) was performed on n = 7 biologically independent PDAC, n = 3 biologically independent CRC and n = 3 biologically independent PTC specimens from patients, together with matched tumour-adjacent tissue. Each specimen was stained and imaged once, and comparable staining patterns were observed across all specimens of a given tumour type; the images shown are representative. CODEX analyses (Fig. 3f–h) were performed on n = 3 biologically independent PDAC specimens, each analysed once within a single multiplexed imaging run, with comparable findings in each; the images shown are representative. Spindle-like HLA-DR+αSMA+COL1+ cells negative for tumour and other immune cell markers were identified by visual inspection. Investigators were not blinded to matrix pseudotime region assignment, which was determined programmatically by the fibre quantification algorithm before inspection.
The in vitro CAF stimulation experiment (Fig. 4d–h and Supplementary Fig. 16) was performed with n = 3 replicate cultures per condition, split from a single primary PDAC CAF isolate and treated and sequenced independently. Differential expression between conditions was assessed using DESeq2 (v1.38.4) with a two-sided Wald test and Benjamini–Hochberg adjustment across the 26,363 genes passing the expression threshold. iCAF signature scores (Fig. 4g,h) are presented as mean ± s.d. of the three replicate cultures, with individual replicate values overlaid, and were compared using two-sided unpaired Student’s t-tests (equal variances assumed). Investigators were not blinded during the in vitro experiments, as all readouts were quantitative and derived computationally from sequencing data.
Ethics statement
FFPE tumour and tumour-adjacent tissue specimens were obtained from the pathology repository of the Stanford University Department of Pathology. No human participants were recruited for the purpose of this study, and all sample collection was performed with the approval of the IRB of the Stanford University School of Medicine. All tissue specimens were coded and de-identified before further processing, in accordance with patient privacy regulations. No animal experiments were performed in this study; all mouse single-cell RNA-seq data analysed here are previously published and were generated under the approvals reported in their original publications. The fresh PDAC surgical specimen used for primary CAF isolation was collected under Stanford IRB protocol IRB-65893, with informed consent obtained preoperatively. All previously published single-cell RNA-seq datasets analysed in this study were obtained from public repositories and were generated under the approvals reported in their original publications.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.