The Tumor Microenvironment transcriptomic Atlas (TMEAtlas)
TMEAtlas is an interactive computational tool from the Longaker Lab that allows users to independently interrogate pan-cancer tumor microenvironment cell atlas data to advance their own research projects.
This site provides interactive visualization of a pan-cancer tumor microenvironment cell atlas with data from 15.4 million cells from 3,381 scRNA-seq tumor samples across 42 solid cancers. It is accompanied by the scTME R package, which annotates new scRNA-seq datasets with the atlas's consensus cell type nomenclature.
How to cite: Januszyk, M., Lu, J.M., Korah, M. et al. Conservation of the tumour microenvironment across organs and species. Nature (2026). https://doi.org/10.1038/s41586-026-11270-4
We welcome feedback via email at januszyk@stanford.edu.
© Stanford University 2026
About
Overview of TMEAtlas
The tumor microenvironment (TME) plays pivotal but conflicting roles in disease progression and therapeutic response. Pan-cancer analysis of cells within the TME, including both the stromal and immune components, are needed to clarify these cell types’ roles given the breadth and diversity of human tumor pathology. In the article 'Conservation of the tumour microenvironment across organs and species' by Januszyk, M., Lu, J., et al., we curate the largest pan-cancer tumor microenvironment cell atlas to date with data from 15.4 million cells, 3,381 samples, and 258 datasets from human and mouse scRNA-seq tumor samples across 42 solid cancers. This TME atlas serves as an interactive web portal for users to independently interrogate these data. Further, this web atlas is accompanied by the scTME software package that will allow users to use our consensus cell type nomenclature to annotate new scRNA-seq datasets (scTME on GitHub).
Conservation of the tumour microenvironment across organs and species
Current understanding of the host tissue response to solid tumours is largely based on investigations of tumour types in isolation. Given the breadth and diversity of human tumour pathology and contributions of tumour microenvironment (TME) cells to disease progression, pan-cancer analysis is needed to bridge major themes of stromal and immune cell evolution. To date, however, efforts to compare the TME response across cancers have been limited in scope. Here we conduct a comprehensive pan-cancer analysis by algorithmically screening thousands of publicly available single-cell transcriptomic datasets for stromal and immune signatures across species and tissue types. We demonstrate conserved hallmarks of host cell transcriptional programming in response to malignancy across 42 cancer types in more than 2,000 human and mouse tumour samples. These findings are codified as an interactive web atlas (tmeatlas.stanford.edu) and software package (scTME) that will serve as practical resources for the systematic interrogation of the TME in tumours.
How to cite
Januszyk, M., Lu, J.M., Korah, M. et al. Conservation of the tumour microenvironment across organs and species. Nature (2026). https://doi.org/10.1038/s41586-026-11270-4
Human Overview
The Human Atlas is composed of 12.8 million cells from 2,855 samples from 190 scRNA-seq datasets of human solid-organ cancers. Cancer-associated fibroblasts (CAFs) are the predominant non-tumor stromal cell in the TME. Our analysis of 328,877 human CAFs resulted in the delineation of four transcriptionally distinct CAF populations: apCAFs, iCAFs, mCAFs, and ssCAFs. The conservation of CAF populations across tumor types suggests a conserved role for CAFs in solid organ tumors.
© Stanford University 2026
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
Mouse Overview
The Mouse Atlas is composed of 2.5 million cells from 526 samples from 68 scRNA-seq datasets of mouse solid-organ cancers. All tumors come from untreated genetically engineered mouse models (GEMMs) or orthotopic cancer cell line models. Our analysis of 225,361 mouse CAFs showed that, as in humans, they segregate into four transcriptionally distinct CAF populations: apCAFs, iCAFs, mCAFs, and ssCAFs. The conservation of these CAF populations across species supports the use of mouse models to study the human TME.
© Stanford University 2026
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
For speed, these plots show a random subsample of the atlas (about 5,000 cells for most populations). The full datasets are available on the Download page.
Michael Januszyk — 2026-08-17
scTME transfers cell type annotations from a pan-cancer tumor
microenvironment atlas onto a new single-cell dataset. The atlas is
built into the package, so no reference needs to be supplied.
Annotations are available at two resolutions: 7 broad parent
populations, or 45 transcriptionally distinct subpopulations.
This vignette annotates a small public dataset and compares the results against that dataset’s own published labels.
scTME is installed from GitHub:
# install.packages("devtools")
devtools::install_github("michaeljanuszyk/scTME")
## Load Packages
library(Seurat)
library(SingleR)
library(SingleCellExperiment)
library(scTME)
The reference atlas loads from the package itself:
data("scTME_atlas", package = "scTME")
The query is a heavily down-sampled (1,000 cell) subset of the Nieto 2021 immune cell atlas (A single-cell tumor immune atlas for precision oncology), which ships with the package:
query_path <- system.file("extdata", "nieto_2021_downsampled1k.rds", package = "scTME")
query_dataset <- readRDS(query_path)
The accompanying manuscript uses a larger 10,000-cell subset of the same dataset, which gives higher fidelity than the 1,000-cell subset used here, at the cost of runtime. The full Nieto 2021 dataset is linked on the Download page of https://tmeatlas.stanford.edu.
Note that this is a Seurat object. SingleCellExperiment objects can also
be passed to scTME.
names(query_dataset@meta.data)
## [1] "orig.ident" "nCount_originalexp" "nFeature_originalexp"
## [4] "metaNames" "cancer" "tissue"
## [7] "lv1_annot" "lv2_annot" "cluster"
The Nieto dataset carries two levels of its own annotations (lv1_annot
and lv2_annot), which are the final assignments made in that study. We
have also added a cluster column, which indexes the sorted order of
the lv2_annot values. Using it as the grouping variable means the
cluster-level examples below test label transfer given correct
groupings, rather than testing de novo clustering; the annotation itself
is still made without reference to the published labels.
table(query_dataset$lv1_annot)
##
## B cells B cells proliferative CD4 effector memory
## 75 3 69
## CD4 naive-memory CD4 recently activated CD4 transitional memory
## 93 45 14
## CD8 cytotoxic CD8 effector memory CD8 pre-exhausted
## 131 21 15
## CD8 terminally exhausted cDC Macro. and mono. prolif.
## 75 4 3
## Macrophages SPP1 Mast cells mDC
## 17 5 19
## Monocytes NK pDC
## 25 45 5
## Plasma B cells T cells naive T cells proliferative
## 21 81 41
## T cells regulatory T helper cells TAMs C1QC
## 65 33 72
## TAMs proinflamatory
## 23
table(query_dataset$lv2_annot)
##
## B cells activated B cells memory
## 5 35
## B cells naive B cells proliferative
## 3 3
## B cells unswitched memory CD4 activated
## 31 26
## CD4 effector memory RA CD4 memory stem cells
## 64 165
## CD4 resident effector memory CD4 transitional memory
## 67 27
## CD8 activated CD8 cytotoxic
## 17 104
## CD8 exhausted CD8 IFN activated
## 98 1
## DC1 DC2 CD1C+
## 1 21
## DC3 LAMP3 DC4 CD1C-
## 3 8
## Macro. and mono. prolif. Macrophages CXCL10
## 5 1
## Macrophages SPP1 Mast cells
## 13 1
## Monocytes tumor infiltrating NK
## 31 66
## pDC Plasma B cells
## 5 20
## T cells proliferative T cells regulatory
## 40 54
## T helper/Th17 TAMs C1QC
## 6 79
table(query_dataset$cluster)
##
## 1 10 11 12 13 14 15 16 17 18 19 2 20 21 22 23 24 25 26 28
## 5 27 17 104 98 1 1 21 3 8 5 35 1 13 1 31 66 5 20 40
## 29 3 30 31 4 5 6 7 8 9
## 54 3 6 79 3 31 26 64 165 67
Now we can go ahead and run scTME. This function accepts up to 7
parameters:
obj: The Seurat or SingleCellExperiment query dataset on which
to run the predictions.ref: (Default: NULL) The reference object powering scTME
(scTME_atlas). If left as NULL, the package’s built-in reference
atlas will load automatically.clusters: (Default: NULL) If NULL, the function evaluates
each cell independently. Alternatively, provide a character array or
column name labeling the groups/clusters in the query dataset (e.g.,
query_dataset$cluster or "cluster") to run faster, cluster-level
predictions.level: (Default: 2) Evaluates the query against all 45
high-resolution subpopulations. It can optionally be set to 1 to
evaluate against the 7 broader parent populations.label.col: (Default: NULL) The specific metadata column in the
reference object containing the cell type labels. If NULL, the
function automatically selects the correct column based on your chosen
level.max.cells: (Default: 50000) To improve runtime, this value
automatically downsamples the query dataset if it exceeds this
threshold.return.details: (Default: FALSE) By default, scTME returns a
vector of the most likely cell type predictions. If set to TRUE, the
program returns a full list object containing additional prediction
strength and confidence metrics.Note that the query downsampling controlled by max.cells is random.
Call set.seed() beforehand if you need a reproducible result.
res = scTME( obj = query_dataset, ref = NULL, clusters = query_dataset$cluster, level=1, return.details = F )
## No reference provided. Loading built-in scTME atlas...
## Log-normalizing reference object...
## Warning: Layer 'scale.data' is empty
query_dataset$prediction <- unname(res)
We can now compare these predictions against those originally assigned
by Nieto. scTME_generateHeatmap() cross-tabulates the two, coloring by
row proportion and printing the raw cell counts.
scTME_generateHeatmap(query_dataset,
pred_col = "prediction",
annot_col = "lv1_annot")
Cluster-level agreement between scTME broad-population predictions and the published labels
Level 2 gives more precise predictions for each query_dataset$cluster.
res = scTME( query_dataset, ref=NULL, clusters = query_dataset$cluster, level=2, return.details = F )
## No reference provided. Loading built-in scTME atlas...
## Log-normalizing reference object...
## Warning: Layer 'scale.data' is empty
query_dataset$prediction = unname(res)
scTME_generateHeatmap(query_dataset,
pred_col = "prediction",
annot_col = "lv2_annot")
Cluster-level agreement for the 45 subpopulations
Setting clusters = NULL annotates each cell independently. This is
slower and noisier than the cluster-level call above, but requires no
prior grouping of the query dataset.
res = scTME( query_dataset, ref=NULL, clusters = NULL, level=2, return.details = F )
## No reference provided. Loading built-in scTME atlas...
## Log-normalizing reference object...
## Warning: Layer 'scale.data' is empty
query_dataset$prediction = unname(res)
scTME_generateHeatmap(query_dataset,
pred_col = "prediction",
annot_col = "lv2_annot")
Cell-level agreement for the 45 subpopulations
To see how the predicted cell types map onto the transcriptomic
landscape, we can visualize the cells in reduced dimensional space.
Here, we project both the original high-resolution annotations and the
new scTME predictions onto the UMAP embeddings using our custom
plotting wrapper.
# Plot the original annotations using the safe wrapper
plot_true <- plotPredictionUmap(query_dataset, reduction = "umap", pred_col = "lv2_annot")
# Plot the scTME predictions using the safe wrapper
plot_pred <- plotPredictionUmap(query_dataset, reduction = "umap", pred_col = "prediction")
# Only attempt to modify and combine the plots if the UMAP actually exists
if (!is.null(plot_true) && !is.null(plot_pred) &&
requireNamespace("patchwork", quietly = TRUE)) {
library(patchwork)
# Override the default titles and remove legends for clean patchwork display
plot_true <- plot_true + ggplot2::ggtitle("True Annotations") + Seurat::NoLegend()
plot_pred <- plot_pred + ggplot2::ggtitle("scTME Predictions") + Seurat::NoLegend()
# Display side-by-side
plot_true + plot_pred
} else {
message("Skipping UMAP visualization because the 'umap' reduction is missing from the object.")
}
UMAP of the query dataset colored by published labels (left) and scTME predictions (right)
The scTME reference atlas is included with the R package and loads automatically when ref = NULL. No separate reference download is needed.
12.86 million cells; about 27 GB to download. Raw counts in split archive parts, cell metadata, and Harmony/UMAP embeddings.
Open human atlas files on ZenodoHuman file list and reassembly instructions
2.55 million cells; about 6.6 GB to download. Compressed raw counts, cell metadata, and Harmony/UMAP embeddings.
Open mouse atlas files on ZenodoAccess the original Nieto et al. 2021 dataset on Zenodo.
Open Nieto 2021 dataset on Zenodo