Pan-cancer classification of single cells in the tumour microenvironment

Abstract

Single-cell RNA sequencing can reveal valuable insights into cellular heterogeneity within tumour microenvironments (TMEs), paving the way for a deep understanding of cellular mechanisms contributing to cancer. However, high heterogeneity among the same cancer types and low transcriptomic variation in immune cell subsets present challenges for accurate, high-resolution confirmation of cells' identities. Here we present scATOMIC; a modular annotation tool for malignant and non-malignant cells. We trained scATOMIC on >300,000 cancer, immune, and stromal cells defining a pan-cancer reference across 19 common cancers and employ a hierarchical approach, outperforming current classification methods. We extensively confirm scATOMIC's accuracy on 225 tumour biopsies encompassing >350,000 cancer and a variety of TME cells. Lastly, we demonstrate scATOMIC's practical significance to accurately subset breast cancers into clinically relevant subtypes and predict tumours' primary origin across metastatic cancers. Our approach represents a broadly applicable strategy to analyse multicellular cancer TMEs.

Fig. 1. Overview of scATOMIC training and classification.

a Hierarchical structure of the pan-cancer tumour microenvironment. The cellular hierarchies in the pan-cancer tumour microenvironment are organized into a flow chart with increasing cell type resolution. Parent nodes represent broad classification branches, and terminal nodes represent specialised cell classes of interest. b Training of classification branches for each parent node (n = 24). The reference datasets are filtered based on transcriptomic-independent information to only include terminal cell types that are found within a particular parental node. Genes that significantly differentiate one cell type from all the others are gathered. Differentially expressed genes (DEGs) with greater specificity to each terminal class, determined by differential expression score (DES), are kept (Methods). A random forest classifier is trained on filtered, library size normalised count matrices to derive a model that provides prediction scores corresponding to the proportion of trees voting for each terminal class within the parental node. Colours on the top of the heatmap illustrate different cell types. c–f Classification of query datasets. c Gene expression count matrices from query tumour biopsies are inputted into the first scATOMIC classification branch model, outputting a cell-by-prediction scores matrix. d Prediction scores (PS) from all blood and non-blood cell subtypes are respectively summed to derive intermediate group score (IGS) distributions associating single cells with their appropriate parental class. e Cells are iteratively interrogated at their next parent nodes’ corresponding models until terminal classification are obtained. Broad classifications occur if the IGS for a cell is lower than the confidence cut-off. In this example, cell 10 is subclassified until a terminal B cell designation is derived. f Differentiating between cancer and tissue-specific non-malignant cells through scoring of bulk RNA-seq derived differentiating gene expression programs (Methods). scATOMIC automatically annotates population 2 as cancer cells, and population 1 as non-malignant. Heatmaps and cell illustrations were created with BioRender.com.

Fig. 2. scATOMIC performs accurately in internal and external validation experiments.

a k-fold cross validation. The reference dataset was randomly split into 5 sub-samples containing equal numbers of each cell type. F1 scores are shown for each cell type in each of the 5 replicates (jitter points). Each fold contained overall ~61,100 cells. Boxplot colours represent the major cell type classes. b External validation in datasets not used for training. scATOMIC was validated on CITE-seq datasets of tumour derived blood cells, datasets of aneuploid cancer cells and stromal cells from primary tumour biopsies. F1 scores are shown for each cell type within individual samples (jitter points). The plot represents n = 357,526 cells from 225 samples. Red dots indicate low-confidence cell type classifications (Methods). Boxplot colours represent the major cell type classes. c scATOMIC outperforms other existing automatic cell type annotators, particularly when applied to identify cancer cells and determine their type. Six existing classifiers were provided the same training/reference and training-independent validation datasets as scATOMIC. Combined F1 scores for each of the three major cell class, blood, cancer, and stroma are shown (jitter points). The plot represents n = 337,790 cells from 221 samples that were given a classification output by all tools. (two-sided Wilcoxon rank sum test comparing scATOMIC to each tool *P < 0.05, **P < 1.1 × 10⁻⁶, ***P < 2 × 10⁻¹⁶, are shown). Boxplot colours represent the different tools. For all plots, boxes and whiskers represent the lower fence, first quartile (Q1), median (Q2), third quartile (Q3), and upper fence. Source data are provided as a Source Data file.

Fig. 3. scATOMIC effectively distinguishes between malignant cells and normal tissue specific cells.

a scATOMIC predictions and inferred ploidy in breast cancer patient CID4066. Cells are coloured by scATOMIC predictions and copy number variation (CNV)-based inferred ploidy. scATOMIC-predicted malignant cells are inferred as aneuploid cells while normal tissue cells are inferred as diploid. b Comparison of scATOMIC cancer predictions and inferred ploidy statues across the training-independent, external validation datasets. Blue bars represent the number of cells predicted as malignant (solid blue) and non-malignant (transparent blue) by scATOMIC. Red bars represent the number of cells inferred as aneuploid (solid red) and diploid (transparent red). Green bars represent agreement rate in each biopsy. Rates do not include cells without a confident ploidy status (that is received an “NA” annotation by CopyKAT). Source data are provided as a Source Data file.

Fig. 4. scATOMIC provides greater cellular resolution than original annotations across tumour datasets.

a Sankey plot comparing original cell type annotations to higher resolution scATOMIC annotations in a recent lung cancer biopsy dataset. scATOMIC identifies lung cancer as the tissue of origin and distinguishes these cells from normal lung tissue cells. scATOMIC identifies subtypes of blood cells. b–g scATOMIC identifies the tumour origin of common cancers and deliver relatively higher resolution in other cell types^,,–. Colours represent the original reported annotations associated with each dataset. The height of each block represents the relative number of cells that received a respective annotation. Source data are provided as a Source Data file.

Fig. 5. Extending the core scATOMIC model to further classify breast cancer subtypes.

a The terminal breast cancer cell node from the core hierarchy of scATOMIC is extended to subclassify breast cancers into their major ER+, HER2+, and triple negative histological subtypes. b Validation of scATOMIC predictions in an external cohort. Pie charts reflecting intra-tumoural breast subtype heterogeneity according to scATOMIC classification are shown for each reported histological subtype. Patient specimens with similar distributions of cell annotations are illustrated together in a single pie chart. c Breast cells from an ER-low tumour (Patient: ER-AH0319) are visualised on UMAP and coloured by scATOMIC predictions. ER + breast cancer cells represent a sub-clonal cancer cell population. d Inferred copy number variation (CNV) profiles of cells from ER-low tumour. Red represents inferred gains, while blue represents inferred losses of genomic regions. The y axis is coloured according to scATOMIC prediction. Colours representing scATOMIC predictions apply to all the panels in the figure. Source data are provided as a Source Data file.

Fig. 6. scATOMIC accurately identifies the tissue of origin in metastatic tumour biopsies.

scATOMIC was applied to 62 metastatic tumours from breast, kidney, lung, ovarian and skin. Metastatic sites included the brain, lungs, GI tract, liver, adrenal glands, lymph nodes, abdomen, and peritoneal cavity. Each pair of dots represents the true tumour origin and the predicted origin. Horizontal connected lines represent correct predictions, while diagonal lines represent incorrect predictions. True tumour origins are coloured by the reported cancer subtype. Circular points represent confident annotations, while triangular points represent low-confidence annotations (Methods). Multi-coloured points represent tumours that received an intermediate scATOMIC annotation. Source data are provided as a Source Data file.