Pan-cancer classification of single cells in the tumour microenvironment

Abstract

Single-cell RNA sequencing reveals 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 >250,000 cancer, immune, and stromal cells defining a pan-cancer reference across 19 common cancer types and employed a novel hierarchical approach, outperforming current classification methods. We extensively confirmed scATOMIC's accuracy on 198 tumour biopsies and 54 blood samples encompassing >420,000 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 analyze multicellular cancer TMEs.

Census: accurate, automated, deep, fast, and hierarchical scRNA-seq cell-type annotation

“…Initial benchmarking compared Census to four other state-of-the-art automated annotation methods: scType 4 , scATOMIC 5 , scibet 6 , and scCATCH 7 . For scType, the primary tissue type as well as ‘immune system’ were chosen for as the tissue type.…”

“…As such, a number of automated cell identification methods have been developed [4][5][6][7][8][9][10][11][12][13] , and cell types in many organ-types have been annotated 14,15 . However, when applied to complex tissues, in practice, most suffer from a number of limitations [16][17][18] .…”

“…Label similarity scores to assess closeness of a predicted label to the original author annotated label were calculated as follows. First, cell-type labels from the original studies and the predicted labels from scType 4 , scATOMIC 5 , scibet 6 , and scCATCH 7 were harmonized to the cell ontology annotations used in the Tabula Sapiens 14 using the closest matching label. Then using the cell-type hierarchy created from the Tabula Sapiens, the label similarity score was calculated for each cell-type prediction as the percent of shared nodes of the shorter of the lineages of either the author annotated label or the predicted cell-label.…”

“…While this approach is suitable for major, well-defined cell-types, it can be challenging to annotate cells from noisy datasets or to identify cell-subtypes for which marker genes are overlapping, poorly expressed, or incompletely described 2,3 . This problem is especially pertinent while analyzing scRNA-seq data from perturbation experiments, disease contexts such as cancer, or treatment conditions.As such, a number of automated cell identification methods have been developed [4][5][6][7][8][9][10][11][12][13] , and cell types in many organ-types have been annotated 14,15 . However, when applied to complex tissues, in practice, most suffer from a number of limitations [16][17][18] .…”

“…We first benchmarked Census against four other state-of-the-art automated annotation methods (scType 4 , scATOMIC 5 , scibet 6 , scCATCH 7 ) using a pancreatic cancer dataset 22 that included 57,530 normal and malignant epithelial, stromal, and immune cells. Annotation performance was evaluated by five metrics: F1 score, balanced accuracy, total accuracy, run time, and "label similarity" scores that we computed using our predefined cell hierarchy to quantify closeness of the predicted label to the study's original annotation (see Methods).…”