Tumor cells are embedded in a multicellular ecosystem, called the tumor microenvironment (TME), which encompasses different types of tumor-infiltrating immune cells, but also normal epithelial cells, endothelial cells, and fibroblasts. The composition of the TME and the multicellular interactions taking place within it profoundly influence tumor progression and are major determinants of patients’ prognosis and response to therapy (Binnewies et al., 2018, Fridman et al., 2012, Fridman et al., 2017). Thus, it is of utmost importance to quantify the presence, abundance, and spatial organization of the various cell subpopulations in the TME.

Deconvolution is a computational technique that allows estimating the proportions of different TME cell subpopulations from bulk-tumor transcriptomics, including RNA sequencing (RNA-seq) data (Finotello & Trajanoski, 2018). “First-generation” deconvolution methods rely on embedded, well-curated signatures describing the transcriptional fingerprints of the cell types of interest. Derived from transcriptomic data obtained from sorted or purified cell types, these signatures cover only a handful of cell types, mainly human immune-cell subsets (Finotello, Rieder, Hackl, & Trajanoski, 2019).

Already in 2017, Racle and colleagues demonstrated that deconvolution signatures can be derived from annotated single-cell RNA-seq (scRNA-seq) data (Racle, de Jonge, Baumgaertner, Speiser, & Gfeller, 2017). Since then, several deconvolution methods that can be trained with ad hoc scRNA-seq data have been developed. These “second-generation” methods can, in principle, learn the transcriptional signatures of any cell (sub)type which has been profiled with scRNA-seq. In parallel, emerging technologies for spatial transcriptomics allow generating RNA-seq data from spatially-resolved spots across whole-tumor tissue slides (Moffitt et al., 2022, Rao et al., 2021). As these spots can cover tissue regions composed of more than one cell (up to about 10 cells, depending on the tissue type and technology), scRNA-seq-informed deconvolution can be applied to quantify the cellular composition of each spot and reveal the spatial architecture of tumors.

Here, we review first-generation, second-generation, and spatial methods for the quantification of the cellular composition of the TME via deconvolution of transcriptomic data (Table 1). We describe their differences, strengths, and limitations, and report the latest results from independent benchmarking studies. We conclude with an outlook on the potential of these methods, as well as the challenges to be overcome to capitalize on the promise of next-generation deconvolution for oncology.