Data-driven inference of crosstalk in the tumor microenvironment

Abstract

Signaling between cancer and nonmalignant (stromal) cells in the tumor microenvironment (TME) is key to tumorigenesis yet challenging to decipher from tumor transcriptomes. Here, we report an unbiased, data-driven approach to deconvolute bulk tumor transcriptomes and predict crosstalk between ligands and receptors on cancer and stromal cells in the TME of 20 solid tumor types. Our approach recovers known transcriptional hallmarks of cancer and stromal cells and is concordant with single-cell and immunohistochemistry data, underlining its robustness. Pan-cancer analysis reveals previously unrecognized features of cancer-stromal crosstalk. We find that autocrine cancer cell cross-talk varied between tissues but often converged on known cancer signaling pathways. In contrast, many stromal cross-talk interactions were highly conserved across tumor types. Interestingly, the immune checkpoint ligand PD-L1 was overexpressed in stromal rather than cancer cells across all tumor types. Moreover, we predicted and experimentally validated aberrant ligand and receptor expression in cancer cells of basal and luminal breast cancer, respectively. Collectively, our findings validate a data-driven method for tumor transcriptome deconvolution and establishes a new resource for hypothesis generation and downstream functional interrogation of the TME in tumorigenesis and disease progression.

A pan-cancer metabolic atlas of the tumor microenvironment

“…TCGA tumor copy number variation data was also downloaded from Xena. Tumor purity (cancer cell fraction) estimates for all TCGA samples were obtained from a previous study (Ghoshdastider et al 2019), where purity was jointly estimated using mutation allele frequencies, copy number profiles, and RNA signatures of each tumor sample.…”

“…In order to estimate the cancer and stromal cell gene expression across tumor types, we performed tumor transcriptome deconvolution as previously reported (Ghoshdastider et al 2019). Briefly, tumor purity was estimated using genomic and RNA data available for each TCGA tumor sample and cancer and stromal compartment expression was estimated using non-negative least squares regression.…”

“…A genome-scale model for human metabolism, RECON3 (Brunk et al 2018), was used to filter and reduce the dataset to 1845 metabolic genes. We used a tumor transcriptome deconvolution approach to estimate expression of metabolic genes in cancer and stromal (non-cancer) cells (Ghoshdastider et al 2019). Briefly, we estimated tumor purity (cancer cell fraction) using genomic and transcriptomic data from each tumor sample, followed by non-negative linear regression to infer the expression of individual metabolic genes in cancer and stromal cells of each tumor type ( Fig.…”

“…Similarly, it is not clear to what extent the metabolic programs of cancer cells might change as they are removed from their natural milieu and cultured in the lab. To address these questions, we used a data-driven tumor transcriptome deconvolution approach (Ghoshdastider et al 2019) to profile and compare cancer and stromal cell metabolic pathways across 7865 tumors and 20 solid tumor types. Using this approach, we explored the distinct metabolic signatures of cancer and stromal cells conserved across most solid tumor types.…”

Toward Systems Biomarkers of Response to Immune Checkpoint Blockers

“…These tools could be exploited to derive biomarkers for immunotherapy by studying the functional effect of cell-cell communication. In a recent study, a curated database of ligand-receptor interactions ( 79 ) was integrated with gene expression data to deconvolute the transcriptional profile of cancer and stromal cells and infer cross-talks in the TME ( 83 ). Interestingly, the authors show that for different cancer types, PD-L1 expression is higher on cancer or stromal cells which nicely correlates with the general responsiveness to immunotherapy.…”