ABSTRACT
The tumour microenvironment comprises complex cellular compositions and interactions between cancer, immune, and stromal components which all play crucial roles in cancer. Various computational approaches have been developed during the last decade that estimate the relative abundance of different cell types in an unbiased manner using bulk tumour RNA data. However, a comparison that objectively evaluates the performance of these approaches against one another has not been conducted. Here we benchmarked six widely used tools and gene sets: Bindea et al. gene sets, Davoli et al. gene sets, CIBERSORT, MCP-counter, TIMER, and xCell. We also introduce ConsensusTME, a consensus approach that uses the union of genes that the six tools used for cell estimation, and corrects for tumour-type specificity. We benchmarked the seven tools using TCGA DNA-derived purity scores (33 tumour-types), methylation-derived leukocyte scores (30 tumour-types), and H&E deep learning derived lymphocyte counts (13 tumour types), and individual benchmark data sets (PBMCs and 2 tumour types). Although none of the seven tools outperformed others in every benchmark, ConsensusTME ranked consistently well in all cancer-related benchmarks making it the top performing method overall. Computational methods that provide robust and accurate estimates of non-cancerous cell populations in the tumour microenvironment from tumour bulk expression data are important tools that can advance our understanding of tumour, immune, and stroma interactions, with potential clinical application if high accuracy estimates are achieved.