Community assessment of methods to deconvolve cellular composition from bulk gene expression

Abstract

Abstract Deconvolution methods infer levels of immune and stromal infiltration from bulk expression of tumor samples. These methods allow projection of characteristics of the tumor microenvironment, known to affect patient outcome and therapeutic response, onto the millions of bulk transcriptional profiles in public databases, many focused on uniquely valuable and clinically-annotated cohorts. Despite the wide development of such methods, a standardized dataset with ground truth to evaluate their performance has been lacking. We generated and sequenced in vitro and in silico admixtures of tumor, immune, and stromal cells and used them as ground truth in a community-wide DREAM Challenge that provided an objective, unbiased assessment of six widely-used published deconvolution methods and of 22 new analytical approaches developed by international teams. Our results demonstrate that existing methods predict many cell types well, while team-contributed methods highlight the potential to resolve functional states of T cells that were either not covered by published reference signatures or estimated poorly by some published methods. Our assessment and the open-source implementations of top-performing methods will allow researchers to apply the deconvolution approach most appropriate to querying their cell type of interest. Further, our publicly-available admixed and purified expression profiles will be a valuable resource to those developing deconvolution methods, including in non-malignant settings involving immune cells.