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

We generated in vitro and in silico admixtures of tumor, immune, and stromal cells and sequenced them along with the purified populations from which they were derived. We used the admixtures as ground truth in a community-wide DREAM Challenge aimed at providing an objective, unbiased assessment of published deconvolution methods and at promoting development of new approaches. We benchmarked seven published methods and found that they could predict most cell types well, though dissecting functional states of T cells remains difficult: few methods predict memory CD4+ T cells or the memory versus naïve compartments of CD8T+ cells and, those that do, show large performance gaps relative to other cell types. We spurred development of 21 novel methods, some of which were able to reliably infer these cellular functional states. Overall design: RNA-seq profiling of 11 purified immune cell populations, 2 purified stromal cell populations, 2 cancer cell lines, and in vitro admixtures of these purified populations.