If you’re using bulk data from The Cancer Genome Atlas or similar large databases and looking for a way to tease apart specific cell types and tumor profiles, you’ll be interested in this new study.

In this DREAM Challenge study, sponsored by NCI’s Cancer System Biology Consortium, the authors summarize 28 bioinformatics and data science methods for deciphering information from bulk gene expression data.

The authors found no single method was best on all cell types and no approach particularly dominated the field. However, the researchers were able to benchmark top-performing methods for specific situations (i.e., such as looking at a certain cell type).

The team noted that a recently developed machine-learning approach (called “Aginome-XMU”) was most accurate in predicting fractions of some cell types, suggesting it’s worth pursuing deep learning motivated methods for deconvolution.

In short, when looking to decipher bulk expression data, the authors recommend you select a method that best addresses the problem you’re trying to solve—tailored to the type of cell you’re examining and the context.

Corresponding author, Dr. Andrew Gentles, of Stanford University, said, “Deconvolving bulk expression data is vital for cancer research, but the various approaches haven’t been well benchmarked. Our results should help researchers select a method that will work best for a particular cell type, or, alternatively, to see the limitations of these methods. Perhaps more importantly, we provide a useful benchmark for testing deconvolution methods.”

He added, “Deconvolution gives us access to information that can help us learn more about the tumor microenvironment, enabling researchers to find actionable therapeutic targets, such as immune checkpoints, which are having spectacular clinical outcomes for some cancer patients.”