Power analysis of cell-type deconvolution methods across tissues

Abstract

Abstract Cell-type deconvolution methods aim to infer cell-type composition and the cell abundances from bulk transcriptomic data. The proliferation of currently developed methods, coupled with the inconsistent results obtained in many cases, highlights the pressing need for guidance in the selection of appropriate methods. Previous proposed tests have primarily been focused on simulated data and have seen limited application to actual datasets. The growing accessibility of systematic single-cell RNA sequencing datasets, often accompanied by bulk RNA sequencing from related or matched samples, makes it possible to benchmark the existing deconvolution methods more objectively. Here, we propose a comprehensive assessment of 29 available deconvolution methods, leveraging single-cell RNA-sequencing data from different tissues. We offer a new comprehensive framework to evaluate deconvolution across a wide range of simulation scenarios and we show that single-cell regression-based deconvolution methods perform well but their performance is highly dependent on the reference selection and the tissue type. We validate deconvolution results on a gold standard bulk PBMC dataset with well known cell-type proportions and suggest a novel methodology for consensus prediction of cell-type proportions for cases when ground truth is not available. Our study also explores the significant impact of various batch effects on deconvolution, including those associated with sample, study, and technology, which have been previously overlooked. The evaluation of cell-type prediction methods is provided in a modularised pipeline for reproducibility (https://github.com/Functional-Genomics/CATD_snakemake). Lastly, we suggest that the Critical Assessment of Transcriptomic Deconvolution (CATD) pipeline can be employed for the efficient, simultaneous deconvolution of hundreds of real bulk samples, utilising various references. We envision it to be used for speeding up the evaluation of newly published methods in the future and for systematic deconvolution of real samples.