Benchmarking reference-based cellular deconvolution algorithms to predict cell proportions

Computational cellular deconvolution enables researchers to estimate the proportions of distinct cell types in bulk RNA-sequencing (RNA-seq) samples using single-cell RNA-seq (scRNA-seq) as a reference. This offers a scalable and cost-effective alternative to physical cell separation. Despite recent advancements in cellular deconvolution methods, rigorous comparative benchmarking across diverse conditions and independent datasets remains limited. Here we present a systematic benchmarking study of ten reference-based deconvolution algorithms: MuSiC, DWLS, BayesPrism, CIBERSORTx, SCDC, BisqueRNA, DISSECT, TAPE, Scaden, and scpDeconv. The algorithms were selected based on architectural diversity and popularity in the field. We evaluate these methods through six experiments using two independent datasets. Under baseline conditions, MuSiC, DWLS, and BayesPrism achieved the strongest performance (mean per-sample Pearson correlation coefficient r>0.95; Lin's concordance correlation coefficient CCC >0.95), while deep learning methods showed greater variability. Depth robustness experiments revealed that SCDC and DWLS were most stable across four sequencing-depth levels. Reference mismatch analysis showed that restricting the scRNA-seq reference to a single developmental stage substantially reduced average performance, with mean Pearson r across stage-restricted references ranging from 0.17 to 0.58 across methods; BayesPrism showed the highest average robustness. Overall, these results provide practical guidance for selecting deconvolution methods under different conditions. The code used to run the experiments on each algorithm is publicly available on GitHub at https://github.com/compbiolabucf/Benchmarking-Deconvolution-Algorithms.