Effective methods for bulk RNA-seq deconvolution using scnRNA-seq transcriptomes

Abstract

Background RNA profiling technologies at single-cell resolutions, including single-cell and single-nuclei RNA sequencing (scRNA-seq and snRNA-seq, scnRNA-seq for short), can help characterize the composition of tissues and reveal cells that influence key functions in both healthy and disease tissues. However, the use of these technologies is operationally challenging because of high costs and stringent sample-collection requirements. Computational deconvolution methods that infer the composition of bulk-profiled samples using scnRNA-seq-characterized cell types can broaden scnRNA-seq applications, but their effectiveness remains controversial. Results We produced the first systematic evaluation of deconvolution methods on datasets with either known or scnRNA-seq-estimated compositions. Our analyses revealed biases that are common to scnRNA-seq 10X Genomics assays and illustrated the importance of accurate and properly controlled data preprocessing and method selection and optimization. Moreover, our results suggested that concurrent RNA-seq and scnRNA-seq profiles can help improve the accuracy of both scnRNA-seq preprocessing and the deconvolution methods that employ them. Indeed, our proposed method, Single-cell RNA Quantity Informed Deconvolution (SQUID), which combines RNA-seq transformation and dampened weighted least-squares deconvolution approaches, consistently outperformed other methods in predicting the composition of cell mixtures and tissue samples. Conclusions We showed that analysis of concurrent RNA-seq and scnRNA-seq profiles with SQUID can produce accurate cell-type abundance estimates and that this accuracy improvement was necessary for identifying outcomes-predictive cancer cell subclones in pediatric acute myeloid leukemia and neuroblastoma datasets. These results suggest that deconvolution accuracy improvements are vital to enabling its applications in the life sciences.

DeMixSC: a deconvolution framework that uses single-cell sequencing plus a small benchmark dataset for improved analysis of cell-type ratios in complex tissue samples

“…In the application to complex retina samples from patients with AMD, DeMixSC identifies subtle yet critical cell-type proportion changes, highlighting its ability to reflect cellular dynamics during disease progressions and facilitate the cell-type-specific gene expression analysis. Most existing single-cell-based deconvolution methods 10–18 are adept at discerning between 7 to 13 cell types from bulk RNA-seq data, DeMixSC aligns with these capabilities as demonstrated in our study. DeMixSC is computationally efficient, completing the analysis of 453 AMD samples in under five minutes, and presents robust convergence against different starting values (see Methods, Extended Data Fig.…”

“…In broader contexts, factors such as library preparation, RNA capture efficiency, reverse transcription protocol, and sequencing depth could serve as potential sources of the technological discrepancy 19, 20, 27 . Thus, we expect the reference matrix derived from sc/snRNA-seq data does not fully represent cell-type-specific expression profiles in bulk samples 10, 22 . Given these discrepancies, performances of existing deconvolution methods will be compromised, as their key assumption about the representative reference is violated.…”

“…Such inconsistencies, which we refer to as “technological discrepancies”, have caused prior deconvolution studies to produce suboptimal estimates of cell-type proportions, particularly when unpaired sc/snRNA-seq data serve as the reference matrix for deconvolving publicly available large bulk cohorts 16, 21, 22 . Several studies have been aware of and attempted to address these issues but ended up with limited success 10, 14, 18 . CIBERSORTx 18 simply implements a batch effect correction step but offers limited improvements in deconvolving complex bulk tissues.…”

“…The ensemble approach of SCDC 14 uses matched bulk and scRNA-seq data from two normal mice breast tissues, which lacks generalizability to patient cohorts. The most recent SQUID 10 builds on top of DWLS 15 with a Bisque-based linear transformation step 23 to align matched bulk and scRNA-seq data, which can distort gene expression profiles, risking overcorrection. Meanwhile, existing benchmarking designs 10, 21, 22, 24 often employ unfavorable datasets such as simulated pseudo-bulk data, cell line mixtures, or publicly available data, none of which are tailored to reveal the negative effect of technological discrepancies.…”

“…The most recent SQUID 10 builds on top of DWLS 15 with a Bisque-based linear transformation step 23 to align matched bulk and scRNA-seq data, which can distort gene expression profiles, risking overcorrection. Meanwhile, existing benchmarking designs 10, 21, 22, 24 often employ unfavorable datasets such as simulated pseudo-bulk data, cell line mixtures, or publicly available data, none of which are tailored to reveal the negative effect of technological discrepancies. Therefore, there is an unmet demand for a comprehensive benchmark study that rigorously illustrates and assesses the impact of technological discrepancies and a robust deconvolution framework that effectively mitigates these discrepancies to enhance analytical accuracy.…”