Penalised regression improves imputation of cell-type specific expression using RNA-seq data from mixed cell populations compared to domain-specific methods

Abstract

RNA-seq studies to infer differential gene expression (DGE) in disease or other conditions often use bulk sequencing of mixed cell populations because single cell or sorted cell sequencing may be prohibitively expensive. However, differential gene expression may be restricted to specific cell populations, meaning that mixed cell studies may lose power. Computational deconvolution can be used to estimate cell fractions from bulk expression data and may also infer average cell-type expression in a set of samples (eg cases or controls), but imputing sample-level cell-type expression is less commonly addressed.In this study, we first assessed the accuracy of cell fractions estimated by three domain-specific methods-CIBERSORTx, bMIND and debCAM/swCAM using a real dataset where mixed peripheral blood mononuclear cells (PBMC) and sorted (CD4, CD8, CD14, CD19) RNA sequencing data as well as flow cytometry data were generated from the same subjects (N=158). Furthermore, we investigated the potential of cross-domain machine learning methods, multiple response LASSO and RIDGE, for imputing sample-level cell-type expression and compared their performance to the three deconvolution methods aforementioned.Estimates of cell fractions by deconvolution methods for CD8, CD14 and CD19 had modest/high correlations r (0.62-0.77) with those from flow data but were less accurate for CD4 (0.43-0.54). All methods appeared to impute sample-level cell-type gene expression well across cell types (median r > 0.85); however, there was a high variation of correlations across subjects per gene. Nevertheless, LASSO/RIDGE exhibited marginally better accuracy than the deconvolution approaches. We considered an alternative measure of accuracy, differential gene expression (DGE) recovery, based on simulating case/control status and comparing detection of DGE between imputed and observed cell type data. We observed higher sensitivity but lower specificity of DGE recovery in LASSO than in deconvolution methods, although overall receiver operating characteristic (ROC) analysis revealed that LASSO/RIDGE had higher area under curves (AUC, median=0.84-0.87 across cell types) than CIBERSORTx (0.62-0.77), bMIND (0.69-0.76) and swCAM (0.64-0.72).We conclude that machine learning methods have the potential to outperform domain-specific methods when suitable training data are available and suggest that further research in this area may optimise machine learning approaches to this problem.