Reference-free deconvolution of complex samples based on cross-cell-type differential analysis: Systematic evaluations with various feature selection options.

Introduction: Genomic and epigenomic data from complex samples reflect the average level of multiple cell types. However, differences in cell compositions can introduce bias into many relevant analyses. Consequently, the accurate estimation of cell compositions has been regarded as an important initial step in the analysis of complex samples. A large number of computational methods have been developed for estimating cell compositions; however, their applications are limited due to the absence of reference or prior information. As a result, reference-free deconvolution has the potential to be widely applied due to its flexibility. A previous study emphasized the importance of feature selection for improving estimation accuracy in reference-free deconvolution. Methods: In this paper, we systematically evaluated five feature selection options and developed an optimal feature-selection-based reference-free deconvolution method. Our proposal iteratively searches for cell-type-specific (CTS) features by integrating cross-cell-type differential analysis between one cell type and the other cell types, as well as between two cell types and the other cell types, and performs composition estimation. Results and discussion: Comprehensive simulation studies and analyses of seven real datasets show the excellent performance of the proposed method. The proposed method, that is, reference-free deconvolution based on cross-cell-type differential (RFdecd), is implemented as an R package at https://github.com/wwzhang-study/RFdecd.