Adaptive Regularized Tri-Factor Non-Negative Matrix Factorization for Cell Type Deconvolution

Abstract

Accurate deconvolution of cell types from bulk gene expression is a fundamental step in characterizing cellular compositions, shedding light on cell-type specific differential expression and physiological states of diseased tissues. However, existing deconvolution methods either require complete cellular gene expression signatures or are entirely reference-free, neglecting partial biological information. Moreover, these methods often disregard varying cell-type mRNA amounts, resulting in biased proportion estimates. Additionally, they fail to utilize useful reference information from external studies such as population cell-type proportions. To address these challenges, we propose an Adaptive Regularized Tri-factor non-negative matrix factorization method for deconvolution, known as ARTdeConv. We rigorously prove the numerical convergence of the algorithm. In benchmark simulations, we demonstrate that ARTdeConv can surpass the performance of state-of-the-art reference-free methods. In addition, estimated proportions by ARTdeConv show a nearly perfect Pearson’s correlation against flow cytometry measurements in a real dataset obtained from a trivalent influenza vaccine study. An R package for our method is available for access on GitHub.