Answers are grounded in this paper and cite their sources.
Complete deconvolution is a class of unsupervised methods that aim to estimate both cell-type proportions and cell-type–specific gene expression profiles (GEPs) from bulk RNA-seq data, without requiring prior reference profiles. Hence, it enables broader applicability to tissues or species for which reference data are limited or unavailable. For many existing methods, nonnegative matrix factorization (NMF) serves as the fundamental computational tool. However, NMF is mathematically ill-posed and a comprehensive understanding of how to overcome the inherent ill-posedness and improve solution accuracy remains lacking. In our previous work, we investigated the necessary requirements for a given bulk RNA-seq dataset to satisfy the solvability conditions in NMF theory. Based on its weak solvability conditions of NMF, a new pipeline of pseudo-bulk tissue data augmented, geometric structure guided NMF model (GSNMF+) has been developed. As a semi-supervised model, it utilizes available single-cell RNA-seq (scRNA-seq) data and has shown obvious improvements in scenarios where the target bulk data contain singular cellular compositions. Nevertheless, when the available scRNA-seq data are not fully consistent with the underlying cell-type-specific GEPs present in the bulk RNA-seq data, it remains a challenge that the deconvolution accuracy will significantly decrease. In this paper, we propose an iterative GSNMF algorithm to address this challenge. First, we propose a new computational algorithm for the GSNMF method (termed as the inner iteration) and provide a theoretical proof of its convergence. We then introduce an iterative process (termed as the outer iteration) to reduce potential discrepancies between available scRNA-seq and the latent cell type specific GEPs in the bulk-tissue data. The proposed methods are evaluated on both synthetic and realistic scRNA-seq data generated bulk-tissue datasets and have shown significant improvements in deconvolution accuracy, particularly in settings where the reference scRNA-seq data are inconsistent.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: