Skip to main content
Back to timeline
Bulletin of mathematical biologySource publication:

When single-cell references don't match, iterative GSNMF still deconvolves bulk tissue, with median cell-proportion error below 0.05 in malaria simulations

Synopsis

Estimating cell-type proportions and cell-type-specific expression profiles together from bulk RNA-seq, without reference profiles, is an ill-posed problem because nonnegative matrix factorization has non-unique solutions. GSNMF+ had stabilized the solve with pseudo-bulk samples built from single-cell data, but its accuracy dropped sharply when the reference single-cell data disagreed with the expression profiles inside the bulk tissue. Iterative GSNMF solves the inner problem with multiplicative updates and repeatedly corrects the reference profiles in an outer loop, improving proportion estimates on synthetic data and on bulk samples built from real Plasmodium falciparum single-cell data, with median errors below 0.05.

Source-provided article image: Iterative Geometric Structure Guided Nonnegative Matrix Factorization Scheme for Bulk Tissue RNA-seq Data Deconvolution.
PubMed

Interpretation

Deconvolution of bulk tissue RNA-seq can now proceed when the reference single-cell data disagree with the expression profiles inside the target tissue, without first reconciling the two. GSNMF+ used single-cell data as auxiliary samples inside the complete deconvolution, and its accuracy fell significantly when the reference data were inconsistent with the target tissue's expression profiles. On synthetic data with reference-to-target expression scale ratios of 1:5 or the reverse, the original GSNMF+ produced target-sample proportions that completely mismatched the ground truth, while the iterative version gave the closest results without any rescaling.

The method splits one deconvolution into an inner and an outer loop: the inner loop solves as before, and the outer loop uses the solved proportions to back out reference expression profiles closer to the target tissue. The inner loop switches to multiplicative updates, an entry-wise rule that adjusts the two nonnegative matrices step by step, replacing the earlier alternating direction method of multipliers, and a proof shows this update converges to a stationary point. On the geometric-structure-constrained nonnegative matrix factorization problem, the alternating direction method of multipliers needed fewer iteration steps for a given error criterion but took longer overall; the outer loop is explicitly described as an empirical, qualitative step with no convergence guarantee.

Perspective

Validation covers computer-generated synthetic bulk data and pseudo-bulk samples built from two independent Plasmodium falciparum single-cell datasets, Dogga et al. and Howick et al., spanning ring, trophozoite and schizont stages and restricted to genes shared by both datasets. Real bulk tissue samples lack known cell proportions, so the current evaluation rests on simulation.

The outer loop is described as an empirical, qualitative procedure with no theoretical support for computational convergence, and the final result may not converge to the true expression-profile matrix. In the P. falciparum simulation the method reproduced the overall picture of more than 90% ring-stage cells with trophozoite and schizont each below 5%, but did not capture fluctuations in some samples where rings fell to 0.6 and the other two stages rose to 0.2; BayesPrism reached an overall Pearson correlation of 0.96 with the ground truth yet missed that overall composition.