Fig. 2: Schematic of the methodology for estimating latent variables in the scRNAseq atlas and then deconvolving them into the clinical-trial patient sample in order to assess their potential as biomarkers that inform clinical outcomes.
From: q-Diffusion leverages the full dimensionality of gene coexpression in single-cell transcriptomics
The two outcomes investigated were progression-free survival (PFS) and overall survival (OS) in accordance with the clinical trial’s protocol. Differential effects were measured by heterogeneity of multivariate Cox proportional hazards. a Hazard-ratio point estimates for patient biomarkers. Under “Marker Effects,” we compare hazard ratios of biomarkers between cet and bev groups. Under “Drug Effects,” we compare hazard ratios of bev to cet between biomarker groups. We test for significant differential effects between groups. In contrast with the q-diffusion results listed here, the structures uncovered by classical NMF and PHATE failed to produce biomarkers with any significant differential effects. Bold: FDR < 0.1; Bold*: FDR < 0.05; Bold**: FDR < 0.01. b Kaplan–Meier estimates of survivals with 95% confidence illustrating the identified differential marker effects under the two treatments. c The qNMF biomarker appears to help bev overall survival (OS) and hurt cet according to a. A number of member genes in the GEP were individually associated with these differential outcomes, as determined by U-tests with FDR < 0.01. Survivals (90% confidence) are stratified by upper and lower quartiles of expression.
