Fig. 1

A Schematic of the eSVD-DE’s matrix factorization, where the observed scRNA-seq data is modeled as a sum of two low-rank matrices, one for the covariates and one for the cells’ latent vectors, with an exponential link function (for the Poisson distribution). B The cells’ latent vectors can be used for diagnostic checks, such as visualization via Isomap. C To account for overdispersion and over-fitting of the dimension reduction, shrink each cell via the negative binomial distribution’s posterior mean. D Represent each individual by a Gaussian distribution among the individual’s cells. E Compute a test statistic analogous to the T-test after aggregating cells from the cases or control individuals in the cohort. C through E are performed for each gene. F Volcano plot, showing a multiple testing cutoff to determine the significant DE genes