Fig. 4
From: C-ziptf: stable tensor factorization for zero-inflated multi-dimensional genomics data
Recovering GEPs from synthetic single-cell RNA-seq data: a UMAP of cells simulated using the Splatter framework, colored by the dominant identity GEP expressed by each cell, b explained variance and gene mode cophenetic correlation of C-ZIPTF factorizations of the simulated data at different ranks, c Silhouette score and inertia of the K-means clustering of gene components resulting from 10 randomly initialized factorization at rank 8, d, e correlation between the ground truth identity (d) and activity (e) GEPs used in the simulation and the corresponding GEPs inferred by C-ZIPTF, f pairwise Pearson correlation between each of the eight latent factors in the gene mode obtained via C-ZIPTF factorization and the original GEPs, g, h the average Pearson correlation between the true GEPs used in the simulation and the inferred GEPs obtained from various factorization methods, results are presented for two different signal intensity levels (0.25 and 0.75), which are indicated by the mean log2 fold change (log2FC) of simulated differentially expressed genes, (g) presents results from simulation done using Splatter, while (h) illustrates results from simulations conducted with scDesign3