Fig. 5

Application of f-scLVM to large-scale scRNA-seq datasets. a The empirical runtime when applying f-scLVM and alternative factor models (RUV, SVA, scLVM, PAGODA) to datasets with increasing size. f-scLVM scales linearly in the number of cells, enabling its use on large datasets with up to 100,000 cells. None of the existing methods could be applied to the largest dataset. Error bars denote plus or minus one standard deviation. b-d Application of f-scLVM to 49,300 retina cells profiled using Drop-Seq. b Visualization of a subset of 2145 cells using a non-linear t-SNE embedding. Colours correspond to cell types identified in [13]. c Analogous t-SNE embedding as in b, but on residual data (“Methods”; Additional file 2: Figure S8). The analysis on the residual dataset revealed additional substructure between cells, including two sub clusters of astrocytes (C1 and C2). d Genes and factors that were differentially expressed (false discovery rate < 10%) between the newly identified astrocyte clusters highlighted in c. The colour code is consistent with panel c; grey dots denote outlying cells. Box plots show median expression and the first and third quartile, whiskers show 1.5 × the interquartile range above and below the box