Fig. 5

Accounting for batch effects in ZINB-WaVE. Upper panels provide two-dimensional representations of the mESC data, with cells color-coded by batch and shape reflecting culture conditions: a Default ZINB- WaVE with only sample-level intercept; b ZINB-WaVE with batch as known sample-level covariate. c Average silhouette widths by biological condition for ZINB-WaVE with and without batch covariate; d average silhouette widths by batch for ZINB-WaVE with and without batch covariate. Although the cells cluster primarily based on culture condition, batch effects are evident in a. Accounting for batch effects in the ZINB-WaVE model (b) leads to slightly better clustering by biological condition (c) and removes the clustering by batch (d). Note the difference in scale between the barplots of c and d. Lower panels provide two-dimensional representations of the glioblastoma data, with cells color-coded by batch: e Default ZINB-WaVE with only sample-level intercept; f ZINB-WaVE with total number of expressed genes as sample-level covariate. Despite the confounding between patient and batch, the addition of a covariate that captures the batch difference allows ZINB-WaVE to remove the batch effect without removing the patient effect