Extended Data Figure 8: Reproducibility of estimated μ, σ2 and α parameters.
From: Single-cell RNA-seq reveals dynamic paracrine control of cellular variation
a–f, Reproducibility of estimated μ, σ2 and α parameters between technical replicates. Scatter plots showing the relation between the estimated α (a), μ (b), and σ2 (c) values for the two unstimulated/LPS 0 h technical replicates. For μ (b) and σ2 (c), estimates are plotted for all genes (farthest on the left), as well as (left to right) for genes only detected in more than 10, 20, 30, 40 or 50 cells, respectively. d–f, show the same plots for the two LPS 4 h technical replicates. g, h, Reproducibility of estimated μ, σ2 and α parameters between biological replicates. Scatter plots showing the relation between the α (g), μ (h), and σ2 (i) values estimates for the two LPS 2 h biological replicates. For μ (h) and σ2 (i), estimates are plotted for all genes (farthest on the left), as well as (left to right) for genes only detected in more than 10, 20, 30, 40 or 50 cells, respectively. j–l, show the same plots for the two LPS 4 h biological replicates. m, n, Relationship between per-gene errors for μ, σ2 and α and the number of cells in which the gene’s expression is detected, or its bulk expression level. Scatter plots displaying the differences in the σ2 (left), μ (middle) and α (right) estimates for each gene between technical replicates for LPS 2 h (m) or LPS 4 h (n) (y axis) as a function of either the number of cells in which the transcript is detected (x axis, for μ and σ2), or the transcript’s bulk expression level (TPM, x axis, for α). Notably, σ2 (left) estimates saturate (denoted by a magenta line and shaded box) after ∼30 detectable events, whereas μ estimates saturate after ∼10. Dashed orange line: 95% confidence interval. o, p, Changes in μ, σ2 and α between time points are significant as compared to null models from both technical and biological replicates. Shown are the cumulative distribution functions (CDF) for shifts in μ (left), σ2 (middle), and α (right) between 2 h and 4 h (red CDF) for the core antiviral (top), peaked inflammatory (middle), and sustained inflammatory (bottom) modules compared to a null model (black CDF) derived from either technical (o) or biological (p) replicates (Supplementary Information). In the vast majority of cases, the changes between time points are significant, as assessed by a Kolmogorov–Smirnof (KS) test (P value of test in the upper left corner of each plot).
