Fig. 2: Benchmarking EVA against Louvain.

a tSNE visualization of a simulated scRNA-Seq dataset consisting of 993 mouse cells drawn from 8 mice encompassing 4 age-groups (1 month, 3 months, 21 months and 30 months) with two mouse replicates per age-group. In the right panel, cells are annotated by perturbation state, where the frequency of cells being in the perturbed (P) state increases from 0.25 in young mice (1 & 3 m) to 0.5 in old mice (21 & 30 m). b Left: Boxplots displaying the ratio of the number of clusters inferred with EVA to the corresponding number inferred with Louvain (nC[EVA]/nC[LV], y-axis) for a range of different purity parameter values (x-axis). A total of n = 100 EVA runs were made at each purity parameter value. Right: Boxplots displaying the modularity (Q), purity (P) and generalized modularity (Z) as a function of purity index parameter a for EVA. Each boxplot contains the values of 100 distinct EVA runs. Boxplot elements indicate median, interquartile range (IQR) and whiskers extend to 1.5 times the IQR. c Nearest neighbor cell-cell graph on which the EVA algorithm is run. Left panel: cells annotated by clusters inferred in one particular EVA run. Middle panel: cells annotated by age-group. Right panel: confusion matrix between the communities inferred with EVA (same run) and age-groups, with the number of cells and one-tailed Binomial test P-value of enrichment shown. Significance is assessed using Bonferroni adjustment at 0.05 level. d Barplot (top panel) displays the number of normal (N) and perturbed (P) cells from each mouse and age-group, using only cells from EVA communities enriched for specific age-groups (same run as in c). Barplot in lower panel displays the ratio N/P for each mouse replicate and age-group. e Violin plot compares the statistical significance (y-axis, -log₁₀P) of P-values from a negative binomial regression of perturbed cell number against age-group as derived from ELVAR (100 runs) against the corresponding statistical significance value derived from an analogous method that uses either the deterministic Louvain algorithm (LVdet) or a non-sequential (non-deterministic, 100 runs) version of Louvain (LVnonseq) in place of EVA. P-values shown are from a one-sided Wilcoxon rank sum test comparing the 100 ELVAR values to the LVdet one, or to the 100 values from LVnonseq. Source data are provided as a Source Data file.