Fig. 1: Flowchart of ELVAR algorithm.

a Given a scRNA-Seq data matrix with cells derived from various conditions (e.g. age-groups), one first derives a cell-cell similarity graph using standard pipelines like Seurat. Cells may also differ in terms of an attribute of interest (e.g. cell state or cell subtype) and the sample replicate it is derived from. b To infer communities from this cell-cell graph, we use an extended Louvain algorithm (EVA) which, unlike the standard Louvain algorithm, takes cell attribute information into account when deriving the communities. In this case, the cell-attribute used in the clustering (the clustering attribute) could be the biological condition it is derived from, in which case the inferred communities will be more enriched for cells of the same condition, as shown. Compared to the standard Louvain algorithm, which aims to maximize the overall modularity Q of the communities, EVA aims to maximize a weighted sum of Q and the overall purity P (a measure of how pure the communities are in relation to the conditions). The a parameter controls the relative importance of Q and P when maximizing the objective function Z. c EVA communities that are significantly enriched for cells from a particular condition are selected for further downstream analysis, thus removing noisy cellular neighborhoods. d For a given condition, cells from all communities enriched for that condition are merged and the distribution of underlying cell-states from each sample replicate are computed. Finally, negative binomial regressions are used to infer if given cell-state fractions (the attribute of interest) vary significantly with condition, whilst taking sampling variability into account.