Extended Data Fig. 2: Topic modelling of scRNA-seq time course.
From: Skin-resident innate lymphoid cells converge on a pathogenic effector state
a, Localized expression of ILC-associated genes. FDL embedding (as in Fig. 1e) with cell profiles (dots), coloured by normalized expression (log-transformed scTransform-corrected counts) of selected genes. b, LDA topic modelling schematic in the context of single-cell gene expression. Topics (top) consist of genes (middle), weighted (bar height, colour gradient) according to their importance in the topic. Cell transcription profiles (bottom rows) are characterized by the contribution (bar length) of each topic (colour); a cell can have multiple topics. c, In our analysis, K = 17 is optimal according to the Bayesian information criterion (BIC). Plot of BIC for selected choices for the number K of topics. d–f, Additional plots for topics 1–3, as in Fig. 1f–k. d, For ‘quiescent-like’ topic 1, plot (left) of the empirical cumulative distribution functions (eCDF) of topic weights, grouped and coloured by time point and cropped to the overall 95th percentile of the topic weights, and FDL plots (right) of cells, coloured by normalized expression (colour bar; low, grey; high, maroon) of topic-associated genes. Analogous plots are shown for ‘ILC2’ topic 2 (e) and ‘ILC3-like’ topic 3 (f). g–v, LDA model results for topics 4–7. For ‘Il2-high’ topic 4, bar plot of top scoring genes, ranked by a score (logarithmic scale) of how well the gene distinguishes this topic from other topics (g); FDL plot (as in a) of cells, coloured by topic weight (colour bar; low, grey; high, teal) (h) or by normalized expression (colour bar; low, grey; high, maroon) of topic-associated genes (i); eCDF plots of topic weights (as in d) (j). Analogous plots shown for topics 5 (k–n), 6 (o–r) and 7 (s–v).
