Extended Data Fig. 10: Methods-related visualizations of the scRNA-seq analysis.

a, b, Axes of variation in gene expression profiles captured by diffusion components (Methods). Diffusion components (DCs) calculated on standardized expression data (Pearson residuals of scTransform-corrected counts) capture axes of variation related to topic modelling. a, For DCs 1–14 (panels), FDL embedding (as in Fig. 1e) of cell profiles (dots), coloured by the coordinate value (colour bar; low, blue; high, red) for each diffusion component. b, For topics 1–4 and 17, scatter plots show cell profiles (dots), plotted by DC 1 versus DC 2 (top), or DC 3 versus DC 4 (bottom), and coloured by topic weight (colour bar: low, grey; high, teal). DCs 2–4 were used to determine start and end positions for the directed-diffusion trajectories (Fig. 2b), because they reflect extremal points of the gene expression spectrum. c, Interpolation validation for Waddington-OT results (Methods). Interpolating cell distributions at intermediate time points by optimal transport (OT) captures the actual distributions better than random interpolation. Entropic OT distance (Wasserstein distance, d) between different pairs (colours) of empirical probability distributions, plotted for consecutive triples of time points (midpoint, x axis; for example, 1 refers to triple of days 0, 1, 2), for cell distributions at previous, current, and next time points (‘previous’, ‘real’ and ‘next’, respectively), and for OT and random interpolations between the previous and next time point distributions. d–g, Topic-9-associated genes and additional information about cell selections for directed-diffusion trajectories in Fig. 2b (Methods). d, FDL embedding (as in Fig. 1e) of cell profiles, coloured by normalized expression (log-transformed scTransform-corrected counts; colour bar: low, grey; high, maroon) for genes that are differentially expressed in cells up-weighted for topic 9, compared with other cells. e, Visualization of cell profiles (dots), coloured by membership in the start (cyan) or end (fuchsia) sets (or grey, otherwise), for each of the directed-diffusion trajectories. Plots show diffusion components (top; DC2, x axis, DC4, y axis for quiescent-to-ILC3-like and ILC2-to-quiescent-like trajectories; DC2, x axis, DC3, y axis for ILC2-to-ILC3-like and cloud-to-ILC3-like trajectories), and FDL embedding (bottom, as in Fig. 1e). f, Visualization of cells explicitly excluded from the potential starting set for the cloud-to-ILC3-like trajectory. For topics 1–3, cells (dots) are plotted by diffusion components (top; DC2, x axis, DC3, y axis) and FDL embedding (bottom), and coloured brightly if their topic weight exceeds the threshold of 0.2, or grey otherwise. Analogous panels for topic 17 showing DC1/DC2 and with threshold 0.08. g, Cell selection for directed-diffusion trajectories in Fig. 2b. For each trajectory (panel), the proportion of sampled paths that a cell occurred in versus the mean normalized pseudotime position for cells (dots) from all time points. Cells are considered part of the trajectory (in orange) if they occur at the very start or end of the paths (0 and 1 on x axis, grey dashed lines), or are contained in sufficiently many sampled paths, as determined by either a global quantile cut-off (blue line) or an adaptive quantile cut-off that depends on the average path position of the cell (red line, spline quantile regression), with different cut-offs chosen for each trajectory. Remaining cells (grey) are excluded from the trajectory.