Fig. 1. Overview of Lamian : a statistical framework for differential pseudotemporal trajectory analysis with multiple samples.
a Using integrated and harmonized scRNA-seqdata across multiple samples, Lamian first groups cells into clusters. b Clustered data is used to infer a pseudotemporal tree structure followed by automatically enumerating all pseudotemporal branches (paths). The uncertainty of tree branches are quantified using a detection rate in bootstrap resampling framework (Module 1), followed by quantifying the variability of branches across samples and identifying differences in the branching structure across conditions (Module 2). c For each tree branch (pseudotemporal path), Lamian can identify two types of differential expression (DE): DE along pseudotime (TDE) and DE associated with sample covariates (XDE) (Module 3). Similarly, Lamian can also identify changes in cell density along pseudotime (TCD) and associated with a sample covariate (XCD) (Module 4). Gene’s or cell abundance’s pseudotemporal patterns are modeled using the combinations of B-spline bases (Φs) to allow non-linear patterns. The combination coefficients are decomposed into effects due to sample covariates (Xsβ, where Xs is the design matrix) and variation among samples with common covariate values (us). Cell-level data ys in sample s are generated from the sample-level curve by adding cell-level random noise ϵs. See Methods and Supplementary Notes for details.