Fit Palantir gene trends over pseudotime — rs_gene

Selects the cells belonging to each branch from the fate probabilities, then fits a landmark Gaussian process with a Matern-5/2 kernel per branch. This is the Mellon-based estimator of the reference, not the legacy GAM.

Whatever expression matrix is handed over is what gets fitted. Nothing here imputes; that decision belongs to the caller.

The defaults are prior-dominated. Palantir's pseudotime is min-max scaled to [0, 1], so a length_scale of 1.0 spans the whole domain and a sigma of 1.0 sits at roughly the signal scale of log-normalised expression. That resolves almost any gene into a smooth monotone or single-peaked curve. Shorten length_scale before believing a bump.

Usage

rs_gene_trends(
  expression,
  pseudotime,
  branch_probs,
  branch_params,
  gene_trend_params
)

Arguments

expression: Numerical matrix of cells x genes. All cells, in the same row order as pseudotime, not just the branch members.
pseudotime: Numerical vector. Pseudotime per cell.
branch_probs: Numerical matrix of cells x fates with the fate probabilities. Rows need not sum to one.
branch_params: List. Parameter list, see params_sc_branch_selection().
gene_trend_params: List. Parameter list, see params_sc_gene_trends().

Value

A list with

trends - List of numerical matrices, one per branch, each of resolution x genes.
grids - List of numerical vectors with the pseudotime grid per branch, running from the branch minimum to its maximum.
branch_cells - List of integer vectors with the cell indices (0-indexed!) selected for each branch.
n_cells - Integer vector with the cell count per branch.
jitter_used - Numerical vector with the jitter each branch's Cholesky needed.

References

Setty, et al., Nat. Biotechnol., 2019.