palantir 1.4.6

v1.4.6

on Python PyPI

Release: pseudotime from several start cells, faster and leaner pseudotime on large data, and a fix for NaN pseudotime when a cell is far from every waypoint. Results for a single start cell are unchanged up to floating-point rounding.

New

• Several early cells. run_palantir and identify_terminal_states accept a list of early cells. Pseudotime then starts from the nearest of them. A warning says when moving the early cells to the diffusion-component boundaries merges some of them (pass use_early_cell_as_start=True to start from the early cells themselves), and when automatic detection finds no terminal state. (#146, #185, #188)

Performance

• Faster, leaner pseudotime on large data, with unchanged results. Shortest paths from the waypoints run in n_jobs processes (serially when the joblib backend is threading), the connectivity check uses scipy.sparse.csgraph instead of networkx, and the refinement works in chunks of cells. (#186, #188)

Fixes

• Pseudotime no longer turns NaN when one cell is far from every waypoint. When a cell was more than ~38 bandwidths from every waypoint, all its waypoint weights underflowed to 0 and normalising them gave 0 / 0, which made the pseudotime of every cell NaN. The weights are now normalised with scipy.special.softmax, which gives the same values up to floating-point rounding (a few ulp) and stays finite. (#187, #188)
• No Pandas4Warning in from_fcs. The deprecated copy keyword of the astype call is dropped. It was inert, so the cast behaves identically on pandas 2 and 3. (#184)

Dependencies

• networkx is no longer a runtime dependency; only the test extras (test, test-base) install it.
• joblib>=1.3 is now required.

Thanks

Thanks to @Marius1311 for #185, #186 and #187, to @xyf5432 for #184, and to @TolgaDuz for raising several start cells in #146.

Full changelog: v1.4.5...v1.4.6