Trajectory inference

Trajectory inference#

Trajectory inference (TI) orders cells along inferred developmental pathways and is one of the workhorses of single-cell analysis. omicverse exposes a single entry point — ov.single.TrajInfer — that dispatches to eleven backends spanning the methods compared in the dynbenchmark study (Saelens et al., 2019, Nat Biotechnol).

This section has three layers:

Layer

Tutorial

When

1. Recommended workflow

t_traj_slingshot

Day-one user. Slingshot end-to-end on a real branching dataset, with PAGA overlay, branch-aware dynamic heatmaps and lineage-resolved trend plots.

2. Unified downstream — fate

t_traj_fate

After any pseudotime backend you can run ov.single.PseudotimeFate(adata, pseudotime_key=...) for macrostates → terminal states → fate probabilities in <1 s on 3.7 k cells (≈300× faster than CellRank). Backend-agnostic — same code for Palantir / Slingshot / SCORPIUS / destiny / ….

3. Method zoo

zoo/index

You want to compare backends side-by-side, or you need a method that handles a specific topology (linear / multifurcating / cycle / convergence). 15 tutorials, one per backend.

Recommendation tree#

        Do you have a known origin / terminal celltype?
                          │
                ┌─────────┴─────────┐
                yes                 no
                 │                   │
        Branching topology?     CytoTRACE2 (stemness gradient)
            │                   or destiny DPT (data-driven root)
   ┌────────┴────────┐
   bifurc / multi    linear
   │                   │
 Slingshot (recommended)  SCORPIUS / TSCAN / Palantir
   Palantir, VIA, URD
   Monocle3, scTour

The eleven backends#

method=

Package

Topology supported

Notes

'palantir'

palantir (vendored)

tree, branching

Reference implementation. Best on continuous differentiation hierarchies.

'diffusion_map'

scanpy

linear, branching

scanpy’s dpt on the standard diffusion map.

'slingshot'

_pyslingshot (vendored)

tree, branching

Recommended — outer-level tutorial. Lineage-resolved pseudotime + branch labels.

'sctour'

sctour

any (latent vector field)

Generative; also yields velocity.

'stavia'

_stavia (vendored)

tree, multifurc, cycle

spatial-aware variant.

'scorpius'

pyscorpius

linear

dynbenchmark-style. Returns 1-D ordering only.

'tscan'

pytscan

tree, branching

mclust-based clustering on first PCs, then MST traversal.

'destiny'

pydestiny-bio

linear, branching

Python port of R destiny. DiffusionMap + DPT, data-driven root via DC1 extremum.

'urd'

pyurd-bio

branching

Flood-fill pseudotime on diffusion-map transitions.

'monocle3'

pymonocle3-bio

tree, branching

UMAP + principal graph; needs explicit root.

'cytotrace'

pycytotrace-bio

gradient

Gene-count correlate of stemness. Single-axis, no branches.

The first 5 (palantir / diffusion_map / slingshot / sctour / stavia) ship inside omicverse. The last 6 — the dynbenchmark zoo — are pip install omicverse[trajectory] extras.

Architecture#

Every backend writes pseudotime under a unified obs schema:

adata.obs['{method}_pseudotime']   float    — universal
adata.obs['{method}_branch']       category — when the backend infers branches
adata.obsm['X_{method}']           latent   — when the backend builds an embedding

Downstream tools (PAGA, dynamic_heatmap, branch_streamplot, dynamic_features) consume any backend’s output via this schema — no if method == ... branching in your downstream code.

Reference#

  • Saelens W, Cannoodt R, Todorov H, Saeys Y. A comparison of single-cell trajectory inference methods. Nat Biotechnol 37, 547–554 (2019). doi:10.1038/s41587-019-0071-9

Contents