Scvelo

scVelo RNA Velocity

When to use

Use for kinetic analysis of aligned spliced/unspliced RNA counts, directional trajectory hypotheses, gene-level phase portraits, and dynamical latent time. Use the model reference for assumptions, transition probabilities, and an optional CellRank handoff.

RNA velocity estimates an expression derivative under a model. Smooth arrows are not observed cell movement, lineage tracing, causal drivers, or proof of future fate. The dynamical estimator is not automatically more accurate than a steady-state estimator on every dataset.

Tested environment and compatibility

The maintained workflow targets scVelo 0.3.4, Scanpy 1.12.4, AnnData 0.13.4, NumPy 2.5.3, pandas 2.3.3, SciPy 1.18.1, Matplotlib 3.11.2, and loompy 3.0.8 on Python 3.13. Use an isolated environment:

uv venv --python 3.13 .venv-velocity
uv pip install --python .venv-velocity/bin/python \
  scvelo==0.3.4 scanpy==1.12.4 anndata==0.13.4 numpy==2.5.3 \
  pandas==2.3.3 scipy==1.18.1 matplotlib==3.11.2 loompy==3.0.8

The packages above were resolved from cache and tested in an isolated uv environment; shell paths above use POSIX syntax. Installation requires network access unless packages are cached. No API key is required.

• Deterministic and dynamical models: native synthetic fitting, graphs, relative times, plots and H5AD round trips are tested on this stack.
• Stochastic model: scVelo 0.3.4's default GLS scalar assignment fails with NumPy 2. Scanpy 1.12 requires NumPy 2, so installing numpy<2 beside current Scanpy is not a solution. A separately validated legacy Scanpy/NumPy environment is required; that legacy stack was not executed in this review. The helper rejects this combination before modifying data, without changing the model.
• pandas 3: scVelo 0.3.4 dynamical fitting and some plots use operations no longer supported by pandas 3. Keep the explicit pandas pin.
• The upstream stable tutorial still contains removed scv.read and older preprocessing calls. Check the 0.3.4 source when a tutorial disagrees with the installed API.

Input contract

1. Obtain spliced and unspliced counts from a velocity-aware quantifier. Keep quantifier/version, reference annotation, counting mode, sample IDs and cell barcode mapping. Upstream quantification is outside the bundled script.
2. Supply cells by genes matrices in adata.layers['spliced'] and adata.layers['unspliced'], with identical cell/gene ordering and unique IDs. Counts must be finite and nonnegative; fractional count estimates are allowed. Never use logged, scaled, residualized, or batch-corrected values as counts. Numerical inspection alone cannot establish that data are raw.
3. Inspect per-library/cluster coverage, doublets, ambient RNA, zero-count cells, and annotation compatibility before fitting. Preserve the original full-gene count file. Retained-gene backups do not preserve filtered-out genes.
4. Resolve barcode prefixes and gene identifiers explicitly when combining files. Do not silently strip library IDs, intersect away most cells, or make duplicated biological IDs unique without understanding why they repeat.

Read H5AD through AnnData. scVelo 0.3.4 has no scv.read or scv.DataFrame:

import anndata as ad
adata = ad.read_h5ad("velocity_counts.h5ad")

For legacy loom input use ad.io.read_loom('counts.loom', X_name='spliced', sparse=True). AnnData 0.13 deprecates loom; convert to H5AD for further work. Its layers[None] aliases X: do not delete it or iterate all layers as if every key were a string. The helper accesses only named velocity layers. Native loom reading and exact, reordered metadata alignment are covered by the tests; AnnData 0.13 write_loom is not used.

Run the maintained workflow

The bundled script accepts local files and never downloads a demonstration dataset on startup:

MPLBACKEND=Agg .venv-velocity/bin/python scripts/rna_velocity_workflow.py \
  velocity_counts.h5ad --mode dynamical --groupby clusters \
  --n-top-genes 2000 --n-neighbors 30 --n-jobs 1 --output-dir velocity_results

# Import a loom and align its raw counts to an existing annotation file:
MPLBACKEND=Agg .venv-velocity/bin/python scripts/rna_velocity_workflow.py \
  counts.loom --processed-h5ad annotated.h5ad --groupby clusters \
  --mode deterministic --output-dir velocity_check

Run these from the skill directory. Input filenames and biological annotations are illustrative; the same CLI and functions are tested on small generated kinetic fixtures. Omit --groupby when annotations are unavailable. Use --no-plots for analysis without computing UMAP.

The helper performs these steps:

1. Validate layers, IDs, model compatibility and grouping; refuse previously generated moments/velocity so preprocessing cannot silently run twice.
2. Back up retained raw layers as spliced_counts and unspliced_counts and rebuild X from raw spliced counts. Filter genes and normalize X and the two count layers on a linear scale.
3. Apply sc.pp.log1p only to X, select highly variable genes with Scanpy, and rebuild PCA and neighbors after subsetting. Reusing stale PCA/neighbors from an unrelated feature set can silently change the velocity model.
4. Compute Ms and Mu with scv.pp.moments(adata, n_neighbors=None) from the explicit Scanpy graph. Moments are dense: budget memory for multiple cells-by-genes arrays, not only the sparse input.
5. For dynamical mode, call recover_dynamics(var_names='all') on the selected genes before velocity(mode='dynamical'); otherwise fit the selected model. Require usable genes and a nonempty velocity graph.
6. Compute velocity coherence and velocity pseudotime; add latent time only for dynamical fits. Rank velocity-associated genes only with at least two groups and at least two cells per group. Record/warn about nonfinite or constant time/coherence outputs in uns['velocity_workflow']['diagnostics']; a constant pseudotime does not support a trajectory ordering. Ranking is exploratory.
7. For plots, recompute UMAP from the rebuilt graph, project velocities, save PNGs directly to the requested directory, then save H5AD with package versions and selected parameters. Existing labels remain annotations, not validated cell identities.

The Python function mutates its argument in place; pass adata.copy() to keep the original object. The CLI writes to output_dir/adata_velocity.h5ad. Version 2 changes the former helper's behavior deliberately: raw layer geometry is rebuilt, missing requested labels are errors, and plots/parallelism can be controlled explicitly. Runtime depends on cells, genes and fit difficulty; there is no universal 10–30 minute expectation.