What it does. RNA velocity analysis with scVelo. Estimate cell state transitions from unspliced/spliced mRNA dynamics, infer trajectory directions, compute latent time, and identify driver genes in single-cell RNA-seq data. Complements Scanpy/scVI-tools for trajectory inference. Part of K-Dense-AI/scientific-agent-skills (AI Scientist skills) (K-Dense-AI/scientific-agent-skills).

Upstream	K-Dense-AI/scientific-agent-skills
Skill file	skills/scvelo/SKILL.md
License	MIT
Author	K-Dense Inc.
Fetched	2026-09-10

Install

• npx skills add K-Dense-AI/scientific-agent-skills --skill scvelo, or copy the skill folder into ~/.claude/skills/scvelo/.
• Raw file: curl -sL https://raw.githubusercontent.com/K-Dense-AI/scientific-agent-skills/HEAD/skills/scvelo/SKILL.md

SKILL.md (verbatim)

name: scvelo
description: RNA velocity analysis with scVelo. Estimate cell state transitions from unspliced/spliced mRNA dynamics, infer trajectory directions, compute latent time, and identify driver genes in single-cell RNA-seq data. Complements Scanpy/scVI-tools for trajectory inference.
license: BSD-3-Clause
compatibility: Requires Python 3.10+ with scvelo, scanpy, and anndata. Verified against scvelo 0.3.4, whose dynamical model and pl.scatter need pandas<3 and whose stochastic estimator needs numpy<2; the deterministic estimator works on current releases.
metadata:
  version: "1.2"
  skill-author: Kuan-lin Huang

scVelo — RNA Velocity Analysis

Overview

scVelo is the leading Python package for RNA velocity analysis in single-cell RNA-seq data. It infers cell state transitions by modeling the kinetics of mRNA splicing — using the ratio of unspliced (pre-mRNA) to spliced (mature mRNA) abundances to determine whether a gene is being upregulated or downregulated in each cell. This allows reconstruction of developmental trajectories and identification of cell fate decisions without requiring time-course data.

Installation: uv pip install scvelo

Key resources:

• Documentation: https://scvelo.readthedocs.io/
• GitHub: https://github.com/theislab/scvelo
• Paper: Bergen et al. (2020) Nature Biotechnology. PMID: 32747759

When to Use This Skill

Use scVelo when:

• Trajectory inference from snapshot data: Determine which direction cells are differentiating
• Cell fate prediction: Identify progenitor cells and their downstream fates
• Driver gene identification: Find genes whose dynamics best explain observed trajectories
• Developmental biology: Model hematopoiesis, neurogenesis, epithelial-to-mesenchymal transitions
• Latent time estimation: Order cells along a pseudotime derived from splicing dynamics
• Complement to Scanpy: Add directional information to UMAP embeddings

Prerequisites

scVelo requires count matrices for both unspliced and spliced RNA. These are generated by:

1. STARsolo or kallisto|bustools with lamanno mode
2. velocyto CLI: velocyto run10x / velocyto run
3. alevin-fry / simpleaf with spliced/unspliced output

Data is stored in an AnnData object with layers["spliced"] and layers["unspliced"].

Standard RNA Velocity Workflow

1. Setup and Data Loading

import scvelo as scv
import scanpy as sc
import numpy as np
import matplotlib.pyplot as plt

# Configure settings
scv.settings.verbosity = 3       # Show computation steps
scv.settings.presenter_view = True
scv.settings.set_figure_params('scvelo')

# Load data (AnnData with spliced/unspliced layers)
# Option A: Load from loom (velocyto output)
adata = scv.read("cellranger_output.loom", cache=True)

# Option B: Merge velocyto loom with Scanpy-processed AnnData
adata_processed = sc.read_h5ad("processed.h5ad")  # Has UMAP, clusters
adata_velocity = scv.read("velocyto.loom")
adata = scv.utils.merge(adata_processed, adata_velocity)

# Verify layers
print(adata)
# obs × var: N × G
# layers: 'spliced', 'unspliced' (required)
# obsm['X_umap'] (required for visualization)

2. Preprocessing

# Filter and normalize. As of scVelo 0.3, filter_and_normalize() only filters
# genes and normalizes per cell -- it no longer takes n_top_genes and no longer
# log-transforms, so the log step and HVG selection come from Scanpy.
scv.pp.filter_and_normalize(
    adata,
    min_shared_counts=20    # Minimum counts in spliced+unspliced
)
sc.pp.log1p(adata)
sc.pp.highly_variable_genes(adata, n_top_genes=2000, subset=True)

# Compute first and second order moments (means and variances)
# knn_connectivities must be computed first
sc.pp.neighbors(adata, n_neighbors=30, n_pcs=30)
scv.pp.moments(
    adata,
    n_pcs=30,
    n_neighbors=30
)

3. Velocity Estimation — Stochastic Model

The stochastic model is fast and suitable for exploratory analysis:

# Stochastic velocity (faster, less accurate)
scv.tl.velocity(adata, mode='stochastic')
scv.tl.velocity_graph(adata)

# Visualize
scv.pl.velocity_embedding_stream(
    adata,
    basis='umap',
    color='leiden',
    title="RNA Velocity (Stochastic)"
)

4. Velocity Estimation — Dynamical Model (Recommended)

The dynamical model fits the full splicing kinetics and is more accurate:

# Recover dynamics (computationally intensive; ~10-30 min for 10K cells)
scv.tl.recover_dynamics(adata, n_jobs=4)

# Compute velocity from dynamical model
scv.tl.velocity(adata, mode='dynamical')
scv.tl.velocity_graph(adata)

5. Latent Time

The dynamical model enables computation of a shared latent time (pseudotime):

# Compute latent time
scv.tl.latent_time(adata)

# Visualize latent time on UMAP
scv.pl.scatter(
    adata,
    color='latent_time',
    color_map='gnuplot',
    size=80,
    title='Latent time'
)

# Identify top genes ordered by latent time
top_genes = adata.var['fit_likelihood'].sort_values(ascending=False).index[:300]
scv.pl.heatmap(
    adata,
    var_names=top_genes,
    sortby='latent_time',
    col_color='leiden',
    n_convolve=100
)

6. Driver Gene Analysis

# Identify genes with highest velocity fit
scv.tl.rank_velocity_genes(adata, groupby='leiden', min_corr=0.3)
df = scv.DataFrame(adata.uns['rank_velocity_genes']['names'])
print(df.head(10))

# Speed and coherence
scv.tl.velocity_confidence(adata)
scv.pl.scatter(
    adata,
    c=['velocity_length', 'velocity_confidence'],
    cmap='coolwarm',
    perc=[5, 95]
)

# Phase portraits for specific genes
scv.pl.velocity(adata, ['Cpe', 'Gnao1', 'Ins2'],
               ncols=3, figsize=(16, 4))

7. Velocity Arrows and Pseudotime

# Arrow plot on UMAP
scv.pl.velocity_embedding(
    adata,
    arrow_length=3,
    arrow_size=2,
    color='leiden',
    basis='umap'
)

# Stream plot (cleaner visualization)
scv.pl.velocity_embedding_stream(
    adata,
    basis='umap',
    color='leiden',
    smooth=0.8,
    min_mass=4
)

# Velocity pseudotime (alternative to latent time)
scv.tl.velocity_pseudotime(adata)
scv.pl.scatter(adata, color='velocity_pseudotime', cmap='gnuplot')

8. PAGA Trajectory Graph

# PAGA graph with velocity-informed transitions
scv.tl.paga(adata, groups='leiden')
df = scv.get_df(adata, 'paga/transitions_confidence', precision=2).T
df.style.background_gradient(cmap='Blues').format('{:.2g}')

# Plot PAGA with velocity
scv.pl.paga(
    adata,
    basis='umap',
    size=50,
    alpha=0.1,
    min_edge_width=2,
    node_size_scale=1.5
)

Complete Workflow Script

import scvelo as scv
import scanpy as sc

def run_rna_velocity(adata, n_top_genes=2000, mode='dynamical', n_jobs=4):
    """
    Complete RNA velocity workflow.

    Args:
        adata: AnnData with 'spliced' and 'unspliced' layers, UMAP in obsm
        n_top_genes: Number of top HVGs for velocity
        mode: 'stochastic' (fast) or 'dynamical' (accurate)
        n_jobs: Parallel jobs for dynamical model

    Returns:
        Processed AnnData with velocity information
    """
    scv.settings.verbosity = 2

    # 1. Preprocessing (scVelo 0.3 dropped log/HVG from filter_and_normalize)
    scv.pp.filter_and_normalize(adata, min_shared_counts=20)
    sc.pp.log1p(adata)
    sc.pp.highly_variable_genes(adata, n_top_genes=n_top_genes, subset=True)

    if 'neighbors' not in adata.uns:
        sc.pp.neighbors(adata, n_neighbors=30)

    scv.pp.moments(adata, n_pcs=30, n_neighbors=30)

    # 2. Velocity estimation
    if mode == 'dynamical':
        scv.tl.recover_dynamics(adata, n_jobs=n_jobs)

    scv.tl.velocity(adata, mode=mode)
    scv.tl.velocity_graph(adata)

    # 3. Downstream analyses
    if mode == 'dynamical':
        scv.tl.latent_time(adata)
        scv.tl.rank_velocity_genes(adata, groupby='leiden', min_corr=0.3)

    scv.tl.velocity_confidence(adata)
    scv.tl.velocity_pseudotime(adata)

    return adata

Key Output Fields in AnnData

After running the workflow, the following fields are added:

Location	Key	Description
adata.layers	velocity	RNA velocity per gene per cell
adata.layers	fit_t	Fitted latent time per gene per cell
adata.obsm	velocity_umap	2D velocity vectors on UMAP
adata.obs	velocity_pseudotime	Pseudotime from velocity
adata.obs	latent_time	Latent time from dynamical model
adata.obs	velocity_length	Speed of each cell
adata.obs	velocity_confidence	Confidence score per cell
adata.var	fit_likelihood	Gene-level model fit quality
adata.var	fit_alpha	Transcription rate
adata.var	fit_beta	Splicing rate
adata.var	fit_gamma	Degradation rate
adata.uns	velocity_graph	Cell-cell transition probability matrix

Velocity Models Comparison

Model	Speed	Accuracy	When to Use
stochastic	Fast	Moderate	Exploratory; large datasets
deterministic	Medium	Moderate	Simple linear kinetics
dynamical	Slow	High	Publication-quality; identifies driver genes

Best Practices

• Start with stochastic mode for exploration; switch to dynamical for final analysis
• Need good coverage of unspliced reads: Short reads (< 100 bp) may miss intron coverage
• Minimum 2,000 cells: RNA velocity is noisy with fewer cells
• Velocity should be coherent: Arrows should follow known biology; randomness indicates issues
• k-NN bandwidth matters: Too few neighbors → noisy velocity; too many → oversmoothed
• Sanity check: Root cells (progenitors) should have high unspliced/spliced ratios for marker genes
• Dynamical model requires distinct kinetic states: Works best for clear differentiation processes

Troubleshooting

Problem	Solution
Missing unspliced layer	Re-run velocyto or use STARsolo with --soloFeatures Gene Velocyto
Very few velocity genes	Lower min_shared_counts; check sequencing depth
Random-looking arrows	Try different n_neighbors or velocity model
Memory error with dynamical	Set n_jobs=1; reduce n_top_genes
Negative velocity everywhere	Check that spliced/unspliced layers are not swapped

Additional Resources

• scVelo documentation: https://scvelo.readthedocs.io/
• Tutorial notebooks: https://scvelo.readthedocs.io/tutorials/
• GitHub: https://github.com/theislab/scvelo
• Paper: Bergen V et al. (2020) Nature Biotechnology. PMID: 32747759
• velocyto (preprocessing): http://velocyto.org/
• CellRank (fate prediction, extends scVelo): https://cellrank.readthedocs.io/
• dynamo (metabolic labeling alternative): https://dynamo-release.readthedocs.io/

Other files in this skill

• references/velocity_models.md
• scripts/rna_velocity_workflow.py

references/velocity_models.md (verbatim)

scVelo Velocity Models Reference

Mathematical Framework

RNA velocity is based on the kinetic model of transcription:

dx_s/dt = β·x_u - γ·x_s   (spliced dynamics)
dx_u/dt = α(t) - β·x_u    (unspliced dynamics)

Where:

• x_s: spliced mRNA abundance
• x_u: unspliced (pre-mRNA) abundance
• α(t): transcription rate (varies over time)
• β: splicing rate
• γ: degradation rate

Velocity is defined as: v = dx_s/dt = β·x_u - γ·x_s

• v > 0: Gene is being upregulated (more unspliced than expected at steady state)
• v < 0: Gene is being downregulated (less unspliced than expected)

Model Comparison

Steady-State (Velocyto, original)

• Assumes constant α (transcription rate)
• Fits γ using linear regression on steady-state cells
• Limitation: Requires identifiable steady states; assumes constant transcription

# Use with scVelo for backward compatibility
scv.tl.velocity(adata, mode='steady_state')

Stochastic Model (scVelo v1)

• Extends steady-state with variance/covariance terms
• Models cell-to-cell variability in mRNA counts
• More robust to noise than steady-state

scv.tl.velocity(adata, mode='stochastic')

Dynamical Model (scVelo v2, recommended)

• Jointly estimates all kinetic rates (α, β, γ) and cell-specific latent time
• Does not assume steady state
• Identifies induction vs. repression phases
• Computes fit_likelihood per gene (quality measure)

scv.tl.recover_dynamics(adata, n_jobs=4)
scv.tl.velocity(adata, mode='dynamical')

Kinetic states identified by dynamical model:

State	Description
Induction	α > 0, x_u increasing
Steady-state on	α > 0, constant high expression
Repression	α = 0, x_u decreasing
Steady-state off	α = 0, constant low expression

Velocity Graph

The velocity graph connects cells based on their velocity similarity to neighboring cells' states:

scv.tl.velocity_graph(adata)
# Stored in adata.uns['velocity_graph']
# Entry [i,j] = probability that cell i transitions to cell j

Parameters:

• n_neighbors: Number of neighbors considered
• sqrt_transform: Apply sqrt transform to data (default: False for spliced)
• approx: Use approximate nearest neighbor search (faster for large datasets)

Latent Time Interpretation

Latent time τ ∈ [0, 1] for each gene represents:

• τ = 0: Gene is at onset of induction
• τ = 0.5: Gene is at peak of induction (for a complete cycle)
• τ = 1: Gene has returned to steady-state off

Shared latent time is computed by taking the average over all velocity genes, weighted by fit_likelihood.

Quality Metrics

Gene-level

• fit_likelihood: Goodness-of-fit of dynamical model (0-1; higher = better)
  • Use for filtering driver genes: adata.var[adata.var['fit_likelihood'] > 0.1]
• fit_alpha: Transcription rate during induction
• fit_gamma: mRNA degradation rate
• fit_r2: R² of kinetic fit

Cell-level

• velocity_length: Magnitude of velocity vector (cell speed)
• velocity_confidence: Coherence of velocity with neighboring cells (0-1)

Dataset-level

# Check overall velocity quality
scv.pl.proportions(adata)  # Ratio of spliced/unspliced per cell
scv.pl.velocity_confidence(adata, groupby='leiden')

Parameter Tuning Guide

Parameter	Function	Default	When to Change
min_shared_counts	Filter genes	20	Increase for deep sequencing; decrease for shallow
n_top_genes	HVG selection	2000	Increase for complex datasets
n_neighbors	kNN graph	30	Decrease for small datasets; increase for noisy
n_pcs	PCA dimensions	30	Match to elbow in scree plot
t_max_rank	Latent time constraint	None	Set if known developmental direction

Integration with Other Tools

CellRank (Fate Prediction)

import cellrank as cr
from cellrank.kernels import VelocityKernel, ConnectivityKernel

# Combine velocity and connectivity kernels
vk = VelocityKernel(adata).compute_transition_matrix()
ck = ConnectivityKernel(adata).compute_transition_matrix()
combined = 0.8 * vk + 0.2 * ck

# Compute macrostates (terminal and initial states)
g = cr.estimators.GPCCA(combined)
g.compute_macrostates(n_states=4, cluster_key='leiden')
g.plot_macrostates(which="all")

# Compute fate probabilities
g.compute_fate_probabilities()
g.plot_fate_probabilities()

Scanpy Integration

scVelo works natively with Scanpy's AnnData:

import scanpy as sc
import scvelo as scv

# Run standard Scanpy pipeline first
sc.pp.normalize_total(adata)
sc.pp.log1p(adata)
sc.pp.highly_variable_genes(adata)
sc.pp.pca(adata)
sc.pp.neighbors(adata)
sc.tl.umap(adata)
sc.tl.leiden(adata)

# Then add velocity on top
scv.pp.moments(adata)
scv.tl.recover_dynamics(adata)
scv.tl.velocity(adata, mode='dynamical')
scv.tl.velocity_graph(adata)
scv.tl.latent_time(adata)

Back to K-Dense-AI/scientific-agent-skills (AI Scientist skills) or Agent skills.