LLMs-Universal-Life-Science-and-Clinical-Skills- sc-velocity

install

source · Clone the upstream repo

git clone https://github.com/mdbabumiamssm/LLMs-Universal-Life-Science-and-Clinical-Skills-

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/mdbabumiamssm/LLMs-Universal-Life-Science-and-Clinical-Skills- "$T" && mkdir -p ~/.claude/skills && cp -r "$T/Skills/Transcriptomics/sc-velocity" ~/.claude/skills/mdbabumiamssm-llms-universal-life-science-and-clinical-skills-sc-velocity && rm -rf "$T"

manifest: Skills/Transcriptomics/sc-velocity/SKILL.md

⚡ Single-Cell RNA Velocity

Quantify transcriptional dynamics by coupling spliced and unspliced matrices using scVelo's stochastic and dynamical models. This skill ingests AnnData/loom counts, performs preconditioning (normalization, moments, phase portrait checks), computes velocities, and returns latent time embeddings with driver gene evidence suitable for trajectory-aware downstream workflows.

When to Trigger

User mentions "RNA velocity," "latent time," "dynamic gene programs," or "spliced/unspliced layers"
Input AnnData includes
```
.layers["spliced"]
```
and
```
.layers["unspliced"]
```
Need to compare lineage directionality following clustering/integration steps

Supported Inputs

Format Notes

.h5ad

Preferred; expects

layers["spliced"]

["unspliced"]

obs

annotations, optional

var

kinetic priors

.loom

Auto-converted to AnnData; verifies

layers/{spliced,unspliced}

Matrices

Provide

--spliced

--unspliced

MTX + metadata to build AnnData wrapper

Typical Workflow

Load + QC

import scvelo as scv
adata = scv.read("sample.h5ad")
scv.pp.filter_and_normalize(adata, min_shared_counts=30, n_top_genes=4000)
scv.pp.moments(adata, n_pcs=30, n_neighbors=30)

Velocity Computation

scv.tl.velocity(adata, mode="dynamical")
scv.tl.velocity_graph(adata)
scv.tl.velocity_confidence(adata)

Latent Time & Drivers

scv.tl.recover_dynamics(adata)
scv.tl.latent_time(adata)
scv.tl.rank_velocity_genes(adata, groupby="leiden")

Visualization & Export

scv.pl.velocity_embedding_stream(adata, basis="umap", color="leiden")
adata.write_h5ad("velocity_annotated.h5ad")

CLI Examples

# Run full pipeline with dynamical model, auto-detect PCA/neighbor params
python omicsclaw.py run sc-velocity --input data/sample.h5ad --mode dynamical --out runs/sc_velocity

# Skip recover_dynamics if kinetics already stored
python omicsclaw.py run sc-velocity --input data/sample.h5ad --mode stochastic --skip-latent-time

# Provide loom inputs and metadata
python omicsclaw.py run sc-velocity --loom data/sample.loom --obs-meta obs.csv --var-meta var.csv

Key Parameters

Flag	Default	Description
`--mode`	`dynamical`	`stochastic` , `steady_state` , or `dynamical` modeling
`--min-shared-counts`	`30`	Genes must appear with this many shared counts during normalization
`--n-top-genes`	`4000`	HVG count for kinetics modeling
`--n-pcs`	`30`	Principal components for moment calculation
`--n-neighbors`	`30`	Neighborhood size for velocity graph construction
`--skip-latent-time`	`False`	Disable latent time + driver gene ranking to save runtime

Outputs

Annotated
```
.h5ad
```
with velocity layers, confidence, latent time, and driver gene tables
Summary report (
```
report.md
```
) capturing runtime, parameter set, model diagnostics, QC thresholds
PNG/PDF figures for streamlines, quiver plots, phase portraits, driver gene heatmaps
```
result.json
```
for orchestrator bookkeeping

Integration Touchpoints

Upstream:

sc-preprocessing

sc-doublet-detection

sc-batch-integration

Downstream:
```
sc-trajectory
```
(uses latent time),
```
sc-cell-communication
```
(velocity-informed ligand prioritization),
```
spatial-velocity
```

Troubleshooting Cheatsheet

Symptom	Fix
Low velocity confidence ( `velocity_confidence` < 0.4)	Increase `min_shared_counts` , rerun normalization/moments, verify layers not log-normalized beforehand
Divergent dynamical fit	Initialize with `mode=stochastic` , inspect phase portraits for genes with insufficient kinetics
Memory pressure (>32 GB)	Use `--subset-gene-file` to preselect driver genes or run on sparse-backed AnnData

Source Reliability

Source	Score	Notes
Theis Lab scVelo repo	0.93	Active upstream repository with peer-reviewed implementation
K-Dense Claude Scientific Skills	0.90	Provides standardized skill scaffolding we mirror for metadata/trigger design