---

name: tooluniverse-scvi-tools description: "Deep learning for single-cell omics using scvi-tools — probabilistic (VAE-based) batch correction and integration with scVI, semi-supervised integration and label transfer with scANVI, CITE-seq RNA+protein joint modeling with totalVI, scATAC-seq analysis with PeakVI, multiome RNA+ATAC joint modeling with MultiVI, spatial transcriptomics cell-type deconvolution with DestVI, and RNA velocity with veloVI. Use when scVI, scANVI, totalVI, PeakVI, MultiVI, DestVI, or veloVI are mentioned by name, or when a request needs deep-learning-based batch correction, reference mapping / scArches-style query-to-reference mapping, denoising single-cell counts, joint modeling of two modalities (RNA+protein, RNA+ATAC, single-cell+spatial), or a variational-autoencoder latent space for single-cell data — as opposed to classic scanpy/Harmony-based integration."

scvi-tools Deep Learning for Single-Cell Omics

Local, guided workflows for training and applying scvi-tools' probabilistic deep-learning models on a user's own AnnData files.

Honesty contract (read first)

This skill trains real neural networks via the real scvi-tools package. It must never fabricate a latent space, a training curve, or a DE result.

1. Preflight before anything. Every script checks import scvi (and anndata/scanpy/numpy) before touching data. If scvi-tools is not installed, the script prints the exact pip install scvi-tools command and exits 0 — it does not estimate, guess, or describe hypothetical training results.
2. Raw integer counts required. Every training script verifies the input layer looks like raw counts and refuses to train on normalized/log data with a clear error, rather than letting scvi-tools fail deep inside training or silently produce meaningless output.
3. Report the device honestly. Every script prints whether it is training on GPU (CUDA/MPS) or CPU before starting, since CPU training can be 10-50x slower — never claim GPU training happened when it did not.
4. Never overwrite the input file. All outputs go to a separate directory/file passed by the caller.

When to Use This Skill

Use when the user:

• Names scVI, scANVI, totalVI, PeakVI, MultiVI, DestVI, veloVI, or scArches specifically
• Wants deep-learning-based batch correction / integration for scRNA-seq with strong or nonlinear batch effects
• Has CITE-seq (RNA + surface protein), multiome (RNA + ATAC), or spatial + single-cell reference data and wants a joint/deconvolved model
• Wants to map a new (query) dataset onto an existing trained reference model (reference mapping / label transfer)
• Wants RNA velocity with model-based uncertainty (not just scvelo's point estimates)

NOT for (use other skills instead):

• Plain scanpy-based QC, normalization, clustering, marker genes, or Harmony batch correction with no deep learning involved -> tooluniverse-single-cell
• Running scVI or scANVI as a shared, remote GPU-backed MCP tool for other callers (not local, guided analysis of your own file) -> setup-scvi-remote-tool / setup-scanvi-remote-tool
• Bulk RNA-seq DESeq2 -> tooluniverse-rnaseq-deseq2
• FASTQ-level QC/trimming -> tooluniverse-fastq-qc

Model Selection

See references/model_selection.md for the full decision tree. Quick version:

Data	Model
scRNA-seq, no labels, want integration	scVI
scRNA-seq, partial/full cell-type labels	scANVI
CITE-seq (RNA + protein)	totalVI
scATAC-seq only	PeakVI
Multiome (RNA + ATAC)	MultiVI
Spatial + single-cell reference	DestVI
RNA velocity (spliced/unspliced)	veloVI
Map new data onto an existing reference	scANVI + load_query_data

Run python scripts/validate_adata.py your.h5ad --batch-key batch --suggest to get a data-driven recommendation instead of guessing.

Workflow

# 1. Check the environment and the data
python scripts/validate_adata.py raw.h5ad --batch-key batch --suggest

# 2. Prepare: QC, HVG selection, stash raw counts in a 'counts' layer
python scripts/prepare_data.py raw.h5ad prepared.h5ad --batch-key batch

# 3. Train the model chosen in step 1
python scripts/train_model.py prepared.h5ad results/ --model scvi --batch-key batch

# 4. Differential expression (posterior-based, not scanpy's rank_genes_groups)
python scripts/differential_expression.py results/scvi_model results/adata_scvi_trained.h5ad \
    de.csv --groupby leiden

# 5. (Optional) Map a new dataset onto a trained scANVI reference
python scripts/transfer_labels.py results/scanvi_model prepared.h5ad query.h5ad results/

For totalVI/MultiVI/DestVI's modality-specific input shapes, and DestVI's two-stage fit, see references/multimodal_workflows.md — do not guess the input format; getting it wrong is the most common failure mode for these three models.

Reference Documentation

• references/model_selection.md — full decision tree, minimum data per model
• references/scrna_integration.md — scVI/scANVI integration details, when to prefer this over Harmony
• references/multimodal_workflows.md — totalVI, MultiVI, DestVI input shapes and the DestVI two-stage fit
• references/troubleshooting.md — install issues, NaN losses, GPU/OOM, key mismatches, silent-failure checklist