Source: https://github.com/aipoch/medical-research-skills

When to Use

Use scvi-tools when you need probabilistic, model-based single-cell analysis beyond standard pipelines (e.g., beyond typical Scanpy workflows), such as:

1. Batch correction and dataset integration for scRNA-seq using a probabilistic latent space (e.g., scVI).
2. Transfer learning / semi-supervised annotation when you have partial labels or want to map new data onto a reference (e.g., scANVI).
3. Uncertainty-aware differential expression where effect sizes and posterior uncertainty matter (Bayesian DE).
4. Multimodal integration across RNA+protein (CITE-seq) or RNA+ATAC (multiome), including paired/unpaired settings (e.g., totalVI, MultiVI).
5. Specialized modalities such as ATAC-seq, spatial transcriptomics deconvolution/mapping, doublet detection, methylation, or RNA velocity.

Key Features

• Unified model API: setup_anndata(...) → Model(adata) → train() → get_*() across model families.
• Probabilistic latent representations for integration, denoising, and downstream clustering/visualization.
• Explicit covariate handling (batch, donor, technical factors) via setup_anndata.
• Bayesian differential expression with posterior-based hypothesis testing and effect-size thresholds.
• Multi-omics models for joint learning across modalities (RNA/protein, RNA/ATAC; paired or unpaired).
• AnnData-first integration with the Scanpy ecosystem for downstream neighbors/UMAP/clustering.
• GPU acceleration via PyTorch (when available).

Model catalogs by modality (for reference):

• scRNA-seq: references/models-scrna-seq.md (scVI, scANVI, AUTOZI, VeloVI, contrastiveVI, …)
• ATAC-seq: references/models-atac-seq.md (PeakVI, PoissonVI, scBasset, …)
• Multimodal: references/models-multimodal.md (totalVI, MultiVI, MrVI, …)
• Spatial: references/models-spatial.md (DestVI, Stereoscope, Tangram, scVIVA, …)
• Specialized: references/models-specialized.md (Solo, CellAssign, MethylVI/MethylANVI, CytoVI, …)

Dependencies

• scvi-tools (latest compatible with your environment)
• python>=3.9
• pytorch>=2.0
• pytorch-lightning>=2.0 (or lightning depending on scvi-tools version)
• anndata>=0.8
• scanpy>=1.9

Installation example:

uv pip install scvi-tools
# Optional GPU extras (package extra name may vary by platform/version)
uv pip install "scvi-tools[cuda]"

Example Usage

A complete runnable example using scVI for batch correction + latent embedding, then Scanpy for neighbors/UMAP/clustering:

import scanpy as sc
import scvi

# 1) Load example data (AnnData)
adata = scvi.data.heart_cell_atlas_subsampled()

# 2) Minimal preprocessing (keep raw counts available)
sc.pp.filter_genes(adata, min_counts=3)
sc.pp.highly_variable_genes(adata, n_top_genes=1200)

# 3) Register AnnData for scVI (raw counts + covariates)
scvi.model.SCVI.setup_anndata(
    adata,
    layer="counts",                 # raw counts layer (not log-normalized)
    batch_key="batch",              # batch column in adata.obs
    categorical_covariate_keys=["donor"],
    continuous_covariate_keys=["percent_mito"],
)

# 4) Train model
model = scvi.model.SCVI(adata)
model.train()

# 5) Extract outputs
adata.obsm["X_scVI"] = model.get_latent_representation()
adata.layers["scvi_normalized"] = model.get_normalized_expression(library_size=1e4)

# 6) Downstream analysis with Scanpy
sc.pp.neighbors(adata, use_rep="X_scVI")
sc.tl.umap(adata)
sc.tl.leiden(adata)

# Optional: uncertainty-aware differential expression
de = model.differential_expression(
    groupby="cell_type",
    group1="TypeA",
    group2="TypeB",
    mode="change",
    delta=0.25,
)
print(de.head())

Model persistence:

model.save("./scvi_model", overwrite=True)
model2 = scvi.model.SCVI.load("./scvi_model", adata=adata)

Implementation Details

• Core approach: deep generative modeling with variational inference (typically VAE-style architectures) to learn a latent representation and a likelihood model for counts.
• Data requirements: models generally expect raw counts (not log-normalized values). Provide counts via layer="counts" or ensure adata.X contains counts.
• Covariate registration: technical factors (e.g., batch_key, donor, QC metrics) are incorporated through setup_anndata, enabling the model to learn representations that reduce unwanted variation.
• Training loop: train() performs amortized inference using neural networks shared across cells; GPU acceleration is used automatically when configured.
• Latent space usage: get_latent_representation() returns batch-corrected embeddings suitable for neighbors/UMAP/clustering in Scanpy.
• Differential expression: differential_expression(...) performs posterior-based comparisons; parameters like:
  • mode="change": composite hypothesis testing on changes
  • delta: minimum effect size threshold
    help control practical significance and uncertainty-aware decisions.
    See references/differential-expression.md for interpretation guidance.
• Model selection by modality: choose the model family based on data type (e.g., scVI/scANVI for scRNA-seq, totalVI for CITE-seq, MultiVI for RNA+ATAC, DestVI for spatial deconvolution). For details, see the corresponding references/models-*.md files.
• Theory background: variational inference, amortized inference, and probabilistic modeling foundations are summarized in references/theoretical-foundations.md.