scvi-tools

scvi-tools (single-cell variational inference tools) is a package for end-to-end analysis of single-cell omics data. It consists of two primary components:

1. Model Interface: Provides an easy-to-use interface for a range of existing probabilistic models for single-cell omics, such as scVI, scANVI, and totalVI.
2. Model Development Tools: Provides tools to build new probabilistic models, powered by PyTorch, PyTorch Lightning, and Pyro.

totalVI (scvi.model.TOTALVI) is a probabilistic generative model designed for CITE-seq data, which contains paired RNA (scRNA-seq) and protein (CITE-seq) expression measurements. It is designed to be scalable to very large datasets (over 1 million cells) and can handle multiple batches and missing protein data.

Key Capabilities:

• Dimensionality reduction
• RNA and protein normalization and denoising
• Differential expression analysis
• Data simulation

Limitations:

• Requires a GPU for efficient inference.
• The balance between RNA and protein in the low-dimensional representation can be difficult to interpret.

MultiVI (scvi.model.MULTIVI) is a multimodal generative model designed to integrate multiome, scRNA-seq, and scATAC-seq data. It is scalable to very large datasets (over 1 million cells) and can be used for downstream tasks like dimensionality reduction, differential expression/accessibility, and the imputation of missing modalities (e.g., predicting gene expression from accessibility data or vice versa).

Note: Fast inference effectively requires a GPU.

scvi-tools is a package for probabilistic modeling and analysis of single-cell, multi, and spatial omics data. It is built on top of PyTorch and AnnData.

Key analysis tasks supported include:

• Dimensionality reduction
• Data integration
• Automated annotation
• Factor analysis
• Doublet detection
• Spatial deconvolution

All model implementations provide a high-level API that interacts with Scanpy and includes standard features like save/load functions and GPU acceleration.

MethylVI (scvi.external.METHYLVI) is a generative model designed for scBS-seq data. It is scalable to very large datasets (over 1 million cells) and is capable of performing various downstream tasks like dimensionality reduction and differential methylation analysis.

Input Requirements: MethylVI requires two count matrices for each methylation context $C$ (e.g., CpG vs non-CpG):

1. Methylated counts ($Y_{mc}$): The number of methylated cytosines in a region for a cell.
2. Total coverage counts ($Y_{cov}$): The total number of cytosines profiled at a given region in a cell.

An optional vector of categorical covariates $S$ (e.g., batch, donor) can also be provided.

Limitations:

• Effectively requires a GPU for fast inference.
• The latent space is not interpretable in the same way as linear methods.

ResolVI is a generative model for single-cell resolved spatial transcriptomics designed to address noise and bias caused by segmentation errors, unspecific background, and limited spatial resolution. It is scalable to datasets exceeding 1 million cells.

Requirements and Limitations:

• GPU Required: Fast inference effectively requires a GPU.
• Input Data: Requires spatially resolved RNA-seq count matrices downstream of cellular segmentation and molecule assignments (e.g., from sequencing or imaging).
• Resolution: Assumes single cells are observed; it is not intended for low-resolution ST like Visium or Slide-Seq.
• Interpretability: The latent space is not interpretable in the same way as linear methods.

Note: Starting with scVI-Tools v1.5, this model is part of scVIVA-Tools and is no longer being maintained within the main scvi-tools repository.

scvi-tools provides several specialized workflows for analyzing and integrating multimodal single-cell data. Key workflows include:

• totalVI: Used for analyzing CITE-seq datasets, integrating CITE-seq and scRNA-seq datasets, and mapping query CITE-seq data to a trained reference model.
• scArches: Used for mapping cells from a query dataset to the latent space of a reference dataset (transfer learning).
• MultiVI: Used for joint analysis of paired and unpaired multi-omic data.
• TotalANVI: A semi-supervised approach for CITE-seq data that leverages partial cell type annotations for label prediction, protein imputation, and differential abundance.
• DiagVI: Used for integrating spatial proteomics with single-cell transcriptomics, or spatial transcriptomics with single-cell transcriptomics.

Decipher is a probabilistic model for single-cell RNA-seq data that learns two distinct latent representations: a low-dimensional, interpretable space v (default dim_v=2) for visualization, and an intermediate space z (default dim_z=10) for detailed cell-state information.

Current Limitations:

• The implementation currently only registers the count matrix (and an optional raw-count layer) via setup_anndata. It does not yet expose condition covariates.
• The implementation is still being aligned with the original method's full feature set.

>>> Decipher.setup_anndata(adata, layer="counts")
>>> model = Decipher(adata)
>>> model.train()