User guide#

scvi-tools is composed of models that can perform one or many analysis tasks. In the user guide, we provide an overview of each model with emphasis on the math behind the model, how it connects to the code, and how the code connects to analysis.

Overview of tasks

scRNA-seq analysis#

Model

Tasks

Reference

scVI

Dimensionality reduction, removal of unwanted variation, integration across replicates, donors, and technologies, differential expression, imputation, normalization of other cell- and sample-level confounding factors

[Lopez et al., 2018]

scANVI

scVI tasks with cell type transfer from reference, seed labeling

[Xu et al., 2021]

LDVAE

scVI tasks with linear decoder

[Svensson et al., 2020]

AUTOZI

for assessing gene-specific levels of zero-inflation in scRNA-seq data

[Clivio et al., 2019]

CellAssign

Marker-based automated annotation

[Zhang et al., 2019]

Solo

Doublet detection

[Bernstein et al., 2020]

scAR

Ambient RNA removal

[Sheng et al., 2022]

contrastiveVI

scVI tasks with contrastive analysis

[Weinberger et al., 2023]

MrVI

Characterization of sample-level heterogeneity

[Boyeau et al., 2024]

ATAC-seq analysis#

Model

Tasks

Reference

PeakVI

Dimensionality reduction, removal of unwanted variation, integration across replicates, donors, and technologies, differential expression, imputation, normalization of other cell- and sample-level confounding factors

[Ashuach et al., 2022]

scBasset

Dimensionality reduction, removal of unwanted variation, integration across replicates, donors, and technologies, imputation

[Yuan and Kelley, 2022]

PoissonVI

Dimensionality reduction, removal of unwanted variation, integration across replicates, donors, and technologies, differential expression, imputation, normalization of other cell- and sample-level confounding factors

[Martens et al., 2023]

Multimodal analysis#

CITE-seq#

Model

Tasks

Reference

totalVI

Dimensionality reduction, removal of unwanted variation, integration across replicates, donors, and technologies, differential expression, protein imputation, imputation, normalization of other cell- and sample-level confounding factors

[Gayoso et al., 2021]

Multiome#

Model

Tasks

Reference

MultiVI

Integration of paired/unpaired multiome data, missing modality imputation, normalization of other cell- and sample-level confounding factors

[Ashuach et al., 2023]

Spatial transcriptomics analysis#

Model

Tasks

Reference

DestVI

Multi-resolution deconvolution, cell-type-specific gene expression imputation, comparative analysis

[Lopez et al., 2022]

Stereoscope

Deconvolution

[Andersson et al., 2020]

gimVI

Imputation of missing spatial genes

[Lopez et al., 2019]

Tangram

Deconvolution, single cell spatial mapping

[Biancalani et al., 2021]

General purpose analysis#

Model

Tasks

Reference

Amortized LDA

Topic modeling

[Blei et al., 2003]

Background#

Glossary#

A Model class inherits BaseModelClass and is the user-facing object for interacting with a module. The model has a train method that learns the parameters of the module, and also contains methods for users to retrieve information from the module, like the latent representation of cells in a VAE. Conventionally, the post-inference model methods should not store data into the AnnData object, but instead return “standard” Python objects, like numpy arrays or pandas dataframes.

A module is the lower-level object that defines a generative model and inference scheme. A module will either inherit BaseModuleClass or PyroBaseModuleClass. Consequently, a module can either be implemented with PyTorch alone, or Pyro. In the PyTorch only case, the generative process and inference scheme are implemented respectively in the generative and inference methods, while the loss method computes the loss, e.g, ELBO in the case of variational inference.

The training plan is a PyTorch Lightning Module that is initialized with a scvi-tools module object. It configures the optimizers, defines the training step and validation step, and computes metrics to be recorded during training. The training step and validation step are functions that take data, run it through the model and return the loss, which will then be used to optimize the model parameters in the Trainer. Overall, custom training plans can be used to develop complex inference schemes on top of modules.

The Trainer is a lightweight wrapper of the PyTorch Lightning Trainer. It takes as input the training plan, a training data loader, and a validation dataloader. It performs the actual training loop, in which parameters are optimized, as well as the validation loop to monitor metrics. It automatically handles moving data to the correct device (CPU/GPU).