PoissonVI

PoissonVI#

PoissonVI [Martens et al., 2023] (Python class POISSONVI) is a variational model for quantitative scATAC-seq fragment counts. It follows the peakVI-style single-cell ATAC workflow, but uses a Poisson observation model to analyze fragment counts rather than binarized accessibility.

The advantages of PoissonVI are:

  • It models quantitative fragment-count information, which can retain signal that is discarded when scATAC-seq matrices are binarized.

  • It exposes familiar peakVI-style tasks, including latent representation learning, normalized accessibility, region factors, and differential accessibility.

  • It supports registered batch and covariate annotations through setup_anndata().

The limitations of PoissonVI include:

  • It expects fragment-count-like input. Read-count matrices with strong even-count artifacts should be converted to fragment counts before training.

  • It is designed for chromatin accessibility regions, so differential_expression() is not implemented; use differential_accessibility() instead.

  • The nonlinear latent space is useful for integration and clustering, but individual latent dimensions are not directly interpretable.

Preliminaries#

PoissonVI takes as input a cell-by-region matrix \(X\) with \(N\) cells and \(M\) genomic regions. The values should be quantitative fragment counts. For 10x scATAC-seq data, the tutorial demonstrates converting read counts to approximate fragment counts with reads_to_fragments() and storing them in a layer before calling setup_anndata().

The registered AnnData object may include a batch key, labels, size factors, and other categorical or continuous covariates:

>>> POISSONVI.setup_anndata(adata, layer="fragments", batch_key="batch")
>>> model = POISSONVI(adata)
>>> model.train()

The setup method checks that the registered matrix has fragment-count-like behavior.

Model Overview#

PoissonVI reuses the variational autoencoder machinery used by other scvi-tools models, with defaults chosen to mirror peakVI where appropriate. A cell-level latent variable \(z_i\) captures biological variation in chromatin accessibility. The decoder maps this latent representation, along with registered covariates, to region-specific Poisson rates.

In the implementation, PoissonVI uses VAE with gene_likelihood="poisson", gene-level dispersion settings that are unused by the Poisson likelihood, layer normalization in both encoder and decoder, and LeakyReLU activation functions.

Inference#

PoissonVI uses amortized variational inference. The encoder approximates the posterior distribution over the latent accessibility representation, and the decoder is optimized to reconstruct the registered fragment-count matrix under a Poisson likelihood.

The model also learns region-specific decoder factors. By default, get_normalized_accessibility() temporarily removes the region-factor bias when returning normalized accessibility, which emphasizes cell-specific accessibility variation. Passing normalize_regions=True reintroduces the region factors and produces estimates closer to the observed input scale.

Tasks#

Here we provide an overview of common tasks. Please see POISSONVI for the full API reference.

Dimensionality Reduction#

The latent representation can be stored in adata.obsm and used with Scanpy for neighbors, clustering, and visualization:

>>> adata.obsm["X_poissonvi"] = model.get_latent_representation()

Normalized Accessibility#

get_normalized_accessibility() returns normalized accessibility estimates for cells and regions:

>>> accessibility = model.get_normalized_accessibility()

The method supports posterior sampling, region subsetting, transform_batch, and importance weights. The learned region factors can be retrieved with get_region_factors().

Differential Accessibility#

Differential accessibility is performed with differential_accessibility(). The method supports both "change" and "vanilla" modes, one-sided or two-sided tests, and optional batch correction.