PoissonVI

Back to Models

Deep generative VAE model for quantitative scATAC-seq fragment counts using Poisson likelihood

Publications

Joint probabilistic modeling of single-cell multi-omic data with totalVI

Gayoso et al.2021

Complexity

★★☆

moderate

Interpretability

★★★

high

Architecture

Poisson VAE

Latent Dim

Fragment Count Modeling

PoissonVI uses Poisson likelihood to model scATAC-seq fragment counts, capturing quantitative accessibility information with full VAE reconstruction

Main Idea

Learn latent representations from scATAC fragment counts by modeling the Poisson-distributed count data with reconstruction

Key Components

Fragment Count Encoder

Encodes quantitative accessibility from fragment counts

Poisson Likelihood

Models fragment count distribution

Batch Correction

Handles technical variation in ATAC experiments

Poisson Decoder

Reconstructs fragment counts from latent representation

Mathematical Formulation

p(x|z,s) = Poisson(λ); λ = exp(decoder(z,s)); X̂ = Decoder(z)

Loss Functions

ELBO

E_q[log Poisson(x|λ(z))] - KL(q(z|x)||p(z))

Data Flow

Fragment Counts → Encoder → Latent Space → Poisson Decoder → Reconstructed Fragment Counts

Architecture Details

Architecture Type

VAE with Poisson Likelihood

Input/Output Types

peak → reconstruction

Key Layers

EncoderPoissonDecoderBatchLayer

Frameworks

PyTorchJAX

Poisson Variational Inference