Back to Models

PeakVI

🧬 scATAC-Specific

Peak Variational Inference

ATAC

Deep generative VAE model for single-cell chromatin accessibility peak analysis

Publications

PeakVI: A deep generative model for single-cell chromatin accessibility analysis

Ashuach et al.2022
Complexity
★★☆
moderate
Interpretability
★★★
high
Architecture
Sparse VAE
Latent Dim
10

Sparse Peak Accessibility Modeling

PeakVI models scATAC-seq peak data using a noise model tailored to the sparsity and binary nature of accessibility with full VAE reconstruction

Main Idea

Learn informative latent representations by modeling peak accessibility with appropriate probabilistic framework and reconstruction

Key Components

Sparse Encoder

Efficiently handles sparse peak matrices

Accessibility Noise Model

Models peak presence/absence with appropriate distribution

Batch Correction

Corrects technical batch effects in ATAC data

Bernoulli Decoder

Reconstructs peak accessibility probabilities

Mathematical Formulation

p(x|z,s) = Bernoulli(ρ); ρ learned from latent; X̂ = Decoder(z)

Loss Functions

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

Data Flow

Peak Data (Sparse) → Encoder → Latent Space → Bernoulli Decoder → Reconstructed Peak Probabilities

Architecture Details

Architecture Type

VAE for Sparse Binary Data

Input/Output Types

peak → reconstruction

Key Layers

SparseEncoderBernoulliDecoderBatchLayer

Frameworks

PyTorchJAX

Tags

vaeatac-seqsparsechromatingenerative