Back to Models

siVAE

✨ Generative Models

Interpretable Deep Generative Model

RNA

VAE with structured latent space for interpretable single-cell modeling

Publications

siVAE: interpretable deep generative models for single-cell transcriptomes

Choi et al.2023
Complexity
★★☆
moderate
Interpretability
★★★
high
Architecture
Structured VAE
Latent Dim
10
Used in LAIOR Framework

Structured Latent Space

siVAE uses structured priors and regularization to learn interpretable latent representations with full VAE reconstruction

Main Idea

Learn interpretable factors by structuring the latent space with domain knowledge and reconstructing expression

Key Components

Encoder

Maps expression to structured latent space

Structured Prior

Incorporates biological structure into latent space

Interpretable Factors

Each dimension corresponds to biological variation

Negative Binomial Decoder

Models count distribution appropriately for reconstruction

Mathematical Formulation

p(x|z) = NB(μ(z), θ); structured prior on z; X̂ = Decoder(z)

Loss Functions

ELBO
Reconstruction + KL divergence

Data Flow

Expression → Encoder → Structured Latent → NB Decoder → Reconstructed Expression

Architecture Details

Architecture Type

VAE with Structured Prior

Input/Output Types

single-cell → reconstruction

Key Layers

EncoderStructuredPriorNBDecoder

Frameworks

PyTorch

Tags

vaeinterpretablegenerativerna