Back to Models

scVI

✨ Generative Models

Single-cell Variational Inference

RNA

Deep generative VAE model for scRNA-seq with probabilistic inference and batch correction

Publications

Deep generative modeling for single-cell transcriptomics

Lopez et al.2018
Complexity
★★☆
moderate
Interpretability
★★★
high
Architecture
Hierarchical VAE
Latent Dim
10

Probabilistic Generative Modeling

scVI learns a low-dimensional latent representation while explicitly modeling zero-inflation and library size effects common in scRNA-seq

Main Idea

Learn interpretable latent representations by modeling the generative process of scRNA-seq data with explicit noise and batch effects

Key Components

Encoder Network

Maps high-dimensional counts to latent space

Decoder Network

Reconstructs counts from latent representation

Zero-Inflation Module

Handles dropout events specific to scRNA-seq

Batch Effect Correction

Learns and corrects batch-specific parameters

Mathematical Formulation

p(x|z,l,s) = ZINB(μ, θ, π); q(z|x) learned via encoder

Loss Functions

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

Data Flow

Count Data → Encoder → Latent Space → Decoder → Zero-Inflated NB Distribution → Reconstructed Counts

Architecture Details

Architecture Type

Variational Autoencoder with Zero-Inflation

Input/Output Types

single-cell → reconstruction

Key Layers

DenseLayerDispersionLayerZeroInflationLayerDecoder

Frameworks

PyTorchJAX

Tags

vaegenerativebatch-correctionrna