FADVI: Factor Disentanglement Variational Inference

FADVI is a deep learning method for single-cell omics and spatial transcriptomics analysis that disentangles batch-related variation, label-related variation, and residual variation using adversarial training and cross-correlation penalties.

Read the documentation for usage and demo.

Features

• Factor Disentanglement: Separates batch effects, cell type effects, and residual variation in single-cell and spatial data
• Integration with scvi-tools: Built on top of the scvi-tools framework for scalable analysis
• Batch Correction: Removes unwanted batch effects (including diverse spatial transcriptomics technologies) while preserving biological signal
• Cell Type Classification: Performs supervised learning for cell type prediction
• Outstanding integration performance: FADVI consistently outperforms state-of-the-art integration methods in benchmarking

Installation

Install from PyPI

pip install fadvi

Quick Start

import scanpy as sc
import scvi
from fadvi import FADVI

# Load your single-cell data
adata = sc.read_h5ad("your_data.h5ad")

# Setup the model
FADVI.setup_anndata(
    adata,
    batch_key="batch",
    labels_key="cell_type",
    unlabeled_category="Unknown",
    layer="counts"
)

# Create and train the model
model = FADVI(adata)
model.train(max_epochs=30)

# Get latent representations
latent_l = model.get_latent_representation(representation="label")
latent_b = model.get_latent_representation(representation="batch")

# Get label predictions
prediction_label = model.predict(prediction_mode="label")

Model Architecture

FADVI uses a variational autoencoder architecture with three latent subspaces:

• z_b: Batch-related latent factors
• z_l: Label-related latent factors
• z_r: Residual latent factors

The model uses adversarial training and cross-covariance penalty to ensure proper disentanglement between these factor subspaces.

Citation

If you use FADVI in your research, please cite our preprint

@article{fadvi2025,
  title={FADVI: disentangled representation learning for robust integration of single-cell and spatial omics data},
  author={Wendao Liu, Gang Qu, Lukas M. Simon, Fabian J. Theis, Zhongming Zhao},
  journal={bioRxiv},
  year={2025}
}