scVI / CITE-seq — Variational Autoencoders for Single-Cell Data

This repository contains a minimal, self-contained implementation of variational autoencoder models for single-cell data, inspired by scVI and its multimodal extension for CITE-seq (RNA + protein).

The code is written in PyTorch and follows a clear separation between data handling, model definition, probabilistic modeling, training, and evaluation.

---

Context

This project was developed as part of the course:

Introduction to Probabilistic Graphical Models
Pierre Latouche — Pierre-Alexandre Mattei
MVA (Master Mathématiques, Vision, Apprentissage)

Course webpage:
https://lmbp.uca.fr/~latouche/mva/IntroductiontoProbabilisticGraphicalModelsMVA.html

The implementation aims to connect:

• probabilistic graphical modeling concepts (latent variables, ELBO),
• variational inference,
• deep generative models applied to real biological data.

---

Models implemented

• SCVI
  RNA-only variational autoencoder with:

  • latent biological variable z
  • latent library size variable l
  • batch-specific priors
  • ZINB likelihood for gene expression

• CiteVI
  Multimodal extension for CITE-seq data:

  • shared latent variable z
  • RNA modeled with ZINB
  • proteins (ADT) modeled with a negative binomial distribution

Both models are trained by maximizing an evidence lower bound (ELBO).

---

Repository structure

src/
├── data.py              # dataset definition and data loaders
├── building_blocks.py   # neural network components (encoders / decoders)
├── distributions.py     # KL terms, likelihoods, ELBOs
├── models.py            # SCVI and CiteVI models
├── training.py          # training loops
└── evaluation.py        # diagnostics and visualization

notebooks/
├── scVI.ipynb
└── citeVI.ipynb

data/
├── README.md            # data download instructions
├── cortex_rna.txt       # RNA-only dataset (not versioned)
└── pbmc_citeseq/        # CITE-seq dataset (not versioned)

Raw datasets are not tracked in the repository.

---

Data

• Cortex scRNA-seq (RNA-only): Used as a benchmark dataset for scVI-style models.
• PBMC 2019 CITE-seq (RNA + protein): Human PBMC dataset combining transcriptomic and surface protein measurements.

Detailed download instructions are provided in data/README.md.

---

Usage

Training and evaluation are demonstrated in the notebooks located in the notebooks/ directory.

Typical workflow:

1. Download the data (see data/README.md)
2. Run the corresponding notebook
3. Train the model using the provided training loop
4. Evaluate the learned latent space via reconstruction metrics and t-SNE

---

Scope and limitations

This repository is intended for:

• educational purposes,
• understanding variational inference in practice,
• illustrating probabilistic modeling choices for single-cell data.

It is not meant to be a production-ready alternative to scvi-tools.

---

References

• Lopez R. et al., Deep generative modeling for single-cell transcriptomics
• Stuart T. et al., Comprehensive Integration of Single-Cell Data, Cell (2019)
• Kingma & Welling, Auto-Encoding Variational Bayes