BioGAN: Enhancing Transcriptomic Data Generation with Biological Knowledge

BioGAN is a novel generative framework that integrates Graph Neural Networks (GNNs) into a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) to generate biologically realistic synthetic transcriptomics data. The model leverages known gene-gene relationships to guide the generation process, aiming to improve the biological plausibility, fairness, and privacy of synthetic omics data.

This project accompanies the manuscript:

"BioGAN: Enhancing Transcriptomic Data Generation with Biological Knowledges" — Francesca Pia Panaccione, Sofia Mongardi, Marco Masseroli and Pietro Pinoli (under review)

---

🧬 Motivation

Modern AI tools in genomics are limited by data scarcity, heterogeneity, acquisition costs, and privacy regulations (e.g., GDPR, HIPAA). BioGAN addresses these issues by:

• Synthesizing realistic transcriptomic profiles conditioned on biological structure.
• Incorporating prior biological knowledge through GNNs to improve fidelity.
• Providing a versatile architecture adaptable to various omics settings.

---

Key Features

• Graph-aware generator: Integrates GNN layers to model regulatory interactions.
• High-dimensional support: Designed for transcriptomic data with tens of thousands of features.
• Robust validation: Multi-faceted evaluation pipeline including:
  • Classification accuracy on downstream tasks
  • Distributional similarity metrics (e.g., Wasserstein distance)
  • Feature-level consistency checks

---

📁 Repository Structure

src/
├── BIOGAN_H2GCN/          # Experimental variant of BioGAN using H2GCN architecture
├── BioGAN_GCN/            # Core BioGAN model implementation based on GCN
├── wpgan/                 # Implementation of Wasserstein GAN with Gradient Penalty
├── metrics/               # Evaluation metrics for synthetic data quality
├── utils/                 # Utility functions (e.g., graph processing, logging)
├── data_loader.py         # Scripts for preprocessing transcriptomic data
├── losses.py              # Custom loss functions (Wasserstein, gradient penalty, etc.)
├── train_model.py         # Main training script for the BioGAN model