GVAE-TME

Graph Variational Autoencoder for Tumor Microenvironment Analysis. Learns spatial and molecular cell representations via a dual-graph GAT encoder with cell-adaptive gating, ZINB expression decoder, and attention-based patient pooling for treatment response prediction.

Quick Start

pip install -r requirements.txt

python train.py --config local   
python train.py --config full  

Outputs saved to outputs/.

Structure

• src/config.py - LOCAL and FULL configs, device detection (cuda/mps/cpu)
• src/model.py - GVAEModel with cell-adaptive gate, dual decoder, response predictor
• src/trainer.py - Two-phase training with reconstruction safeguards
• src/analysis.py - Rare cell detection, clustering, prediction
• src/data_utils.py - Graph construction from AnnData
• train.py - End-to-end pipeline

Key Features

• Cell-adaptive hybrid graph fusion (learned gate blends molecular and spatial graphs per-cell)
• ZINB expression decoder + adjacency decoder
• KL-based rare cell detection
• Attention-based patient pooling for treatment response
• Two-phase training prevents latent collapse
• Patient-level train/val/test splits (no data leak in Phase 2)
• Contrastive loss with adaptive negative mining

Requirements

• Python 3.9+
• PyTorch 2.0+
• PyTorch Geometric 2.3+
• scanpy, anndata, leidenalg
• See requirements.txt for full list