How chenzrg/gesubnet works
Main Components
a. Data Ingestion Layer
• “data/” directory with four cancer datasets (BRCA, GBM, LGG, OV)
• File formats: zipped gene‐expression data
b. Data Loading & Preprocessing
• dataloader.py
• utils.py (feature normalization, graph construction helpers)
c. Model Definitions
• model/GAT.py, model/GCN.py, model/NeoGNN.py, model/SELM_NN.py
d. Training Orchestration
• GeSubNet_main.py (experiment entry point)
• trainer.py (training loop, evaluation, checkpointing)
• GeSubNet_para_log.py (hyperparameter logging)
e. Persistence & Logging
• log/ (model checkpoints, intermediate figures, result spreadsheets)
f. Case Analysis & Reporting
• Case Analysis/ (post-training network visualizations, example outputs)
Relationships & Data Flows
• Data files ─► dataloader.py ─► PyTorch DataLoader ─► trainer.py
• trainer.py ─► model/*.py (calls forward/backward on chosen GNN)
• trainer.py ─► utils.py (for loss calculation, metrics)
• trainer.py ─► logs (checkpoints: GRN.pt; matrices: PDF; results: XLSX)
• GeSubNet_main.py ties together dataloader, trainer, model selection, and logging
• Post-training, Case Analysis module reads saved models and results to generate figures
Architectural Patterns & Principles
• Modular separation: data loading vs. model logic vs. training orchestration
• Pipeline pattern: clear, linear flow from raw data to final outputs
• Configuration-driven: GeSubNet_para_log.py captures hyperparameters, enabling reproducible experiments
• Persistence layer: file-system based checkpoints for reproducibility and analysis
Technologies & Frameworks
• Python 3.x
• PyTorch (torch, torchvision, torchaudio)
• Standard Python libs (numpy, pandas, matplotlib, networkx, etc., as in requirements.txt)
• File I/O (zip, PDF generation, Excel via pandas/xlsxwriter)