SMOGONET

We present SMOGONET (Spiking Multi-Omics Graph cOnvolutional NETwork), an adaptation of MOGONET that incorporates principles from spiking neural networks as a spiking framework for multi-omics data integration and classification, with a specific focus on breast cancer diagnosis.

REPOSITORY STRUCTURE:

• SMOGONET/: main folder containing the code for the SMOGONET model.
• requirements.txt: file containing the dependencies needed to run the code.
• reference/: folder containing referecne papers.
• dumpdrawer/: folder containing initial testing codes and files to get accustomed to the material.
• MOGONET/: folder containing original MOGONET implementation by Wang et al.
• SpikingGCN/: folder containing code from the Spiking GCN by Zhu et al.
• SMOGONET.pdf: scientific paper style report on the work done.

USAGE:

The code has been developed using python 3.96, no other versions have been tested.

Dependencies are listed in the requirements.txt file. To install them, run:

pip install -r requirements.txt

Training SMOGONET:

A full training run (until convergence) can be performed running SMOGONET/run.py (-W ignore to suppress some warnings from orginal implementation of MOGONET).

Average accuracy, F1 and F1-macro scores over 10 run are printed at the end of the training.

python -W ignore SMOGONET/run.py

Training SMOGONET with EA initialization:

A full run of the EA for model parameters initialization can be performed running SMOGONET/run_ea.py.

A full training run of the best model found by the EA is then performed and the average accuracy, F1 and F1-macro scores over 10 run are printed at the end of the training.

python -W ignore SMOGONET/run_ea.py

Notebooks:

• SMOGONET/test_joined.ipynb: notebook to train the joined architecture, 100 epochs pretraining for the SGCNs and 500 epochs for the SVCDN. The best model is saved in the /models folder and can be loaded to just perform testing run.
• SMOGONET/test_ea.ipynb: notebook to perform EA for model parameters initialization, the best model the goes through a full training run like in test_joined.

AUTHORS:

• Gabriele Quaranta
• Giulio Maselli
• Mattia Pecce
• Alperen Bitirgen

REFERENCES:

Wang et al, "MOGONET integrates multi-omics data using graph convolutional networks allowing patient classification and biomarker identification", 2021.

• https://github.com/txWang/MOGONET

Zulun Zhu et al, "Spiking Graph Convolutional Networks," 2022.

• https://github.com/ZulunZhu/SpikingGCN

TODOS:

Documentation:

• Update readme
• Project report

Utils:

• Modify 'compute_spiking_node_representation' to have the encoding function as parameter
• Clean imported utils from original mogonet

Models Architecture:

• Use the SGCN spiking ouput as they are for SVCDN instead of dec->ccdt->enc

Training / Testing:

• Define train/test functions in own .py not ipynb for single model
• [ ] Rn all the data is passed at once, add batch processing (original MOGONET uses full batch)
• Save/load single models
• Save best ea model checkpoints
• Make .py to run joined from terminal
• Make .py to run ea from terminal

Possible Experiments:

• EA for model params ?