Informed, but Not Always Improved: Challenging the Benefit of Background Knowledge in GNNs

Paper: https://arxiv.org/abs/2505.11023

Installation

Dependencies are listed in the env.yml file. You can create a conda environment with:

conda create --yes -f env.yml -n dev
conda activate dev

Dependencies for individual models are listed in their respective folders. A separate conda environment is recommended for each model.

Models

The cancer subtype detection models [1,2] are located under models directory. Please download the datasets from the respective repositories.

Li24 [1]

Repo: https://github.com/NabaviLab/Multimodal-GNN-for-Cancer-Subtype-Clasification

conda create --yes -f models/li24/env.yml -n li24
conda activate li24
cd models/li24
bash experiment_job.sh

MPK-GNN [2]

Repo: https://github.com/Xiaoshunxin/MPK-GNN

conda create --yes -f models/mpk-gnn/env.yml -n mpk-gnn
conda activate mpk-gnn
cd models/mpk-gnn
bash experiment_job.sh

Synthetic Experiments

The synthetic data generation and experiments are located under src/kill_gnn/synth directory.

conda activate dev
python src/kill_gnn/synth/experiment.py

References

[1] B. Li and S. Nabavi, “A multimodal graph neural network framework for cancer molecular subtype classification,” BMC Bioinformatics, vol. 25, no. 1, p. 27, Jan. 2024, doi: 10.1186/s12859-023-05622-4.

[2] S. Xiao, H. Lin, C. Wang, S. Wang, and J. C. Rajapakse, “Graph Neural Networks With Multiple Prior Knowledge for Multi-Omics Data Analysis,” IEEE J. Biomed. Health Inform., vol. 27, no. 9, pp. 4591–4600, Sept. 2023, doi: 10.1109/JBHI.2023.3284794.