ICoN

Integration using Co-attention across biological Networks

Generate protein embedding

1. To set up the environment, run the commands in icon_env.txt file.

2. To run ICoN with the best hyperparameters on yeast networks, run (from project folder)

   python code/main_icon.py code/config/icon_best_yeast.json
   

3. You can change the hyperparameters or the networks you are integrating by editing or creating a new config file in code/config.

   i. To run co-attention based ablation study on ICoN use the config file: icon_no-coattn_yeast.json

   ii. To run noise induction module based ablation study on ICoN use the config file: icon_no-noise_yeast.json

   iii. To run ICoN on yeast networks each corrupted with 30% false positive and false negative edges, use: icon_best_noisy03_yeast.json

   iv. To run ICoN on yeast networks each corrupted with 50% false positive and false negative edges, use: icon_best_noisy05_yeast.json

Evaluate generated embedding

We utilized BIONIC-evals [1] to evaluate embeddings generated by ICoN and compare ICoN with other unsupervised biological network integration models. The details for reproducing the figures in the paper are outlined below.

To evaluate ICoN in terms of downstream tasks: i. gene module detection, ii.gene coannotation prediction, and iii. gene function prediction we utilized BIONIC-evals.

We have provided the datasets (embeddings generated from 1 run of models), standards, config, and script used in creating figures in the manuscript of ICoN. To reproduce these figures (for one run) follow the instructions below:

1. First install BIONIC-evals following the instructions given in BIONIC-evals

2. Place our provided <script> folder inside <BIONIC-evals/bioniceval>.

3. Now replace the following folders in <BIONIC-evals/bioniceval> with our provided folders here: i. datasets ii. config iii. standards

   Note: We have provided some files in .zip format. Please extract them before proceeding.

i. Comparative analysis between ICoN and other network integration models (and input networks):

1. Run BIONIC-evals with <config/single_runs/yeast.json>
2. Then run:

   python paper_plots.py <bionic_eval_results_folder>
   

ii. Ablation study of ICoN:

Co-attention

1. Run BIONIC-evals with <config/single_runs/ablation_nocoattn.json>
2. Then run:

   python ablation_study_coattn.py <bionic_eval_results_folder>
   

Noise induction module

1. Run BIONIC-evals with <config/single_runs/ablation_nonoise.json>
2. Then run:

   python ablation_study_noise.py <bionic_eval_results_folder>
   

iii. Co-attention coeffcient:

Run:

python co_attention_weights-lineplot.py <bionic_eval_datasets_folder>

iv. Robustness to noise:

1. Run BIONIC-evals with <config/single_runs/noisyinput_icon_bionic_union.json>
2. Then run:

   python noise_robustness.py <bionic_eval_results_folder>
   

References

[1] Duncan Forster and congyoua (2022) “duncster94/BIONIC-evals: v0.2.0”. Zenodo. doi: 10.5281/zenodo.6964943.