This repository contains the implementation of our feature learning algorithm, part of a manuscript submission for IEE-BIBM 2022.

• numpy>=1.23.0
• torch>=1.9.0
• torch-geometric>=2.0.3
• optuna>=2.10.1
• scikit-learn>=1.1.2
• scipy>=1.8.1
• lifelines>=0.27.0
• tqdm>=4.64.0

  All requirements are listed in requirements.txt and can be installed together using the below pip installation command

pip install -r requirements.txt

In this repository, we have provided all the raw data and a sample version of the processed data used for training the model.

• data/raw: Contains raw (unprocessed) data used including all the nodes and edges used in the heterogeneous knowledge graph.
• data/processed: Contains a sample version of processed train/val/test data that can be used for training our multimodal algorithm.

The following scripts can be used to train the model on IPF-specific data provided in the data directory.
However, these scripts can be used to train other disease-related data by simply plugging-in the new input files.
Please refer to the jupyter notebook process_data.ipynb for generating the processed data file.

First, the script pre_train_kg_attn.py can be used for unsupervised training of the KG input, using the below syntax.

python  pre_train_kg_attn.py 
        --dim_size 32 
        --num_neg 10.0 
        --num_epochs 100 
        --prop_val 0.2 
        --prop_test 0.2 
        --patience 10 
        --outdir gold

• --dim_size: Hidden dimensions of the KG-encoder i.e., the dimensionality of final trained features.
• --num_neg: Number of negative edges samples for each positive edge for the link-prediction objective.
• --prop_val and --prop_test: Proportion of validation edges and test edges respectively from the input KG.
• --num_epochs: Maximum number of training epochs.
• --patience: Tolerance value to be used during the early-stopping mechanism.
• --outdir: The IPF network type indicating which set of PPI edges to be used for the KG.

Next, the script save_pretrain_embeddings.py can be used to generate and save the gene embeddings coming out of the unsupervised pretraining of the KG.

Finally, the python script train_IPF_mse_pre.py can be used to train the final multimodal learner. The below syntax can be used to train on the IPF-specific data.

python train_IPF_mse_with_pre.py 
        --model KGNetAttn 
        --save_path saved/training/gold 
        --indir data/processed/ 
        --emb_file saved/pretrain/gold/embedding_dim_32_neg_10.0.npy

• --model: The feature aggregation model/mechanism to be used for combining the individual unimodal gene features for generating multimodal feature matrix.
• --indir: Path to folder containing the processed data files (train/val/test data).
• --emb_file: Path to file containing the pre-trained gene feature inputs from heterogeneous KG input modality.
• --save_path: Path to output folder for saving the final model checkpoint and test metrics.