LOVENet: Large Optimized Vector Embeddings Network for Drug Repurposing

This codebase provides a implementation for the paper LOVENet: Large Optimized Vector Embeddings Network for Drug Repurposing at NeurIPS 2023 Workshop.

Project Overview

A new framework maximizing the synergistic effects of knowledge graphs and large language models (LLMs) to discover novel therapeutic uses for pre-existing drugs. Specifically, our approach fuses information from pairs of embedding from Llama2 and heterogeneous knowledge graphs to reveal complex relations between drugs and diseases.

Project Structure

The repository comprises various Python scripts, each catering to a specific aspect of the LOVENet model:

• LLM/representation.py: Creating the embedding and feature representation from LLM.
• GNN/main.py: The main script that orchestrates the workflow of the model.
• GNN/model.py: Contains the architecture of the GNN and fusion with LLM representations.
• GNN/layer.py: Defines the custom layers used in the neural network.
• GNN/utils.py: Provides utility functions for data handling and auxiliary tasks.
• GNN/args.py: Handles parsing of command-line arguments.
• GNN/load_data.py: Loads and preprocesses the data for the model.

Getting Started

First to extract the embedding of drug and disease run the following:

1. LLM Embedding:

   • Execute representation.py in LLM/.
   • Set inputs:
     • path: Directory ("B_data/").
     • drug_disease_path: Drug-disease relations file.
     • disease_path: Disease ID to name mapping file.
     • drug_path: Drug ID to name mapping file.
     • directory_path: Saved model directory.

2. Run the main code First the following arguments should be specified:

python main.py --device_id 1 \
               --dataset /B-dataset \ # Target dataset
               --llm_rep_path /llm_rep_final_dataset.pkl \ # LLM representation extracted in step one
               --Disease_mapper /Disease_B_mapper.pkl \ # Disease ID to name mapping file
               --Drug_mapper /Drug_B_mapper.pkl \ # Drug ID to name mapping file
               --saved_path /B-dataset \
               --seed 0 \
               --print_every 1 \
               --nfold 5 \
               --epoch 200 \
               --batch_size 2048 \
               --learning_rate 0.005 \
               --weight_decay 0.0 \
               --check_metric auc \
               --k 15 \ # Top k similarities used as related drug disease
               --aggregate_type BiTrans \
               --topk 1 \ # Selecting k prediction result
               --hidden_feats 64 \
               --num_layer 2 \
               --dropout 0.2 \
               --batch_norm \
               --skip False \
               --mil False \
               --ins_predict False

Prerequisites

• Python 3.10.12
• Ubuntu 22.04.1 LTS
• CUDA Version 12.0
• NVIDIA RTX 3090 Ti GPU

Installing Dependencies

Run pip install -r requirements.txt to install the required Python libraries.

Contact

For any queries regarding LOVENet, please contact the authors:

Sepehr Asgarian - sasgarian@klick.com Sina Akbarian - sakbarian@klick.com Jouhyun Jeon - cjeon@klick.com