• TreatAgent: A LLM Agent for Drug-Disease Treatment Prediction

  If you like our project, please give us a star ⭐ on GitHub for the latest update.

  

  🎯 Overview

  This system combines advanced computational methods with LLM to predict whether a given drug molecule can effectively treat a specific disease. It employs a multi-agent architecture where specialized expert agents analyze different aspects of drug-disease compatibility and synthesize their findings into a comprehensive assessment.

  🏗️ Architecture

  Pipeline

  

  Core Components

  • Main Pipeline (main.py): Orchestrates the entire evaluation process
  • Multi-Agent System (agent_system.py): Conduct a comprehensive analysis
  • ADMET Analysis (tool1.py): Molecule-level drug property assessment
  • DTI Scoring (tool2.py): Drug-target interaction analysis
  • Disease Risk Assessment (tool3.py): Clinical context evaluation

  Expert Agents

  Agent	Role	Analysis Focus
  Agent 1	ADMET Expert	Molecular properties, bioavailability, safety
  Agent 2	DTI Expert	Drug-target interactions, mechanism validation
  Agent 3	Clinical Expert	Therapeutic development, market potential
  Agent 4	Synthesis Expert	Multi-criteria decision making, final scoring

  🚀 Quick Start

  Installation

  # Clone the repository
  git clone https://github.com/Evfidiw/DrugAgent
  cd TreatAgent
  
  # Install dependencies
  pip install -r requirements.txt
  
  # Set API and URL in config.py
  export API_VALUE="your_actual_api_key"
  export URL_VALUE="your_proxy_url"

  Basic Usage

  python main.py --method multiagent
  or
  sh scripts/gpt4o.sh

  Bootstrapping

  # Modify the path and the model
  python bootstrap.py

  📊 Data Requirements

  Data Extraction

  Extract filtered dataset from data/raw_data.csv, which can be obtained from Clinical Trial Outcome Prediction.

  python extract_dataset.py

  This creates data/extracted_single_disease_smiles.json with:

  • Single disease per sample
  • Single SMILES string per sample
  • Filtered for high-quality entries

  Based on this, we randomly selected 1/10 of the samples, obtaining a small dataset data/inputs.json. Then, we used bootstrap sampling to statistically analyze the experimental results on these samples.

  📊 Experiments

  

  🔗 Acknowledge

  • Clinical Trial Outcome Prediction
  • DeepPurpose
  • ADMET-AI
  • ChEMBL