arer90/ANO_solubility_prediction

Automated Network Optimizer (ANO) for Enhanced Prediction of Intrinsic Solubility in Drug-like Organic Compounds: A Comprehensive Machine Learning Approach

Overview

This repository presents a novel approach to predicting aqueous solubility of drug-like organic compounds using our Automated Network Optimizer (ANO) framework. By integrating advanced machine learning techniques with automated feature selection and hyperparameter optimization, we achieve state-of-the-art prediction accuracy for intrinsic solubility (logS).

System Requirements

Dependencies

• Python 3.12 or later
• TensorFlow 2.15.0 (Linux/MacOS/WSL)
• TensorFlow 2.15.0-GPU (Windows)
• RDKit 2024.3.1
• pandas 2.2.1
• scikit-learn 1.4.1.post1
• seaborn 0.13.2
• matplotlib 3.8.3
• optuna 3.5.0

Repository Structure

Jupyter Notebooks

1. 1_standard_ML.ipynb

   • Comprehensive evaluation of traditional ML approaches
   • Random Forest, XGBoost, and SVM implementations
   • Baseline performance metrics and comparative analysis

2. 2_solubility_fingerprint_comparison.ipynb

   • Detailed analysis of molecular fingerprint methods
   • Evaluation of ECFP, MACCS, and custom fingerprints
   • Performance comparison across fingerprint types

3. 3_ANO_with_feature_checker.ipynb

   • Implementation of ANO framework
   • Automated feature importance analysis
   • Real-time feature selection optimization

4. 4_ANO_feature.ipynb

   • Optimal physicochemical feature search using ANO

5. 5_ANO_structure.ipynb

   • Hyperparameter optimization using ANO

6. 6_ANO_network_[fea_struc].ipynb

   • Network architecture optimization based on optimal physicochemical features

7. 7_ANO_network_[struc_fea].ipynb

   • Network architecture optimization based on optimal hyperparameters

8. 7_Solubility_final_HPO_proving.ipynb (Bug fixing...)

   • Performance validation of final ANO model

9. 8_solubility_xai.ipynb

   • Model explainability analysis
   • Permutation importance and SHAP evaluation
   • Correlation analysis between physicochemical features and logS
   • Implementation of Lipinski's Rule of 5

Core Python Modules

• basic_model.py

  • Foundation architecture for fingerprint analysis
  • Modular design for easy extension
  • Comprehensive validation methods

• feature_search.py

  • Feature search implementation for ANO (used in 4_ANO_feature.ipynb)

• feature_selection.py

  • Feature selection implementation for ANO (used in 5_ANO_structure.ipynb, 6_ANO_network_[fea_struc].ipynb, 7_ANO_network_[struc_fea].ipynb)

• learning_model.py

  • ANO learning model implementation
  • Used in deep learning and feature optimization notebooks (used in 3_ANO_with_feature_checker, 3_solubility_descriptor_deeplearning, 4_ANO_feature, 5_ANO_structure.ipynb, 6_ANO_network_[fea_struc].ipynb, 7_ANO_network_[struc_fea].ipynb)

Key Innovations

• 49 carefully selected chemical descriptors for target dataset
• Fast and efficient selections of chemical descriptors and hyperparameters in machine learning models

Model Availability

Pre-trained models and complete results are available at: https://huggingface.co/arer90/ANO_solubility_prediction/tree/main

Version

Current Version: 1.0.2 (2024.11)

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If you use this work in your research, please cite:

@article{ANO2024solubility,
  title={Prediction of intrinsic solubility for drug-like organic compounds using Automated Network Optimizer (ANO) for physicochemical feature and hyperparameter optimization},
  author={Chung, Young Kyu and Lee, Seung Jun and Lee, Jonggeun and Cho, Hyunwoo and Kim, Sung-Jin and Huh, June},
  journal={ChemRxiv},
  year={2024},
  doi={10.26434/chemrxiv-2024-mp291}
}