Deep learning integration of molecular and interactome data for protein-compound interaction prediction

Area of Science:

• Computational Chemistry
• Bioinformatics
• Machine Learning

Background:

• Virtual screening computationally predicts protein-compound interactions, aiding seed compound discovery.
• Current machine learning methods use either molecular structure or network data, with limited integration.
• Combining molecular and network data for interaction prediction remains underexplored.

Purpose of the Study:

• To develop a deep learning method integrating protein features, compound features, and multiple interactome data types.
• To evaluate the method's performance against state-of-the-art approaches using benchmark datasets.
• To assess the method's robustness in predicting interactions with unseen proteins and compounds.

Main Methods:

• Developed a deep learning framework integrating molecular structure data (protein/compound features) and interactome data (PPIs, etc.).
• Created three benchmark datasets of varying difficulty for comprehensive evaluation.
• Applied statistical tests (Wilcoxon signed-rank test) to confirm performance improvements.

Main Results:

• The proposed deep learning method significantly outperforms existing machine learning techniques in protein-compound interaction prediction.
• Integration of molecular structure and multi-interactome data synergistically enhances prediction accuracy.
• The method demonstrates superior robustness in predicting interactions involving novel proteins and compounds.

Conclusions:

• Integrating diverse data sources (molecular structure and interactome networks) is crucial for accurate protein-compound interaction prediction.
• Deep learning offers a powerful framework for leveraging multi-modal data in drug discovery and virtual screening.
• The developed method provides a more reliable tool for identifying potential seed compounds.