Molecular persistent spectral image (Mol-PSI) representation for machine learning models in drug design

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for molecular modeling

Background:

• Effective molecular representations are crucial for AI-driven drug design.
• Existing methods struggle to capture comprehensive structural, physical, chemical, and biological properties.
• Advanced models like 3D CNNs and graph neural networks require robust input features.

Purpose of the Study:

• To develop a novel, efficient, and effective molecular representation for AI-based drug design.
• To introduce the Molecular Persistent Spectral Image (Mol-PSI) as a 2D image-based molecular descriptor.
• To enhance protein-ligand binding affinity prediction using Mol-PSI and a CNN model.

Main Methods:

• Proposed Mol-PSI, an equal-sized 2D image representation for molecular structures and interactions.
• Developed a parallel Convolutional Neural Network (CNN) architecture tailored for Mol-PSIs.
• Validated the model on three standard protein-ligand binding affinity datasets (PDBbind-v2007, PDBbind-v2013, PDBbind-v2016).

Main Results:

• Mol-PSI provides a unique one-to-one image representation, suitable for deep learning models.
• The CNN model utilizing Mol-PSIs achieved superior performance in protein-ligand binding affinity prediction.
• Results surpassed traditional machine learning models across all tested PDBbind databases.

Conclusions:

• Mol-PSI is a powerful and versatile molecular representation for AI-based drug design.
• This approach facilitates improved accuracy in predicting molecular interactions.
• Mol-PSI holds significant potential for broad applications in molecular data analysis and pharmaceutical research.