Spectra-descriptor-based machine learning for predicting protein-ligand interactions.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Machine learning models are vital for identifying drug lead compounds.
• Current models lack reliable, physicochemical-based molecular descriptors.
• Existing methods struggle to capture complex spatial and electronic molecular information.

Purpose of the Study:

• Introduce the Fragment Integral Spectrum Descriptor (FISD) as a novel physicochemical descriptor.
• Enhance virtual screening models for drug discovery.
• Improve the representation of molecules and proteins in machine learning.

Main Methods:

• Developed the Fragment Integral Spectrum Descriptor (FISD) using spatial and electronic structure information.
• Integrated FISD with a classical neural network model.
• Validated the model's performance against conventional descriptors and complex models.
• Applied FISD to predict and screen potential ligands for protein targets.

Main Results:

• FISD combined with a classical neural network achieved performance comparable to complex models.
• The novel descriptor effectively utilizes spatial and electronic molecular data.
• Successful prediction and screening of potential binding ligands for two protein targets were demonstrated.
• FISD shows broad applicability and practicality in virtual screening.

Conclusions:

• FISD represents a significant advancement in molecular and protein representation for machine learning.
• This descriptor accelerates the drug discovery process by improving virtual screening efficiency.
• FISD offers a more reliable and physicochemical-based approach to molecular descriptors.