PT-Finder: A multi-modal neural network approach to target identification

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Efficient bioactive compound target identification is critical but challenging and costly.
• Machine learning offers a promising approach for predicting compound-protein interactions.
• Current methods often rely on ligand structural similarity, limiting scope.

Purpose of the Study:

• To develop a machine learning model for predicting protein targets of chemical compounds.
• To create a user-friendly application for rapid and accurate target identification.
• To accelerate the drug discovery and repurposing pipeline.

Main Methods:

• A multi-modal neural network was constructed using protein sequences and active inhibitors.
• The model was trained on a comprehensive library of proteins and ligands.
• Performance was validated using accuracy and MPRAUC metrics.

Main Results:

• The developed model achieved 82% accuracy and an MPRAUC of 0.80.
• PT-Finder, an offline application, was created based on the trained model.
• PT-Finder can predict targets for hundreds of compounds in seconds.

Conclusions:

• PT-Finder provides a fast, accurate, and user-friendly solution for protein target identification.
• The tool significantly accelerates drug discovery workflows by enabling rapid prediction.
• Freely accessible source code and application facilitate broader adoption and research.