A deep learning framework combining molecular image and protein structural representations identifies candidate drugs...

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Artificial intelligence (AI) and deep learning show potential for identifying drugs for human diseases, including pain.
• Predicting compound-protein interactions (CPIs) is crucial for drug discovery but remains challenging.

Purpose of the Study:

• To present an interpretable deep-learning framework, LISA-CPI, that integrates ligand molecular image representations and receptor 3D structures for accurate CPI prediction.
• To leverage AI for identifying novel therapeutic agents for pain management.

Main Methods:

• Developed LISA-CPI, a framework combining unsupervised deep learning (ImageMol) for ligand representation and AlphaFold2 (Evoformer) for protein structure analysis.
• Trained and evaluated LISA-CPI on a dataset of 104,969 ligands and 33 G-protein-coupled receptors (GPCRs).

Main Results:

• LISA-CPI demonstrated a significant improvement (approximately 20% reduction in mean absolute error) over state-of-the-art models in predicting experimental CPIs.
• Identified potential repurposable drugs like methylergometrine and candidate metabolites such as citicoline for pain treatment targeting human GPCRs.

Conclusions:

• The integration of molecular image and 3D protein structural data within a deep learning framework provides a powerful computational tool for drug discovery.
• LISA-CPI offers a promising approach for developing treatments for pain and other complex diseases.