A modeling framework for embedding-based predictions for compound-viral protein activity.

Area of Science:

• Computational Biology
• Drug Discovery
• Machine Learning

Background:

• Drug repurposing accelerates the identification of treatments for emerging viral diseases like COVID-19.
• De novo drug design is a lengthy, costly, and inefficient process.
• Machine learning offers a powerful approach to predict compound-viral protein interactions for drug repurposing.

Purpose of the Study:

• To develop and validate a machine learning framework for predicting compound-viral protein activity.
• To identify existing compounds with potential therapeutic efficacy against SARS-CoV-2 proteins.
• To rank potential drug candidates for COVID-19 treatment.

Main Methods:

• Utilized deep learning-induced vector embeddings for compounds and viral proteins.
• Developed a consensus framework for predicting compound-viral protein activity.
• Employed molecular docking simulations to assess binding affinities of identified compounds.

Main Results:

• Achieved high prediction accuracy with a mean Pearson correlation of 0.916 and R2 of 0.840.
• Identified a ranked list of 47 compounds targeting key SARS-CoV-2 proteins (PL-PRO, 3CL-PRO, Spike).
• The identified compounds include antivirals, anticancer agents, antibiotics, and investigational drugs with favorable binding energies.

Conclusions:

• The machine learning framework effectively predicts compound-viral protein activity and aids in drug repurposing.
• The study identified promising drug candidates for COVID-19 treatment, warranting further investigation.
• The developed framework and identified compounds are publicly available, facilitating further research.