Computational Prediction of Compound-Protein Interactions for Orphan Targets Using CGBVS

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Artificial Intelligence (AI) and Machine Learning (ML) are increasingly used for in silico prediction of Compound-Protein Interactions (CPI).
• Chemical genomics-based virtual screening (CGBVS) is an AI technique utilizing support vector machine (SVM) for CPI prediction, known for accuracy and ease of use.

Purpose of the Study:

• To evaluate the efficacy of CGBVS in identifying ligands for orphan G protein-coupled receptors (GPCRs) lacking prior ligand information.
• To develop a method for assessing the applicability of CGBVS for predicting GPCR ligands.

Main Methods:

• Employed CGBVS with pairwise kernel-based SVM for predicting Compound-Protein Interactions (CPI).
• Utilized G protein-coupled receptor (GPCR) family data, specifically testing prediction on a virtual orphan GPCR with omitted ligand information during training.
• Assessed prediction accuracy using an applicability index and compared performance based on different protein and compound descriptors, including Multiple Sequence Alignment (MSA).

Main Results:

• Prediction accuracy varied significantly across different GPCRs.
• Models incorporating Multiple Sequence Alignment (MSA) as protein descriptors demonstrated superior overall prediction accuracy.
• The choice of protein descriptors had a greater impact on accuracy than the choice of compound descriptors.
• Prediction accuracy was positively correlated with the amount of available ligand information for related GPCRs.

Conclusions:

• CGBVS is a viable method for predicting ligands of orphan GPCRs.
• Multiple Sequence Alignment (MSA) is a key feature for improving prediction accuracy in CGBVS for GPCRs.
• The availability of related ligand data significantly influences the success of CGBVS in predicting GPCR ligands.