Deep learning-based classification model for GPR151 activator activity prediction.

Area of Science:

• Biochemistry
• Computational Chemistry
• Pharmacology

Background:

• G protein-coupled receptor 151 (GPR151) is implicated in various physiological and pathological processes.
• GPR151 activators show therapeutic potential for metabolic disorders, necessitating further research.
• Accurate activity prediction models are crucial for efficient drug discovery and virtual screening.

Purpose of the Study:

• To develop a reliable, learning-based method for predicting GPR151 activator activity.
• To enhance the efficiency of virtual screening in the early stages of drug discovery.
• To explore novel molecular feature extraction techniques for improved predictive performance.

Main Methods:

• A novel molecular feature extraction algorithm inspired by the bag-of-words model was introduced.
• Mol2vec was employed for diverse molecular feature extraction.
• Three feature selection algorithms and three deep learning models were utilized for activity prediction.
• Five different classifiers were evaluated, including SVM.

Main Results:

• The proposed deep learning-based feature extraction method demonstrated superior performance compared to traditional algorithms.
• The optimal model, Mol2vec-CNN, significantly improved classification accuracy and stability.
• The SVM classifier achieved the highest accuracy (0.92) and F1 score (0.76).

Conclusions:

• The developed deep learning model is effective for GPR151 activator activity prediction.
• The feature extraction algorithm outperforms traditional methods in activity prediction tasks.
• The model shows promising utility in the pre-screening phase of drug virtual screening.