MEF-AlloSite: an accurate and robust Multimodel Ensemble Feature selection for the Allosteric Site identification...

Area of Science:

• Computational biology and bioinformatics
• Structural biology
• Drug discovery and medicinal chemistry

Background:

• Allosteric regulation controls protein function and offers advantages over orthosteric ligands for drug development.
• Identifying allosteric sites is crucial for novel drug design and understanding biological mechanisms.
• Current machine learning methods like PASSer struggle with accuracy using only 3D structural data for allosteric site identification.

Purpose of the Study:

• To develop a robust and accurate computational model for identifying allosteric binding sites.
• To improve upon existing machine learning approaches by integrating diverse feature sets.
• To enhance the understanding of the relationship between protein features and allosteric site function.

Main Methods:

• Collected 9460 diverse features from literature to characterize protein pockets.
• Developed the Multimodel Ensemble Feature Selection for Allosteric Site Identification (MEF-AlloSite) model.
• Employed multimodal feature selection on a small training set (90 proteins) to enhance predictive performance.

Main Results:

• MEF-AlloSite achieved a 1-6% higher mean of average precision and ROC AUC compared to state-of-the-art methods (PASSer2.0, PASSerRank).
• Statistical analysis (Student's t test, Cohen's D) confirmed the significance of MEF-AlloSite's performance improvement.
• Analysis of selected features provided insights into the complex nature of allosteric regulation in proteins.

Conclusions:

• Integrating amino acid-based information with 3D structural data significantly improves allosteric site prediction accuracy.
• MEF-AlloSite represents a robust and accurate advancement in computational allosteric site identification.
• This method holds promise for accelerating the discovery of novel therapeutics targeting allosteric sites.