ALADDIN: Docking Approach Augmented by Machine Learning for Protein Structure Selection Yields Superior Virtual...

Area of Science:

• Computational chemistry
• Structural bioinformatics
• Drug discovery

Background:

• Protein flexibility and solvation complicate molecular docking.
• Ensemble docking uses multiple protein structures but benefits from relevant structure selection.

Purpose of the Study:

• To develop a robust machine learning approach for selecting optimal protein structures for docking.
• To improve the accuracy of structure-based virtual screening for flexible protein targets.

Main Methods:

• An integrated machine learning and docking approach (ALADDIN) was developed.
• ALADDIN utilizes random forest classifiers to identify the most suitable protein structure from an ensemble for each compound.
• Performance was evaluated against single-structure docking, ensemble docking, and similarity-based docking.

Main Results:

• ALADDIN demonstrated superior performance on three out of four targets, showing higher AUC values compared to other methods.
• Improvements in AUC ranged up to 0.15, 0.11, and 0.16 over baseline methods.
• Performance was limited for cytochrome P450 3A4, similar to other tested approaches.

Conclusions:

• ALADDIN offers a robust method for selecting protein structures in docking, particularly for flexible targets.
• This approach is highly applicable for structure-based virtual screening of kinases, viral enzymes, and anti-targets.
• ALADDIN enhances the efficiency and accuracy of drug discovery pipelines for challenging protein targets.