A Ligand-Based Virtual Screening Method Using Direct Quantification of Generalization Ability.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning applications

Background:

• Ligand-based virtual screening is crucial for drug discovery.
• Conventional machine learning methods struggle with imbalanced and high-dimensional chemical data.
• Overfitting is a significant challenge in virtual screening model development.

Purpose of the Study:

• To introduce a novel machine learning algorithm, local beta screening (LBS), for efficient ligand-based virtual screening.
• To address the limitations of conventional methods in handling imbalanced and high-dimensional datasets.
• To accurately assess overfitting risk without relying on resampling techniques.

Main Methods:

• Developed the local beta screening (LBS) algorithm.
• Quantified generalization ability using a refined loss function for direct overfitting risk assessment.
• Validated LBS robustness through simulation and real-world datasets.
• Applied LBS to screen for HIV-1 integrase multimerization activators.

Main Results:

• LBS demonstrated superior screening accuracy and model interpretability compared to conventional algorithms.
• The algorithm effectively handled imbalanced and high-dimensional data without resampling.
• Experimental validation confirmed LBS's predictive power, identifying six active compounds.
• The most potent identified compound exhibited an EC50 of 0.71 µM.

Conclusions:

• LBS offers an accurate and efficient approach for ligand-based virtual screening, particularly for challenging datasets.
• The algorithm mitigates overfitting risks, enhancing the reliability of virtual screening results.
• LBS successfully identified novel activators of HIV-1 integrase multimerization, demonstrating its practical utility in drug discovery.