Estimation of the applicability domain of kernel-based machine learning models for virtual screening.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Virtual screening relies on quantitative structure-activity relationship (QSAR) models but struggles with diverse chemical datasets.
• Current applicability domain methods often use vector descriptors, limiting compatibility with structured kernel-based QSAR models.
• Defining the chemical space where QSAR models are reliable is crucial for accurate predictions.

Purpose of the Study:

• To propose and evaluate novel methods for estimating the applicability domain of kernel-based QSAR models.
• To address the limitations of existing applicability domain approaches for structured kernel methods.

Main Methods:

• Developed three distinct approaches for applicability domain estimation tailored for kernel-based QSAR.
• Utilized support vector regression with three different structured kernels on three virtual screening tasks.
• Quantitatively assessed applicability using a score for each compound in the screening dataset.

Main Results:

• The proposed methods successfully differentiated reliable prediction regions within the chemical space.
• Evaluating model performance on subsets defined by applicability scores showed clear separation.
• Omitting compounds with low applicability scores significantly improved virtual screening performance.

Conclusions:

• The developed applicability domain formulations effectively identify compounds unsuitable for reliable QSAR prediction.
• Excluding unreliable predictions enhances virtual screening efficiency by reducing search space.
• The identified compounds would likely not be predicted as active by the model anyway, thus not hindering discovery.