Isometric Stratified Ensembles: A Partial and Incremental Adaptive Applicability Domain and Consensus-Based...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Quantitative structure-activity relationships (QSAR)

Background:

• Traditional QSIR methods struggle with imbalanced datasets, leading to biased classification.
• Accurate classification is crucial for hit ranking and selection in early drug discovery.
• Existing applicability domain methods may not fully capture model performance variations.

Purpose of the Study:

• To introduce and validate a novel QSIR strategy: isometric stratified ensembles (ISE).
• To improve classification performance and reliability for imbalanced chemical datasets.
• To enhance hit ranking and selection by defining local performance regions.

Main Methods:

• Developed a 2D mapping of classification statistics (consensus and applicability domain).
• Employed internal cross-validation to define isometric local performance regions.
• Utilized recursive decorrelated variable selection and class ratio balancing during training.

Main Results:

• ISE demonstrated superior performance compared to existing QSIR tools on imbalanced PubChem datasets.
• Achieved a global AUC of 0.82 for colloidal aggregators, increasing to 0.88 in the highest confidence stratum.
• The method effectively identified high-confidence regions for improved hit selection.

Conclusions:

• ISE offers a robust strategy for QSIR analysis, particularly for imbalanced datasets.
• The approach enhances classification accuracy and provides reliable applicability domain assessment.
• ISE facilitates better hit ranking and selection in virtual screening campaigns.