QuantileBootstrap

QuantileBootstrap: Quantile-based model validation framework for computational drug discovery

QuantileBootstrap implements quantile-based data partitioning and rank-based loss functions to validate predictive models on structure-activity datasets, emphasizing generalization across activity distribution levels and out-of-sample ranking performance of high-activity molecules.

Key Features:

• Quantile-Based Data Partitioning: Constructs training and testing sets using quantile splits of the activity distribution function to evaluate model performance across distinct activity levels and reduce overfitting associated with random partitioning.
• Rank-Based Loss Functions: Incorporates two rank-based loss functions that penalize errors in predicting the ranks of high-activity molecules in out-of-sample data.
• Algorithm Benchmarking: Evaluates neural networks, random forests, support vector machines (regression), and ridge regression across 25 high-quality structure-activity datasets.

Scientific Applications:

• Computational Drug Discovery: Assesses predictive models for small-molecule activity, supporting selection of algorithms with improved generalization and reduced overfitting in structure-activity modeling.

Methodology:

The framework partitions data using activity distribution quantiles and evaluates models with rank-based performance metrics that penalize extrapolation errors on structurally diverse molecules outside the training set, providing systematic validation of predictive power in structure-activity datasets.