Implements genetic algorithm-based variable selection for building
quantitative structure-activity relationship (QSAR) models. The package provides
a workflow for selecting optimal predictor subsets from large
descriptor spaces using leave-one-out cross-validation (LOOCV) with Q2 as the
fitness criterion. Features include automatic handling of multicollinearity via
variance inflation factor (VIF) thresholding, customizable genetic algorithm
operators, and diagnostic tools for model evaluation. Supports both training set
optimization and external validation, plus nested (double) cross-validation for
unbiased performance estimation and predictor stability diagnostics. Built-in
visualization functions include Q2 curves and Williams plots to assess model
applicability domain. The method is demonstrated in papers predicting
antibacterial activity by Araya-Cloutier et al. (2018)