The authors present a GA optimization technique for cosine-based k-nearest neighbors classification that improves predictive accuracy in a class-balanced manner while simultaneously enabling knowledge discovery. The GA performs feature selection and extraction by searching for feature weights and offsets maximizing cosine classifier performance. GA-selected feature weights determine the relevance of each feature to the classification task. This hybrid GA/classifier provides insight to a notoriously difficult problem in molecular biology, the correct treatment of water molecules mediating ligand binding to proteins. In distinguishing patterns of water conservation and displacement, this method achieves higher accuracy than previous techniques. The data mining capabilities of the hybrid system improve the understanding of the physical and chemical determinants governing favored protein-water binding.