Exploring predictive biomarkers from clinical genome-wide association studies via multidimensional hierarchical mixture models