Using distance covariance for improved variable selection with application to learning genetic risk models