A search for optimal fastness/accuracy compromises in developing predictive models for estimating CYP2D6 inhibition properties of drug candidates