Learning interpretable disease self-representations for drug repositioning