A large descriptor set and a probabilistic kernel-based classifier significantly improve druglikeness classification