De Novo Design of Molecules Towards Biased Properties via a Deep Generative Framework and Iterative Transfer Learning