Magicmol - Fast and variable synthetic accessibility molecule generative model with reinforcement learning

Abstract

The flourishment of machine learning and deep learning methods have boosted the development of cheminformatics, especially when it comes to the application of drug discovery and new materials exploration. Lower time and space expenses make it possible for scientists to search the enormous chemical space for novel molecules. During this process, the generative model plays a significant role in quickly generating novel molecules, including predefined properties. Recently, some work combines reinforcement learning strategies with a generative model to optimize the property of generated drug-like molecules, which notably improved a batch of critical factors when it comes to drug discovery. However, a common problem of these RNN (recurrent neural network) based methods is that the novelty and validity of generated molecules depend on the “teacher-forcing strategy.” which is learned from an extensive database. Thus, there is inevitable to meet this situation that several generated molecules have difficulty in synthesizing even if owning higher desired properties such as binding score. In this paper, we proposed a new pipeline called Magicmol to directly change the synthetic accessibility of RNN-based generative models using reinforcement learning without introducing extra parameters. With proper strategies, the model can generate either hard-to-synthesis or easy-to-synthesis molecules with high validity. And our result shows a significant shift in the distribution of molecules with different synthetic accessibility when compared with the original model ,and we concluded some possible applications of Magicmol.