Scaffold-driven molecular generation via reinforced RNN with centroid distance evaluation

De novo molecular design is learning from existing data to propose a new chemical structure that meets the desired properties. Still, it is difficult to de novo design a variety of novel molecules with desirable properties because the different properties of molecules cannot be balanced using a simple generative method. To solve this problem, this study proposes a reinforced Pareto-optimized molecular scaffold clustering generation method, ScaRL-P. Molecular scaffold information can help us identify molecules of the same pattern, cluster according to the core characteristics of the scaffold, and screen out the molecules with ideal properties. In addition, this study uses Pareto optimization to construct a three-dimensional Pareto frontier - biological activity, diversity, and in-cluster reward value, and cooperate with molecular scaffold clustering to obtain the dominant molecules. The multi-dimensional frontier obtained by adjusting the Pareto frontier in reinforcement learning modeling is transformed into the final reward return, which is provided to the agent in reinforcement learning to learn a molecular generation strategy that is close to the optimal attribute distribution.ScaRL-P demonstrated superior performance in binding affinity and optimization for three protein objectives (KOR, PIK3CA, and JAK2), outperforming several GPC benchmark methods.