De novo drug design based on Stack-RNN with multi-objective reward-weighted sum and reinforcement learning

In recent decades, drug development has become extremely important as different new diseases have emerged. However, drug discovery is a long and complex process with a very low success rate, and methods are needed to improve the efficiency of the process and reduce the possibility of failure. Among them, drug design from scratch has become a promising approach. Molecules are generated from scratch, reducing the reliance on trial and error and prefabricated molecular repositories, but the optimization of its molecular properties is still a challenging multi-objective optimization problem.

In this study, two stack-augmented recurrent neural networks were used to compose a generative model for generating drug-like molecules, and then reinforcement learning was used for optimization to generate molecules with desirable properties, such as binding affinity and the logarithm of the partition coefficient between octanol and water. In addition, a memory storage network was added to increase the internal diversity of the generated molecules. For multi-objective optimization, we proposed a new approach which utilized the magnitude of different attribute reward values to assign different weights to molecular optimization. The proposed model not only solves the problem that the properties of the generated molecules are extremely biased towards a certain attribute due to the possible conflict between the attributes, but also improves various properties of the generated molecules compared with the traditional weighted sum and alternating weighted sum, among which the molecular validity reaches 97.3%, the internal diversity is 0.8613, and the desirable molecules increases from 55.9 to 92%.