Design And Optimization Of Drug Molecular Based On Reinforcement Learning

In recent decades,with the emergence of new diseases,drug development has become extremely important.However,the process of drug development is cumbersome and has a very low success rate.When a new disease appears,it often needs to invest a lot of manpower and material resources,and it takes a lot of time before there may be relevant treatment drugs,which is extremely unfavorable to the patients,especially serious diseases.Therefore,it is urgent to use some methods 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 premade molecular repositories,but the optimization of molecular properties is still challenging and hinders the implementation of re-architecting drug approaches,so it is necessary to develop a new framework for drug optimization.This study is mainly based on the classic algorithm of Reinforcement Learning（RL）-REINFORCE algorithm-to optimize multi-target attributes and multi-targets.Its main achievements are as follows:1.De novo drug design based on Stack-RNN with multi-objective reward-weighted sum and reinforcement learning.In order to achieve greater internal diversity in the optimized molecular compounds,we first proposed the use of two stack-augmented recurrent nets（Stack-RNN）to jointly generate a framework of drug-like molecules.It also adds a memory storage network to increase the internal diversity of its molecular compounds.Secondly,in order to solve the problem that the properties of compounds are extremely biased to a certain attribute due to the possible conflicts between attributes in the process of multi-objective optimization,a reward method was proposed to assign different weights to molecular optimization by using the reward values of different attributes.Finally,dynamic exploration strategy was used to alleviate the dilemma of exploration and development in the process of reinforcement learning optimization.In order to verify the effectiveness of the method,data of two drugs were collected respectively for experimental verification.Meanwhile,compared with two different reward mechanisms（traditional weighted sum and alternate weighted sum）and two different models（ORGANIC and Drug Ex）,their properties were improved.2.Pareto multi-target drug design and optimization based on Stack-RNN and reinforcement learning.In order to further optimize drug molecules so that they can be combined with multiple targets,we proposed to use two two-layer Stack-RNN as the generation model to alternately sample molecules,and add the Pareto multi-objective sequencing method to deal with the contradictions that may occur in multi-target drug design.To verify the availability of the method,we collected three target data of adenosineA₁,A_2A and h ERG for experimental verification,and compared them with the weighted sum method.