Navigating Drug Design and Toxicity Profiling Using Reinforcement Learning in Pharmaceutical Research

Drug design is an extensive process that leads to minimal yield concerning the valid molecules that are accepted. With the incorporation of Reinforcement learning, this phenomenon can be optimized for druglikeness and toxicity. RL can automate this strenuous and laborious process by introducing self-learning capabilities that iteratively update the policy value of the environment, contributing to creating a cost-efficient and time-saving mechanism for drug design in pharmaceutical sciences. In this study, the actor-critic algorithm is used to choose an action and compute the Quantitative Estimate of Druglikeness (QED) value. The reward function is based on the QED value and toxicity. The primary drug chosen is Aspirin. The RL agent was successfully able to generate nine unique drugs that have similar druglikeness with an average QED of 0.41 with the highest being 0.60 and an average toxicity level of 0.09. The production of these diverse molecules could help avoid the high toxicity that otherwise causes side effects in Aspirin.