End-to-end Deep Reinforcement Learning for Targeted Drug Generation

The long period of time and the enormous financial costs required to bring a new drug to the market are a clear impediment to the development of new drugs. Deep Learning techniques at early stages of drug discovery can help to select candidate drugs with biological properties of interest, reduce the enormous research space of drug-like compounds and minimize these issues. This study aims to perform generation of targeted molecules by training the recurrent neural network to learn the building rules of production of valid molecules in the form of SMILES strings and optimize it to produce molecules with bespoke properties through Reinforcement Learning. The fitness of the newly generated molecules is obtained by a second neural network model. To demonstrate the effectiveness of the method, we trained the proposed model to design molecules with high inhibitory power for the k-opioid receptor (KOR). The optimized model was able to generate molecules with a stronger affinity for KOR, maintaining the percentage of valid molecules and, with satisfactory internal and external diversities based on Tanimoto similarity over 95%.