GAN Based Approach for Drug Design

Deep Learning models have been a tremendous breakthrough in the field of Drug discovery, greatly simplifying the pre-clinical phase of this intricate task. With an intention to ease this further, we introduce a novel method to generate target-specific molecules using a Generative Adversarial Network (GAN). The dataset consists of drugs whose target proteins belong to the class of Tyrosine kinase and are specifically active against some of the growth factor receptors present in the human body. An Autoencoder network is used to learn the embeddings of the drug which is represented in the SMILES format and the deep neural network GAN is used to generate structurally valid molecules using drug-target interaction as the validating criteria. The model has successfully produced 39 novel structures and 15 of them show satisfactory binding with at least one of the target receptors.