REINFORCED SEQ2SEQ ADVERSARIAL AUTOENCODER FOR DE NOVO MOLECULAR DESIGN

Abstract

Subject of Research.The modern models of deep training for generation of target small organic molecules are studied. The studies were carried out on two datasets of 250,000 drug-like molecular compounds from the ZINC database and 23,000 kinase molecular structures collected manually from the open accessed ChemBL database. Method.We propose the model of a deep neural network based on the concepts of adversarial learning and reinforcement learning. The model controls the molecular validity of the generated structures through the use of a recurrent seq2seq autoencoder and an external generator. The presence of an external generator gives the model flexibility in the choice of architecture, and also allows for the input conditions for the generation. Main Results. Comparative experiments have shown that the proposed model is better than its closest competitors in experiments with pre- and post-training in terms of generating valid and unique molecular structures. Additional chemical analysis of generated structures demonstrates the best quality of the introduced model in comparison with the other competitor models. Practical Relevance.The proposed model can be used by medical chemists as an intelligent assistant for development of new drugs.

Keywords

• and also allows for the input conditions for the generation. Main Results. Comparative experiments have shown that the proposed model is better than its closest competitors in experiments with pre- and post-training in terms of generating valid and unique molecular structures. Additional chemical analysis of generated structures demonstrates the best quality of the introduced model in comparison with the other competitor models. Practical Relevance. The proposed model can be used by medical chemists as an intelligent assistant for development of new drugs.
• machine learning
• deep learning
• reinforcement learning
• generative adversarial networks
• drug design and development