Introduction:A special place in the development of new drugs is computer modeling of potential drug candidates. At this stage, the molecular structure of a drug is created and virtually validated. Molecular structures are created mostly by bioinformatics specialists and medical chemists. Therefore, the process of creating and virtual testing of molecules is long and expensive.Purpose:Developing a model of a deep generative adversarial neural network and its reinforcement environment for generating targeted small organic molecular structures with predetermined properties, as well as reward functions for molecular diversity.Results: The developed deep neural network model called ATNC is based on the concepts of adversarial learning and reinforcement learning. The model uses a recurrent neural network with external memory as a generator of molecular structures, and a special neural network block for selecting the generated molecules before their real estimation by the environment. A new objective reward function of internal clustering by diversity is proposed, which allows the model to generate more diverse chemistry. Comparative experiments have shown that the proposed ATNC model is better than its closest competitor in terms of generating unique and more complex valid molecular structures. It has also been demonstrated that the the molecules generated by ATNC match to the a priori distributions of the key molecular descriptors of the training molecules. Experiments were conducted on a large dataset of 15 000 drug-like molecular compounds collected manually from the ChemDiv collection.Practical relevance:The proposed model can be used as an intelligent assistant in developing new drugs by medical chemists.