Генерации графов заданной структуры: от глубоких нейронных сетей к квантовым моделям (на примере создания новых лекарств)
Year: 2026
Type: article
Abstract: The discovery of new chemical compounds with specified properties is a challenging problem in drug development. Many studies encode molecules as string representations derived from molecular graphs; however, this approach is computationally expensive and does not readily extend to general (non-molecular) graphs. Advances in graph deep learning make it possible to train generative models directly on graph representations, avoiding costly search in the discrete and extremely large space of chemical structures. MolGAN is a family of implicit models for generating small molecular graphs that combines generative adversarial networks (GANs) with reinforcement learning (RL) to produce molecules with target chemical properties. This review considers four variants: the baseline MolGAN (up to 9 atoms), Large MolGAN (up to 20 atoms) with a graph-expansion mechanism that reduces the generation of disconnected graphs, a MolGAN variant based on WGAN as a more stable alternative to standard GAN training, and a hybrid MolGAN incorporating quantum computing modules. The paper describes the model architectures, compares their performance on established benchmarks, and discusses limitations and directions for future research.
Authors: Ф.Т. Адылова, Р.Р. Давронов
Language: Russian
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: hybrid
OpenAlex ID: W7134195884