BEGAN: Boltzmann-Reweighted\nData Augmentation for Enhanced GAN-Based\nMolecule Design in Insect Pheromone Receptors

Identifying\nsmall molecules that bind strongly to target proteins\nin rational molecular design is crucial. Machine learning techniques,\nsuch as generative adversarial networks (GAN), are now essential tools\nfor generating such molecules. In this study, we present an enhanced\nmethod for molecule generation using objective-reinforced GANs. Specifically,\nwe introduce BEGAN (Boltzmann-enhanced GAN), a novel approach that\nadjusts molecule occurrence frequencies during training based on the\nBoltzmann distribution exp­(−Δ<i>U</i>/τ),\nwhere Δ<i>U</i> represents the estimated binding free\nenergy derived from docking algorithms and τ is a temperature-related\nscaling hyperparameter. This Boltzmann reweighting process shifts\nthe generation process toward molecules with higher binding affinities,\nallowing the GAN to explore molecular spaces with superior binding\nproperties. The reweighting process can also be refined through multiple\niterations without altering the overall distribution shape. To validate\nour approach, we apply it to the design of sex pheromone analogs targeting Spodoptera frugiperda pheromone receptor SfruOR16,\nillustrating that the Boltzmann reweighting significantly increases\nthe likelihood of generating promising sex pheromone analogs with\nimproved binding affinities to SfruOR16, further supported by atomistic\nmolecular dynamics simulations. Furthermore, we conduct a comprehensive\ninvestigation into parameter dependencies and propose a reasonable\nrange for the hyperparameter τ. Our method offers a promising\napproach for optimizing molecular generation for enhanced protein\nbinding, potentially increasing the efficiency of molecular discovery\npipelines.