GEP-DNN4Mol: automatic chemical molecular design based on deep neural networks and gene expression programming

The inverse design of molecules has attracted widespread attention in the field of chemical molecular design. However, existing methods fail to address the diversity of the generated molecules. In this work, we propose a molecule generation method called GEP-DNN4Mol to generate molecules with good diversity and desired properties in the exploration of vast chemical space. GEP-DNN4Mol leverages a special gene expression programming algorithm as a generator for molecular generations, uses a deep neural network as an evaluator to guide the update of the generator by extracting the molecular features of the generated molecules, and couples with SMILES and SELFIES molecular representations. The experimental results show that the proposed approach outperforms the state-of-the-art methods in the performance of generated molecules and the efficiency of exploration in chemical space. The molecules generated by GEP-DNN4Mol have advantages in terms of total validity, high novelty, and good diversity.

The online version contains supplementary material available at 10.1007/s13755-025-00344-8.