Machine Learning-Guided Generation of Novel Pseudo-Natural Products: Applications to Accelerate Drug Discovery

Natural products (NPs) are a critical source for drug discovery, and artificial intelligence (AI) is utilized to improve the efficiency of NP-based drug discovery. However, the existing AI-driven models typically generate a library of pseudo-natural products that only covers a small portion of the chemical space and the compounds were also restricted by poor drug-likeness profiles. Herein, the GPT1 is developed to generate diverse pseudo-natural products with excellent validity, uniqueness, and novelty while retaining molecular features similar to the training set. Subsequently, the Augmented Hill-Climb (AHC) strategy is employed to generate synthetically accessible compounds with enhanced drug-likeness. Using the integrated NPDL-GEN model (GPT1 + AHC), compounds <b>G1</b>-<b>G5</b> were obtained, exhibiting significantly improved drug-likeness profiles. Furthermore, the pseudo-natural products <b>H1</b>-<b>H3</b> generated via transfer learning also possess potent anti-inflammatory activities. Thus, our developed machine learning models can accelerate NP-based drug discovery.