Abstract 4464: A large language model for diverse 2D molecular generation (2DMG) and its application in focused library generation and rational design

Abstract AI-driven large language models for molecular generation are gaining significant interest in structure-based drug design. However, challenges such as dataset quality, molecular diversity, and synthesizability still persist. To address these challenges, we introduce StoneWise 2DMG, an advanced model that sparks creativity in medicinal chemistry experts, accelerating the drug discovery process by generating diverse 2D molecular structures. The model is pre-trained on a vast dataset comprising 30 million commercial compound structures (covering most global commercial compounds), 8 million patent-derived compound structures (spanning from the 1960s to 2023), 11 million building block structures, and billions of easily synthesizable molecular structures. This large-scale training enables the generation of drug-like molecules suitable for focused library generation and rational design. The 2DMG model can be utilized for focused library generation. By analyzing the structures of molecular glues that have been published or advanced into clinical trials, we selected glutarimide as a seed molecule for focused library design. Utilizing the 2DMG model, we are able to rapidly generate a diverse, rational, synthesizable, and drug-like molecular library. This molecular glue library is currently under synthesis in the laboratory and will be employed for high-throughput screening in the future. We then extended the application of our model to tackle a more advanced challenge in rational drug design. In a case study involving ABBV-CLS-484 (a molecule targeting both tumor and immune cells; PDBID: 7UAD), we analyzed the key binding fragment, thiadiazolidin-3-one, in the crystal structure. The model then used this critical fragment as a seed to generate a focused library of over 100, 000 compounds, which were docked and further selected by medicinal chemistry experts. Ten compounds were chosen for synthesis, and four showed acceptable activity, making them suitable for lead optimization. In summary, StoneWise 2D Molecular Generation model, trained on billions of data points, offers a robust platform for generating diverse and structurally rational molecules. The model supports focused library design for high-throughput or virtual screening and facilitates rational design to explore and expand chemical space, thus accelerating the drug discovery process. Citation Format: Zihao Liu, Qingbo Xu, Zhenyu Wu, Hongbo Zhang. A large language model for diverse 2D molecular generation (2DMG) and its application in focused library generation and rational design [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 4464.