Integrating Multimodal AI with Encoded Medicinal Knowledge and Physics Simulations: A Unified Platform for Accelerated Drug Discovery from Patent Analysis to Potency Validation
Huimin Cheng, Yuliang Wu, Xiaowen Niu, Yuhang Wu, Tian Zhou, Junjie Zou, Huobin Wang, Chunwang Peng, Zhiqiang Liu, Rui He, Lijie Peng, Yongpan Chen, Zhang Zhang, Mingjun Yang, Jian MaAbstract
Traditional drug discovery faces challenges from fragmented data, tacit knowledge dependence, and inefficient design-test cycles. We present an integrated AI workflow combining multimodal data curation, encoded medicinal chemistry rules, and physics-based simulations to establish a closed-loop from patent analysis to candidate validation. Application across six patents for four targets successfully deciphered SAR and identified representative molecules. A retrospective case study on RET kinase inhibitors demonstrated the workflow’s efficiency: focusing on patent-derived SAR enabled rapid discovery of Cpd-31 (RET IC50 = 0.48 nM, 95.88% tumor inhibition at 10 mg/kg) with only three initial compounds synthesized. This data-centric approach streamlines early drug discovery by providing objective molecular design foundations, significantly reducing synthetic efforts and hypothesis-driven exploration.