Multimodal profiling of Pepcan-CB1 receptor structure-activity relationships: integrating molecular dynamics simulations, biological profiling, and the deep learning model MuMoPepcan.

In machine learning of drug discovery, the scale of accessible data is often strictly limited, while few-shot learning in wet-lab experimental data limits the accuracy of machine learning algorithms. Cannabinoid receptors are involved in various important physiological activities, and pepcans are key components of the endocannabinoid system. Herein, we proposed a combined dry-wet lab experimental framework that incorporated molecular dynamics simulation (MDS) data into peptide biological activity prediction. We validated our hypothesis on cannabinoid receptors type 1 (CB1) and pepcans: (1) In the study, we synthesized 45 pepcan peptides to establish a bioactivity dataset and identified RD-pepcan-11 as an lead analgesic compound by Bioscreening, with systematic characterization of its CB1 selectivity and pharmacodynamics.; (2) Millions of conformational data were generated by MDS and a CB1-pepcans conformation dataset was constructed; (3) Combining wet-lab data and MDS data, a deep learning model - MuMoPepcan was developed, reducing prediction errors to within the error range of wet-lab experiments. This study not only identified novel high-potential pepcans - RD-pepcan-11, but also demonstrated that MDS can serve as an effective data augmentation method to scale up drug-receptor datasets, thereby improving model generalizability and performance.