Machine learning-powered discovery of a novel berberine derivative inducing SCD-dependent ferroptosis in osteosarcoma

Abstract Background Despite decades of therapeutic development, osteosarcoma survival remains poor. Although berberine (BBR) shows anti-tumor activity, its efficacy is limited. We addressed this through structural modification and machine learning-guided discovery, developing a novel derivative: 9-O-methoxyethylberberrubine bromide (B1). Methods In vivo, subcutaneous and orthotopic models were established in BALB/c nude mice using 143B cells. Treatment groups received daily B1 (0.1-5 mg/kg) or berberine (5, 50 mg/kg); a positive control group received doxorubicin (1 mg/kg). Tumor growth was assessed by volume and weight; tissue necrosis, proliferation, and apoptosis were analyzed. In vitro, human osteosarcoma cells (143B, U2OS, HOS) and human bone marrow mesenchymal stem cells (hBMSCs) were treated with B1, and anti-proliferation was evaluated via CCK-8, EdU, colony formation, and transwell assays. We integrated machine learning into our proteomic discovery pipeline to prioritize critical targets. Proteomic sequencing was followed by multi-algorithm feature selection including least absolute shrinkage and selection operator (LASSO), Ridge, Elastic Net, mRMR, and univariate filtering. Mechanistic validations employed molecular docking, thermal shift assays, surface plasmon resonance (SPR), co-immunoprecipitation, ubiquitination assays, and lipidomics. single-cell RNA sequencing compared malignant osteosarcoma cells with normal bone microenvironment components. Results B1 exhibited dose-dependent anti-tumor effects superior to BBR. Machine learning-driven integration of proteomic profiles unanimously nominated Sterol CoA desaturase (SCD) as the key target across all feature selection algorithms, showing both maximal relevance and minimal redundancy. Mechanistically, B1 acts as a molecular glue that recruits the E3 ligase neural precursor cell expressed, developmentally down-regulated 4-like (NEDD4L) to SCD, inducing its ubiquitination and degradation. Single-cell RNA sequencing confirmed significant overexpression of SCD in malignant osteosarcoma cells, further highlighting its therapeutic relevance. Computationally prioritized SCD targeting disrupted lipid metabolism, causing saturated lipid accumulation, mitochondrial damage, and oxidative stress. This ultimately promoted glutathione peroxidase 4 (GPX4)-mediated lipid peroxidation and ferroptosis. Resistance to B1 occurred with SCD overexpression, while arachidonic acid supplementation partially restored tumor survival. Conclusions By incorporating machine learning into drug target discovery, we established B1 as a ferroptosis inducer targeting the NEDD4L-SCD axis. Our study provides both a robust therapeutic strategy against chemoresistant osteosarcoma and a compelling blueprint for AI-augmented oncology drug development. Graphical Abstract