Sepsis is a life-threatening condition driven by dysregulated immune responses to infection with excessive M1 macrophage polarization-driven cytokine storm which plays a key role in the early progression of sepsis. Targeting macrophage polarization represents a promising therapeutic strategy to improve sepsis outcomes. Conventional drug discovery is hampered by high costs, long timelines and low success rates, posing significant challenges to the identification of novel M1 polarization inhibitors. In this study we constructed a novel transformer-variational autoencoder (TVAE) that integrated complementary molecular fingerprints (extended-connectivity fingerprints, ECFP; molecular ACCess system keys, MACCS keys; 4-point pharmacophore fingerprints, 4-PP) into probabilistic latent distributions to screen for M1-polarization inhibitors. From 5516 natural products, TVAE combined with experimental validation identified ombuin as the top candidate. In vitro, ombuin (10 μM) potently suppressed LPS-induced M1 polarization and pro-inflammatory cytokine (IL-6, TNF-α) release. In cecal ligation and puncture (CLP)-induced mouse sepsis model, administration of ombuin (15, 45 mg/kg, i.p.) significantly improved survival and ameliorated systemic inflammation by modulating the balance of M1/M2 macrophage polarization. By performing LiP-MS assay, we demonstrated that ombuin bound to and activated aldehyde dehydrogenase 2 (ALDH2), thereby suppressing NF-κB p65 nuclear translocation, a key event underlying NF-κB-driven M1 macrophage polarization. Collectively, our AI-driven pipeline efficiently discovers immunomodulatory agents and positions ombuin as a promising lead for sepsis therapy.
© 2026. The Author(s) under exclusive licence to The Shanghai Institute of Materia Medica, Chinese Academy of Sciences and Chinese Pharmacological Society.