A multimodal computational workflow identifies OR2J3 as the key olfactory receptor for terpenoids in Citrus aurantium 'Changshanhuyou' pericarp

Citrus aurantium 'Changshanhuyou' pericarp (CP) exhibits unique and complex olfactory characteristics, yet the underlying olfactory recognition mechanisms remain unclear. In this study, CP samples were first grouped based on sensory differences using an electronic nose (E-nose). A novel dual-model screening strategy was then established by integrating traditional machine learning (Orthogonal partial least squares discriminant analysis, OPLS-DA) with deep learning (Autoencoder, AE) for the identification of key volatile organic compounds (VOCs). Molecular docking and molecular dynamics (MD) simulations were employed to systematically decipher the olfactory perception mechanisms of the selected key VOCs. The results revealed a clear shift in CP odor from green and pungent in early stages to mild and fruity in later stages, with terpenoids playing a central role in this transition. The dual-model approach identified 53 high-confidence key VOCs, 17 of which were terpenoids, including D-Limonene (relative content 6.29-55.10%). Molecular docking showed that the Class II olfactory receptors (ORs), particularly OR2J3, exhibited strong binding affinities for terpenoids (-4.3 to -6.9 kcal/mol). Key hydrophobic residues (e.g., ALA9, LEU171, and PRO170) located in transmembrane domains TM3-TM6 formed a hydrophobic pocket. Subsequent 100 ns MD simulations further elucidated that OR2J3 formed stable complexes with characteristic terpenes (D-Limonene, γ-Terpinene, δ-Elemene and Germacrene B). Crucially, a significant reduction in solvent accessible surface area (ΔSASA≈-4.31 nm 2 ) and low structural fluctuation (RMSD < 0.5 nm, RMSF < 0.2 nm) confirmed that hydrophobic effects are the primary thermodynamic driving forces. In contrast, the OR51 displayed higher conformational flexibility (RMSD < 0.6 nm, RMSF < 0.7 nm). This work provides a robust dual-model framework for screening key aroma compounds and first identifies OR2J3 as the core receptor target for CP’s characteristic aroma. This multimodal framework advances multiscale analytical methodologies for critical VOC screening and receptor-level aroma perception studies in citrus species, providing novel insights into the molecular basis of complex olfactory phenomena. • Machine learning (ML) and Autoencoder (AE) identified 35 and 28 key metabolites. • AE-derived metabolites match e-nose experimental data significantly better than ML. • Four identified terpenoids were verified by molecular docking and dynamics simulation. • Class II ORs demonstrated more stable binding preferences for terpenoids.