Discussion

The present study systematically explored serum metabolic signatures associated with maximal voluntary ventilation (MVV) among native high-altitude Tibetans based on a rigorous multi-stage metabolome-wide analytical framework. After standardized sample screening, optimized clinical confounder adjustment, stable feature selection, and dual internal and external model validation, three serum metabolites, including N-Undecanoylglycine, Isosorbide Dinitrate, and Hypoglycin B, were identified as independent correlates of pulmonary ventilatory function in this hypoxic-adapted population. These findings provide preliminary metabolomic insights into individual differences in ventilatory capacity under chronic high-altitude hypoxia, complementing existing genetic and environmental studies on high-altitude pulmonary adaptation.

A core strength of this study lies in its standardized and robust analytical strategy tailored for high-dimensional metabolomic data. Unlike conventional single-step metabolite screening approaches that are prone to false-positive results and poor reproducibility (Broadhurst and Kell 2006), the current study integrated univariate tiered threshold filtering, elastic net regularization, and a predefined fallback screening rule to balance the efficiency and stability of metabolite mining (Chong et al. 2018, Xia et al. 2009). Elastic net regression (α = 0.5) was adopted to overcome the limitations of single-penalty regularization, effectively tolerating metabolite multicollinearity and retaining potentially meaningful weak-effect metabolic features, which is highly suitable for untargeted metabolomic data with ultra-high dimensionality (Chamlal et al. 2024). In addition, this study optimized clinical confounding adjustment based on population characteristics: 20 original clinical variables were rationally integrated, screened, and filtered by collinearity diagnosis to obtain 17 independent covariates, avoiding biased association results caused by redundant or highly collinear clinical factors (Liu et al. 2024). Strict dataset partitioning was also implemented, with all metabolite screening and model training procedures completed exclusively in the training set to eliminate data leakage, while independent external validation was used to objectively evaluate model generalizability (Joeres et al. 2026).

Among the three identified metabolites, N-Undecanoylglycine showed a significant negative correlation with MVV. As a medium-chain fatty acid glycine conjugate, acyl glycines are closely linked to lipid metabolism imbalance and oxidative stress response in respiratory tissues (Isa et al. 2024, Tan et al. 2010, Newgard 2017). Under chronic high-altitude hypoxic stress, abnormal accumulation of N-Undecanoylglycine may reflect disrupted lipid homeostasis in the respiratory system, partially explaining individual declines in ventilatory reserve function (Murray et al. 2018). By contrast, Isosorbide Dinitrate and Hypoglycin B were positively associated with MVV. Notably, Isosorbide Dinitrate is an exogenous drug-derived metabolite rather than an endogenous metabolic molecule, which acts as a classic nitric oxide donor to improve microcirculation and alleviate hypoxic vasoconstriction in lung tissues, potentially contributing to preserved pulmonary ventilation function (Murray et al. 2018, Beall et al. 2012). Hypoglycin B, a characteristic amino acid derivative, may participate in hypoxic metabolic adaptation and energy metabolism regulation in plateau residents, supporting the stability of pulmonary physiological function under long-term hypoxic exposure (D’Alessandro et al. 2016, Chicco et al. 2018). Collectively, these three metabolites participate in multiple biological processes including lipid metabolism, microcirculation regulation, and hypoxic adaptation, suggesting that MVV variation in native high-altitude Tibetans is modulated by complex multi-dimensional metabolic patterns (Lewis et al. 2008).

Notably, the metabolite-based predictive model exhibited a noticeable decrease in predictive performance from internal cross-validation to external validation. The model achieved a moderate predictive R2 of 0.191 in ten-fold cross-validation of the training set, whereas the external validation R2 dropped to 0.090. This attenuation is a common phenomenon in population-based metabolomic predictive studies and can be attributed to multiple objective limitations (Tzoulaki et al. 2014, Ganna et al. 2014). First, the single-center cohort and moderate sample size (n = 168) inevitably limited the statistical power and population representativeness of the study, making the model susceptible to subtle population heterogeneity and random noise during external validation (Button et al. 2013). Second, although comprehensive clinical confounding adjustment was performed based on collected baseline indicators, the current study only recorded histories of hypertension and dyslipidemia without detailed information on individual medication use. This limitation makes it impossible to correct for the potential confounding effect of exogenous drug exposure, which may interfere with the association signal of the drug-derived metabolite Isosorbide Dinitrate (Suhre et al. 2011). Third, due to the limited number of final effective metabolites, formal metabolic pathway enrichment analysis could not be conducted, which restricted the systematic interpretation of the underlying metabolic mechanisms linking metabolites and pulmonary ventilation function (Xia and Wishart 2010).

In addition, several methodological limitations of the current multi-stage screening framework should be acknowledged. The preset fallback strategy of retaining top five metabolites with the smallest univariate P values when elastic net yields no valid features ensures the integrity of analytical workflow, but it may introduce marginal weak-effect metabolites and increase the uncertainty of model fitting (Meinshausen et al. 2010). Meanwhile, bootstrap stability verification confirmed the repeatability of the screening process, but the relatively small validation set sample size weakened the persuasive power of external generalization results (Steyerberg et al. 2010). Furthermore, this study only focused on the cross-sectional correlation between serum metabolites and MVV, and cannot clarify the causal direction of metabolic variation and pulmonary function differences under chronic hypoxia (Lawlor et al. 2008).