Result

Baseline characteristics of the study population

After rigorous quality control, a total of 168 participants were ultimately enrolled in the present study from an initial pool of 202 high-altitude Tibetan adult samples. Twenty clinical covariates were initially collected for potential confounding adjustment. Prior to statistical modeling, multicollinearity screening was performed with a variance inflation factor (VIF) threshold of 10. Three covariates with severe multicollinearity were excluded, and the remaining 17 independent clinical variables were adopted for all subsequent analyses. All statistical analyses were conducted in R 4.3 with a fixed random seed (12345) to ensure full result reproducibility. The enrolled population was stratified by MVV quartiles and randomly divided into a training set (n = 119) and an independent validation set (n = 49) at a 7:3 ratio. Baseline characteristics of the two subsets, including demographics, anthropometrics, lifestyle factors, sleep indicators, and medical histories, were summarized in Table 1. Inter-group balance was evaluated using Wilcoxon rank-sum tests, Chi-square tests, or Fisher’s exact tests as appropriate. All covariates exhibited no significant difference between the two groups (all P > 0.05), indicating reasonable and unbiased dataset partitioning. Detailed baseline comparison statistics for all original covariates are provided in Supplementary File1. Baseline profiles were presented with raw original values, while Z-score standardization was applied to continuous phenotypic and metabolomic variables before regression analysis. The overall analytical procedure is illustrated in Fig. 1.

Fig. 1: Analytical workflow for identifying MVV-associated metabolites. After preprocessing (Z-score standardization, VIF screening), 168 participants were split into training (n = 118) and validation (n = 50) sets. A four-step screening pipeline (residualization, univariate regression, Elastic Net, multivariable regression) identified three key metabolites. Model validation included 10-fold CV, bootstrap, sensitivity analysis (exclusion of exogenous metabolites), and independent validation

Multi-stage screening of MVV-associated serum metabolites

All metabolite‑screening procedures were strictly restricted to the training dataset. A three‑tier hierarchical screening pipeline integrating univariate regression, tiered‑threshold filtering, elastic‑net regularization (α = 0.5, lambda.1se), and a pre‑specified fallback strategy was implemented to select candidate metabolites. In the first tier, univariate linear regression was performed for each metabolite against MVV residuals adjusted for clinical covariates; no metabolite satisfied the false‑discovery‑rate‑corrected significance threshold of FDR < 0.05. We then relaxed the threshold to nominal univariate P < 0.05, yielding 109 candidate metabolites for elastic‑net regularization; however, this model returned no metabolites with non‑zero coefficients. Given that ventilatory capacity (MVV) is shaped by multifactorial physiological regulation and individual metabolite‑MVV marginal associations may be weak, we further expanded the screening margin by adopting a looser univariate threshold of P < 0.01, which retained 255 metabolites (Supplementary File 1). Nevertheless, elastic‑net regression still failed to output features with non‑zero coefficients under this enlarged candidate pool. Following our pre‑defined analytical protocol, the top‑five metabolites with the smallest univariate P‑values were therefore selected as candidate features for subsequent multivariable regression.

After full adjustment for the 17 retained clinical confounders within multivariable linear‑regression models, three serum metabolites were identified as independently and significantly associated with MVV (P < 0.05), and these association results are visualized in Fig. 2. Specifically, N‑Undecanoylglycine was negatively associated with MVV (β = −5.237, P = 0.002), whereas Isosorbide Dinitrate (β = 0.245, P = 0.007) and Hypoglycin B (β = 4.655, P = 0.006) exhibited positive associations with MVV.

Fig. 2: Associations between three identified serum metabolites and maximal voluntary ventilation (MVV). Dot plot illustrating effect sizes (β) and corresponding 95% confidence intervals derived from multivariable linear regression fully adjusted for 17 clinical covariates. The metabolites correspond to X52235 (N‑Undecanoylglycine, β = −5.237, P = 0.002), X53153 (Isosorbide Dinitrate, β = 0.245, P = 0.007), and X54386 (Hypoglycin B, β = 4.655, P = 0.006). Metabolite IDs are shown in the plot, with full metabolite names annotated in the figure legend

Predictive performance, screening stability and sensitivity analysis of the metabolite‑based model

To quantitatively evaluate the predictive performance and robustness of the final multivariable metabolite model, ten‑fold cross‑validation (10‑CV) was performed within the training set, yielding a cross‑validated R2 of 0.191. The full model achieved a training R2 of 0.424 and an independent validation R2 of 0.090, with a slightly lower 10‑CV R2 of 0.191. Key predictive metrics for both the primary model and sensitivity‑analysis model are summarised in Table 2. The evident attenuation of predictive accuracy from internal cross‑validation to external independent validation indicated limited generalization ability of the model, which may be attributable to the relatively modest sample size and the complex, multifactorial nature of pulmonary ventilatory variation in high‑altitude native populations. Bootstrap resampling analysis was further applied to assess the stability of metabolite screening. The top‑ten metabolites ranked by bootstrap inclusion frequency are visualized in Supplementary Figure S2, and raw bootstrap outputs are provided in Supplementary File 2. The top three candidate metabolites exhibited moderate bootstrap inclusion frequencies (X53153: 0.205; X54386: 0.125; X52235: 0.110), suggesting that individual metabolite‑MVV marginal associations were relatively weak and susceptible to subtle sampling variation. This moderate stability reflects the inherent complexity of MVV regulation, which is modulated by extensive physiological and environmental factors rather than dominated by a small number of strong metabolic signals.

Model | Training-setR2 | 10-fold CV R2 | Validation-set R2
Primary model(three metabolites) | 0.424 | 0.191 | 0.090
Sensitivity model (excluding X53153) | 0.577 | 0.169 | 0.189

Considering that one of the three identified metabolites (Isosorbide Dinitrate, X53153) is an exogenous drug‑related metabolite and detailed medication information was unavailable in the present cohort, we performed a sensitivity analysis by excluding this exogenous metabolite and repeating the entire multi‑stage screening and modeling pipeline. Corresponding performance metrics from the sensitivity analysis are also presented in Table 2. Notably, the sensitivity model yielded improved predictive performance, with a training R2 of 0.577 and an independent validation R2 of 0.189 after excluding the exogenous feature. Detailed regression coefficients and P‑values obtained from the sensitivity‑analysis multivariable model are provided in Supplementary File 3. The retained stable endogenous metabolite associations and improved external predictive capacity demonstrated that the core metabolic signatures associated with MVV were not dependent on exogenous drug interference, further supporting the reliability of the primary findings.

In addition, given that only three independent metabolites were identified, formal pathway enrichment analysis was not conducted, as such a small set of molecular markers cannot provide statistically robust biological pathway inference. Collectively, although the final metabolite‑based model exhibited moderate predictive power, the relatively low explanatory capacity may objectively reflect the genuine weak metabolic susceptibility of ventilatory phenotypic variation in high‑altitude Tibetan populations, rather than insufficient model fitting.