Identifying Metabolomic Biomarkers of Lung Function Decline in People with HIV

Abstract

Background:

Pulmonary complications in people living with HIV (PWH) have shifted away from infectious disease and towards chronic disease. HIV is an independent risk factor for chronic obstructive pulmonary disease (COPD), with PWH developing COPD younger and declining faster in pulmonary function. As an accelerated decline is associated with greater mortality, there is a need to identify individuals at high risk of longitudinal decline.

Setting:

59 adults with HIV enrolled from the Pittsburgh Lung HIV study cohort.

Methods:

Targeted metabolite profiling was performed on baseline bronchoalveolar lavage fluid (BALF, n=35) and serum samples (n=54) using liquid chromatography-high resolution mass spectrometry. Longitudinal pulmonary function tests (median 3 measurements over 2.95 years with a follow-up interval of 1.34 years) were used to determine rates of decline. Predictive modeling and feature selection algorithms identified baseline clinical and metabolomic factors associated with longitudinal decline across forced expiratory volume, forced vital capacity, and diffusing capacity of the lung.

Results:

Predictive models found the BALF metabolome to successfully predict outcomes more consistently than serum. Key BALF metabolites such as elevated carnitine and reduced pyruvate predicted greater risk of longitudinal decline. Low serum citrate levels were a robust predictor of decline across multiple tests. Probabilistic graphical models supported direct relationships between these metabolites and lung function decline.

Conclusion:

Baseline metabolomic profiling, especially using BALF, can help identify PWH at risk for accelerated lung function decline. Key metabolic pathways related to glucose oxidation, fatty acid metabolism, and amino acid metabolism underlie observed lung function changes.

Introduction

Human immunodeficiency virus (HIV) infection is a global disease affecting an estimated 39 million people worldwide¹. In the era of antiretroviral therapy (ART), life expectancy for people living with HIV (PWH) on ART is comparable to individuals without HIV², and the primary cause of pulmonary complications has shifted away from infectious diseases and towards chronic, non-infectious disease, including chronic obstructive pulmonary disease (COPD)³. COPD, characterized by persistent airflow limitation, is a leading cause of death globally⁴ and is a common, chronic, non-infectious disease in PWH⁵. HIV independently increases the risk of developing COPD by 14% after adjusting for smoking⁶, and PWH often develop non-infectious pulmonary diseases at a younger age than people without HIV⁷. Prior studies have shown that PWH experience faster longitudinal decline in two measures of airflow limitation, forced expiratory volume in 1 second (FEV1) and forced vital capacity (FVC), compared to people without HIV^8,9, and that a faster rate of lung function decline is associated with increased mortality¹⁰. As such, there is a need to identify mechanisms underlying HIV-associated lung function decline to aid in the development of new therapies and identification of high-risk individuals to improve outcomes for PWH.

Metabolomics offers a promising avenue for addressing these challenges. Metabolites, as the output of interactions among biological processes and environmental factors, are attractive as potential biomarkers and provide insight into mechanisms underlying complex disease phenotypes¹¹. COPD is a heterogeneous disease in both presentation and biological mechanisms, and metabolomic profiling has been successfully applied to identify disease subtypes and explore disease mechanisms¹².

In this study, we performed targeted metabolomic profiling in the bronchoalveolar lavage fluid (BALF) and serum of PWH with liquid chromatography high resolution mass spectrometry (LC-HRMS). By profiling both local (BALF) and systemic (serum) biofluids, we aimed to identify metabolomic signatures predictive of longitudinal lung function decline. We applied an array of machine learning techniques, including elastic net regression¹³, random forests¹⁴, and causal probabilistic graphical modeling¹⁵, to develop predictive models of longitudinal lung function decline and robustly select features significantly associated with longitudinal outcomes. These findings have the potential to inform biomarker discovery, improve risk stratification, and identify pathways for therapeutic interventions in COPD among PWH.

Methods

Study Participants

We enrolled 59 PWH from the Pittsburgh Lung HIV study cohort^16,17 who were 38 to 66 years old and were participating in ongoing studies of lung disease in HIV. Individuals had previously undergone bronchoscopy. Those with other active infections, FEV1 below 30% predicted, uncontrolled systemic diseases, or allergies to medications used during the procedure were excluded from bronchoscopy. Some individuals in this cohort were previously analyzed by Wendt et al¹⁸ at baseline. Demographic and baseline clinical data were collected by standardized participant interview including age, gender, smoking history, current antiretroviral therapy (ART), and history of prior pneumonia. Peripheral CD4+ counts and plasma HIV RNA levels (viral load) were confirmed by chart review.

Pulmonary Function Testing

All participants performed spirometry before and after bronchodilator administration (400mcg of albuterol administered with a spacer), including forced expiratory volume in 1 second (FEV1) and forced vital capacity (FVC)^19–21 and diffusing capacity for carbon monoxide (DLCO). Spirometry and DLCO were performed following American Thoracic Society (ATS)/European Respiratory Society (ERS) standards²². Pulmonary function test (PFT) and DLCO measurements that passed quality control criteria (Grade A, B, or C) were included in the final analysis. Hankinson and Neas equations were used to determine percent predicted values of spirometry and DLCO^22,23, respectively. Pulmonary function analyses were performed prior to widespread use of Global Lung Function Initiative (GLI) reference equations, but a recent study found little difference in use of these reference equations versus GLI in PWH²⁴. DLCO was corrected for hemoglobin and carboxyhemoglobin²².

Longitudinal Identification of Lung Function Decline

PFTs were conducted at baseline and during longitudinal follow-up. Individuals were followed for a median duration of 2.95 years (IQR: 2.23 years), with PFTs conducted at a median of 3 time points (IQR: 2 time points) for a median follow-up interval of 1.34 years (IQR: 0.81 years). To measure longitudinal lung function decline, a linear regression model was fit for each individual to determine the average rate of change of FEV1 %predicted, FVC %predicted, FEV1/FVC, and DLCO %predicted over time. For each pulmonary function test, lung function decline is represented by a binary variable where decline is defined as a negative average rate of change.

Construction and Validation of Predictive Models

To predict longitudinal lung function decline using baseline clinical features and BALF and serum metabolites (Table S1 for full list), we constructed predictive models using three distinct methods: (1) FCI-Max²⁵, a method for learning causal probabilistic graphical models¹⁵, (2) elastic net regression¹³, and (3) random forests¹⁴. By utilizing these three approaches, we enable our predictive models to capture a variety of different types of relationships between clinical and metabolomic features and lung function decline. Each method has different hyperparameters that must be tuned to construct the final predictive model. We utilized cross-validation to tune these hyperparameters and to quantify the predictive performance of our models on unseen data. Due to smaller sample size, leave-one-out cross-validation (LOOCV) was used in the analysis of the BALF metabolites, while 10-fold cross-validation was used in the analysis of the serum metabolites. For each method and measure of lung function decline, the hyperparameters with the highest AUC in cross-validation were selected for the final predictive model. Further discussion of the targeted metabolomic profiling, causal probabilistic graphical models, and the cross-validation performed for each method are available in the Supplementary Methods.

Selection of Features Predictive of Lung Function Decline

Each method for learning predictive models provides distinct approaches to selecting features that are related to the prediction target. For the probabilistic graphical models, the set of relevant features was selected as the Markov blanket of the measure of lung function decline. The Markov blanket is the set of features that render a target variable conditionally independent of all other variables in a dataset²⁶. For elastic net regression, the set of relevant features is the set of features with non-zero coefficients. Finally, for random forests, predictive features were selected based on permutation feature importance¹⁴, with features with a mean decrease in accuracy greater than zero selected. However, these three different predictive modeling methods select distinct sets of predictive features. We select the set of clinical and metabolomic features identified by all three methods as the final set of features consistently predictive of each measure of longitudinal lung function decline. These features are then used to construct multivariate logistic regression models of longitudinal lung function decline.

Statistical Analysis

For each measure of longitudinal lung function decline that our models could predict significantly better than random, we constructed a multivariate logistic regression model that regressed on the set of consistently selected features identified as described above. We tested the significance of the association of each feature with longitudinal lung function decline using a Wald test on the regression coefficients. All p-values resulting from these tests of association were adjusted to control the false discovery rate (FDR) through the Benjamini-Hochberg procedure²⁷.

Results

Characteristics of the Cohort

Of the 59 participants, 31% were women, 66% were Black and 32% were White, and the median age was 52 years (Table 1). 66% of participants had a history of cigarette smoking and 46% were current smokers. Median post-bronchodilator FEV1% predicted was 90%, median post-bronchodilator FVC% predicted was 92%, median post FEV1/FVC was 78%, and median DLCO% predicted was 78%. FVC% predicted decline occurred in the fewest participants, at 44%, while FEV1/FVC decline occurred in the greatest number of participants, at 68%. Serum metabolites were measured in 54 individuals, while BALF metabolites were measured in 35 individuals.

Table 1:

Clinical and spirometry characteristics of the enrolled PWH from the Pittsburgh Lung HIV study cohort, pp = %predicted.

Characteristic	Value (n=59)
Baseline FEV1pp, Median (IQR)	90.3% (77.9%−101.7%)
Baseline FVCpp, Median (IQR)	92.3% (84.1%−102.3%)
Baseline FEV1/FVC, Median (IQR)	78.3% (70.6%−83%)
Baseline DLCOpp, Median (IQR)	77.6% (64.3%−84.2%)
FEV1pp Decline, Number (%)	29 (49.2%)
FVCpp Decline, Number (%)	26 (44.1%)
FEV1/FVC Decline, Number (%)	40 (67.8%)
DLCOpp Decline, Number (%)	27 (45.8%)
Current Smoker, Number (%)	27 (45.8%)
Gender: Female, Number (%)	18 (30.5%)
Gender: Male, Number (%)	41 (69.5%)
Race: Black, Number (%)	39 (66.1%)
Race: White, Number (%)	19 (32.2%)
Race: Other, Number (%)	1 (1.7%)
Age (years), Median (IQR)	52 (46–56.5)
CD4+ Counts, Median (IQR)	673 (325.5–900.5)
CD4+ Counts: Missing, Number (%)	24 (40.7%)
Viral Load, Median (IQR)	39 (19–39)
Viral Load: Missing, Number (%)	24 (40.7%)
Asthma, Number (%)	7 (11.9%)
Asthma: Missing, Number (%)	7 (11.9%)
Pneumonia, Number (%)	10 (16.9%)
Pneumonia: Missing, Number (%)	7 (11.9%)
Current ART, Number (%)	49 (83.1%)
Current ART: Missing, Number (%)	7 (11.9%)
IV Drugs Ever, Number (%)	8 (13.6%)
IV Drugs Ever: Missing, Number (%)	7 (11.9%)
Marijuana Last 6 Months, Number (%)	18 (30.5%)
Marijuana Last 6 Months: Missing, Number (%)	7 (11.9%)
BALF Metabolomics Measured, Number (%)	35 (59.3%)
Serum Metabolomics Measured, Number (%)	54 (93.1%)

Metabolomic and clinical features were associated with baseline lung function

For both the individuals with BALF (n=35) and serum (n=54) metabolomics, we identified clinical features and metabolites that were significantly associated with baseline pulmonary function in PWH (FDR<0.1) using the Grow-Shrink Markov blanket procedure with the BIC score²⁸ (Figure 1). The BALF metabolomics data identified metabolites linearly associated with baseline DLCO% predicted, FEV1% predicted, and FVC% predicted (Fig 1A). Carnitine was positively associated with both FEV1% predicted and FVC% predicted, while pyruvate, creatinine, and lactate were negatively associated with DLCO% predicted, FEV1% predicted, and FVC% predicted respectively. In individuals with serum metabolomics, we identified clinical features and metabolites linearly associated with baseline DLCO% predicted, FEV1/FVC, and FVC% predicted (Fig 1B). Both age and current smoking status were negatively associated with FEV1/FVC. We also identified three amino acids associated with baseline lung function: aspartate was positively associated with DLCO% predicted, while proline and alanine were negatively associated with DLCO% predicted and FVC% predicted respectively.

Figure 1:

Forest plot depicting standardized regression coefficients for linear regression models of baseline PFTs. Models are constructed from BALF metabolomics dataset (left) and serum metabolomics dataset (right). Only baseline PFTs with at least one feature identified by the Grow-Shrink procedure are shown. (pp = %predicted).

Metabolomic and clinical features successfully predicted longitudinal lung function decline

For both the BALF and serum metabolomic datasets, predictive models for each of the four measures of longitudinal lung function decline were constructed using three methods (elastic net, random forests, and FCI-Max). The ability of these models to predict lung function decline was assessed by cross-validation, with model performance measured by the area under the receiver-operator characteristic curve (AUC). Predictive models with AUCs significantly greater than 0.5 at an FDR p-value < 0.05 were able to successfully predict lung function decline in PWH (Figure 2). The BALF metabolomic dataset successfully predicted longitudinal decline in all four PFTs with at least one model, while the serum metabolomic dataset was only able to successfully predict longitudinal decline in FEV1/FVC and FEV1% predicted with at least one model. Across both datasets and all lung function tests, elastic net regression was able to best predict longitudinal outcomes.

Figure 2:

Receiver operating characteristic (ROC) curve of predictive models able to predict longitudinal PFT decline significantly better than random at FDR<0.05. Only ROC curves for PFTs that were significantly predicted by at least one model were shown for both BALF (left) and serum (right) metabolites. Three types of predictive models were learned for each prediction task: (1) causal probabilistic graphical models (Causal), (2) elastic net logistic regression (Elastic Net), and (3) random forests (Random Forest), indicated by color and line type in the ROC plots. (PP: %predicted).

Identification of BALF metabolites predictive of longitudinal lung function decline

To assess the association of the set of consistently predictive BALF metabolomic and clinical features with longitudinal lung function decline, we performed multivariate logistic regression using the features presented in Table 2. For each pulmonary function test, a subset of these features was significantly associated with longitudinal decline (FDR<0.1). In the prediction of FEV1/FVC decline, the strongest significant predictor was the baseline measurement of the FEV1/FVC, with higher baseline levels of FEV1/FVC indicating an increased risk of longitudinal decline. Elevated levels of carnitine were significantly associated with increased risk of longitudinal decline in both FEV1/FVC and FVC% predicted. Additionally, elevated levels of valine were significantly associated with an increased risk of longitudinal DLCO% predicted decline. Finally, elevated hydroxyproline was significantly associated with better longitudinal lung function outcomes for FEV1% predicted, and elevated pyruvate was significantly associated with better outcomes for FVC% predicted.

Table 2:

Standardized coefficients of multivariate logistic regression models of PFT decline constructed with clinical and BALF metabolomic features selected by all three predictive models for each measure. Significant features (FDR < 0.1) are denoted in bold, pp = %predicted.

PFT	Feature	Coefficient (Std. Err.)	FDR p value
FEV1/FVC Decline	Baseline FEV1/FVC	2.029 (0.807)	0.09
	Carnitine	1.304 (0.663)	0.09
	Leucine	0.921 (0.631)	0.177
	Acetylcarnitine	−0.284 (0.602)	0.637
FEV1pp Decline	Hydroxyproline	−0.797 (0.403)	0.09
FVCpp Decline	Carnitine	4.279 (1.786)	0.09
	Pyruvate	−1.521 (0.731)	0.09
	Hydroxyproline	−2.364 (1.282)	0.102
	Hypoxanthine	−0.674 (0.595)	0.238
DLCOpp Decline	Valine	0.842 (0.413)	0.09
	Baseline DLCOpp	0.866 (0.497	0.112

Identification of serum metabolites predictive of longitudinal lung function decline

We applied the same approach to identify subsets of clinical and serum metabolomic features that were predictive of longitudinal pulmonary function decline (Table 3). For three PFTs (FEV1/FVC, FEV1% predicted, and FVC% predicted), a subset of these features was significantly associated with longitudinal decline (FDR<0.1). First, low baseline DLCO% predicted and high baseline FVC% predicted were significantly associated with increased risk of FEV1/FVC and FVC% predicted decline respectively. Additionally, elevated levels of the metabolites hydroxyproline and taurine were significantly associated with an increased risk of longitudinal FEV1/FVC decline. Elevated levels of serum citrate were significantly associated with a decreased risk of pulmonary function decline in all three PFTs (FEV1/FVC, FE1% predicted, and FVC% predicted; Table 2).

Table 3:

Standardized coefficients of multivariate logistic regression models of PFT decline constructed with clinical and serum metabolomic features selected by all three predictive models for each measure. Significant features (FDR < 0.1) are denoted in bold, pp = %predicted.

PFT	Feature	Coefficient (Std. Err.)	FDR p value
FEV1/FVC Decline	Hydroxyproline	1.343 (0.448)	0.033
	Taurine	1.268 (0.529)	0.047
	Baseline DLCOpp	−1.178 (0.504)	0.047
	Citrate	−0.912 (0.391)	0.047
	Hypoxanthine	0.375 (0.462)	0.418
FEV1pp Decline	Citrate	−1.663 (0.843)	0.083
	Leucine	−0.478 (0.342)	0.217
	Baseline FEV1/FVC	−0.38 (0.309)	0.262
	Baseline DLCOpp	−0.267 (0.302)	0.411
FVCpp Decline	Citrate	−2.661 (1.065)	0.047
	Baseline FVCpp	0.703 (0.326)	0.062
	Hexose	−0.977 (0.566)	0.127

The direction of the association for hydroxyproline, a significant predictor of longitudinal lung function decline in both the BALF and serum, disagreed between the two metabolomic datasets. In the BALF, elevated hydroxyproline was associated with better longitudinal lung function, while in the serum, hydroxyproline was associated with worse longitudinal lung function. While potentially contradictory, an analysis of the correlation between metabolites measured in both the BALF and serum metabolomics revealed that they were largely uncorrelated (Fig S1), apart from proline, glutamine, and lactate (FDR<0.05). Specifically, hydroxyproline had a correlation of 0.096, (p=0.61), suggesting that BALF levels of hydroxyproline were independent of serum levels, and their association with longitudinal lung function decline may result from distinct mechanisms.

Probabilistic graphical models identify direct links between metabolomic biomarkers and longitudinal lung function decline

Probabilistic graphical models enable us to explore associations among clinical and metabolomic features that are robust to conditioning on all other subsets of features in a dataset. In the probabilistic graphical models learned for pulmonary function decline, clinical features, and BALF metabolomics (Fig S2A), we saw adjacencies between carnitine and the longitudinal decline of FEV1/FVC, FEV1% predicted, and FVC% predicted. Additionally, pyruvate was adjacent to FVC% predicted decline. Finally, the amino acids hydroxyproline and leucine were linked to pulmonary function decline. In both the probabilistic graphical models for pulmonary function decline measures that were successfully predicted by clinical features and serum metabolomics (FEV1% predicted decline and FEV1/FVC decline, Fig S2B), we observed that citrate was directly linked to pulmonary function decline. Additionally, we observed adjacencies between metabolites related to amino acid metabolism (taurine, hydroxyproline, serine, and dimethylarginine) and pulmonary function decline. Though limited sample size makes the identification of the graphical models unreliable, we use bootstrapping to assess the stability of individual adjacencies.

Discussion

By combining baseline BALF and serum metabolomics measurements with longitudinal pulmonary function in PWH, we demonstrated that baseline clinical and metabolomic features were able to successfully predict longitudinal lung function decline. Additionally, by integrating the feature selection results of three distinct machine learning methods, we identified a subset of clinical and metabolomic features significantly associated with longitudinal lung function decline. Finally, by utilizing probabilistic graphical models, we learned interpretable networks of robust conditional dependence relationships between metabolomic and clinical features and longitudinal lung function decline.

Our predictive models demonstrated that BALF metabolites were more consistently able to predict longitudinal lung function decline than serum metabolites; BALF metabolites successfully predicted longitudinal decline in DLCO% predicted, FEV1/FVC, FEV1% predicted, and FVC% predicted, while serum metabolites were only able to successfully predict decline in FEV1/FVC and FEV1% predicted. These findings suggest that the BALF metabolome is more informative than the serum metabolome for predicting lung function decline and may better reflect metabolic processes in the lung associated with that decline. Prior studies have demonstrated similar results, with significantly more BALF than serum metabolites associated with contemporaneous measures of lung function and COPD phenotypes in matched samples²⁹. Wendt et al. directly assessed the ability of paired BALF and serum metabolomic samples to discriminate between obstructive lung disease (OLD) and matched controls in PWH and found that only the BALF metabolome was able to distinguish between the two groups¹⁸. Additionally, experimental studies in mice models have demonstrated that many metabolic changes induced in the lung by cigarette smoke exposure are not reflected in paired serum samples³⁰.

Applying our feature selection technique to identify BALF metabolomic features consistently associated with lung function decline revealed potential signatures of metabolic dysregulation in the lung. The opposing effects of carnitine and pyruvate on the likelihood of pulmonary function decline suggest that a shift in the fuel sources utilized in cellular respiration is predictive of long-term longitudinal lung function decline in PWH. Carnitine, which aids in transferring long-chain fatty acids into the mitochondria³¹, plays a key role in fatty acid oxidation. Dysregulation in lipid metabolism has previously been demonstrated in individuals with COPD^32–34, and cigarette smoke exposure increases fatty acid oxidation in mice³⁵. Additionally, previous analysis of PWH with and without OLD found that an increase in acylcarnitines and decrease in phosphatidylcholines was associated with worse baseline FEV1 %predicted, suggesting that dysregulation in lipid metabolism disrupts the formation and/or maintenance of pulmonary surfactant¹⁸. On the other hand, pyruvate is an intermediate product in the metabolic pathway through which glucose undergoes oxidation. Experiments in mouse alveolar type II cells demonstrated that cigarette smoke exposure leads to a decrease in respiration metabolizing glucose³⁶. The same study additionally demonstrated that cigarette smoke exposure resulted in an increase in respiration utilizing the fatty acid palmitate and suggested a mechanism by which increased fatty acid utilization in respiration disrupts surfactant biosynthesis pathways, leading to lung function decline³⁶. Our analysis of longitudinal decline in PWH supports these results, with elevated pyruvate predicting better outcomes, while elevated carnitine predicted poorer outcomes.

Although serum metabolites were less informative for lung function decline overall, our analysis did identify factors consistently associated with decline in lung function. Low serum citrate levels were significantly associated with an increased risk of pulmonary function decline across multiple measures. This observation may be due to previously reported metabolic changes in skeletal muscle function³⁷, which include reduced strength, endurance, mitochondrial density, oxidative to glycolytic ratio, and aerobic enzyme activity. In particular, the activity of citrate synthase, the enzyme responsible for catalyzing the reaction that produces citrate in the citric acid cycle, is lower in the skeletal muscle tissue of individuals with COPD both at rest³⁸ and after exercise³⁹. This metabolic adaptation of skeletal muscles away from oxidative respiration may result in a decrease in citrate production in individuals with COPD, suggesting that low levels of serum citrate correspond to greater systemic oxidative stress and an increased likelihood of longitudinal pulmonary function decline.

The causal probabilistic graphical models learned by FCI-Max recapitulate many of the observations we made regarding features linked to longitudinal lung function decline, providing evidence that these associations are robust even when conditioned on any subset of features in our dataset. The observation of BALF carnitine and pyruvate levels in the graphical models of pulmonary lung function decline provides additional evidence that they may represent a shift away from glucose oxidation and towards fatty acid oxidation. In the graphical models of serum metabolites and longitudinal lung function decline, we observed that citrate is directly connected to the two measures of lung function decline that were successfully predicted. These adjacencies represent a robust conditional dependence between serum citrate and pulmonary function decline, suggesting that it can be an effective serum biomarker. Additionally, the graphical models of both the BALF and serum metabolites found direct links between certain amino acids (leucine, hydroxyproline, taurine, serine, and dimethylarginine) and lung function decline, suggesting that changes in amino acid metabolism play a role in lung function decline. A prior analysis of serum metabolites in individuals with COPD and healthy controls in the ECLIPSE⁴⁰ study cohort revealed shifts in the concentration of several amino acids, including hydroxyproline, leucine, serine, and arginine, in patients with COPD compared to healthy controls and among COPD subtypes⁴¹. Finally, baseline measurements of pulmonary function were linked to the decline of DLCO% predicted and FEV1/FVC, which could be explained by a reduced capacity for further lung function decline in individuals with low pulmonary function at baseline.

A key limitation of our study is the limited sample size, which reduces the power of both our predictive models and statistical tests. Additionally, the lack of an external validation cohort limits the generalizability of our findings in different patient populations. We addressed these challenges using cross-validation, which allowed us to estimate the performance of our models out of the training sample, and using bootstrapping to demonstrate the stability of the direct associations learned by our graphical models. Furthermore, while causal discovery algorithms aim to recover causal interactions, they cannot establish cause-effect relationships in real-world settings. Instead, these models should be interpreted as networks of associations inferred under more stringent criteria than traditional correlation networks or undirected graphical models. Consequently, this study is exploratory and hypothesis-generating, with future research in larger, more diverse cohorts required to validate our findings and assess generalizability of identified biomarkers. Notably, our observation that the BALF metabolome outperforms the serum metabolome in predicting longitudinal lung function highlights the importance of focusing on metabolites in the local context of the lung in future studies.

Our work has demonstrated success in the prediction of longitudinal lung function decline based on clinical and metabolomic data. Further studies using a larger cohort measuring metabolomics and longitudinal lung function decline may allow for the construction of more accurate predictive models of lung function decline in PWH.