Opposite effects of chronic HIV infection and antiretroviral medication on organismal and organ-specific biological aging

Abstract

People with HIV (PWH) on combination antiretroviral therapy have an elevated risk for aging-related non-AIDS comorbidities. We assess whether HIV infection accelerates biological aging in two independent cohorts of PWH using six organ-specific and three organism-wide aging clocks derived from plasma proteomics of healthy individuals. Proteomic age acceleration significantly correlates with DNA methylation age and is linked to comorbidities and mortality. HIV infection accelerates systemic biological aging, with Mendelian randomization demonstrating causality between organ aging and inflammatory or metabolic complications. Accelerated aging in PWH is further related to the total HIV reservoir, and specific antiretroviral drugs reduce age acceleration. These data reveal important causal effects between chronic HIV infection, antiretroviral medication, biological aging and age-associated diseases, highlighting targets for improving health span in PWH.

People living with HIV (PLWH) have an increased risk for aging-related comorbidities. Here, the authors develop a proteomics-based immune aging framework for PLWH and demonstrate that immune aging is accelerated in HIV infection, is closely linked to total viral reservoir burden, and is modulated by antiretroviral therapy.

Introduction

Many studies in the last two decades focused on the physiological and molecular mechanisms underpinning biological aging, with the aim of identifying therapeutic targets to slow or even reverse this process. Several ‘hallmarks of aging’ have been proposed, including telomere attrition, epigenetic alterations, cellular senescence, stem cell exhaustion, and chronic inflammation¹. This has led to an increased interest in understanding the host and environmental factors that modulate the kinetics of aging, in order to try to slow these processes. Interestingly, an increasing body of evidence suggests that severe infections may impact the aging process. Indeed, after severe infections such as sepsis², tuberculosis³, or COVID-19⁴, metabolic and epigenetic scars can lead to dysregulation of immune responses leading, on the one hand, to systemic inflammation (inflammaging), and on the other hand, to poor responses to microbial stimulation (immune paralysis). Both inflammaging and immune paralysis are associated with biological aging. These effects likely contribute to the increased susceptibility to infectious and cardiovascular complications after sepsis, tuberculosis, or post-viral syndromes⁵. A comprehensive understanding of the effects and mechanisms through which chronic infections impact biological aging and age-dependent complications is however missing.

Human immunodeficiency virus (HIV) is a lentivirus that causes chronic infection, leading to loss of CD4 T-lymphocytes and eventually to severe opportunistic complications. With the advent of combination antiretroviral therapy (cART), people with HIV (PWH) can now achieve long-term viral suppression, leading to substantially increased life expectancy, close to individuals without HIV. However, PWH using cART are at increased risk to develop chronic inflammatory non-AIDS comorbidities, such as cardiovascular diseases, liver steatosis and fibrosis, and cancer^6,7, a pattern suggestive of premature aging processes^8–10. Therefore, understanding which biological processes underly premature aging in well-treated PWH may provide important insights into age-related comorbidities, the effect of chronic inflammation on aging, and future therapies to slow the aging process.

Progress in the methodologies used to characterize the aging process led to the development of molecular scores that mirror biological aging and longevity¹¹, with the epigenetic scores based on DNA methylation being the best known¹². More recently, plasma proteomics has emerged as a promising tool for studying accelerated aging, with initial efforts focusing on whole-body aging clocks^13,14. The latest advances extend this approach to the development of organ-specific aging clocks based on plasma proteomic signatures¹⁵. In the present study, we employed these molecular tools to investigate organismal and organ-specific aging in PWH treated with long-term cART.

The study is based on the 2000HIV project (n = 1850), a prospective longitudinal cohort of virally suppressed PWH using cART, integrating in-depth multi-omics data and clinical measurements¹⁶. We developed whole-body and organ-specific proteomic aging clocks using blood plasma proteomic data from a healthy cohort, we validated them in an independent PWH cohort (200HIV), and applied them thereafter to the entire 2000HIV cohort. We systematically analyzed the associations between proteomic age with DNA methylation age and with key clinical features, such as HIV disease stage, cART, comorbidities, and medication use. We used Mendelian randomization to identify causality relationship between aging scores and comorbidities. Finally, we assessed the effect of the HIV reservoir and cART on the biological aging of PWH.

Results

Plasma proteomic age scores are robust predictors of chronological and biological aging in PWH

To evaluate whether proteomic data can successfully predict biological age in healthy individuals and PWH, plasma proteins were analyzed in various cohorts using different Olink panels. 1254 proteins overlapped and were available in plasma samples across three independent cohorts: 200FG (general population), 200HIV (PWH cohort), and 2000HIV (PWH cohort). These proteins were subsequently used to train the proteomic aging clock models (Fig. 1a). Specifically, data from 200FG cohort served as the training set, while data from 200HIV and 2000HIV were used as independent test set and exploration set. Four machine learning algorithms—LASSO, elastic net, ridge regression, and LightGBM with bootstrap aggregation—were employed to train the models, and their performances were systematically compared. We observed that LASSO demonstrated comparable performance to elastic net and ridge regression, and they in turn overperformed LightGBM on both the training and test sets (Fig. 1c–e). Given that LASSO inherently incorporates feature selection and keep less features compared to elastic net, making it more effective at retaining proteins with high importance, we ultimately selected the LASSO to construct subsequent organ-specific proteomic aging clocks.

Fig. 1. Overview of the study design and analytic approaches.

a 1254 plasma proteins from 98 healthy participants in 200FG cohort were used for training models to predict chronological age using bootstrap aggregated LASSO regression model. Important features were selected to calculate the conventional age of PWH on 205 participants in 200HIV cohort. b Mutually exclusive organ-specific protein selected base on GTEX bulk organ RNA-seq were used to train in total 9 organ-specific aging clocks base on 200FG cohort. Models were tested using 200HIV and 2000HIV cohorts. As comparison, 5 DNA methylation age clocks were calculated for participants in 200HIV cohort. To test the validity of these aging models, the age gaps were associated with multiple measures of body index, HIV progression, cART, comorbidities, medication, genetic variance, and HIV reservoir. Source data are provided as a Source Data file.

Next, we utilized tissue RNA-seq expression data from the GTEx project to identify organ-specific proteins for the construction of organ-specific proteomic aging clock models (Fig. 1b). In total, we developed nine proteomic aging clocks: six organ-specific aging clocks for the artery, brain, intestine, lung, liver, and pancreas, along with three additional general clocks—a conventional aging clock (using all 1254 proteins), an all-organ aging clock (using 513 all-organ-specific proteins) (Supplementary Data 5), and an organismal aging clock (using 741 non-organ-specific proteins). These models were compared to assess their performance and usefulness as a proxy in predicting biological age. All nine proteomic aging models demonstrated a significant correlation between the predicted age and chronological age (Fig. 1g and Fig. S7). However, given the relatively poor performance of the models for the pancreas and lung, we excluded these two from subsequent downstream analyses.

To further validate the biological relevance of our proteomic age models, we used DNA methylation data from the 2000HIV cohort to calculate five well-established DNA methylation age scores: HorvathAge, HannumAge, PhenoAge, GrimAge, and GrimAge2, and calculated their correlations with the seven proteomic age scores (Fig. 2a). We observed a high correlation between the five DNA methylation ages, with an average Pearson correlation coefficient of r = 0.85, highlighting the consistency across different methylation age clocks. Notably, the conventional age, all-organ age, and organismal age clocks also exhibited high inter-correlations (r average = 0.86), and subsequently also showed strong correlations with the five DNA methylation ages (r average = 0.78), supporting the feasibility of predicting age using proteomic data from circulating proteins in PWH. Interestingly, we found that the artery age (r = 0.58–0.68) and brain age (r = 0.67–0.84) models demonstrated moderate to high correlations with the DNA methylation ages and the conventional age clock. In contrast, the liver (r = 0.33–0.43) and intestine (r = 0.40–0.46) age models exhibited weaker correlations with both the DNA methylation ages and the conventional age clock. Similarly, we compared the age gap derived from different age clocks (Fig. 2b). The distribution of proteomic age gap is depicted in Fig. 1b. In addition to observing moderate correlations among the age gaps of various DNA methylation clocks, we found that the artery age gap exhibited only weak correlations with the age gap calculated based on the other clocks (r average = 0.25). In contrast, the age gap for the liver and intestine clocks showed moderate-to-high correlations (r = 0.65), suggesting common mechanisms in the aging processes of these organs.

Fig. 2. Proteomic age acceleration correlates with epigenetic aging and clinical phenotypes in people with HIV.

a Pairwise Pearson correlation of proteomic age and DNA methylation age from participants in 2000HIV cohort. b Pairwise Pearson correlation of proteomic age gap and DNA methylation age gap from participants in 2000HIV cohort. Colors represent the magnitude and direction of the correlation coefficients, as indicated by the color scale. c Box plots depict the distribution of age gaps estimated by proteomic and DNA methylation aging clocks. Box plots show the median (center line), interquartile range (box), and whiskers indicating the minimum and maximum values. Colors are used to distinguish different age gaps. Forest plot displaying the associations between age gap and physical health indicators, including IMT(Intima media thickness), liver stiffness measurement, BMI (body mass index), Diastolic blood pressure (first), Creatinine, CAP liver score, eGFR (estimated Glomerular Filtration Rate), HEIGHT, and recent eGFR smaller than 90 (d) HIV stages related factors, including VL_LATEST(most recent viral load), AIDs malignancy ever (ever have an AIDS-defining malignancy), CD8_LATEST (most recent CD8 T cell count), Mutation before ART(HIV mutations conferring ART-resistance before start ART), HIV_STAGE, CD4_LATEST (most recent CD4 T cell count), CD4CD8_LATEST (most recent CD4/CD8 ratio), CD4_NADIR (lowest CD4 T cell count ever), CD4CD8_PRECART (most recent CD4/CD8 ratio before start ART) (e), current use of cART (f), comorbidities (g), and current medication usage (h) in PWH. For each of the associations between age gap and clinical parameters, the mean beta ±95% confidence intervals derived from linear models were plotted. All statistical tests were two-sided. False discovery rate–adjusted P values (q values) were calculated using the Benjamini–Hochberg procedure, and significant associations are indicated. Statistics were derived from individual study participants in the indicated cohorts. Sample size (n = 1850) refers to the number of participants included per group. Each data point represents one individual. Source data are provided as a Source Data file.

People with HIV show accelerated biological aging

Next, we assessed whether chronic HIV infection influences biological aging scores. The different age gaps calculated in PWH either by proteomic or methylation scores were strongly associated with chronological age, so we applied correction of chronological age on age gaps in all downstream analysis (Fig. S8). Importantly, we observed a trend of age acceleration in PWH in four out of the five methylation age gaps examined, with the exception of the PhenoAge gap (Fig. 2c). A similar acceleration of the aging process in PWH was apparent when assessing proteomic age scores, as shown by the conventional age acceleration that encompasses both the ‘all-organ age gap’ and the ‘organismal age gap’, as well as the brain age acceleration. In contrast, other organ-specific proteomic scores did not differ significantly from chronological age or displayed negative age acceleration, as in the case of intestine and liver, suggesting a divergent pattern of aging dynamics across different organs.

Although we lacked statistical power to obtain significant differences, we observed a general tendency that the conventional age acceleration in PWH of Asian descent was higher compared to European or African ancestries, both in younger and older populations (Fig. S9a). Additionally, sex-specific differences in organ-specific age acceleration were evident across different age strata. Males with PWH exhibited higher intestine and organismal age acceleration compared to females, whereas females demonstrated higher artery and liver age gaps (Fig. S9b). These findings highlight the heterogeneity in organ-specific age gaps in PWH, which may be linked to distinct clinical characteristics and comorbidity susceptibilities.

Next, we leveraged clinical data from the 2000HIV cohort to further examine the relationship between proteomic age and biological aging. Specifically, poorer physical health indicators (Fig. 2d), HIV-related parameters (Fig. 2e), different comorbidities (Fig. 2g), and medication usage (Fig. 2h) were all significantly correlated with age advancement. Notably, we also observed a significant impact of current use of antiretroviral medication (Fig. 2f) on age advancement, suggesting their potential impact on biological aging in PWH. For organ-specific age, the age advancement of liver and intestine showed the strongest associations with body mass index (BMI) and the following hepatic parameters: CAP liver score (steatosis parameter), liver stiffness measure (LSM) and alanine aminotransferase (ALAT) concentrations (Fig. S10a, b, c, e). The age advancement of all organ-specific ages was significantly correlated with creatinine (CREAT) concentration (Fig. S10d), arguing for an important impact of kidney function on biological aging.

PWH were also classified based on lowest CD4 T cell count, ranging according to the level of immunosuppression (Stage 1: ≥500 CD4 cells/µL, Stage 2: 200–499 CD4 cells/µL, and Stage 3: <200 CD4 cells/µL or an AIDS-defining diagnosis). We found that a more advanced HIV stage was significantly associated with conventional age, organismal age, and all-organ age advancement, but no significant association with organ-specific age scores (Fig. 3b). This association was also reflected in key HIV parameters, including latest CD4 T cell count, lowest recorded CD4 T cell count, latest CD8 T cell count, most recent CD4/CD8 T cell ratio, and the CD4/CD8 T cell ratio before initiating cART (Fig. 3a, Fig. S11a, b, c, d). Although all PHW were virally suppressed because of cART or spontaneous HIV control, some individuals (3.24%) have low level viremia (>40 HIV-RNA copies/ml). Interestingly, we found that the latest viral load was significantly associated with the advancement of organismal age, all-organ age, and conventional age, while no strong associations were observed with organ-specific aging clocks (Fig. 3c). Also, consistently undetectable viral load over the past three years was significantly associated with less organismal age acceleration, which indicate that chronic HIV infection with episodes of low-level viremia contributes to accelerated ageing (Fig. S11e). These findings strongly suggest that chronic HIV infection contributes to accelerated systemic biological aging. The importance of an undetectable plasma viral load is also underlined by comparing elite controllers (n = 21) that spontaneous control HIV infection with persistent undetectable plasma viral loads without cART and normal progressors, using cART. Elite controllers showed a trend of age deceleration on organismal age, although the difference did not reach statistical significance, most likely due to limited number of individuals in the elite controller group (Fig. 3d). Furthermore, in elite controllers we also observed a trend of less age acceleration in multiple proteomic and DNA methylation aging clocks, including intestine age, artery age, liver age, HorvathAge, PhenoAge, GrimAge, GrimAge2 (Fig. S12).

Fig. 3. Associations between immune aging, HIV disease parameters, and viral reservoir size.

Forest plot displaying the associations between age gaps and latest CD4CD8 ratio (a), HIV stages (b), and latest viral load (c). The significant association (FDR) is shown. d Box plots illustrate the distribution of organismal age gaps between ART naive elite controllers and normal progressors on ART stratified by chronological age group (18–35, 36–60, ≥60 years). Box plots show the median (center line), interquartile range (box), and whiskers indicating the minimum and maximum values. Colors are used to distinguish different groups. e Forest plot displaying the association between HIV total reservoir (orange) and intact reservoir (blue) with age gaps. The significant association (FDR) is shown. For each of the associations between age gap and clinical parameters, the mean beta ±95% confidence intervals derived from multivariable linear models were plotted. All statistical tests were two-sided. False discovery rate–adjusted P values (q values) were calculated using the Benjamini–Hochberg procedure, and significant associations are indicated. Sample size (n = 1850) indicates the number of individuals included in each regression or association analysis. Analyses were performed at the level of individual participants; no technical replicates were used. Models were adjusted as described in “Methods”. Source data are provided as a Source Data file.

Although HIV replication can be spontaneously controlled in elite controllers or by cART in the other infected individuals, HIV persists in latently infected cells, mostly resting memory CD4 + T cells. Around 95% of the reservoir consists of defective proviral HIV DNA that, unlike intact proviral DNA, does not contribute to viral replication after cART interruption, but may still result in HIV RNA transcripts or proteins that elicit a host response. To more precisely analyze the effect of HIV on the aging process, we investigated whether the total or the intact HIV reservoir size is associated with biological aging¹⁷. Interestingly, we observed that all conventional, all-organ, organismal, and brain proteomic and DNA-methylation age (except HannumAge) acceleration scores were significantly associated with total HIV reservoir (Fig. 3e). This is highly overlap to the age gaps we reported in Fig. 2c. In contrast, proteomic and epigenetic age scores were not associated with the intact HIV reservoir levels.

Mendelian randomization demonstrates a causal relationship between biological aging and co-morbidities

To assess whether the impact of HIV infection on inflammatory aging scores is biologically relevant, we investigated whether age acceleration is associated with increased susceptibility to complications. Exploring organ-specific aging scores, we found that brain age advancement was significantly associated with central nervous system (CNS) complications, including multiple sclerosis, epilepsy, Alzheimer’s disease, Parkinson’s disease, and other CNS disorders (Supplementary Data 12), which represent well-established morbidities in people with HIV (PWH) (Fig. 4a).Furthermore, artery age advancement showed a strong correlation with cardiovascular disease, particularly deep venous thromboembolism (VTE) (Fig. 4b, c). Importantly, we found that all-organ age advancement was significantly correlated with mortality within two years follow-up period (β = 4.68, 95% CI [1.70, 7.66], FDR = 0.021). In addition, the conventional age gap also showed a moderate correlation with two-year mortality (β = 3.25, 95% CI [0.37, 6.13], P = 0.027) (Fig. 4d). Although most associations did not reach statistical significance due to the limited number of death cases (n = 24) and different death causes, including malignancy and infection (Supplementary Data 6), these data strongly suggest that proteomic age advancement may serve as a predictive survival biomarker in PWH. Collectively, our findings confirm relation between organ-specific age clock and organ-specific morbidities, while all-organ clock relates more closely with mortality.

Fig. 4. Associations between biological aging, clinical outcomes, and causal inference analyses.

Forest plot displaying the associations between age gaps and diagnosis of CNS disease (central nervous system disease) (a), cardiovascular disorder (other) (b), venous thromboembolism (c), and 2-year mortality (d) Significant associations are indicated based on false discovery rate (FDR) correction. e Forest plots showing the causal effect as estimated using each of the SNPs on their own, and comparing against the causal effect as estimated using the methods that use all the SNPs. Plots are presented for those results with IVW P value < 0.05 in the forward MR. f Scatter plots illustrating the distribution of SNP estimates of age advancement of intestine on GWAS cardiovascular disease outcome. Different colors represent different MR methods. Plots are presented for those results with IVW P value < 0.05 in the forward MR. For each of the associations between age gap and clinical parameters, the mean beta ±95% confidence intervals derived from linear models were plotted. The two-sided P values are reported in Supplementary Data 4 and Supplementary Data 9, no adjustment for multiple comparisons was applied. Scatter plot showing the relationship between LBP (g), CD14 (h), and FABP2 (i) expression and intestine age. Each dot represents an individual, and the red line indicates the mean fitted linear regression model with 95% confidence interval shading. Regression coefficients (β) and two-sided P values are shown, no adjustment for multiple comparisons was applied. Statistics were derived from individual participants. Sample size (n = 1850) indicates the number of individuals per group or analysis. Source data are provided as a Source Data file.

To further investigate whether organ-specific, all-organ, conventional and organismal age advancement causally impact disease outcomes, we performed Mendelian Randomization (MR) analysis. For this, we used 119 independent SNPs as instrumental variables (IVs) significantly associated with the exposure (age gap) at P < 1 × 10^–5, and extracted the summary statistics of GWAS performed in cardiovascular diseases (stroke, coronary artery disease, myocardial infarction, etc), diabetes, liver fibrosis or steatosis. We tested for causal relationship in exposures with more than eight variants, performing sensitivity analyses. A strong causal relationship between age advancement and disease outcomes, which met all the above-described criteria (see “Methods”), was observed between intestinal age advancement and cardiovascular disease (β = 0.0005, P_IVW = 0.03) (Fig. 4e, Supplementary Data 7, Supplementary Data 8)¹⁸. Scatterplots showed a linear regression line for the positive associations between age gap and risk of corresponding disease (Fig. 4f). The causal estimates obtained from the rest of MR methods (MR-Egger, weighted median, weighted mode, and simple mode, simple median, maximum likelihood and inverse variance weighted with fixed effects) showed similar direction to those from the primary IVW method (Supplementary Data 9). Of note, there were no evidence of significant heterogeneity or pleiotropy (P < 0.05, Supplementary Data 4). Furthermore, MR-PRESSO detected no outliers (P for global test of pleiotropy >0.05) and the leave-one-out sensitivity analyses showed no single SNP had a substantial impact on the results.

To further validate the relationship between intestinal age advancement and cardiovascular disease, we selected circulating protein markers associated with intestinal integrity measured in the 2000HIV cohort (LBP, CD14, and FABP2) and tested them for association with the intestinal age (Fig. 4f–h). In line with the hypothesis tested, we found a significant positive association between the biomarkers associated with intestinal integrity and age advancement. Overall, Mendelian randomization analyses indicate that intestinal age advancement is linked to cardiovascular disease, and we hypothesize that translocation of gut bacterial products due to loss of mucosal integrity may contribute to systemic inflammation and accelerated atherosclerosis.

Differential effect of specific ART drugs on biological and specific organ age acceleration as measured by proteomic clock in PWH

As our data shows the impact of HIV on biological aging, we next assessed a possible effect of the different antiretroviral drugs on biological aging. As virologically suppressed chronic HIV infection is shown to have important impact on biological aging, and biologic age gap in our PWH was correlated to comorbidities, also during 2 year follow up, we next assessed whether there is a difference in the effect of specific antiretroviral medication on age acceleration in PWH. Cumulative drug exposure was used to investigate the impact of ART on aging, as the effect of the medication may be dose and time related, and in order to account for treatment switches. The cumulative drug exposure and all age acceleration scores were explored by regression analyses (Fig. 5a).

Fig. 5. Antiretroviral drug exposure and modulation of biological aging.

a Heatmap displays the regression beta effect size representing the association between ART drugs and age gaps. ART drug classes include integrase strand transfer inhibitors (INSTI), non-nucleoside reverse transcriptase inhibitors (NNRTI), nucleoside reverse transcriptase inhibitors (NRTI), and protease inhibitors (PI). Color intensity represents the magnitude and direction of the regression beta coefficients. Statistical significance of each association was determined by linear regression models. Scatter plot showing the relationship between cumulative exposure to rilpivirine (RPV) (b), nevirapine (NVP) (c), tenofovir disoproxil fumarate (TDF) (d), ritonavir (RTV) (e) and conventional age gap. Each dot represents an individual and the purple line indicates the mean fitted linear regression model with 95% confidence interval shading. Regression coefficients (β) and two-sided P values are shown, no adjustment for multiple comparisons was applied. f The table summarizes significant associations between cumulative exposure to RPV, lamivudine (3TC), emtricitabine (FTC), and tenofovir alafenamide (TAF) and age gaps, after controlling for the effects of all possible co-administered antiretroviral drugs. For each drug, the corresponding drug type, estimated beta effect size, 95% confidence interval, and two-sided P value are reported, no adjustment for multiple comparisons was applied. Statistics were derived from individual study participants. Sample size (n = 1850) indicates the number of individuals included per analysis. Source data are provided as a Source Data file.

We observed that exposure duration to certain nucleoside reverse transcriptase inhibitors (NRTIs), particularly lamivudine (3TC), was significantly associated with decreased age acceleration, as estimated by either proteomic aging in the intestine or DNA methylation-based models such as HannumAge, GrimAge, and GrimAge2. In contrast, other NRTIs—including tenofovir disoproxil fumarate (TDF), tenofovir alafenamide (TAF), and emtricitabine (FTC)—exhibited divergent effects on organ-specific aging trajectories. Notably, TDF and FTC exposure duration was significantly associated with accelerated DNA methylation age. Interestingly, TDF demonstrated a stronger association with deceleration of conventional aging metrics, compared to TAF, which is a prodrug of TDF. TDF and TAF both result into the active compound tenofovir diphosphate, but TDF results in higher plasma concentrations while TAF leads to higher intracellular levels in certain cells such as lymphocytes and hepatocytes. Remarkably, increased liver ageing score was seen with accumulating TAF exposure. Furthermore, stavudine (D4T), an NRTI with well-documented toxicities, showed significant associations with age acceleration in multiple organs, including the intestine and liver, although it was also paradoxically associated with less age acceleration in the brain. These findings highlight the heterogeneous and organ-specific effects of antiretroviral therapy drugs on biological aging. Among non-nucleoside reverse transcriptase inhibitors (NNRTIs), both rilpivirine (RPV) and nevirapine (NVP) were significantly associated with lower biological aging across multiple organ-specific measures. Within the class of integrase strand transfer inhibitors (INSTIs), dolutegravir (DTG) was notably linked to lower age scores of the intestine and liver. As for protease inhibitors (PIs), cobicistat (COBI) showed significant associations with decelerated aging scores in the artery and overall organismal age, as well as in methylation-based measures such as GrimAge and GrimAge2.

Considering the cumulative drug exposure time of all agents, only TDF, RPV and NVP showed a significant association with conventional age deceleration (Fig. 5b–d,). Ritonavir (RTV), a protease inhibitor, showed significant conventional age acceleration (Fig. 5e). Given that antiretroviral drugs are typically administered in combination, we cataloged both the currently and cumulative ART use in our cohort (Supplementary Data 10 and Supplementary Data 11). To disentangle the independent contributions of individual drugs, we constructed multivariable regression models to assess the specific effects of selected ART agents—including dolutegravir (DTG), bictegravir (BIC), doravirine (DOR), nevirapine (NVP), rilpivirine (RPV), elvitegravir (EVG), darunavir (DRV), lamivudine (3TC), emtricitabine (FTC), tenofovir alafenamide (TAF), and tenofovir disoproxil fumarate (TDF)—on aging outcomes. We then highlighted the most significant associations (Fig. 5f). Notably, 3TC exhibited the strongest association with deceleration of both organismal and intestinal aging scores. Another NRTI, FTC, was also significantly linked to deceleration of organismal age scores. Among the NNRTIs, RPV showed consistent associations with deceleration across conventional, multi-organ, and artery-specific aging metrics. Intriguingly, TAF was associated with acceleration of both all-organ and brain aging scores, consistent with the differential aging effects observed earlier when compared to TDF.

Collectively, our findings indicate that certain classes of cumulative ART exposure—particularly NRTIs and NNRTIs—are associated with reduced age acceleration in PWH. These associations became even more pronounced for specific agents such as RPV, 3TC, and FTC after adjusting for the effects of ART co-administration, suggesting independent, protective roles against biological aging. These results provide a rationale for future randomized clinical trials to investigate the impact of specific ART regimens on biological age gaps as surrogate outcomes, and on non-AIDS comorbidities as definitive clinical endpoints in PWH.

Discussion

An increasing number of recent studies argue that biological age may be a better predictor of health compared to chronological age¹⁹. Recently, organ-specific aging scores have been developed based on plasma proteomics, deepening our understanding of the aging processes¹⁵. Many factors may contribute to premature aging, such as chronic low-grade inflammation, oxidative stress and mitochondrial dysfunction. All these processes are disturbed in PWH, even when virally suppressed because of effective cART²⁰. In addition, lifestyle and environmental factors such as smoking, alcohol use, physical inactivity, and socioeconomic stress are also known to influence the heterogeneity of biological aging observed among PWH. In the present study, we applied a comprehensive analysis to predict whole-body and organ-specific aging using plasma proteomic data and epigenetic age scores, providing a systematic evaluation of factors influencing biological aging in PWH. This analysis resulted several findings: 1. people with HIV show premature biological aging, both at the level of the entire organism, as well as at the level of specific organs; 2. age acceleration is related to the total HIV reservoir size, confirming the role of HIV infection in the aging process; 3. organ and systemic proteomic aging clocks are associated with comorbidities and mortality, with a causal relationship between premature intestinal aging and CVD.; and 4. the cumulative use of certain antiretroviral drugs, in particular some belonging to the class of reverse transcriptase inhibitors, is associated with decreased systemic and organ-specific aging in PWH, while others show opposite effects in line with their well-known toxicity profile.

The first major observation of our study is that PWH show premature biological aging. This is in line with a recent study in which we have shown that especially young PWH show accelerated inflammatory aging¹⁴, and accompanies studies that showed an increased epigenetic age in HIV-infected individuals²¹. Our finding that the total HIV reservoir size was associated with biological aging scores, argues for a direct role of HIV infection or more specifically of viral transcripts that originate from defective proviral DNA, but still capable to induce inflammation¹⁴. The intactness of the viral reservoir, that mostly reflect only 5% of the total reservoir size, does not seem to be important for age acceleration. Our finding of lower age acceleration in elite controllers, that are known to have a small HIV reservoir size that is mostly locked up in transcriptionally silent DNA regions, also underscores the role of HIV infection itself for impacting biological aging.

The relevance of the association between the HIV infection and biological aging is apparent at several levels. First, Mendelian randomization demonstrates that accelerated aging in PWH is not a mere epiphenomenon, but is a cause of severe co-morbidities. From this perspective, the causal link demonstrated between intestinal aging and cardiovascular diseases is especially relevant: it is tempting to speculate that intestinal aging in PWH is associated with increased leakage of microbial products, leading to systemic inflammation and subsequently cardiovascular complications. Indeed, this hypothesis is strengthened by the association of gut translocation biomarkers with intestinal aging. Second, increased organ age acceleration was associated with higher risk for co-morbidities and mortality during follow-up, demonstrating the importance of biological age acceleration for health span.

Another strong argument for the relevance of chronic infection with HIV for biological aging is provided by the counter-regulatory effects of certain antiretroviral agents, notably the reverse transcriptase inhibitors, both NRTI as well as NNRTI, on age acceleration. Both the current use as well as the cumulative drug use were associated with age deceleration. Earlier studies have suggested that some anti-HIV drugs have anti-aging effects in PWH, especially dolutegravir²². We now demonstrated that this effect is shared by several groups of anti-HIV medication, which argues that down-regulation of HIV expression and proliferation itself has a beneficial anti-aging effect. These findings raise additional important questions to be investigated in future studies: do other chronic (viral) infections have similar effects on biological aging? Could antiretroviral medication exert anti-aging effects also in non-HIV individuals, for example by inhibition of retroelements known to be activated during the aging process? Our data warrant more studies on the anti-aging effects of ART, which opens a direction of investigation in the field of aging research.

This study has also limitations. First, due to variations in cohorts and batch effects in the proteomic data, we had only a partial overlap in plasma proteins between the healthy and HIV datasets. Furthermore, we only considered the mutual organ specific proteins rather than any overlapping highly expressed protein across organs. This constrained the number of proteins available for training of the organ-specific aging models, leading to the omission of key organs such as the heart and kidney. In addition, as our training set was relatively small (comprising 98 healthy participants), the model may not fully capture the heterogeneity encountered in real-world clinical settings: this limitation should be taken into consideration as well. Second, this is a cross-sectional study, albeit large, and future longitudinal studies with long follow-up are warranted to strengthen the conclusions drawn here. Third, the sex imbalance within the PWH cohort presents another limitation. Although sex was included as a covariate in our analyses, this imbalance may still introduce potential biases in downstream analyses. Finally, the majority of the volunteers in the cohorts studied are individuals of European descent, and therefore the conclusions of the studies should only be cautiously extrapolated to other populations. Future studies should validate these findings in non-European populations.

In conclusion, we systematically revealed the impacts of HIV infection and antiretroviral drugs on biological aging in PWH. Proteomic and epigenetic aging scores demonstrate predictive power for future all-cause mortality and showed significant associations with comorbidities, highlighting their potential as biomarkers for aging and disease risk. Importantly, these aging measures may also help to guide ART optimization, especially for individuals at higher risk of treatment-related or aging-associated complications.

Methods

Human cohorts

Our study investigated two independent cohorts of PWH, the 200HIV and 2000HIV cohorts, as well as a population-based cohort of individuals without HIV from the general population (the 200FG cohort) recruited within the Human Functional Genomics Project (HFGP). The 200HIV cohort is a cross-sectional cohort of PWH receiving combination antiretroviral therapy, established to study host immune function and inflammation during treated HIV infection. The 2000HIV cohort is a large, well-characterized observational cohort of PWH in long-term clinical follow-up, with extensive longitudinal clinical, immunological, and molecular data collected during suppressive antiretroviral therapy. The 200FG cohort comprises healthy volunteers from the general population recruited to investigate inter-individual variation in immune responses. Detailed information regarding cohort recruitment, inclusion criteria, and follow-up procedures has been previously described¹⁴. In this study, we utilized proteomic data from 98 volunteers from the 200FG cohort, 205 volunteers from the 200HIV cohort, and 1850 volunteers from the 2000HIV cohort, along with DNA methylation data, genetic data, and clinical information (Table 1). The study included adults with HIV enrolled in the 200HIV and 2000HIV cohorts. Participant sex was self-reported. The number of participants, sex distribution, and age range are provided in Table 1 and in the Nature Portfolio Reporting Summary. Written informed consent was obtained from all participants prior to inclusion. Participants did not receive financial compensation for study participation. This study was complied with all relevant ethical regulations. The 2000HIV study protocol was approved by an accredited medical research ethics committee, the Independent Review Board Nijmegen (NL68056.091.81) and published at clinicaltrials.gov (ID: NCT03994835).

Table. 1.

The demographic characteristics of the volunteers from the three cohorts: 200FG, 200-HIV and 2000-HIV

		Description	2000HIV	200HIV	200FG
basic	N		1850	205	98
	Age mean (±-sd)	Mean age at the time of baseline visit	50.7 ± 17.2	77.7 ± 8.9	73.2 ± 10.8
	Age range	Range of age at the time of baseline visit	19 − 89	36 − 104	27 − 101
	Female %	Percent of female	15%	7.80%	24.50%
	BMI(kg/m2) (+-sd)	Body mass index	25.53216 ± 4.253038	24.1 ± 3.530286	–
	ethnicity	Ethnical background (self-report)
	Asian		84 (4.54%)	–	–
	Black		187 (10.10%)	–	–
	Hispanic		49 (2.65%)	–	–
	Mixed	Two grandparents of at least two different ethnical ancestries	130 (7.03%)	–	–
	Native American		3 (0.16%)	–	–
	White		1395 (75.41%)	205 (100%)	98 (100%)
	Unknown		2 (0.11%)	–	–
Clinical features	IMT(mm) (+-sd)	What is the Intima media thickness?	0.6746048 ± 0.1408462	–	–
	Liver stiffness measurement(kPa) (+-sd)	Liver stiffness measurement	4.82263 ± 2.074682	–	–
	RRSYST1(mmHg) (+-sd)	First blood pressure (systolic)	131.4512 ± 16.07886	–	–
	Diastolic blood pressure (first)(mmHg) (+-sd)	First blood pressure (diastolic)	80.61456 ± 9.751658	–	–
	Creatinine(μmol/L)(+-sd)	What was the most recent value of Creatinin?	91.06284 ± 28.91067	–	–
	CAP liver score(dB/m)(+-sd)	Controlled attenuation parameter	250.5757 ± 53.37944	–	–
	eGFR (ml/min/1,73 m2)(+-sd)	What was the most recent eGFR?	67.06714 ± 22.8976	–	–
	Recent eGFR smaller than 90 (ml/min/1,73 m2)	Specify the most recent eGFR if EGFR was 90 or <90
	≤90		1181 (63.84%)	–	–
	>90		665 (35.95%)	–	–
	Unknown		4 (0.22%)	–	–
	Height(cm) (+-sd)	Body height	177.7562 ± 8.888835	–	–
HIV related	HIV duration (years)		13.70517 ± 8.198396	8.47 ± 6.611545	–
	VL_LATEST(copies/ml)	What was the most recent viral load?
	>40		60 (3.24%)	–	–
	Undetectable		1790(96.76%)	–	–
	AIDs malignancy ever	Did patient ever have an AIDS-defining malignancy?
	Yes		113 (6.11%)	–	–
	No		1730(93.51%)	–	–
	Unknown		7 (0.38%)	–	–
	CD4_LATEST(10^9 cells/L)(+-sd)	What is the most recent CD4 T cell count at baseline?	0.740951 ± 0.302558	–	–
	CD4_NADIR(10^9 cells/L)(+-sd)	What was the lowest CD4 T cell count ever?	0.290259 ± 0.2170506	–	–
	CD8_LATEST(10^9 cells/L)(+-sd)	What is the most recent CD8 T cell count at baseline?	0.9340733 ± 0.4613754	–	–
	CD4CD8_LATEST(+-sd)	What is the most recent CD4/CD8 ratio	0.9497539 ± 0.49561	–	–
	CD4CD8_PRECART(+-sd)	What was the most recent CD4/CD8 ratio before start ART?	0.3473684 ± 0.2699624	–	–
	Mutation before ART	Were there HIV mutations conferring ART-resistance before start ART?
	Yes		161 (8.70%)	–	–
	No		567 (30.65%)	–	–
	Unknown		1122 (60.65%)	–	–
	HIV_STAGE	CDC stage at the time of start ART
	Stage 1	≥500 CD4 + T cellen/µL	241 (13.03%)	–	–
	Stage 2	200–499 CD4 + T cellen/µL	846 (45.73%)	–	–
	Stage 3	<200 CD4 + T cellen/µL or AIDS defining diagnosis	710 (38.38%)	–	–
	Unknown		33 (1.78%)	–	–
	RESIDUAL_VIR	During the 3 years prior to baseline, was the viral load always unquantifiable? Quantifiable viral load is VL > 40 copies/ml
	Yes		1488(80.43%)	–	–
	No		350(18.92%)	–	–
	Unknown		12(0.65%)	–	–
	Elite controller	Primary comparison between elite controllers (EC) and non-controlling PLHIV on ART (Non-EC)
	EC_persistent		60(3.24%)	–	–
	EC_transient		54(2.92%)	–	–
	Non-EC		1697(91.73%)	–	–
	Unknown		39(2.11%)	–	–
cART	CART.NRTIs	Participant uses a NRTI (such as abacavir, lamivudine, emtricitabine, but NOT tenofovir) as antiretroviral therapy?
	Yes		1758(95.03%)	–	–
	No		92(4.97%)	–	–
	CART.NtRTIs	Participant uses a NtRTI (such as tenofovir alafenamide or tenofovir disproxil) as antiretroviral therapy?
	Yes		1286(69.51%)	–	–
	No		564(30.49%)	–	–
	CART.NNRTIs	Participant uses NNRTI (non-nucleoside reverse transcriptase inhibitor) as antiretroviral therapy?
	Yes		727(39.30%)	–	–
	No		1123(60.70%)	–	–
	CART.Integrase_inhibitors	Participant uses an integrase inhibitor as antiretroviral therapy?
	Yes		1013(54.76%)	–	–
	No		837(45.24%)	–	–
	CART.Protease_inhibitors	Participant uses a protease inhibitor as antiretroviral therapy?
	Yes		176(9.51%)	–	–
	No		1674(90.49%)	–	–
	CART.Other	Use of booster regimes (cobicistat or ritonavir)
	Yes		368(19.89%)	–	–
	No		1482(80.11%)	–	–
co-Comorbidity	Parkinsons disease	Was participant ever diagnosed with Parkinsons
	Yes		4(0.22%)	–	–
	No		1846(99.78%)	–	–
	DM2	Was participant ever diagnosed with diabetes mellitus type 2
	Yes		88(4.76%)	–	–
	No		1762(95.24%)	–	–
	COPD	Was participant ever diagnosed with COPD
	Yes		58(3.14%)	–	–
	No		1792(96.86%)	–	–
	Inflammatory arthritis	Was participant ever diagnosed with inflammatory arthritis
	Yes		24(1.30%)	–	–
	No		1826(98.70%)	–	–
	CVD.Other	Was participant ever diagnosed with any other cardiovascular disease (Other than hypertension, myocardial infarction, stroke, cardial arrhythmia, peripheral arterial vascular disease, thromboembolism, pulmonary embolism, or angina pectoris)
	Yes		80(4.32%)	–	–
	No		1770(95.68%)	–	–
	Hypertension	Was participant ever diagnosed with hypertension
	Yes		459(24.81%)	–	–
	No		1391(75.19%)	–	–
	Angina pectoris	Was participant ever diagnosed with angina pectoris
	Yes		57(3.08%)	–	–
	No		1793(96.92%)	–	–
	Asthma	Was participant ever diagnosed with asthma
	Yes		98(5.30%)	–	–
	No		1752(94.70%)	–	–
	Liversteatosis	Was participant ever diagnosed with Liversteatosis
	Yes		198(10.70%)	–	–
	No		1652(89.30%)	–	–
	Deep venous thromboembolism	Was participant ever diagnosed with deep venous thromboembolism
	Yes		65(3.51%)	–	–
	No		1785(96.49%)	–	–
	CNS disease	Was participant ever diagnosed with a central nervous system disease/disorder
	Yes		155(8.38%)	–	–
	No		1695(91.62%)	–	–
	Endocrine metabolic disorders	Endocrine (hormonal) or metabolic disorders diagnosed
	Yes		563(30.43%)	–	–
	No		1287(69.57%)	–	–
	RESP	Was participant ever diagnosed with a pulmonary disease
	Yes		222(12%)	–	–
	No		1628(88%)	–	–
	CVD	Cardiovascular disorder diagnosed
	Yes		595(32.16%)	–	–
	No		1255(67.84%)	–	–
	Metabole syndrome	Does subject meet criteria of metabole syndrome
	Yes		502(27.14%)	–	–
	No		1348(72.86%)	–	–
	MH_TR_GE	Was participant ever diagnosed with a gastroenterology disease
	Yes		439(23.73%)	–	–
	No		1411(76.27%)	–	–
Medication use	MED_STAT.Evolocumab	Did participant use Evolocumab in the last two weeks
	Yes		3	–	–
	No		1847	–	–
	MED_ANTIINFLAM.Other	Did participant use other anti-inflammatory medication in the last two weeks (Other than NSAID or corticosteroids)
	Yes		8	–	–
	No		1842	–	–
	MED_ANTIDIAB.Insulin	Did participant use insulin in the last two weeks (Both rapid and long acting)
	Yes		25	–	–
	No		1825	–	–
	MED_ANTIDIAB.Suderivate	Did participant use sulfonylureum derivative medication in the last two weeks
	Yes		21	–	–
	No		1829	–	–
	MED_ANTIDIAB.Other	Did participant use other antidiabetic medication in the last two weeks (Other than insulin, metformin or SU derivatives)
	Yes		6	–	–
	No		1844	–	–
	MED.Antidiabetics	Did participant use antidiabetic medication in the last two weeks
	Yes		79	–	–
	No		1771	–	–
	MED_PSY.Other	Did participant use other psychotropic medication in the last two weeks (Other than benzodiazepines, antipsychotics, lithium, TCA, SSRI or stimulants)
	Yes		31	–	–
	No		1819	–	–
	MED_ANTIDIAB.Metformin	Did participant use metformin in the last two weeks
	Yes		64	–	–
	No		1786	–	–
	MED_AC.VKA	Did participant use vitamin K antagonist medication in the last two weeks
	Yes		25	–	–
	No		1825	–	–
	MED_PLATINH.Clopidogrel	Did participant use clopidogrel in the last two weeks
	Yes		43	–	–
	No		1806	–	–
	AB_1M	Has patient used any antibiotics the previous 30 days?
	Yes		81	–	–
	No		1769	–	–
	MED.Antihypertensiva	Did participant use antihypertensive (=blood pressure lowering) medication in the last two weeks
	Yes		402	–	–
	No		1448	–	–
	MED.Anticoagulants	Did participant use anticoagulant medication in the last two weeks
	Yes		206	–	–
	No		1644	–	–
	MED_AC.Plateletinhibitors	Did participant use platelet inhibitory medication in the last two weeks
	Yes		159	–	–
	No		1691	–	–
	MED_PLATINH.ASA	Did participant use acetylsalicylic acid in the last two weeks (Such as ascal or carbasalaatcalcium)
	Yes		121	–	–
	No		1728	–	–
	MED_STAT.Rosuvastatin	Did participant use rosuvastatin in the last two weeks
	Yes		184	–	–
	No		1666	–	–
	MED.Cholesterol_lowering	Did participant use cholesterol lowering medication in the last two weeks
	Yes		384	–	–
	No		1466	–	–
	MED.Other	Did participant use other medication that do not fit in the above mentioned categories in the last two weeks (Medication that does not fall in the category of anticoagulant, antidiabetic, cholesterol lowering, antihypertensive, anti-inflammatory, psychotropic, vitaminD or antimycotic)
	Yes		744	–	–
	No		1106	–	–
	MED.No_other	Did participant use no other medication than their HIV medication in the last two weeks (except for potentially cART)
	Yes		538	–	–
	No		1312	–	–
2-year-mortality		Dead after 2 years follow up
	Yes		24(1.30%)	–	–
	No		1826(98.70%)	–	–

Proteomic profiling of circulating plasma proteins

Plasma proteins were measured using a proximity extension assay coupled with next generation sequencing as a readout method by OLINK Proteomics AB (Uppsala Sweden)²³. Protein measurements are delivered as Normalized Protein expression (NPX) values, which is Olink’s relative protein quantification unit on log2 scale. Olink has developed a built-in quality control (QC) system using internal controls to control over technical performance of assays and samples.

Plasma proteins from the 2000HIV and 200FG cohort were measured in three batches. The first batch (n = 692 samples) was measured using the library Olink® Explore 1536 consisting of 1472 proteins divided into four 384-plex panels focused on inflammation, oncology, cardiometabolic and neurology proteins (panels I). The second batch (n = 692 samples) was measured using the Olink® Explore Expansion 1536 consisting of 1472 proteins divided into four 384-plex panels focused on additional inflammation II, oncology II, cardiometabolic II and neurology II proteins (panels II). The third batch was measured using the full library (Olink® Explore 3072) consisting of 3500 proteins divided into eight 384-plex panels focused on inflammation, oncology, cardiometabolic and neurology proteins (panels I and II). Detailed panel information can be found in Supplementary Data 1. For this study, we used the first and third batch (1500 proteins). Bridging normalization was performed to remove batch effect between panels by following the next steps for each protein¹: we first calculated the median of the bridging samples for each protein in the two batches²; subsequently, we calculated the median difference keeping one batch as a reference³; finally, we subtracted the median difference from each protein in the non-reference batch. Limit of detection (LOD) values per protein were re-adjusted by the same adjustment factor as the respective protein measurements after bridging normalization.

After removing batch effects using bridging normalization, standard QC per protein and sample was performed prior to statistical data analysis. In each of the four panels from the Olink® Explore 1536 platform, IL6, TNF, CXCL8 proteins were measured as technical duplicates for QC purposes. Strong correlations were observed between the technical duplicates among panels (Spearman rho correlation r > 0.9), and therefore, we selected the measurements from the inflammatory panel. Next, we excluded proteins with lower limit of detection (LOD) >= 25 of the samples, resulting in 1306 proteins in the 2000HIV and 200FG cohort for follow-up analysis. In addition, to detect outliers, we performed principal component analysis (PCA) using the NPX values. Outliers were defined as those samples falling above or below four standard deviations (SD) from the mean of principal component one (PC1) and/or two (PC2). After removing outliers, 1850 and 98 samples remained in the 2000HIV and 200FG cohort, respectively. Proteomic profiling of the 200HIV was performed using the library Olink Explore 1536 platform. QC per protein and sample was performed as described above. After excluding proteins with LOD > = 25 and samples based on PCA (Fig. S1a), 1254 protein measurements of 205 samples remained for follow-up analysis of the 200HIV cohort.

Model benchmarking

We compared the performance of four machine learning methods—LASSO, elastic net, ridge regression, and LightGBM on the training set 200FG (n = 98) and age was predicted into two independent cohorts of PWH, the 200HIV (n = 205) and 2000HIV cohort (n = 1,850). For all models, the normalized expression levels of 1254 proteins measured using the Olink platform were used as input features to predict the chronological age of each sample. During model training, the training set was subjected to 500 rounds of bootstrap resampling, for each bootstrap iteration hyperparameter tuning was conducted using five-fold cross-validation implemented through the caret package²⁴ in R. The final predicted age values are derived by averaging the outputs from all 500 optimized models. The final model selection was based on the average R2 and average RMSE in both training set and validation cohorts. The robustness test of all bootstrap aggregated LASSO aging models was shown in Figs. S2 and S3. Given the relatively small size of our training dataset, we sought to enhance the credibility of our aging clock by benchmarking its performance against previously published proteomic aging clocks trained using LASSO regression (Lehallier et al., 2019²⁵; Hamilton et al., 2023¹⁵) (Fig. S4a). The predicted proteomic ages exhibited a strong Spearman correlation with those from the reference models (Fig. S4b, c), providing robust support for the reliability of our approach.

Organ-specific protein selection

We used normalized bulk tissue RNA-seq expression data from the GTEx database²⁶ to identify organ-specific proteins, as previously described¹⁵. Since the GTEx dataset includes a large number of sub-organs, we consolidated these sub-organs into corresponding main organs followed by the same categories as previous research and defined the expression level of each gene in a main organ as the highest expression value observed among its sub-organs. Considering that the LASSO model inherently applies stringent feature selection, we aimed to retain as many input protein features as possible. To define organ-specific genes, we used proteins for which the expression level of a gene in one organ was at least twice as high as its expression in any other organ. These organ-specific genes were then mapped to our Olink proteomic data. As a result, we identified 513 organ-specific proteins that successfully mapped to our dataset, accounting for approximately 41% of the total proteins analyzed. Based on the distribution of organ-specific protein counts (Fig. S1f), we first selected the brain, lung, artery, liver, pancreas, and intestine (at least with over 15 features) as target organs.

Organ-specific aging clock training and age gap calculation

We applied the bootstrap LASSO model using the expression levels of organ-specific proteins identified for the artery, brain, intestine, lung, liver, and pancreas in the training set (200FG) as input features to predict their biological age. The predicted values for the training set were calculated as the mean of the outputs from 500 bootstrap iterations. As a supplementary analysis, we also trained three additional aging clock models for comparison: a conventional age clock with all proteomic expression, an all-organ age clock with all organ-specific protein, and an organismal age clock with all non-organ specific protein. After evaluating the average performance across 500 bootstrap-trained models (Figs.2,3) and corresponding out-of-bag (OOB) (Supplementary Data 2) estimates for each organ, we selected the brain, artery, liver, and intestine for downstream analyses based on their consistently robust predictive accuracy. These models were included to evaluate and contrast their predictive performance with the organ-specific aging clocks.

The predicted age obtained from each model was designated as the organ-specific age. The trained models were then applied to the 200HIV and 2000HIV datasets to calculate the organ-specific age for PWH. The age gap for each sample in each model was defined as the difference between the predicted age and the chronological age:

Age gap = Predicted age—Chronological age

To account for baseline age gaps in healthy individuals and their variation across age groups, we adjusted the age gap in the 2000HIV cohort by subtracting the mean age gap of the corresponding age group in the 200FG dataset, as follows:

Corrected age gap≤35=Age gap−E[(Predicted age−Chronological age)∣FG200 ≤ 35]

Corrected age gap36 − 60=Age gap−E[(Predicted age−Chronological age)∣FG200 36 − 60]

Corrected age gap>60=Age gap−E[(Predicted age−Chronological age)∣FG200 > 60]

where E[⋅]represents the expected value (mean) within each respective age group in the 200FG dataset (≤35 years, 36–60 years, and >60 years).

This adjusted 2000HIV age gap was used in all subsequent analyses.

DNA methylation profiling

DNA methylation profiling was performed on 1914 samples as previously described^16,27. DNA was extracted from EDTA whole blood by the Radboudumc Genetics Department using the ChemagicStar automated configuration (Hamilton Robotics) with magnetic polyvinyl alcohol (M-PVA) bead-based technology. DNA concentration and purity (260/280 nm ratio) were assessed using a NanoDrop spectrophotometer. Samples were normalized to 50 ng/µL in TE buffer and randomly assigned to plates. High-quality samples were analyzed using the Illumina Infinium MethylationEPIC BeadChip array (manifest B5).

DNA methylation data processing

Standard sample- and probe-level quality control (QC) procedures were applied. Raw IDAT files from the 2000HIV cohort were processed using the minfi package in R (v4.2.0)²⁸. Samples with gender mismatches or poor quality were excluded. Probes were removed if they had >10% missing values (detection P > 0.01), mapped to sex chromosomes, overlapped with common SNPs (MAF > 5% in European populations), or mapped to multiple genomic loci. Stratified quantile normalization was applied²⁹. Methylation β-values were calculated as β = M / (M + U + 100), where M and U represent the methylated and unmethylated signal intensities, respectively.

DNA methylation age calculation

Normalized methylation β-values were used to estimate DNA methylation age (DNAm age) with five blood-based epigenetic clocks: HorvathAge¹², HannumAge³⁰, PhenoAge³¹, and GrimAge/GrimAge2³², using the DNAm age calculator (https://dnamage.clockfoundation.org; accessed October 2024). Preliminary age advancement scores were obtained by subtracting chronological age from DNAmage. To remove the confounding effect of chronological age, we regressed DNAm age on chronological age and used the resulting residuals—termed residual DNAm age gaps (Fig. S5)—for downstream analyses.

Linear modeling and meta-analyses

We examined the associations between corrected organ age gaps and various factors, including physical health indicators, HIV stages, comorbidities, medication usage, and antiretroviral therapy (ART) in the 2000HIV cohort. To assess these relationships, we employed linear models controlling for age, sex, and ethnicity, using the following formula (detailed Ethnicity information was recorded in Table 1):

Corrected age gap ∼ Variable of interest + Age + Sex + Ethnicity

For accumulative ART analysis, we employed linear models controlling for age and all ART co-administration:

Corrected age gap ∼ ART accumulation + Age + co-administration ART

Additionally, we applied the Benjamini-Hochberg method to adjust for multiple testing burden where appropriate (indicated as q-value). Meta-analyses with multiple linear regression model were conducted in R using glmnet²⁴ package to compare and aggregate effect sizes and confidence intervals across multiple age models.

Genotyping, quality control and imputation

DNA was extracted from each participant’s whole blood. The Illumina Infinium Global Screening Array was used for genotyping all participants of multiple ancestries in the 2000HIV cohort. Prior to imputation, QC for raw variants and samples was performed using PLINK v1.90b³³. Genetic variants with a call rate genotype missingness of more than 5% and those deviating from Hardy-Weinberg equilibrium (HWE) with a P value < 10^-6 were excluded from the dataset. The HWE exact test was performed with variants stratified by ancestry. Samples with a call rate <97.5% and those that showed a heterozygosity rate that deviated more than three standard deviations (SD) from the mean heterozygosity rate per self-reported ancestry were excluded. Genetic variants that passed QC were converted from GRCh37 to GRCh38 genomic build using the UCSC liftOver tool³⁴. Next, TOPMed Freeze5 was used on genome build GRCh38 to align strands to the TOPMed reference panel. We used the McCarthy group tools for alignment (https://www.well.ox.ac.uk/~wrayner/tools/). After QC, 582,404 variants from 1864 individuals were retained for the imputation procedure. The filtered raw variants were uploaded to the TOPmed Imputation server and imputed against the TOPMed (version r2 on GRCh38) reference panel. The imputed variants were filtered using BCF tools stratified by ethnicity, excluding variants with low imputation quality scores (R2 < 0.3 or ER2 < 0.7) or MAF < 1%. This yielded 10,810,841 variants from 1864 members of the 2000HIV multi-ancestry cohort.

Quantitative trait locus mapping on age advancement

We performed quantitative trait locus (QTLs) mapping using the imputed genetic data and age advancement scores of 2000HIV samples of European ancestry. After imputation, 1331 samples of European ancestry had both genetic and age advancement scores. First, a standard post-imputation QC was performed using PLINK v1.90b. During QC per SNP, SNPs that deviate from HWE with a P value < 10^-6 and MAF below 5% were excluded. We mapped the age advancement scores to genotype data using a linear model with age, sex and the first five genetic PCs as covariates to account for population stratification. Furthermore, we evaluated genomic inflation factors (λGC) to assess potential population stratification within our cohort. A λ value close to 1 indicates no detectable inflation of test statistics and therefore minimal confounding due to population structure. The calculated λ values were as follows: λ_artery = 0.992, λ_brain = 0.982, λ_conventional = 1.007, λ_intestine = 1.009, λ_liver = 1.021, λ_organismal = 1.002, λ_pancreas = 1.011, λ_lung = 1.017, and λ_{all_organ} = 1.011. All values are close to 1, confirming the absence of significant population stratification in our QTL analyses. as a covariate. QTLs associated to age advancement were selected to perform mendelian randomization as described in detail below.

Mendelian randomization

To determine causality between organ-specific, all-organ, conventional and organismal age advancement and certain diseases, two sample Mendelian randomization (MR) was performed using the R package TwoSampleMR version 0.6.9 with default settings. For each exposure (organ-specific, all-organ, conventional, and organismal age advancement), genetic variants used as instrumental variables (IVs) were extracted from the summary statistics of QTLs on age advancement identified in this study and publicly available genome-wide association studies (GWASs) on disease outcomes of interest in the OpenGWAS database³⁵. An overview of studies extracted from the Open GWAS database included in the Mendelian randomization analyses are provided in the Supplementary Data 3 (n = 63 studies). All GWAS studies on disease outcomes used in this study were from populations of European ancestry to eliminate demographic stratification bias.

The IVs were selected based on the following criteria: SNPs were significantly associated with the exposure (age advancement) in the 2000HIV cohort (P < 1 × 10^–5) and, a stringent clumping was performed to extract independent SNPs. Genetic variants were clumped using a linkage disequilibrium (LD) r2 < 0.001 and a window size of 10,000 kb using the 2000HIV cohort of European ancestry as a reference for clumping. SNP proxies were added automatically by the TwoSampleMR R package. In total, 148 independent SNPs were selected as instrumental variables after clumping. SNPs associated to exposure were from GRCh38 to GRCh37 genomic build before MR.

Only if there were still at least six SNPs remaining per exposure, MR was performed. In addition, SNPs that were associated with the age advancement of more than 5 exposures using the 2000HIV cohort of European ancestry were considered pleiotropic and excluded (Fig. S6). Twenty-three (n = 23) SNPs showed pleiotropic effects and removed, resulting in 119 unique independent SNPs. Further, the strength of each SNP was assessed by F-statistic using the formula F = β2/Se2, whereas β and Se are the coefficient and standard error of exposure respectively. SNPs with F-statistics <10 were regarded as weak IVs and should be discarded in the following analyses. None of the SNPs were discarded due to F-statistics. Finally, Steiger-filtering was used to exclude SNPs, which explain more variance in the outcome than the exposure, as these SNPs are likely to be invalid instruments (which either act though horizontal pleiotropy or proxy a reverse causal pathway from outcome to exposure). We harmonized exposure and outcome data to ensure that effect estimated corresponded to the same allele for each SNP (Supplementary Data 4).

We used the inverse variance weighted (IVW) method as the main approach to evaluate the potential causal relationship between age advancement and disease outcomes by combining the β-values and the standard errors of the causal estimate from them (forward MR-IVW). Seven additional effective methods, including MR-Egger, weighted median, weighted mode, and simple mode, simple median, maximun likelihood and inverse variance weighted with fixed effects were also applied to evaluate the possible causal relationship comprehensively. For MR results with significant IVW (P < 0.05), sensitivity analyses were performed to evaluate if the causal estimates are robust to violations of MR underlying assumptions.

First, we performed the mendelian randomization pleiotropy residual sum and outlier test (MR-PRESSO) to detect potential outlier variants³⁶. Furthermore, the MR-Egger regression was used to evaluate the bias generated by gene pleiotropy, of which the intercept is an indicator. In addition, the Cochran’s Q statistics was applied to quantify the heterogeneity between SNPs. Thirdly, we used the leave-one-out analysis to verify whether there are SNP outliers that strongly affect the results by eliminating each SNPs and then re-calculating the causal estimates using the IVW method on the rest. Finally, to evaluate the possibility of reverse causality between age advancement and diseases, we performed MR-IVW in the other direction (reverse MR; GWAS diseases used as an exposure and age advancement as outcome). We precluded results for which less than five SNPs were available as instrumental variables per exposure. For the reverse MR, we used the exposure/outcome pairs that reached P nominal significance in the forward MR-IVW analyses (P < 0.05). We excluded any results that had a nominal significant IVW results in the reverse MR while passing all the sensitivity analyses. A strong causal relationship between the age advancement and disease was considered when the following criteria were met¹: the IVW method demonstrated a significant difference (forward MR, P < 0.05)²; the seven MR methods provided consistent estimations (forward-MR)³; the Cochran’s Q test, MR-Egger test, and MR-PRESSO global test had no significance (forward MR, P > 0.05) and⁴; the reverse MR showed no significant difference (MR-IVW P > 0.05).

All statistical analyses were performed using R statistical software (version 4.2.0). LD clumping was performed using rtracklayer::liftOver in R software. QTL mapping was performed using the MatrixEQTL package³⁷. Mendelian randomization was performed u using the ‘TwoSampleMR’, ‘MR-PRESSO’, ‘ieugwasr’ R packages.

Quantification of total and intact HIV-1 DNA from CD4 + T cells

Total and intact HIV-1 DNA levels were measured in CD4 + T cells from 1850 PWH by digital PCR (dPCR) using the Rainbow proviral HIV-1 DNA assay. Starting from 40 Mio cryopreserved PBMCs, CD4 + T-cells were enriched by negative selection using EasySep Human CD4 + T-cell isolation kit on the Robosep-S (Stemcell Technologies, Vancouver, Canada). Genomic DNA (gDNA) was extracted using the QiaAmp DNA mini kit on the Qiacube (Qiagen, Hilden, Germany) with two elution steps of 50 µL. DNA concentrations were determined using 2 µL of the eluted DNA with the SpectraMax Quant AccuBlue HiRange dsDNA Assay Kit by using the SpectraMax i3x (Molecular Devices, San Jose, California, United States). Samples were stored at −20 °C prior to dPCR quantification. HIV-1 DNA levels were quantified in triplicate by dPCR using the Rainbow proviral HIV-1 DNA assay on the QIAcuity Four platform (Qiagen, Hilden, Germany). Depending on the DNA concentration, either 18 µL of eluted gDNA was used per replicate for samples with concentrations below 90 ng/µL, or 10 µL for those exceeding 90 ng/µL. HIV-1 DNA levels were normalized by measuring the reference gene RPP30 in duplicate and reported per million CD4 + T-cells. For normalization and DNA shearing assessment, a 1/100 dilution was made for each sample and 5 µL was used as input. Total HIV-1 DNA levels were assessed by the RU5 region in the Rainbow assay and were highlighted when the result was below the limit of detection (i.e. 10 copies per well). Intactness levels were obtained by the presence of at least two (psi and env) and maximum five target regions (RU5, psi, gag, pol and env) in the Rainbow assay. Automatic thresholds were calculated with the Rainbow Shiny tool and adapted if needed per sample when a threshold crossed the double-positive population. In case the intactness result was a zero-value, results of individual targets were checked. If positive partitions were observed in all target regions, the intactness result was undetectable and was artificially calculated as one intact copy in the total cell input for that sample. If there were no positive partitions in psi and/or env due to presumable signal failure, no intactness level was reported.

Statistical analysis

In this study, we reported false discovery rate (FDR) values for all meta-analyses and P values for individual regression analyses. Statistical significance was indicated as follows: ***: FDR or P < 0.001, **: FDR or P < 0.01, *: P < 0.05, NS: not significant.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization: M.G.N., A.vd.V., Y.Z., V.M. Methodology: Y.Z., V.M., N.V., M.B., W.V., A.G., L.vE., J.S., M.B.e, M.D. Investigation: Y.Z., V.M., N.V., M.G.N., A.vdV. Visualization: Y.Z., V.M., N.V. Funding acquisition: M.G.N., A.vd.V. Project administration: M.G.N., A.vdV. Supervision: M.G.N., A.vdV. Writing—original draft: Y.Z., V.M., N.V., M.G.N., A.vdV. Writing—review & editing: Y.Z., V.M., N.V., M.G.N., A.vdV., C.R., T.O., L.J., C.J.X., Y.L., L.V.

Peer review

Peer review information

Nature Communications thanks Na He, Sulin Wu and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The proteomic datasets analysed in this study are part of the ongoing 2000HIV project and have been deposited in the Radboud University institutional data repository under the 10.34973/qk29-f305. Genetic data analysed in this study have been deposited in the Radboud University institutional data repository under the 10.34973/p96d-kz55. In accordance with the 2000HIV consortium data governance policy and ethical approvals, individual-level clinical and molecular data are currently under embargo and cannot be publicly released until formal project completion. The full dataset will become openly accessible in January 2027. Prior to this date, access to the minimum dataset required to interpret, verify, and extend the findings of this study may be granted to qualified researchers upon request, subject to approval by the 2000HIV data access committee and compliance with institutional and ethical regulations. Requests should be directed to the corresponding authors (Prof. Dr. Andre van der Ven and Prof. Dr. Mihai G. Netea). Access requests are typically reviewed within a reasonable timeframe, and approved data access is granted for research purposes in accordance with the approved request. Source data are provided with this paper.

Code availability

All R scripts used for data preprocessing, model training, and age estimation rely exclusively on publicly available packages, and no custom proprietary code was developed. Code for generating proteomic aging clocks has been deposited in Zenodo at 10.5281/zenodo.18328533. The code is provided under the MIT License.

Competing interests

M.G.N. is a scientific founder of TTxD, Lemba, Biotrip and Salvina. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-69412-1.