Unsupervised analysis shows shared phenotypes of HIV and age

As a preliminary exploration of this dataset, we ran an unsupervised analysis to identify age-associated methylation sites and their relation to HIV infection. Analysis of a previous methylome-wide screen of 538 healthy subjects (Hannum et al., 2012) identified as many as 61,592 methylation sites associated with age at a 1% false-discovery rate (FDR, likelihood ratio test in multivariate regression model with Benjamini-Hochberg correction). Validation of these sites in whole blood from a second control cohort from the European Prospective Investigation into Cancer and Nutrition (Riboli et al., 2002) (EPIC, N = 662) confirmed 26,927 of these sites as strongly associated with age (Figure 1A, Table S2).

Open in a separate window

Figure 1

Shared epigenetic signature of HIV infection and aging

A, Discovery and validation of CpG methylation markers associated with age. B, Distribution of t-statistics measuring association of each methylation marker with HIV status. Colors indicate groups of markers identified in (A): Gray, all markers; yellow, age-associated markers from discovery phase; violet, subset of age-associated markers confirmed in validation. C, Principal component (PC) analysis of the validated age-associated markers, in which the first PC (y-axis) is positively associated with both age (x-axis) and disease status (HIV+, green; HIV−, blue). D, Potential relationships among HIV infection, epigenetics, disease, and aging. Black: known; Dashed gray: potential; Green: connections explored in this study. See also Table S2 and Table S3.

Among these validated age-associated sites, we found a striking association with methylation in the HIV+ patients relative to healthy controls (P < 10⁻¹⁰⁰, Figure 1B). Further analysis of these sites found a positive association of the first principal component with both age and HIV status (Figure 1C, Table S3, association by multivariate linear model P < 10⁻⁸). These findings support a link between HIV infection and aging (Rickabaugh et al., 2015), as quantitatively measured by epigenomic profiling (Figure 1D).

Benchmarking and refinement of epigenetic aging models

Given the shared effects of HIV and aging, we sought to determine whether HIV causes the same biological aging signature as previously found in cohorts of uninfected individuals (Hannum et al., 2012; Horvath et al., 2013, Marioni et al., 2015). We tested aging models from both our group (Hannum et al., 2012) and Horvath (Horvath, 2013) in three independent datasets derived from whole blood samples (Hannum et al., 2012; Riboli et al., 2002; this study; Table S4). Although the Hannum and Horvath modeling efforts were based on different methodologies and training data, we found they made very similar predictions (r = 0.9, Pearson’s correlation, Figure 2A) and furthermore that a consensus of the two models outperformed either model individually (Figures 2B and 2C, Table S4). For this reason, we used this consensus model for all remaining analyses.

Open in a separate window

Figure 2

Epigenetic models accurately predict age and indicate advanced aging for HIV-infected individuals

A, Scatter plot comparing the ages predicted using the Hannum et al. and Horvath models on healthy controls (n = 1,246 from HIV−, Hannum et al. and EPIC datasets). Red points indicate patients that were discarded due to disagreement between the two aging models (n = 66). B–C, Accuracy of the consensus model (y-axis) to predict true chronological age (x-axis) in datasets from Hannum et al. (n = 497, B) or EPIC (n = 637, C). Panels (A–C) show patients between 25 and 68 years old. D, Scatter plot of predicted biological age (consensus aging model) versus chronological age for HIV− healthy controls. E, Scatter plot of biological time versus chronological time since HIV onset for infected subjects. F, Violin plots showing the distribution of residuals from regression of biological versus chronological age. Three groups are shown: HIV− controls, short-term HIV+ infected individuals, and long-term HIV+ infected individuals. Note that the red circle indicates an outlier, which is not used to fit the violin profile, but is used in all statistical assessments. A–E, Black dashed lines indicate diagonal (y = x). r, Pearson’s correlation coefficient. ** indicates P < 10⁻⁵. See also Table S4.

A potential issue with these models arises in the fact that methylation profiles from whole blood are influenced by cell composition, and different cell types have different methylation states (Jaffe and Irizarry, 2014). These differences might be particularly pronounced in HIV-infected patients, some of whom have low CD4⁺ T cell counts (Trono et al., 2010). To understand the sensitivity of epigenetic aging models to cell type composition, we downloaded two datasets profiling sorted cells across shared sets of individuals (Absher et al., 2013, GSE59250; Reynolds et al., 2014, GSE56046). Among these sorted cell datasets, we saw good concordance of epigenetic age predictions with chronological age (Figure S2A–F). Epigenetic age was reproducible across different cell types profiled from the same patients, with high agreement of age estimates (r > 0.77–0.88) and moderate but very significant agreement of age advancement (Pearson’s r > 0.45–0.68; P < 0.0001 for all associations, Figure S2G–J).

While we therefore expect the contribution of cell composition to be minimal, we nonetheless developed an algorithm to individually normalize each methylation profile using methylation-derived cell type information. In brief, we used a previously reported method (Jaffe and Irizarry, 2014) to reliably predict blood composition (Figure S3) and adjust out the expected contribution of cell-type specific effects. This procedure greatly limited the effects of age and HIV induced blood composition changes in downstream analyses (Experimental Procedures, Figure S4).

HIV+ individuals have advanced DNA methylation age

We next used this consensus aging model to calculate the ‘biological age’ of each individual in our cohort (Table S4). For uninfected controls, the calculated biological age had a very high concordance with chronological age (Figure 2D, Pearson’s r = 0.94). In contrast, the HIV+ patients had a biological age advancement of 4.9 years on average (P < 10⁻⁸ by Student’s t-test, 95% confidence interval 3.4 – 7.1 years, Figures 2E and 2F). These results were consistent with our previous unsupervised analysis (Figures 1B and 1C) in suggesting that HIV infection leads to advanced aging. Furthermore, we found that the age advancement of HIV+ individuals was negatively correlated with the ratio of CD4⁺/CD8⁺ T lymphocytes (Spearman’s rho = −0.2, P < 0.02). CD4⁺ T cells are a major indicator of immune integrity (Leung et al., 2013; Serrano-Villar et al., 2014) and are inversely associated with morbidity and mortality, including from non-AIDS defining diseases (El-Sadr et al., 2006); similarly, the CD4/CD8 ratio predicts non-AIDS morbidity (Leung et al., 2013; Serrano-Villar et al., 2014). This finding links biological aging of HIV-infected individuals to a clinical measure of disease progression, and it raises the possibility that patients with stable immune responses may be less affected by the advanced aging phenotype. Taking into account a recently estimated 4.2% increase in mortality risk per year of biological age advancement using the Hannum model (Marioni et al., 2015), the changes observed in HIV+ patients result in an expected total mortality risk increase of 19%.

Age advancement is independent of HIV duration

Notably, patients more recently infected with HIV (< 5 years) had no significant difference in age advancement from those patients with chronic (>12 years) infection (P > 0.5, Mann-Whitney U Test; Figure 2F). Similar findings emerged from a regression analysis of the chronological versus biological time since infection: the slope did not differ from one (0.98 ± 0.06, standard error) whereas the y-intersect was significantly positive (5.2 ± 0.9; Figures 2E and 2F). These findings lend support to the theory that age advancement occurs early in the course of disease as a consequence of acute infection or reaction to drug treatment (Guaraldi et al., 2011; Smith et al., 2012). The lack of an increase of age advancement with disease duration seems to contradict alternative views that HIV-mediated aging occurs through cumulative effects of latent virus (Appay and Rowland-Jones, 2002) or chronic therapeutic intervention (Torres and Lewis, 2014). We did however observe less variation in age advancement within the chronically infected HIV+ individuals (Figure 2F, P < 0.002, Bartlett’s test relative to recently infected group), perhaps reflecting the comparative stability of infection and immune response on long-term cART therapy (Luz et al., 2014; Rosenblatt et al., 2005).

Age advancement is independent of cellular composition

While the direct effects of cell type composition on the whole blood methylome were corrected by the adjustment described above (also see Experimental Procedures), we considered that it was still possible that changes in cell-type composition could lead to downstream, indirect changes in the epigenomes of all blood cells. If this were the case, cell-type associated changes could be responsible for the observed increase in biological age in the HIV+ cases. To assess this possibility, we constructed a multivariate linear model in which cell-type composition variables and HIV status were used to predict biological age as measured by the methylome (Table 1). In this model, the presence of HIV was associated with an age advancement of 3.8±1.1 years, while the presence of natural killer cells accounted for additional increases in biological age. In an even more conservative test, we modeled age advancement with cell-type composition variables alone and found that the unexplained variation in this model still had a significant association with HIV infection (P = 0.02, Likelihood Ratio Test, Table 1). Thus, even in a very conservative analysis, HIV infection still has association with advanced aging that is independent of cell composition.

We also sought to experimentally assess if the observed age advancement due to HIV infection was also observed in purified cell populations. Using standard calculations of statistical power, we estimated that a sample of 48 patients, balanced approximately between cases and controls, would have 81% power to detect the same aging advancement effect as our primary screen at p < 0.01. Accordingly, this number of subjects was prospectively recruited from the University of Nebraska Medical Center (Experimental Procedures, Table S5). Whole blood was separated immunomagnetically to isolate pure populations of neutrophils and CD4⁺ T cells.

As in whole blood, unsupervised analysis showed a clear effect of HIV in age-associated methylation markers (Figure 3A–B). Application of epigenetic models of aging in these pure-cell datasets showed good concordance of predicted age with chronological age in both cell types (Figure 3C–F). In neutrophils, the Hannum model predicted a 2.5 year increase in age due to HIV infection (P < 0.03, 95% CI 0.6–5.0 years, Figure 3E) whereas the Horvath model showed a smaller effect of 0.4 year (P > 0.05). In contrast, CD4⁺ T-cells had a much stronger and more consistent HIV response in both models, with the consensus aging model showing an increase of 5.7 years in the HIV+ subjects (P < 10⁻⁵, 95% CI 3.4–7.9 years, Figure 3F). These data indicate that the effect of epigenetic age advancement is not merely an artifact of changing blood composition, but likely reflects true aging signals. The stronger effect size within CD4⁺ T-cells (Figure 3G–H) suggests that these cells may be exposed to more age-like stress than neutrophils, although further work is needed to understand how disease may affect aging rates across different cell types and tissues.

Open in a separate window

Figure 3

Age advancement in validation cohorts of purified cells

A–B, Unsupervised principal component (PC) analysis of methylation patterns in purified blood cell types, in which the first PC is positively associated with both age (x-axis) and disease status (HIV+, green; HIV- blue). (A) New CD4+ T-cell cohort across 5999 markers that are age-associated in CD4⁺ T-cells (GSE59250). (B) New neutrophil cohort across markers probes that are age-associated in neutrophils (GSE65097). C–F, Control (C–D) and HIV+ (E–F) subjects for sorted cell validation datasets comparing chronological age to the Hannum et al. epigenetic aging model in neutrophils (C, E) and consensus aging model in CD4⁺ T-cells (D, F). G–H, Violin plots showing age advancement in the two sorted cell datasets. For B, in initial analysis the first PC heavily reflected an outlier point, which was removed for this analysis after which the PC was recalculated. See also Figure S2.

HIV and aging have shared and distinct methylation patterns

Having identified a large effect of HIV infection on age associated methylation signals, we then sought to better understand the wider changes instigated by HIV. We identified 81,361 CpG markers associated with HIV infection (Benjamini-Hochberg corrected P < 0.01; likelihood ratio test using a multivariate linear model, Table S2). Of these, 2569 upregulated markers and 1769 downregulated markers were also associated with aging, a 3.2 and 1.4-fold enrichment over random expectation, respectively (Figure 4A, Fisher’s Exact Test P < 10⁻³⁰, Table S6). We found that markers associated with both HIV and aging were enriched in DNAse hypersensitivity sites and transcriptional start sites, suggesting methylation changes in DNA regions under active regulation. These CpG markers were also enriched in binding sites for polycomb repressive complex (PRC2) (Figure 4B), a switch that tightly regulates genes required for differentiation and renewal, and in Drosophilia is linked to longevity (Siebold et al., 2010). These findings reinforce previous reports that PRC2 targets are irreversibly repressed by methylation during the aging process (Beerman et al., 2013; Deaton and Bird, 2011; Teschendorff et al., 2010). Interestingly, markers associated with HIV but not aging had a very different functional enrichment profile (Figure 4B), indicating additional mechanism(s) for epigenetic alteration associated with HIV.

Open in a separate window

Figure 4

HIV and aging have shared and distinct methylation patterns

A, Overlap table comparing the set of CpG markers associated with HIV and the set of validated age-associated markers (see Figure 1A). Numbers indicate probe counts in each overlap, colors correspond to odds ratio of overlap compared to background. B, Odds ratios of enrichment for a panel of genomic features, evaluated in sets of markers associated with age, HIV, or both. PRC2, polycomb repressive complex 2 binding sites; DHS, DNase hypersensitivity sites; TSS, transcription start sites. C, Distribution of methylation states for the CpG marker sets defined in (A). See also Table S2 and Table S6.

We have previously reported that age-associated markers in older subjects tend away from a fully methylated or unmethylated state and instead move towards disorder (with a methylation fraction of 50% representing complete disorder) (Hannum et al., 2012). We found that HIV-infected patients displayed a similar trait: among markers associated with HIV, 66% tended towards disorder, compared with 70% of age-associated markers (Experimental Procedures, Figure S5). Furthermore, whereas age-associated markers tended to have a low methylation fraction that increased with age, HIV-associated markers were more equally balanced between low and high methylation states (Figure 4C).

Selection criteria and subject recruitment

HIV+ subject samples were obtained from CHARTER as a Resource (www.charterresource.ucsd.edu). The CHARTER study was comprised of HIV-infected participants at varying stages of disease and with differing histories of antiretroviral treatment, with a focus on neuromedical and neurobehavioral assessments (Heaton et al., 2010). We requested information on subjects for which DNA had been obtained. Demographic and clinical data were filtered for non-Hispanic white males (to match the control group) who were free of Hepatitis C virus, not diabetic, on cART, and adherent to therapy. Two groups were selected for study, those more recently infected by HIV (but after the acute infection stage, 0.8 – 5.0 years of infection) and those chronically infected (>12.0 years). As a control, 44 non-Hispanic white males without HIV were recruited from the San Diego area.

For validation samples 35 HIV+ subjects along with 25 healthy controls were recruited prospectively for the purpose of this study to match the characteristics of the primary cohort. Cells were purified using immunomagnetic separation and DNA was extracted from purified cell populations. While most subjects used had both neutrophils and CD4⁺ T-cells profiled, differing DNA yield for some subjects prohibited profiling of both cell types for some patients.

Clinical and demographic data are presented in Table S1 and Table S5.

Sample collection and methylation analysis

DNA was purified from whole blood samples using PaxGene collection tubes (Qiagene) and FlexiGene DNA extraction kits (Qiagen). Methylation analysis was performed using Infinium HumanMethylation450 BeadChip Kits (Illumina). 500ng of DNA was bisulfite converted using EZ DNA Methylation Kits (Zymo Research) and subsequently processed for HumanMethylation450 BeadChips following manufacturer’s instructions. Following hybridization, BeadChips were scanned using the Illumina HiScan System.

Data pre-processing

All methylation data for HIV+ and HIV− subjects were deposited in the Gene Expression Omnibus (GEO) under accession number GSE67705. For the Hannum et al. (Hannum et al., 2012) and EPIC (Riboli et al., 2002) studies, raw data were obtained from GEO accessions GSE40279 and GSE51032. All data were processed through the Minfi R processing pipeline (Aryee et al., 2014). Cell counts were estimated by the estimate Cell Counts function in Minfi using flow sorted cell populations made available by Houseman et al. (Houseman et al., 2012). To limit variability in methylation levels due to differing cell type composition in the whole blood samples, methylation levels were adjusted for each CpG marker as follows:

• Average methylation levels for each cell type were obtained from the Houseman et al. (Houseman et al., 2012) flow sorted blood dataset.

• A theoretical methylation level was assessed for each patient by assuming their blood to be a mixture of these pure cell populations at the estimated cell type proportions.

• The difference of each patient’s methylation level from the average was assessed.

• This difference was subtracted from the original raw dataset.

We followed the protocol established to be optimal by Marabita et al. (Marabita et al., 2013) first quantile normalizing the data and then performing beta-mixture quantile (BMIQ) normalization (Teschendorff et al., 2013). To limit batch effects, all arrays across the three studies were normalized together. For use in the Horvath methylation age model, raw data were normalized to a gold standard reference distribution following the protocol provided in the manuscript (Horvath, 2013). The sole deviation from the Horvath protocol was an additional cell composition adjustment performed in a similar manner as described above, after BMIQ normalization. While the cell composition adjustment was not part of either the Horvath or Hannum et al. processing pipeline, recent work (Jaffe and Irizarry, 2014) has shown cell type composition to be a key confounding factor in methylation analysis.

Benchmarking the aging models

Aging models were assessed using the Hannum et al. and EPIC (Riboli et al., 2002) datasets. While the Hannum et al. dataset was used to train both epigenetic aging models, the EPIC data were made available after the time of construction of both models and thus provide an independent assessment of performance. We limited analysis to patients between the ages of 25 and 68 years of age for better comparison to the HIV cohort. Among the HIV and EPIC cohorts, we saw slightly better performance of the Hannum model (which was trained using only whole blood data) than the Horvath model (trained in a variety of tissues), but when a simple average of these two models was taken (the ‘consensus model’), we found better performance than either separately (Table S2).

Epigenetic model concordance filter

One key drawback of current models of molecular age is the lack of a confidence measure in model prediction for any particular individual. To address, this we utilized the concordance between the two models as an additional filter of data quality. Despite general agreement between the models, in a number of subjects biological age predictions varied by more than 20%. This analysis resulted in the filtering of three HIV+ cases and two HIV− controls in our primary cohort (Table S1).

Linear scaling of epigenetic age

For both models and across all datasets a linear scaling factor existed when comparing chronological verses biological age. In order to properly compare the performance of the models and to best calibrate them to our dataset, we performed a linear adjustment to all model fits for the control data to a unit slope with a zero intercept. Note that this affected the model error when compared in an absolute sense, but did not affect the correlation between biological and chronological age. In HIV+ patients, we adjusted this particular cohort to the regression fit of the matched controls.

Screening for differentially methylated markers in response to HIV infection

For the results described in Figure 4 and ​and5,5, we ran a multivariate linear model to test for differentially methylated markers in response to HIV infection. This model used predicted cell type composition and age as covariates. Significance was assessed via a likelihood-ratio test for the improvement of a model fit with HIV as the variable of interest.

Disorder of Methylation in Response to HIV and Aging

We observed increasing disorder of the methylome by both aging and HIV infection. To assess the possibility that the increasing age advancement might be explained by increasing disorder, we conducted a principal component analysis on 7967 age-associated markers that trended away from < 50% of sites methylated with age(Figure S5A). This analysis produced a similar result to that shown in Figure 1C, in which the first principal component of the cohort was still associated with both with age and HIV infection. Furthermore, we observed that only 231 of 436 (53%) markers used in the two aging models tended towards methylation values of 50% of sites methylated (i.e. disorder, binomial P = 0.2) and found that while there is a general entropy increase in HIV+ patients across the entire methylome (Figure S5B), there is no such effect in the markers used in the biological age models (Figure S5C).

Identification of differentially methylated regions

We identified 25,491 markers that were associated with HIV, trended away from disorder, and were not associated with age. For the discovery and visualization of differentially methylated regions (Figure 5A), we calculated a rolling statistic on the density of ‘hits’ in 200 marker windows. From this analysis it was clear that a genomic region encompassing the HLA and histone gene clusters was enriched for markers in our query set, and post-hoc analysis confirmed a strong enrichment in the genomic interval traditionally assigned to the HLA region (~29MB – 33MB on chromosome 6, odds ratio = 1.3, P < 10⁻¹⁰). For further refinement and visualization of this signal we conducted a similar scan statistic on a section of chromosome 6 in Figure 5D. In this targeted analysis we relaxed our criteria and looked for regions of consistent increases or decreases in methylation in HIV+ verses HIV− subjects. This analysis showed a number of ‘peaks’ of hypomethylation both in the histone gene region as well as near the HLA genes.

This work was supported by the National Institute of Mental Health (P30 MH062261), the National Cancer Institute (P30 CA023100) and the California Institute for Regenerative Medicine. The CHARTER study is supported by the National Institutes of Health (HHSN271201000030C). We acknowledge the work of the UNMC Flow Cytometry Research Facility in data acquisition for this publication. We wish to thank Roman Sasik and Aaron Chang for advice on methylome analysis, and Cherie Ng for constructive discussions and comments on the manuscript.