Age-associated changes in the circulating human antibody repertoire are upregulated in autoimmunity

Background The immune system undergoes a myriad of changes with age. While it is known that antibody-secreting plasma and long-lived memory B cells change with age, it remains unclear how the binding profile of the circulating antibody repertoire is impacted. Results To understand humoral immunity changes with respect to age, we characterized serum antibody binding to high density peptide microarrays in a diverse cohort of 1675 donors. We discovered thousands of peptides that bind antibodies in age-dependent fashion, many of which contain di-serine motifs. Peptide binding profiles were aggregated into an “immune age” by a machine learning regression model that was highly correlated with chronological age. Applying this regression model to previously-unobserved donors, we found that a donor’s predicted immune age is longitudinally consistent over years, suggesting it could be a robust long-term biomarker of humoral immune ageing. Finally, we assayed serum from donors with autoimmune disease and found a significant association between “accelerated immune ageing” and autoimmune disease activity. Conclusions The circulating antibody repertoire has increased binding to thousands of di-serine peptide containing peptides in older donors, which can be represented as an immune age. Increased immune age is associated with autoimmune disease, acute inflammatory disease severity, and may be a broadly relevant biomarker of immune function in health, disease, and therapeutic intervention.


exogenous spiking of endogenous age-associated interferants
To test if antibody-peptide binding was impacted by typical interferants, we used serum from four healthy donors with and without an interfering substance. We spiked interferants at a single high concentration: triglycerides, rheumatoid factor (RF), conjugated bilirubin, human anti-mouse antibody, hemoglobin, and unconjugated bilirubin (Methods). We then compared the log10 ratios of intensities measured with and without individual interferants to determine the impact the interferant had on peptide-array binding. Interestingly, we found that RF had an impact, though weak, on age-associated binding (p < 0.06, enrichment ratio of ~1.9). We examined a crystal structure of RF binding to its natural antigen, the antibody Fc region (1,2). We observed that the heavy chain contains multiple di-serine and 'SSV' motifs ; however, most of these motifs were in the antibody Fab, not Fc, region and those in Fc were (1) not bound by RF in published crystal structures (PDB: 1ADQ) and (2) previous reports indicate that RF-shared epitope across RA patients and healthy donors does not include the di-serine motifs (3)(4)(5)(6). Furthermore, murine immunization by human IgG resulted in similar epitopes as found in RA-patient rheumatoid factor, which also did not include binding to serine motifs. Additionally, the secondary antibody used, which binds to IgG Fc, was not associated with age-associated peptide features ( Figure 4S1B), which indicates that generic serum antibody reactivity to IgG Fc is not associated with age.

Cytokine concentrations do not improve prediction of chronological age
We next assessed the ability to predict age from antibody binding and cytokine concentration data. Cytokines levels were measured by Luminex assay and quantified using the manufacturer provided protocol and software (Life Technologies Corp., Carlsbad, CA). A custom cytokine panel was used to measure CD40L, EGF, eotaxin (CCL11), GM-CSF, IFN-alpha, IFN-gamma, IL-1 beta, IL-1RA, IL-2R, IL-6, IP-10 (CSCL10), MCP-1 (CCL2), TNF-alpha, TNF-RI and TNF-RII. CRO and ACDC levels were measured separately. Four of the cytokines tested, IP10, eotaxin, sIL2Ra, and sTNFR1 were positively correlated with age while sCD40L was negatively correlated with age ( Figure 3S4A). Two cytokines, sIL1Ra and CRP positively correlated with BMI ( Figure 3S4B).
Given the prominent association of chronic inflammation biomarkers with age and BMI, we sought to understand if these markers could improve the antibody binding-based regression model. Binding and cytokine data were normalized by mean-centering after taking the log transform (Methods). To avoid explicit feature filtering, we used machine learning regression methods that encouraged sparsity to limit model complexity (Methods). We considered a number of machine learning models; however, we ultimately performed chronological age regression with a model that was a weighted linear combination of normalized fluorescent intensities (Methods).
When we combine cytokines as independent variables with all antibody-binding peptide features, we find cytokines do not add predictive capacity ( Figure 3S4C). Chronological age predictions by serum cytokine levels and antibody profile generally agree with a Pearson correlation of r = 0.35 indicating a related but independent prediction ( Figure   3S4D). Furthermore, we find that if we first develop an antibody binding score and combine this with each cytokine (which dramatically reduces dimensionality) and use ridge regression or elastic net, there is no improvement of prediction on an independent test set (not shown).
As we lack a gold standard for humoral immunosenescence, we moved ahead with antibody binding rather than cytokine measurements. This is because the cytokine markers, even in aggregate, were significantly less predictive of chronological age.
Furthermore, adding cytokine markers on top of antibody binding did not add any predictive value.
Thus, while cytokines are important measurements of innate immunity, we focused attention on antibody binding measurements via peptide microarrays to better understand immunosenescence in the adaptive immune response.