Detection of a novel, integrative aging process suggests complex physiological integration

Abstract

Many studies of aging examine biomarkers one at a time, but complex systems theory and network theory suggest that interpretations of individual markers may be context-dependent. Here, we attempted to detect underlying processes governing the levels of many biomarkers simultaneously by applying principal components analysis to 43 common clinical biomarkers measured longitudinally in 3694 humans from three longitudinal cohort studies on two continents (Women's Health and Aging I & II, InCHIANTI, and the Baltimore Longitudinal Study on Aging). The first axis was associated with anemia, inflammation, and low levels of calcium and albumin. The axis structure was precisely reproduced in all three populations and in all demographic sub-populations (by sex, race, etc.); we call the process represented by the axis "integrated albunemia." Integrated albunemia increases and accelerates with age in all populations, and predicts mortality and frailty--but not chronic disease--even after controlling for age. This suggests a role in the aging process, though causality is not yet clear. Integrated albunemia behaves more stably across populations than its component biomarkers, and thus appears to represent a higher-order physiological process emerging from the structure of underlying regulatory networks. If this is correct, detection of this process has substantial implications for physiological organization more generally.

Fig 1. Biomarker loading order and stability for PCA1 across datasets and subsets.

Loading importance is calculated as the loading divided by the sum of the absolute values of all loadings. These values are ordered from high (red, on bottom) to low (magenta, on top) for the first 20 loadings; remaining loadings are grouped together as “Other” in white. Accordingly, hemoglobin has the strongest loading, then hematocrit, then albumin, etc. The order and colors are derived from the full analysis combining the first visits of individuals in all three datasets (top-left panel, left column, “All”) and applied to all other columns in the figure. Stability of loadings is indicated by conservation of loading heights across bars. (For an example of unstable loadings, see Fig. 5.) For each panel, the loadings for the full dataset are at left. Numbers indicate subset sample sizes. For all panels except BLSA, the 43-variable set is used; for BLSA there was insufficient sample size to perform PCA on subsets with 43 variables, so the 34-variable analysis is presented.

Fig 2. Biomarker loading order and stability for PCA2 across data sets and subsets.

Loading importance is calculated as the loading divided by the sum of the absolute values of all loadings. These values are ordered from high (red, on bottom) to low (magenta, on top) for the first 20 loadings; remaining loadings are grouped together as “Other” in white. Accordingly, CRP has the strongest loading, then triglycerides, then IL-6, etc. The order and colors are derived from the full analysis combining the three data sets (left column, top-left panel “All”) and applied to all other columns in the figure. Stability of loadings is indicated by conservation of loading heights across bars. For each panel, the loadings for the full data set are at left. Numbers indicate subset sample sizes. For all panels except BLSA, the 43-variable set is used; for BLSA there was insufficient sample size to perform PCA on subsets with 43 variables, so the 34-variable analysis is presented. The poor performance of BLSA is because several of the key variables in PCA2 are missing in the 34-variable set.

Fig 3. Correlations among the first three PCA axes (“pca1”, “pca2”, and “pca3”).

Correlations were calculated from the full merged 43-variable data set or from each of the separate data sets (WHAS: w; InCHIANTI: i; BLSA: b). The loadings from each subset-based PCA are reapplied to the full data set to calculate the scores used in the correlations. Ellipses below the diagonal indicate correlations visually: blue when positive, red when negative, and darker and narrower when stronger. Correlation coefficients are above the diagonal.

Fig 4. Biomarker loading order and stability for PCA3 across data sets and subsets.

Loading importance is calculated as the loading divided by the sum of the absolute values of all loadings. These values are ordered from high (red, on bottom) to low (magenta, on top) for the first 20 loadings; remaining loadings are grouped together as “Other” in white. Accordingly, neutrophils have the strongest loading, then AST, then lymphocytes, etc. The order and colors are derived from the full analysis combining the three data sets (left column, top-left panel “All”) and applied to all other columns in the figure. Stability of loadings is indicated by conservation of loading heights across bars. For each panel, the loadings for the full data set are at left. Numbers indicate subset sample sizes. For all panels except BLSA, the 43-variable set is used; for BLSA there was insufficient sample size to perform PCA on subsets with 43 variables, so the 34-variable analysis is presented.

Fig 5. Biomarker loading order and stability for PCA25 (the 25^th axis, chosen randomly as an example of an unstable axis) across data sets and subsets.

Loading importance is calculated as the loading divided by the sum of the absolute values of all loadings. These values are ordered from high (red, on bottom) to low (magenta, on top) for the first 20 loadings; remaining loadings are grouped together as “Other” in white. Accordingly, magnesium has the strongest loading, then glucose, then eosinophils, etc. The order and colors are derived from the full analysis combining the three data sets (left column, top-left panel “All”) and applied to all other columns in the figure. Stability of loadings is indicated by conservation of loading heights across bars. For each panel, the loadings for the full data set are at left. Numbers indicate subset sample sizes. For all panels except BLSA, the 43-variable set is used; for BLSA there was insufficient sample size to perform PCA on subsets with 43 variables, so the 34-variable analysis is presented.

Fig 6. Correlations among versions of PCA1 calculated from the full merged 43-variable dataset or from data subsets.

The loadings from each subset-based PCA are reapplied to randomly selected visits of the full dataset to calculate the scores used in the correlations [27,28]. Ellipses below the diagonal indicate correlations visually: blue when positive, red when negative, and darker and narrower when stronger. Correlation coefficients are above the diagonal. For all panels except BLSA, the 43-variable set is used; for BLSA there was insufficient sample size to perform PCA on subsets with 43 variables, so the 34-variable analysis is presented.

Fig 7. Age-adjusted biomarker loading order and stability for PCA1 across data sets and subsets.

Loading importance is calculated as the loading divided by the sum of the absolute values of all loadings. These values are ordered from high (red, on bottom) to low (magenta, on top) for the first 20 loadings; remaining loadings are grouped together as “Other” in white. Accordingly, hemoglobin has the strongest loading, then hematocrit, then albumin, etc. The order and colors are derived from the full analysis combining the first visits of individuals in all three data sets (top-left panel, left column, “All”) and applied to all other columns in the figure. Stability of loadings is indicated by conservation of loading heights across bars. For each panel, the loadings for the full data set are at left. Numbers indicate subset sample sizes. For all panels except BLSA, the 43-variable set is used; for BLSA there was insufficient sample size to perform PCA on subsets with 43 variables, so the 34-variable analysis is presented.

Fig 8. Age trajectories of PCA1 in the three data sets, based on Bayesian mixed models.

(a) 34-variable data set; (b) 43-variable data set; (c) ages 20–50 only. In (a) and (b), BLSA and InCHIANTI are based on fixed quadratic models with a random (individual) intercept, while WHAS is based on a fixed linear model with a random intercept. Linear models with random intercept were used in (c). Credibility intervals are based on calculating, independently for each age, which of the 1000 iterations’ trajectories were in the 2.5^th and 97.5^th percentiles. Note that the better fit of the linear model for WHAS appears to be due to the more limited age range for this dataset.

Fig 9. Stability across datasets of correlations between age and PCA axes or biomarkers.

Biomarkers are sorted from highest to lowest |loading| in PCA1. Colored boxes indicate significant correlations (p<0.05) with darker shading indicating stronger correlation (blue = positive, red = negative). Note that many biomarkers are significantly correlated in opposite directions in different datasets. Boxes are absent for variables with insufficient sample size in the respective dataset. The weak correlation between age and PCA1 in the youngest subgroup is expected given the non-linear relationship with age (Fig. 8).

Fig 10. Cross-dataset comparison of pairwise correlations among the main biomarkers determining PCA1.

Anemia is based on the first axis of a PCA analysis of anemia-related markers. Ellipses indicate correlations visually: blue when positive, red when negative, and darker and narrower when stronger. Correlation coefficients appear in ellipses and non-significant correlations are marked with an “X”.

Fig 11. Point estimates with 95% CIs for regression models predicting health outcomes based on PCA1.

Blue indicates WHAS and red indicates InCHIANTI. Solid lines are for 43-variable models and dashed lines for 34-variable models. W or I for WHAS or InCHIANTI respectively; 34 or 43 for the number of variables; F65+ if the model only included women aged 65+ (all models for WHAS); Base if the model was a cross-sectional analysis of baseline values, and Pred for models predicting a change in the parameter at the subsequent visit. Poisson regressions not shown; see S1 Results.