Kidney biomarker insights through factor analysis

With so complex an organ as the kidney, serving numerous functions and subject to diverse insults, dependence on two biomarkers in clinical nephrology, reflecting primarily two glomerulocentric aspects of kidney health/function (eGFR and albuminuria) has hindered the path to individualized care.

Numerous biomarkers have been studied in CKD and AKI, and tubular biomarker validation is particularly urgent, given the importance of tubulointerstitial health.¹ As yet no single biomarker has dethroned eGFR and albuminuria or joined them in a triumvirate: biomarker combinations may be a path forward.² Measuring a multiplicity of kidney biomarkers could have two benefits. First, individual markers may characterize different aspects of kidney health/function, aspects identified in experimental cellular/molecular studies, potentially enabling comprehensive non-invasive characterization. Second, combining multiple biomarkers into summary scores could reduce noise and improve accuracy. So, can a multimarker panel provide insight into distinct underlying kidney health/function processes and disease outcomes?

In this issue of AJKD, Bullen et al³ investigate a tubular biomarker panel in CKD using innovative analyses relating 10 biomarkers (8 urine, 2 serum) to kidney outcomes (eGFR change, AKI). They perform a nested observational analysis of participants with CKD from SPRINT⁴ (2,351 people with baseline eGFR 20–59 ml/min/1.73m²).

The 8 urinary biomarkers are α1m, β2m, IL-18, KIM-1, NGAL, YKL-40, MCP-1, and UMOD; the two serum markers are FGF23 and iPTH (posited to reflect tubular function by way of bone/mineral metabolism). The group has previously published related analyses: one investigated how the 8 urine biomarkers individually related to subsequent AKI (lower UMOD and higher α1m were associated with higher AKI risk).⁵ The other used factor analysis to derive distinct factors from the biomarkers—the factors applied in the present study—but investigated cardiovascular outcomes and mortality.⁶

Factor analysis is a classical multivariate technique that allows derivation of distinct summary variables (factors) from correlated variables.⁷ These factors, in appropriate contexts, can be interpreted to represent unobserved latent variables that give rise to the data, and Bullen et al treat factors as such. The primary goal is not dimension reduction (10 biomarkers reduced to 4 factors in this case), but summarizing variable interrelationships in an interpretable way that can be interrogated for validity—in this study by relating the factors to kidney outcomes.⁷ Given the evidence basis for underlying distinct kidney processes giving rise to the measurable biomarkers,¹ factor analysis seems a desirable technique. The factor derivations and score calculations are presented in a prior paper,⁶ but since the present paper relates the factor scores to kidney outcomes it is the one in which the construct validity of underlying latent kidney dimensions is tested. The subgroupings of biomarkers arrived at are: NGAL, IL-18 and YKL-40 (termed the “tubule injury/repair” factor); KIM-1 and MCP-1 (“tubule injury/fibrosis”); α1m and β2m (“tubule reabsorption”); and UMOD, FGF23, and iPTH (“tubule reserve/mineral metabolism”). This derivation and naming nicely combines data-driven (factor derivations) and pathophysiology-driven (interpreting latent variables) approaches, with interpretation and naming of the latent variables informed by experimental and observational evidence about individual biomarkers.¹ Each participant was assigned a score on each of the 4 factor dimensions, and the relationship of factor scores with kidney outcomes was assessed (Figure 1).

Figure 1. Steps involved in the factor derivation, interpretation, and evaluation.

Notes: The factor rotations used an oblique technique (Promax), so factors were not orthogonal, but are represented as such in panel 1 for easier visualization. Values of factor loadings (which we graph in panel 1) were obtained from supplementary material from a prior publication by the same group; biomarker names in gray have small factor loadings.⁶ The x and y axes in panel 1 are labeled with the factor names (for interpretability), although the names would not be applied until the step represented in panel 2.

Abbreviations: α1m, α−1 microglobulin, β2m; β−2 microglobulin; IL18, interleukin-18; KIM1, kidney injury molecule-1; NGAL, neutrophil gelatinase associated lipocalin; YKL-40, chitinase-3-like protein; MCP-1, monocyte chemoattractant protein-1; UMOD, uromodulin; min. met., mineral metabolism; inj./fib., injury/fibrosis; inj./rep., injury/repair; res./min.met., reserve/mineral metabolism.

The main outcomes were longitudinal eGFR change and AKI hospitalization or emergency room visit (over ~4 years). Longitudinal eGFR change is an appropriate metric, given its sensitivity and usefulness as a surrogate metric for patient-centered kidney outcomes.⁸ AKI in this setting, however, is more challenging, but the stringent criteria applied for AKI adjudication in SPRINT hopefully resulted in meaningful diagnoses. Models were adjusted for the usual covariates (most importantly baseline eGFR and albuminuria).

Mean eGFR change was 1.5% decrease per year, and a small number of AKI events occurred. In unadjusted and minimally adjusted models, each of the 4 factors was associated with eGFR decline. After adjusting for baseline eGFR and albuminuria, only “tubule reserve/mineral metabolism” (UMOD, iPTH, and FGF23) was associated with eGFR: eGFR decline grew by 0.58% per year for each standard deviation (SD) increase in the factor score (note that iPTH and FGF23 loaded positively on this factor, and UMOD negatively [Figure 1]). The other 3 factors added minimal information about eGFR change, with point estimates of eGFR decline from 0.06% to 0.17% per year (per factor score SD). “Tubule injury/repair” and “tubule injury/fibrosis” were associated with higher risk of AKI on adjusted analyses. Finally, the authors performed an exploratory analysis among 122 participants who had AKI and enough eGFR measurements to define pre- and post-AKI slopes. “Tubule injury/repair” and “tubule injury/fibrosis” were both associated with faster eGFR decline after AKI events, but not before. While certainly suffering from multiple hypothesis testing, this finding is interesting.

Biomarker studies must be interpreted in the context of the population studied and thus the underlying processes that may be driving the outcomes. The present subcohort from SPRINT had relatively benign CKD, without diabetes and with minimal albuminuria (in addition to older age⁹). Thus, the findings that some factors were not related to the outcomes should be interpreted not as strikes against the underlying biomarkers—nor against the factor conception—but as suggesting that these biomarkers (and the underlying constructs) are not relevant for this group. Biomarkers cannot be useful in all circumstances, and the heterogeneous nature of kidney disease entails a heterogeneous toolbox of markers. Important context is provided by a recent study (Puthumana et al)¹⁰ evaluating a subset of the discussed biomarkers in humans and mice. In this study, MCP-1, UMOD, and YKL-40 were measured in a very different population: people 3 months post-hospitalization (1,538 participants, half selected for AKI, the other half for no AKI). This differs from the SPRINT sub-study selection vastly, of course, but differs little in biomarker measurement and goals: steady state assessment to provide insights into long-term kidney outcomes and the pathophysiology of kidney health/function. The post-hospitalization study’s 3 biomarkers each come from a separate factor in the SPRINT sub-study (no “tubule reabsorption” component). Notably, Puthumana et al found that YKL-40 and MCP-1 were substantially correlated (r = 0.62), much more so than observed in the SPRINT sub-study (r = 0.19),⁶ supporting the relevance of heterogeneity and suggesting that the “tubule injury/repair” and “tubule injury/fibrosis” dichotomy may not generalize. Each of the 3 biomarkers was associated with eGFR decline, and with higher risk of a composite kidney outcome over 4 years (regardless of hospital AKI). These 3 markers were additionally evaluated using 2 mouse models (repair, atrophy) providing additional information to gauge the validity of the 4 factor conception offered by Bullen et al. The mRNA expression of the mouse analogs to each of these 3 biomarkers correlated strongly with atrophy and fibrosis.⁴ The MCP-1 and YKL-1 mRNA analogs were higher in the setting of fibrosis, whereas the UMOD mouse mRNA analog was higher in the repair setting, and substantially lower in the atrophy/fibrosis setting. This, combined with the absence of correlation between UMOD levels and the other two markers in the human cohort, provides further support for distinct “tubule reserve” and “tubule injury” factors (but not for the subdivision of the “tubule injury” factor into separate “repair” and “fibrosis” factors, although this distinction may have been apparent if the additional biomarkers were measured).

So did the SPRINT sub-study factors add value, compared to studying the biomarkers individually? We cannot tell based on the results in this and prior papers whether factors added statistical value (would UMOD alone have been a more potent predictor than its factor?). Regardless, we believe that factor analysis added value in highlighting crucial qualitative aspects of interpreting biomarkers in the context of pathophysiological evidence, which can drive novel questions, research, and understanding. In innovatively applying an old statistical technique to a new problem, working towards bridging the translational divide between biomarker panels and pathophysiologic understanding, and moving towards non-invasive tubular health assessment, the authors, we believe, have provided an important service. While we suspect that the inherent linearities in factor analysis and heterogeneity of kidney disease mean that machine learning techniques will ultimately be superior in squeezing maximal predictive and pathophysiological value out of biomarker panels, we congratulate Bullen et al for their important contribution.