Predicting Heart Failure: The Promise of Proteomics

This article refers to ‘Proteomic prediction of incident heart failure and its main subtypes’ by V. Emilsson et al., published in this issue on pages 87–102. The advent of high-throughput proteomics has revolutionized our ability to study the association of circulating protein levels with prevalent and incident disease on a large scale. Besides offering new insights into disease mechanisms, proteomics allows for the identification and validation of protein biomarkers for disease diagnosis, prognosis, and response to treatment, offering new opportunities for a more personalized approach to disease prevention and treatment. Nonetheless, the integration of proteomic advancements into clinical practice remains an ongoing endeavour. A main reason for this delay is that enhancement of predictive accuracy at the individual level for a specific outcome with the use of proteomics does not directly translate into actionable inference for the participant population. Machine learning methods, taking advantage of an ever-increasing computing capacity, have expedited the analysis of high-dimensional proteomic datasets, with remarkable efficiency. However, unless (i) a model that incorporates proteomic markers can be applied to a wider population or (ii) a proteomics-based discovery (‘hit’) can be tied to a testable pathophysiological hypothesis that can lead to new therapeutic targets, translating the advances in proteomics into actionable next steps for disease prevention and treatment is challenging. In this issue of the Journal, Emilsson and colleagues1 report their findings on the ability of serum proteins to predict future heart failure (HF) events, with reduced or preserved ejection fraction (HFrEF and HFpEF, respectively), using samples and data from 5457 participants of the longitudinal Age, Gene/Environment Susceptibility Reykjavik Study (AGES-RS). This population-based cohort included older participants (66 to 96 years old at the time of blood collection), enrolled between 2002 and 2006. After a median of 5.5 years, there were 440 cases with HF, and in the subset with echocardiographic data available, 167 cases of HFrEF and 188 of HFpEF. The left ventricular ejection fraction (LVEF) cut-off used for HFrEF was ≤40% and for HFpEF ≥50%. Using a rigorous statistical approach, the investigators identified 10 serum proteins that improved the predictive properties (in terms of concordance gain, using Harrell's C-index) of the Framingham risk score (FRS) and of a model incorporating FRS components and additional risk factors, including diabetes and calcium scores, for new HF events. The predictive gain with proteomic markers was driven by N-terminal pro-B-type natriuretic peptide (NT-proBNP), cardiac troponin I, and matrix metalloproteinase-12 (MMP12), all of which emerged consistently in the statistical iterations. Conceivably, fewer proteomic predictors added reliably to the prediction of incident HFrEF and HFpEF events, as the number of events was smaller. Protein predictors performed better close to the time of incident HF events. The performance of these proteomic predictors was externally evaluated in the Cardiovascular Health Study (CHS). The study by Emilsson and colleagues1 has several key strengths. The authors deserve to be applauded for comprehensive, detailed work on a complex dataset. In terms of both sample size and number of evaluated serum proteins, this study stands as one of the largest population-based cohort studies on incident HF to date. The population-based nature of the cohort increases the generalizability of the findings, with the caveat that only participants of European ancestry were included. Participants have undergone detailed phenotyping of several health domains at baseline, and new HF events have been individually adjudicated, indicating a high quality of ascertainment for covariates and outcomes of interest. The authors used a rigorous statistical approach (least absolute shrinkage and selection operator [LASSO] regression with bootstrap) to select predictors, which is important considering the large number of predictors. In addition, the authors tested the incremental value of proteins over a robust set of known HF risk factors and cardiovascular risk prediction models and presented detailed data on discrimination (C-index), which are not always presented in similar studies. The results were strengthened by the replication of the cumulative performance of proteomic predictors in an external cohort study (CHS). On the downside, the study focused on older adults (over age 65), which may limit generalizability of findings to younger populations and thus the extrapolation of the predictive ability of proteomic markers to a demographic group more amenable to prevention efforts. Several observations from the excellent study of Emilsson et al.1 merit further discussion. The most striking finding is that the lion's share of added predictive value among serum proteins was taken by NT-proBNP, which equaled that of all other proteins combined (for incident HF prediction), regardless of adjustment model or ejection fraction. In fact, when a comprehensive prediction model that included calcium scores was used as baseline, only NT-proBNP, cardiac troponin I, and MMP12 seemed to add substantial discriminatory ability for all incident HF. For incident HFrEF, nuclear factor erythroid 2-related factor 1 (NFE2L1) emerged as a potential additional predictor, as did tissue inhibitor of metallopeptidase 4 (TIMP4) for incident HFpEF. Proteins improved discrimination, with a clinically meaningful C-index of the full model (0.84 in the derivation cohort, 0.82 in CHS), but the resources required to implement such a model are not trivial. To some extent, these discoveries lead us back to the starting point. First, there is pattern of NT-proBNP dominating predictive value in proteomic studies. In previous work,2-5 the added predictive value from other proteins, when reported, was relatively small and inconsistent (Table 1). This represents a fundamental struggle between seeking impact at the population level and personalizing predictions. An additional challenge specific to HF is that HF is a syndrome, not a single disease process. Therefore, besides ‘generic’ markers of high intracardiac pressures or myocardial damage, i.e. natriuretic peptides or troponins, respectively, which are likely to be progressively elevated in stage B to D HF,6 identifying proteins that reflect activation of specific pathways is more challenging. Aberrations in individual protein levels may signal activation of pathways that warrant further investigation for certain subgroups at risk. However, adjusting for common risk factors and ‘common pathway’ proteins can eliminate these potential markers from a rigorously regularized (e.g. LASSO) model. This problem is compounded in large community-based cohorts, where the goal is to identify participants with stage A (or 0) HF at high risk for progression to clinical HF. In this scenario, there is also considerable temporal distance between the exposure (protein levels at the index visit) and the event of interest (new HF) and a multitude of potential mediating pathways, including e.g. metabolic derangements, ischaemic heart disease, and primary cardiomyopathic processes to name a few. The different proteomic methods and selection of proteins in the various studies compound the problem (Table 1). On the other hand, simply improving the properties of a prediction model for incident HF with the addition of proteomics would be welcome. However, this application comes with another set of challenges, related to reproducibility, transportability, and practical applicability of such a model. For example, studies on incident HF have included different protein panels (e.g. cardiovascular- and/or inflammation-oriented),2-5 and thus both proteomic ‘hits’ and model improvement were based on availability (Table 1). Nine proteins associated with incident HF after adjustment for risk factors. When NT-proBNP was added, MMP12, OPG, U-PAR and TIM-1 remained significant predictors for HF. 38 proteins were associated with incident HF, in four main network clusters: - Inflammation and apoptosis - Extracellular matrix remodelling, angiogenesis, and growth - Blood pressure regulation - Metabolism Time to event analysis (Cox proportional hazards) for incident events. Logistic regression with LASSO for C-statistic. Pathway enrichment analysis. C-statistic improved from 0.649 to 0.699 with protein predictors, with NT-proBNP adding the most; the remaining proteins improved C-index by 0.028. No pathway was enriched for HF. Second, how can we interpret the strong statistical association between e.g. NFE2L1 and HFrEF? Previous work has linked activation of nuclear factor erythroid 2-related factor 2, a transcription factor that regulates detoxifying and antioxidant gene expression in the liver, with protective cardiac and renal effects.7, 8 However, in Emilsson et al.,1 NFE2L1 was associated with increased incident HFrEF risk. Although there is no work directly linking NFE2L1 to myocardial effects to date, NFE2L1-knockout mice express reduced levels of lipolytic genes in adipocytes, leading to adipocyte hypertrophy, severe inflammation, and insulin resistance, suggesting that NFE2L1 has a protective role in adipocytes.9 This discordance highlights another core challenge with proteomics: ‘hits’ based on availability (e.g. NFE2L1 was simply part of the protein panel, rather than a priori hypothesis, can lead to mechanistic dead ends). For actionable next steps, additional work is needed. One approach to sort out causative associations is to supplement the analysis with Mendelian randomization using genetic instrumental variables.4, 10 Another approach would targeted proteomic profiling using established methods of protein quantification.11 Serial assessment of proteomics could also help disentangle causative associations.12 Proteomic predictors demonstrated stronger performance in the immediate period before a new HF event (approximately 6 months), but their contribution diminished somewhat over an extended period. One can view this time-dependent prediction ‘gradient’ in different ways. On one hand, the early predictive performance of proteomics may point to subclinical HF, a group of patients suitable for timely diagnostic and therapeutic interventions. On the other hand, it may suggest that we simply need better screening approaches for stage B HF at the population level, not necessarily with proteomics. In this line of thought, several therapies are recommended by contemporary guidelines for symptomatic (stage C) HFrEF, including beta-blockers, mineralocorticoid receptor antagonists (MRAs), sodium–glucose cotransporter 2 inhibitors (SGLT2i), and angiotensin receptor–neprilysin inhibitors (ARNIs). However, for patients with LVEF ≤40% without symptoms (stage B HF), angiotensin-converting enzyme inhibitors or angiotensin receptor blockers rather than ARNIs are still recommended and MRAs and SGLT2i are not indicated. The line between stages B and C HF is often blurred and it is conceivable that readily available, robust proteomic markers of high incident HF risk or discriminatory biomarkers of (yet unrecognized) HF may be utilized to identify patients with earlier stages of HF (stages A and B) that may ultimately benefit from upstream treatment with established HF therapies. While identification of novel proteomic pathways may provide mechanistic insights into the development of HF, risk assessment, and HF treatment options, implementation will be a major challenge, even if these assays become more widely available and easier to use in the clinical setting. Hypertension, the factor with the strongest attributable risk for HF, is easy to detect and many effective therapies are available at low cost. Yet only 25–50% of people with hypertension have their blood pressure controlled to levels associated with lower HF and mortality risk.13 Thus, it will be critical for clinicians to use novel tools such as ‘omics’ data and machine learning-based risk models to create diagnostic and treatment pathways that can be successfully implemented on a large scale. In summary, the important study by Emilsson and colleagues1 brings to the forefront a number of pressing questions when it comes to integration of proteomics to strategies for alleviating the burden of HF at the population level. Conflict of interest: none declared.