Proteomic architecture of frailty across the spectrum of cardiovascular disease

Abstract While frailty is a prominent risk factor in an aging population, the underlying biology of frailty is incompletely described. Here, we integrate 979 circulating proteins across a wide range of physiologies with 12 measures of frailty in a prospective discovery cohort of 809 individuals with severe aortic stenosis (AS) undergoing transcatheter aortic valve implantation. Our aim was to characterize the proteomic architecture of frailty in a highly susceptible population and study its relation to clinical outcome and systems‐wide phenotypes to define potential novel, clinically relevant frailty biology. Proteomic signatures (specifically of physical function) were related to post‐intervention outcome in AS, specifying pathways of innate immunity, cell growth/senescence, fibrosis/metabolism, and a host of proteins not widely described in human aging. In published cohorts, the “frailty proteome” displayed heterogeneous trajectories across age (20–100 years, age only explaining a small fraction of variance) and were associated with cardiac and non‐cardiac phenotypes and outcomes across two broad validation cohorts (N > 35,000) over ≈2–3 decades. These findings suggest the importance of precision biomarkers of underlying multi‐organ health status in age‐related morbidity and frailty.


| INTRODUC TI ON
With improvements in cardiac intervention and prevention during the past three decades, individuals who would have previously succumbed to acute, non-communicable diseases (cardiovascular disease [CVD], oncologic) now survive to an older age with multiple advanced chronic conditions (Ijaz et al., 2022).This changing clinical landscape challenges the routine application of high-risk therapy in higher risk individuals specifically in age-related conditions, like CVD (Jha et al., 2017;Leon et al., 2010;Smith et al., 2011;Reardon et al., 2017;Waksman et al., 2018), where an interplay between cardiac and non-cardiac physiology impact outcomes.In this context, understanding how frailty-an impaired ability to maintain homeostasis during physiologic stress (Clegg et al., 2013)-modifies treatment response is critical.Despite associations of several frailty measures with clinical outcomes (Clegg et al., 2013;Guralnik et al., 1994Guralnik et al., , 1995;;Ijaz et al., 2022), there remains significant heterogeneity in how frailty is assessed among older adults, including those with CVD (Rohrmann, 2020), with concerns around how best to reproducibly define and quantify frailty across centers and conditions as major limitations to widespread adoption (Rockwood & Howlett, 2018).While efforts to define molecular correlates of chronological aging abound (Ahadi et al., 2020;Basisty et al., 2020;Emilsson et al., 2018;Lehallier et al., 2019Lehallier et al., , 2020;;Sebastiani et al., 2021;Tanaka et al., 2018), their application in tissues accessible clinically (e.g., blood) has largely been limited to an epidemiologic context (Landino et al., 2021;Liu et al., 2022;Sathyan et al., 2020;Tanaka et al., 2020), without a clear ability to define the impact of circulating biochemistry on downstream, posttherapy outcome (Ferrucci & Fabbri, 2018;Ramonfaur et al., 2022).
Given the potential for early identification of "accelerated" aging and molecular intervention (Sinha et al., 2014), identifying pathways of human frailty related to poorer tolerance of intervention may prioritize adjunctive avenues of therapy and investigation to enhance resilience in this growing population.
Here, we hypothesized that biological pathways of frailtyrevealed through integrating 12 measures of frailty with comprehensive proteomic profiling-would identify older individuals at high risk of mortality despite intervention.We studied 809 individuals with symptomatic, severe aortic stenosis (AS) undergoing transcatheter valve implantation (TAVI)-an age-related cardiovascular condition in which frailty has had prognostic implication (Kiani et al., 2020).We quantified 979 circulating proteins alongside 12 measures of frailty encompassing body composition, cognition, nutrition, patient-centered assessment of well-being, functional measures, and biochemistry.
We developed, validated, and characterized proteomic signatures of three composite axes of frailty against post-TAVI mortality, and explored the generalizability of our findings and their age dependence across multiple studies (35,559 community-dwelling adults from Iceland (Ferkingstad et al., 2021); human studies across the life-course (Lehallier et al., 2019); and 1894 community-dwelling individuals in the Framingham Heart Study [FHS]).Ultimately, we sought to define a proteomic architecture of frailty in structural heart disease and characterize its broad relevance to multi-organ phenotypes, function, and outcome to inform future studies of risk and therapy.

| Study populations
To derive proteomic correlates of frailty in advanced heart disease, we studied 809 individuals with severe AS from a multicenter prospective cohort study (Perry et al., 2022;Stein et al., 2022) where frailty measures were systematically collected, split into two samples: (1) a derivation sample (N = 233) that had complete data on 12 measures of frailty and (2) a validation sample (N = 576) comprised of the remainder of our multicenter AS cohort that did not have complete data on the 12 frailty measures (Table 1).Both samples had follow-up for vital status.
Overall, the AS cohort had a median age 83 years (range 46-100 years, TA B L E 1 Baseline characteristics of the aortic stenosis cohort.44% women), with a high prevalence of coronary artery disease (70%) and diabetes (nearly 40%).The derivation and validation samples were largely comparable, with some imbalance in diabetes prevalence (43% in validation vs. 28% in derivation), body mass index, and in some measures of self-reported health status (e.g., KCCQ-12 and EQ-VAS).
The FHS cohort had lower prevalent cardiometabolic morbidity in FHS relative to our AS cohort, consistent with its younger mean age and being a community-based population.As we conducted original analyses in FHS, the study population is reported in Table S2.For detailed cohort characteristics of the other replication cohorts, the reader is directed to the parent publications (Eiriksdottir et al., 2021;Ferkingstad et al., 2021;Lehallier et al., 2019).

| Multidimensional frailty measures are classified into three broad phenotypic groups
Distribution of frailty measures in our derivation sample is in Table 1 with correlations in Figure S1.Given the physiologic and statistical relatedness across frailty measures, we used principal component analysis (PCA) to identify composite axes of frailty (Figure 2).The top three principal components (PCs) explained ≈49% of variance in the frailty phenome studied (loadings for each of the three PCs in Figure 2a).The first PC ("axis") was weighted predominantly on patient-reported metrics of well-being, including PHQ-2, EQ-VAS, KCCQ-12, and MNA-SF (hereafter called "patient-reported outcomes").In addition, consistent with known phenotypic dimorphism by sex, we observed higher body composition and physical functional scores for men relative to women (Figure S3).

| Proteomic correlates of frailty identify older adults at high risk for mortality after cardiac intervention
To identify proteomic signatures of frailty, we next used linear regression methods (both ordinary and LASSO) across the proteome as independent variables with each individual frailty measure or each composite frailty axis (from the PCA above) as the dependent variable in separate models (results in the Data File S1).Hemoglobin and albumin were related to the greatest number of proteins, followed by gait speed, nutrition, and KCCQ-12.LASSO regressions for each of the three frailty axes selected 191 unique proteins, with fewer proteins selected in models for patient-reported outcomes than for body composition or physical function.LASSO-based protein signatures of each frailty axis (protein "score" for that phenotypic axis, see Section 4) had variable model fits, with model fits generally poorest for patient-reported outcomes (fit for hold-out folds during LASSO optimization shown in Figure S4a; fit across entire derivation sample shown in Figure S4b).To validate these protein scores of frailty, we imputed missing frailty data in the validation sample (using multivariate imputation by chained equations, see Section 4) to correlate frailty axes with the protein scores.This demonstrated similar relations as the derivation sample: a poor relation in models for patient-reported outcomes (Spearman ρ = 0.17), moderate correlations for body composition (Spearman ρ = 0.40), and physical function (Spearman ρ = 0.41).Given the need for complete data in PCA, the use of imputation for data missingness was restricted only to test replication of association of protein scores to the composite axes of frailty.We did not observe effect modification by sex on the relationship between individual proteins and frailty axes after FDR adjustment (Benjamini-Hochberg) for multiple testing of interaction terms.Each protein score was related to the frailty measures most heavily loaded in the parent frailty axis from which it was derived (Figure S5).Accordingly, each protein score exhibited a similar age and sex relation as the parent frailty axes (maximum Spearman |ρ| = 0.29 for age across all protein scores).
We next assessed the relation of each protein score and frailty axis from which it was derived with all-cause mortality.Across a median 3.2 years of follow-up (in derivation sample; 25th-75th percentile 1.3-3.6 years), each of the three frailty axes had point estimates for post-TAVI mortality in a protective range, with only physical function significantly related to mortality after clinical risk adjustment (Figure 3).Protein scores of frailty exhibited similar estimates for mortality in both derivation and validation samples, generally robust to multivariable adjustment at a median 2.9 years follow-up (25th-75th percentile 1.2-3.9years).Of note, in sensitivity analyses, associations with mortality were robust to adjustment for simpler biomarkers canonically associated with cardiovascular mortality (NT-proBNP, hemoglobin, albumin (Ibrahim & Januzzi Jr., 2018; Table S3).

| The proteome implicates both known and novel pathways of human frailty
We used proteins associated with the 12 frailty measures in linear models for pathway analysis, respectively (at a 5% FDR).The proteins identified implicated broad pathways of innate and adaptive immunity (e.g., cytokine signaling and TNF), canonical cell growth and signaling pathways (e.g., PI3K-Akt signaling), and organ fibrosis and metabolism (e.g., extracellular matrix remodeling and turnover, glycosylation; Figure S6).In addition to proteins with known relation to body composition (e.g., leptin and insulin-like growth factor binding proteins), several novel proteins with roles in adipose tissue metabolism and inflammation were identified (in association with frailty axes), including PLIN1 (higher expression related to increased adipocyte size, altered lipid handling, and improved whole-body glucose tolerance Kern et al., 2004) (Crunkhorn, 2020;Siddiqui et al., 2022), MSTN (Schafer et al., 2016), IL6 (Strassmann et al., 1992), and FABP4 (Kim et al., 2013;Lee et al., 2017)

| Chronological age does not fully account for broad variability in the frailty proteome
Given relevance of implicated pathways across the life-course (e.g., immunity, cell growth, and metabolism), we next sought to quantify the extent to which proteins related to frailty axes were explained by age.In the AS cohort, age only accounted for a small fraction of the total variability in protein scores (Figure 4a), with sex, and BMI accounting for more of the variability, and protein scores were weakly related to age (Figure 4b).S7).These life-course patterns may largely have been established by the time advanced heart disease (AS) requiring intervention had developed (purple line demonstrating age range of AS cohort, Figure 4c), accounting for the low variation explained by age in our AS sample.

| The frailty proteome and clinical risk
Given the physiologic relevance of frailty-implicated pathways across multiple organs in advanced CVD, we next studied relations of the frailty proteome to health status and disease-free longevity.
In proteins associated with any frailty axis in our AS studies that were measured in a large Icelandic cohort (70 proteins ;Ferkingstad et al., 2021), we found (1) a limited effect of age on protein concentration and (2) associations between proteins and multi-organ morbidity generally in a directionally plausible manner (Figure 5a).Of note, queried proteins were strongly related to metabolic-inflammatory phenotypes not directly cardiac (glycemic control, body composition, inflammatory markers, malignancy).
We next examined the relation of protein scores with causespecific mortality in FHS (Table 2).At a median 26 years after proteomics (25th-75th percentile 19-27 years, 755 deaths, 211 CVD-related), a higher physical function protein score was associated with lower all-cause mortality in FHS (Figure 5b; Table 2).
Given the strong association between all-cause mortality and the proteomics of physical function, we next sought to examine whether that mortality association would be driven by non-cardiovascular (versus cardiovascular) causes.We carried that score forward into competing risk models for CVD versus non-CVD mortality in FHS, where we found that the proteomics of physical function were associated with non-CVD mortality in FHS (Figure 5b; Table 2).

| DISCUSS ION
Here, we quantify 979 circulating proteins in 809 older individuals with severe AS to identify a proteomic "fingerprint" of frailty defined across 12 measures (Afilalo et al., 2017) spanning physical function, cognition, nutrition, biochemistry, self-reported wellbeing, and body composition.We determined proteomic correlates of frailty measures, specifying canonical pathways of organ function (e.g., inflammation, cell growth and senescence, cachexia) as well a host of mediators of tissue-specific biology not previously widely reported in human frailty (e.g., myogenesis, adipose tissue inflammation, and lysosomal metabolism).Protein scores of three major frailty axes defined by integrating 12 frailty measures and proteins were strongly related to mortality after cardiac intervention, independent of clinical risk.Despite reported statistically significant age association in epidemiologic cohorts with a broader range of age (Ferkingstad et al., 2021), BMI and sex accounted for significantly greater variability in protein scores of frailty axes than age in older patients with AS (Figure 4a).Across eight decades of life (≈20-100 years), we observed heterogeneous patterns of abundance of frailty-related proteins across age (Figure 4c), with patterns wellestablished by advanced age.Across a large number of individuals, frailty-related proteins were associated with a broad array of noncardiac comorbidities and outcomes, including directionally consistent associations with mortality in long-term follow-up in thousands of community-dwelling individuals (Figure 5).In FHS, we observed a significant association between the protein score corresponding to physical function with all-cause and non-CVD mortality over two decades.Collectively, these findings extend the growing aging literature toward the cardiovascular space and emphasize the potential for proteomic studies in the context of advanced CVD to identify functional, prognostic pathways of risk for interrogation in advanced heart disease.
Separating "biological" from "chronological" aging using molecular information has been the subject of a large body of work in aging research (Ahadi et al., 2020;Basisty et al., 2020;Emilsson et al., 2018;Lehallier et al., 2019Lehallier et al., , 2020;;Sebastiani et al., 2021;Tanaka et al., 2018).Approaches that generate molecular "clocks" using epigenetic (Horvath, 2013), transcriptional (Peters et al., 2015;Shavlakadze et al., 2019), genomic (Singh et al., 2019), proteomic (Tanaka et al., 2018), and metabolomic (Cheng et al., 2015) information have been advanced to identify relevant pathways of and individuals with "accelerated" aging ultimately connected to longevity, including some reports of causespecific mortality (Eiriksdottir et al., 2021).While these studies have illuminated mechanisms and biomarkers of aging, most do not study individuals with CVD at older ages, where varying degrees of multi-organ frailty (beyond chronologic aging itself) may play a critical role (Collard et al., 2012).Given the prognostic relevance and reversibility of frailty (Chang et al., 2004;Guralnik et al., 1994Guralnik et al., , 1995;;Pandey et al., 2023;Perera et al., 2006;Puthoff, 2008;Volpato et al., 2008) Our study directly addresses these limitations by employing molecular discovery in a common clinical situation where frailty is routinely considered and prognostic (AS) (Kiani et al., 2020).Unlike prior cohort-based reports (Liu et al., 2022;Sathyan et al., 2020;Walston et al., 2002), our cohort had a dramatically higher rate of CVD and diabetes (≈70% and ≈40% overall, respectively measures extending prognostic multi-organ structure-function (the "Essential Frailty Toolset"; Afilalo et al., 2017) to guide discovery is a fundamental strength to move beyond clinical gestalt in frailty assessment (Ijaz et al., 2022).Furthermore, the inability of age to capture a large variation in the frailty proteome at the time of TAVI (relative to comorbidity) in our older population highlights importance of discovery in a clinical CVD context.Despite statistical age associations in up to 80% of the quantified proteome in a large age range in epidemiology, the reported effect sizes are small (Ferkingstad et al., 2021).We observed similarly weak relations with age when examining proteins related to frailty in our analysis (Figure 5a), suggesting that mechanisms beyond chronological age are likely involved in the biology of frailty.In an older population undergoing clinical cardiac intervention, it is possible that an age-related alteration in the proteome may already be prevalent/established, with inter-individual differences determined by comorbidity (Figure 4).This notion broadly underscores the potential importance of patient-level heterogeneity and human molecular studies to prioritize targets for therapeutic or mechanistic discovery in frailty.Indeed, "anti-aging" pharmacology directed at metabolism may impact the proteome decades earlier to "prepare" organs for intervention (metformin and GDF-15; Coll et al., 2020); SGLT2 inhibition and PLIN1 (Yang et al., 2021), RNA therapeutics (Fitzgerald et al., 2017;Solomon et al., 2019).
Biologically, our results implicated broad pathways relevant to both cardiac and non-cardiac physiology in aging around a theme of host inflammatory response, cell growth and senescence, and cachexia.Several proteins related to body composition and muscle function specified known pathways (leptin signaling, IGFBPs, GDF-15, IL6, MSTN), concordant with prior human observations and canonical mechanisms of human frailty.For example, our results are broadly consistent with a reported fall in myostatin (MSTN) with age, a relation to greater lean mass and grip strength (in men) (Bergen III et al., 2015), and a decreased muscle oxidative capacity and force generation in MSTN-null mice (Amthor et al., 2007).Moreover, our integrative approach facilitated discovery of an array of molecules with novel, emerging roles across a broad tissue biology relevant to aging, including adipose tissue metabolism and inflammation (e.g., PLIN1, Kern et al., 2004), activin signaling (Carlsson et al., 2009;Goebel et al., 2022), and PTX3 (Kocyigit et al., 2014) ease who display greater frailty (Pandey et al., 2023).The novelty of this approach is the application of broad molecular characterization to frailty at the point of its clinical utility for CVD, addressing heterogeneity in how frailty is assessed in clinical practice (Cooper et al., 2021).Certainly, direct clinical application of proteomics as an actionable biomarker requires demonstration of its reversibility with intervention and advancing from a broad "omic" space with relative quantification (as done in nearly all molecular studies of aging and frailty) to a more precise, select panel with absolute quantification.
Importantly, for some therapies where earlier application in patients with more advanced multimorbidity is currently standard (e.g., TAVI), proteomics cannot be viewed as a "gatekeeper" to intervention, but as a barometer for rapid stratification of individuals in need of more aggressive pharmacologic or rehabilitative therapy (in addition to TAVI) to limit poor outcome (Kitzman et al., 2021).Indeed, our results suggest high residual risk post-TAVI captured by the frailty proteome-a unique opportunity to intervene more aggressively after acute CVD has been addressed.By analogy, application of these results to interventions with high morbidity and resource utilization (e.g., ventricular assist and transplant) may offer additional pre-intervention opportunities to target individuals at high risk for adjunctive intervention.While not clinically available, some proteins identified in our analysis have published pharmacologic modifiers (He et al., 2021;Vandeghinste et al., 2018).While our results suggest both cardiac and non-cardiac implications of the frailty proteome in large populations (FHS), utilization of these or similar signatures to parse cardiac from non-cardiac morbidity after "correction" of cardiac output deficits (e.g., with ventricular assist) is a striking potential for future work.Finally, as our proteomic platforms broaden, these clinical opportunities may be met by potential molecular targets for intervention to improve frailty, similar to what has been attempted in other spaces in heart disease (e.g., RNA-based therapies Fitzgerald et al., 2017;Solomon et al., 2019).
Several limitations of our study merit comment.Our AS sample included several non-continuous, non-normal exposures and some differences in covariates between discovery and validation subsamples (Table 1), potentially biasing discovery.In particular, the validation sample is biased to include individuals with greater BMI and diabetes due to CT based measures of adiposity being unavailable in participants at the extremes of waist circumference.
In addition, while PCA-based frailty axes were internally consistent with our clinical experience and were related to outcome, a larger sample size will likely be needed to examine potential sex differences in how the proteome relates to the frailty phenome.
We recognize that matching frailty-related proteins from our AS cohort to other studies with aptamer-based proteomics offers unique challenges, including differences in specificity profile (Katz et al., 2022).While increased variance due to SomaScan-Olink platform differences may lead to null association (due to poor concordance for some proteins; Katz et al., 2022), we observed a consistent association with outcomes and phenotypes as well as a reasonable degree of correlation (median Spearman r = 0.58) between the two proteomic platforms on previously published data (Katz et al., 2022), where available, with a caveat that some proteins were negatively correlated.While the cross-sectional relation of protein scores with frailty measures in FHS may be limited by survival bias (proteins measured ≈7 years prior to frailty measures), the longitudinal Cox regression results replicated, suggesting these protein scores capture meaningful clinical outcomes.
Lack of racial diversity across samples included here is a significant limitation and reflects the limited diversity in transcatheter registries (Alkhouli et al., 2019).The use of molecular features in blood testing may facilitate broader implementation of objective measures of frailty to address racial disparities (Usher et al., 2021).
Ultimately, broader studies across race, frailty, and disease states with an eventual aim of absolute quantification of risk will be essential to personalize frailty assessment rapidly with prognostic and potential mechanistic implications. In

| ME THODS
An overview of study design and statistical methods is shown in Figure 1 and Figure S1.

| Study population
The discovery cohort comes from a multicenter, prospective cohort study of participants with symptomatic, severe AS undergoing TAVI (Perry et al., 2022;Stein et al., 2022).A key strength of this cohort is the systematic, prospective assessment of frailty (defined below).Severe AS was defined according to American Society of Echocardiography guidelines (peak velocity ≥4 m/s, mean gradient ≥40 mm Hg, or indexed aortic valve area <0.6 cm 2 /m 2 ; Baumgartner et al., 2017) et al., 2016).Cardiovascular disease in FHS was defined as prior myocardial infarction, coronary death, angina pectoris, coronary insufficiency, heart failure, stroke or transient ischemic attack, or intermittent claudication (D'Agostino Sr. et al., 2008).The Institutional Review Board at each institution approved each study.

| Discovery cohort (AS/TAVI)
Venous blood was collected in AS samples prior to TAVI, processed within 30 minutes, and stored at −80°C.Plasma proteins were quantified using the Olink Explore 1536 panel (Olink, Uppsala, Sweden) in three batches (Assarsson et al., 2014).Proteins related to frailty axes in the discovery cohort were matched to proteins in the replication cohorts using UniProt identifier.
We excluded 258 proteins from the Oncology panel across all batches due to a technical issue in the first batch that limited accuracy of these proteins.We used median normalization approaches to perform batch correction (with batch 3 as the referent, given most samples were run in this batch).We excluded 154 proteins if >25% of reported values were below the reported level of detection and excluded 145 proteins with a coefficient of variation greater than 40%, yielding 979 proteins available for analysis.Protein levels (in normalized protein expression units, log 2 scale) were mean-centered and standardized to unit variance for modeling.

| Replication cohorts
Aptamer-based proteomics (SomaScan) was used in all replication studies (Ferkingstad et al., 2021;Lehallier et al., 2019;Nayor et al., 2020).Published data were used for Icelandic participants (Ferkingstad et al., 2021) and the U.S. and European cohort studies of aging and age-related disease (VASeattle, PRIN06, PRIN09, and GEHA; Lehallier et al., 2019).For FHS, proteomics was performed in two batches as described (Nayor et al., 2020).FHS investigators accounted for batch effects as previously described (Nayor et al., 2020), via log-transforming and standardizing proteins in each batch separately, pooling batches, and subsequently rank normalizing the entire FHS sample.Plate-adjusted standardized residuals were subsequently used for regression to address batch effects comprehensively.

| Frailty assessment
The discovery (AS) cohort prospectively assessed measures of frailty in all participants.Since there is not one universally accepted definition of frailty, for this analysis we selected elements of the Fried frailty phenotype combined with variables included in the Afialo toolset that was developed specifically for the TAVI population then conducted a PCA to define axes (or dimensions) of frailty.We did not include categorical measures of frailty (such as exhaustion or unintentional weight loss from the Fried frailty phenotype) due to their heavy weightings in the discovery cohort (e.g., almost all participants reporting no unintentional weight loss) and incompatibility with PCA.We included 12 separate measures of frailty including questionnaire-based assessments, functional assessments, and biochemical and radiographic measures (Table S1).Three global assessments of frailty and quality of life were assessed via questionnaire, and included Katz Index of Independence in Activities of Daily Living (ADL) score (Katz et al., 1970), EuroQol Visual Analogue Scale (EQ-VAS; Nancy Devlin & Janssen, 2020), and the Kansas City Cardiomyopathy Questionnaire summary score (KCCQ-12;Green et al., 2000).Physical frailty was assessed by average handgrip strength (by dynamometer), average gait speed (5-meter walk time), visceral fat area indexed to height 2 , and psoas muscle area indexed to height 2 .For participants who were unable to perform the gait speed test, a value of 0 was imputed.Pre-TAVI computed tomography (CT) scans were used to measure psoas muscle area index and visceral fat area index using OsiriX software (Rosset et al., 2004).Bilateral psoas muscle area and visceral fat area were measured manually on a single 3 mm slice at the L4 level in the transverse plane.individually as a summary measure of its composite axis in downstream models.

| Identifying proteomic correlates of frailty
To identify proteins related to frailty within the derivation sample, we used linear regression with individual proteins as independent variables and each of the 12 measures of frailty as dependent variables, with adjustments for age and sex.A false discovery rate (FDR; Benjamini-Hochberg method) was used to control type 1 error.
This regression approach was repeated with frailty axes scores (from PCA) as dependent variables.Proteins associated with the 12 measures of frailty (FDR <0.05) were selected and mapped to both KEGG pathways and the Reactome database (clusterProfiler in R; Wu et al., 2021) respectively.Given our proteomic coverage did not cover all circulating proteins (N = 979), we used statistical tests only to select pathways for visualization (p values were generated by hypergeometric tests and adjusted by Benjamini-Hochberg method).
Proteins associated with any of the three frailty axes (generated by PCA) with an FDR <0.10 were examined in external datasets.
To construct parsimonious models for frailty, we next used LASSO regression (caret in R Kuhn, 2008) with frailty axes scores as dependent variables and all proteins (standardized to mean = 0, variance = 1) as penalized independent variables.LASSO models were developed in the subset of participants in the discovery (AS) cohort included in the PCA (derivation sample as mentioned above).
Cross-validation (10 folds, with 5 repeats) was used to optimize model hyperparameters (e.g., lambda).Resulting models were then used to create protein scores for each of the frailty PCs, by taking the sum of the product of each regression coefficient and protein level for each individual.These protein scores represent a bloodbased proteomic "fingerprint" of frailty for downstream analyses.Rosow-Breslau questions (ability to do heavy work, ability to walk a half mile), Katz ADLs, grip strength (Jamar dynamometer), gait speed (4-meter walk at usual pace), and time to complete five chair stands (Liu et al., 2016).In addition, we included relations with visceral and subcutaneous adiposity as a measure of body composition (Fox et al., 2010).We used Cox regression to relate each of the 3 protein scores with all-cause mortality with adjustments for sex and age in minimally adjusted models, with further adjustments for BMI, smoking status, diabetes, anti-hypertensive medication treatment, total and HDL cholesterol, systolic blood pressure, and prevalent CVD.We then used a competing risk model (Fine-Gray) to evaluate for CVD versus non-CVD mortality for protein scores that were associated with all-cause mortality in standard Cox models (D'Agostino et al., 2008;Fine & Gray, 1999).R (versions 4.2.1 and 4.2.2) was used for analyses.A two-tailed next studied the relation of our composite proteomic frailty axes scores in the 1894 FHS participants to frailty measures, and cause-specific mortality.Recalibration efforts (described in Statistical methods) were excellent (Spearman ρ range 0.89-0.92),with resulting scores in FHS demonstrating similar sex-based differences and limited association with age (Pearson |r| 0.06-0.11; Figure S8).Proteins used to recalibrate the scores from our discovery cohort (Olink) to the FHS (SomaScan) demonstrated a moderate correlation (median Spearman ρ = 0.58 [25%-75%: 0.21-0.71]),where available, in published data(Katz et al., 2022).Protein scores for body composition and physical function (at FHS Exam 5) exhibited generally concordant relation to visceral and subcutaneous fat or measures of physical function/frailty, respectively, at a median 6.9 years later (for frailty measures), though with mitigation of effect size after age-and sex-adjustment for several measures, consistent with the broader age range in FHS (Table , islet cell function (ADGRG1Duner et al., 2016), muscle cell physiology (ITGA11Grassot et al., 2014), EFNA1 (Alonso-Martin et al., 2016), LRRN1   (McKellar et al., 2021), lysosomal metabolism (CTSLWu et al., 2011), extracellular matrix handling and fibrosis (SDC1Yang & Friedl, 2016), among others that specify frailty mechanisms not necessarily specific to the heart.These broad mechanistic implications are consistent with our phenotype and outcome associations across thousands of individuals in Iceland and FHS for a broad array of metabolicinflammatory conditions (Figure5a) and non-cardiovascular death (Figure5b) that are neither fully nor directly reversible with cardiaconly intervention.These results are consistent with all-cause mortality in a subset of the Icelandic population across a broader age range (≈22,000 individuals, ≈20-100 years old), where several proteins related to lower physical function in our study (both canonical, e.g., GDF-15, and more novel, e.g., MZB1, ASGR1) were associated with increased mortality(Eiriksdottir et al., 2021).From a clinical perspective, these results are compelling given recent reports suggesting potentially greater benefit to physical rehabilitation interventions in individuals with advanced heart dis-

E 1
Graphical abstract and study diagram.and cause-specific mortality ► Relation of proteome -multi-organ morbidity and all-cause mortality How can we use molecular characterization to identify pathways of frailty in advanced heart disease?How is the frailty proteome related to cardiac vs. noncardiac morbidity and outcomes?
Cognitive (Mini-Cog total score;Borson et al., 2003), psychosocial (Patient Health Questionnaire-2 [PHQ-2]Kroenke et al., 2003), and nutritional measures (Mini Nutritional Assessment-Short Form [MNA-SF]) were included as additional metrics of frailty(Rubenstein et al., 2001).Finally, we included hemoglobin and albumin given their inclusion in the Essential Frailty Toolset and association with post-TAVI outcomes(Afilalo et al., 2017).F I G U R E 2 PCA of12 frailty measures identifies 3 composite axes of frailty.(a) Loadings of the three composite axes of frailty using PCA with varimax rotation.(b) Heatmap of study participants in the derivation sample (columns) demonstrates heterogeneity in PC scores and individual measures of frailty.Statistical analysis 4.4.1 | Summarizing 12 frailty measures in composite phenotype measuresWe observed correlations among measures of frailty (FigureS1), prompting an approach to generate composite axes of frailty using PCA.We conducted PCA (with varimax rotation; using psych in R;Revelle, 2022) on participants with complete data on all 12 measures of frailty ("derivation" sample, N = 233).Frailty measures were mean-centered and standardized (mean = 0, variance = 1) for PCA.Principal components were selected by examination of a scree plot (generating 3 PCs summarizing 12 component frailty) and labeled based on loadings for that PC.Each PC score was used F I G U R E 3 Protein-based scores are independently related to all-cause mortality.Forest plots of Cox regression for all-cause mortality using phenotype and protein-based scores in derivation (a) and validation (b) samples.
We examined the association of protein scores with frailty measures and long-term CVD and non-CVD outcomes in the FHS Offspring cohort.We used recursive feature elimination on 367 proteins common to FHS and the AS cohort (determined by matching on UniProt identifier) in linear models (in caretKuhn, 2008, with a 5% tolerance) to recalibrate scores developed using Olink data in our discovery sample (AS) for FHS (given the differences in proteomic coverage).Recalibration fit using this approach was good (Spearman ρ range = 0.89-0.92across the 3 scores).The refitted models were then applied to FHS by summing the product of each regression coefficient (from recursive feature elimination) and protein level for each individual.We used logistic andF I G U R E 5The frailty proteome, systemic multimorbidity, and cause-specific mortality.(a) Heatmap of proteins associated with any frailty axis (FDR <0.10) also measured in >35,000 Icelanders (Icelandic Cancer Project and deCODE).Fill values are from age and sex linear adjusted models for each phenotype/outcome.The annotation bar presents the protein's relation with age in Icelanders and the protein's relation with frailty axes.Phenotype names are as provided by the parent study investigators(Ferkingstad et al., 2021).(b) Cumulative incidence curves for all-cause mortality, cardiovascular mortality, and non-cardiovascular mortality stratified by tertiles of protein score of physical function in FHS.These are for visualization of the survival association (unadjusted).The adjusted hazard ratio for a continuous marker (from Cox regression) is reported.
relate protein scores (from FHS Exam 5) to frailty measures collected at FHS Exam 7 (≈7 years later), including

a Model B b HR (95% CI) p value HR (95% CI) p value
Adjusted for sex, age, BMI, smoking status, diabetes, anti-hypertensive medication treatment, total cholesterol/HDL cholesterol, systolic blood pressure, and prevalent CVD.
(Yazdanyar & Newman, 2009)imates in this age range seen clinically(Yazdanyar & Newman, 2009).In this context, the use of common frailty TA B L E 2 Protein scores of frailty are associated with all-cause mortality and non-cardiovascular mortality.Cox regression models for allcause mortality and Fine-Gray competing risk models for CVD and non-CVD mortality.Variable Model A All-cause mortality (N = 1894; deaths = 755) a Adjusted for sex and age.b . All participants in the co- Jha et al., 2017)and validate findings, we sought to replicate our findings using published data from multiple human cohorts.To examine age-related changes in the frailty proteome, we analyzed data from (1) 171 individuals across the lifespan (age 21-107 years) with previously reported plasma proteomics (aptamer-based assay, So-maScan, Somalogic) from several U.S. and European cohort studies of aging and age-related disease (VASeattle, PRIN06, PRIN09, and GEHA;Jha et al., 2017); (2) reported cross-sectional associations of a Hazard ratio is expressed per 1 standard deviation increase in score.Full adjustment includes age, sex, body mass index, smoking history, diabetes, coronary artery disease, and eGFR.
Groothuis-Oudshoorn, 2011).We applied the PCA model from the derivation sample in the validation sample (with imputed data) and correlated the resulting composite PC-based phenotypes with the protein scores from LASSO.While we recognize that the CT measures may not be fully missing at random (potentially limit-Proteomics of frailty are weakly related to age and appear to manifest decades prior to advanced age.(a) Stacked bar plot of the proportion variance explained in phenotype (frailty axes scores) and protein scores by age, sex, BMI, eGFR, diabetes, and smoking history (residuals not shown).Models for explanatory variance for phenotype scores are from the AS derivation sample with complete Lehallier et al., 2019)measures (N = 233); models for protein scores pool all 809 participants (AS derivation and AS validation samples).(b)Scatterplotsdemonstrate relation between phenotype or protein scores with age, with correlation (Spearman).(c)Age-relatedchanges in plasma proteins modeled by loess (based on Z-scores of protein levels) from 171 individuals (age range: 21-107 years;Lehallier et al., 2019).Proteins were selected based on association (FDR <0.10) with one of the frailty axes in the AS cohort and availability in the N = 171 sample.Of note, no proteins were associated with patient-reported outcomes (PC1) in linear models (see File S1), so this is not shown in the heat bar at the top of the heatmap.The 25th-75th percentile of age in the AS cohort is shown in purple beside the age heat bar, suggesting any agerelated changes in the proteome may already be established by the time of AS intervention.*Proteins/genes with FDR <0.10 in linear models of frailty axes in the human AS cohort.frailty axis also present in the Iceland dataset, matched by UniProt) with (1) age (from reported linear regression for protein as a function of age and sex) and (2) a subset of 44 reported phenotypes selected from a total 373 phenotypes based on relevance to aging across multiple systems (linear/logistic regression for phenotype as a function of protein, age, and sex).