Plasma proteomic profile of frailty

Abstract Frailty is a state of decreased physiological reserve and increased vulnerability to adverse outcomes in aging, and is characterized by dysregulation across various biological pathways. Frailty may manifest biologically as alteration in protein expression, possibly regulated at genetic, transcriptional and epigenetic levels. In this study, we examined the proteomic profile associated with frailty defined by an established cumulative frailty index (FI). Using the SomaScan® assay, 4265 proteins were measured in plasma, of which 55 were positively associated and 88 were negatively associated with the FI. The proteins most strongly associated with frailty were fatty acid‐binding proteins, including fatty acid‐binding protein (FABP) (p = 1.96 × 10−19) and FABPA (p = 8.10 × 10−16), leptin (p = 1.43 × 10−14), and ANTR2 (p = 7.95 × 10−20). Pathway analysis with the top 143 frailty‐associated proteins revealed enrichment for proteins in pathways related to lipid metabolism, musculoskeletal development and function, cell‐to‐cell signaling and interaction, cellular assembly, and organization. Frailty prediction model constructed with elastic net regression utilizing 110 proteins demonstrated a correlation between predicted frailty and observed frailty (r = 0.57, p < 2.2 × 10−16). Predicted frailty was also more strongly correlated with chronological age (r = 0.54, p < 2.2 × 10−16) than observed frailty (r = 0.37, p = 1.2 × 10−15). This study identified novel proteins and pathways related to frailty that may offer improved frailty phenotyping and prediction.


| INTRODUC TI ON
Frailty is a late life phenotype, which is associated with low physiologic reserve and increased vulnerability to adverse outcomes such as disability, hospitalization, and death (Fried, Darer, & Walston, 2003;Fried, Ferrucci, Darer, Williamson, & Anderson, 2004). Frailty is a multidimensional construct and involves several components, including physical, psychological, cognitive, and social domains (Fried et al., 2001;Gobbens, van Assen, Luijkx, & Schols, 2012). The complexity of this clinical syndrome has made it difficult to elucidate its biology. Although both genetic and proteomic approaches have been applied, previous studies have been inconclusive regarding the biology of frailty. The main limitations in previous proteomic studies were the fewer number of proteins analyzed as well as the small sample sizes. A study that compared six frail to six non-frail older adults found 31 out of the 226 proteins examined to be highly expressed in frail participants compared to non-frail participants including angiotensinogen (ANGT), kininogen-1 (KG), and antithrombin III (AT) (Lin et al., 2017). Another small study that focused on glycoproteins identified an association of haptoglobin, transferrin, and fibrinogen with frailty (Shamsi et al., 2012). A number of studies targeting candidate proteins in pathways related to oxidative stress and inflammation revealed an association between frailty with Interleukin-6 and lipoprotein phospholipase A2 (Ershler & Keller, 2000;Liu et al., 2016).
However, only limited conclusions can be drawn from these prior studies in relationship to the biology of frailty as they were based on a candidate pathway approaches and were restricted to few participants as well as proteins. To date, no large-scale proteomic study has been carried out in regard to frailty. An additional challenge is to distinguish the biological antecedents of frailty from aging. Since frailty is strongly associated with chronological age, both may share a common biological signature (Xue, 2011).
Frailty is a multidimensional concept that stems from imbalance in multiple biological pathways. Thus, this complex clinical phenotype may be better interrogated by employing an unbiased approach focused on high-throughput proteomic or genomic analysis. To elucidate the proteomic signature associated with frailty, we examined the cross-sectional association between 4265 proteins and frailty in 880 community-residing Ashkenazi Jewish (AJ) older adults participating in the LonGenity Study (Lehallier et al., 2019;Sathyan et al., 2018). We employed an unbiased approach using the SomaScan assay, which is a highly multiplexed, sensitive, quantitative, and reproducible proteomic tool that can assess thousands of proteins simultaneously in a single blood sample (Candia et al., 2017). The principle of SOMAmer ® reagents is based on aptamer technology that uses single stranded DNA-based protein affinity reagents. We defined frailty using the established cumulative frailty index (FI) proposed by Rockwood et al. that includes a diverse range of deficits to capture the complex and multidimensional nature of the frailty phenotype (Searle, Mitnitski, Gahbauer, Gill, & Rockwood, 2008).
Further, we developed a model to predict frailty based on proteomic markers. Establishing the proteomic signature of frailty using an unbiased approach and a comprehensive frailty definition may provide new insights into pathways and underlying biology of frailty in aging.

| Study population
Among the 880 eligible individuals in the LonGenity cohort who had both phenotypic and proteomic data available, 448 were offspring of parents with usual survival (OPUS) and 432 were offspring of parents with exceptional longevity (OPEL). Demographic and clinical characteristics are summarized in Table 1. The mean FI score for the eligible study sample was 0.163 (standard deviation [SD] = 0.086).
The mean frailty scores for OPEL and OPUS were 0.151 ± 0.079 and 0.175 ± 0.091, respectively. The mean age of the participants was 75.35 ± 6.56 years, and 54.8% were women. The sample was highly educated with mean years of education being 17.52 ± 2.88 years.

| Association analysis with frailty
There were 143 proteins that were significantly associated with the cumulative FI ( Figure 1). Of these, 55 proteins were positively associated with the FI, while 88 proteins were negatively associated with the FI.

| Decreased expression of proteins with frailty
Top proteins that were negatively associated with FI were an- ( Figure 1). The proteins with the most negative associations with frailty are listed in Table 3.
Additionally, the analysis revealed significant associations be- Gender stratified analysis showed a significant difference in the number of proteins associated with frailty in male and female participants. There were 88 proteins associated with frailty in the 482 female participants, while only 16 proteins were associated with frailty in 398 male participants (Table S1 and S2; Figure S1). Top hit proteins, including FABP, ANTR2, NELL1, and FABPA, were associated with frailty in both genders (Tables S1 and S2). Proteins that were significantly associated only in males include BCMA (B-cell maturation antigen alias tumor necrosis factor receptor superfamily member 17), TSP2 (thrombospondin-2), contactin-4, F177A (family with sequence similarity 177 member A1), CD248 (endosialin), and EFS (embryonal Fyn-associated substrate) (Table S1).

| Pathway analysis
Pathway analysis using IPA revealed that proteins related to lipid metabolism were the top "molecular and cellular functions" associated with frailty (

| Frailty prediction using proteomic markers
We generated a proteomic signature of frailty using an elastic net regression by fitting in a cluster or subset of proteins from 4265 proteins that best predicted frailty. For this purpose, we divided our cohort into training and validation sets, with each group consisting of 440 unique participants. Elastic net regression applied to the training set selected 110 proteins out of the total number for proteomic frailty predictors (Table S5). Of these, 24 were associated with the frailty phenotype in the analysis above. The correlation between the predicted FI and observed cumulative FI in the validation cohort was r = 0.57 (p < 2.2 × 10 −16 ) ( Figure 2). In our prediction model, the correlation between predicted FI and cumulative FI did not differ by sex, and both fall in line with overall correlation of 0.57.
Association analysis results for all the 4265 SOMAmers with frailty-as well as gender-based stratified results are provided in Tables S6-S8.

| D ISCUSS I ON
The present study aimed to decipher the proteomic signature of frailty. To our knowledge, this is the first large-scale proteomic study using the SomaScan Platform approach to elucidate the molecular phenotype of frailty at the proteomic level. The results of untargeted proteomic approach offer new insights into the pathogenesis and biomarkers of frailty.
The study identified a number of proteins that were positively as well as negatively associated with the clinical frailty phenotype.
Top hit proteins that were positively associated with frailty (FABP, FABPA, and leptin) pointed toward a role for the lipid metabolism pathway in frailty. This was also confirmed by the pathway analysis as well as post hoc analysis including markers of lipid metabolism.
Interestingly, the top two identified proteins belonged to the fatty acid-binding protein family (~15 kDa proteins), which binds a hydrophobic ligand (fatty acids) in a reversible and noncovalent manner.
Higher percentage of saturated fatty acids intake have been shown to be associated with higher levels of frailty (Jayanama, Theou,

TA B L E 3 (Continued)
F I G U R E 2 Correlation of observed cumulative frailty index and predicted frailty index using proteomic data. Frailty prediction using Elastic net regression method in 440 participants in the validation set. Correlation of predicted frailty using proteomic markers and cumulative frailty index was 0.57 Godin, Cahill, & Rockwood, 2019). The top hit FABP, also referred to as FABP-H, is a protein coded by FABP3 gene (Chr 1p32-1p33). It is expressed mainly in the heart and skeletal muscle and is involved in intracellular long-chain fatty acid transport similar to other fatty acid-binding proteins. H-FABP has been shown to be a highly sensitive biomarker for acute coronary syndrome, including myocardial infarction, and it also predicts mortality after such an event (Kilcullen et al., 2007). The second top most hit, fatty acid-binding protein adipocyte (FABPA), is coded by FABP4 (chr 8q21.13) gene and is expressed mainly in adipocytes and macrophages. FABPA is closely linked with obesity and metabolic syndrome (Xu et al., 2006).
It is associated with lipolysis and also acts as an adipokine playing a causative role in insulin resistance and atherosclerosis. Studies have also shown that Fabp-deficient mice are protected against metabolic diseases, and have extended health span with protection against glucose intolerance and insulin resistance (Charles et al., 2017). These mice were also protected against inflammation and loss of adipose tissue integrity (Charles et al., 2017). Additionally, molecular inhibition of FABPA was found to be a successful therapeutic intervention against atherosclerosis and diabetes mellitus type 2 in a mouse model (Furuhashi et al., 2007). Leptin was the third top most protein to be positively associated with frailty. Leptin is a hormone produced mainly by the adipose cells and is involved in the regulation of body fat. Hence, it plays an important role in maintaining body weight and energy balance (Havel, 2000). However, this balance is lost in con-  (Fried et al., 2001). However, further studies are warranted to determine ANTR2's mechanistic role in frailty. NELL1 was another top hit F I G U R E 3 Correlation of observed cumulative frailty index and predicted frailty index with chronological Age. Higher correlation was observed with the predicted frailty and chronological age compared to actual cumulative frailty index and chronological age protein negatively associated with frailty. Another top hit protein negatively associated with frailty was NELL1, under expression of which has been associated with inadequate skeletal mineralization and age related osteoporosis. NELL-1 improved bone mineral density in a rat model and bone formation in a sheep model (James et al., 2015).
Other well-studied proteins previously associated with frailty and aging were also related to frailty in our study. These include higher levels of CRP and lower levels of KLOTHO (Shardell et al., 2017;Soysal et al., 2016). Recent studies, including those utilizing SomaScan array, have strengthened the role of MIC-1/GDF-15 protein, a stress-induced cytokine from the TGF-B family, with aging and associated traits (Tanaka et al., 2018;Wiklund et al., 2010).
Studies have shown up-regulation of GDF-15 in cardiovascular diseases (cardiomyopathies, heart failure, atrial fibrillation, and stroke) and with type 2 diabetes, where higher levels associated with fasting glucose, insulin resistance index, and glycated hemoglobin (Adela & Banerjee, 2015;Berezin, 2016).
Prior studies have demonstrated a higher prevalence of frailty among women compared to men (Fried et al., 2001). Women also have longer lifespans compared to men (Austad, 2006). We found a greater number of frailty-associated proteins in females compared to males. These observations might suggest that there are more pathways leading to frailty in women compared to men. Frailty-related proteins exclusive to males like thrombospondin-2 (TSP-2) play an important role in myocardial matrix integrity (Schroen et al., 2004).
Increased expression of TSP-2 predicted cardiac mortality in 992 elderly men even after adjustment for other cardiovascular risk factors (Golledge, Clancy, Hankey, & Norman, 2013). Expression of TSP-2 rises in response to cardiac hypertrophy, which may lead to cardiac failure (Schroen et al., 2004). The greater number of female-specific frailty-associated proteins suggests the possibility of homeostasis disturbance that results in dysregulated protein networks to be more prevalent in females. These observations might be underlying basis for the observed "male-female health survival paradox," which is characterized by higher mortality rates in men despite higher rates of frailty and medical comorbidities in women (Kingston et al., 2014). We created a proteomic signature of frailty in our LonGenity cohort that achieved a correlation of 0.57 with actual frailty. A higher correlation between predicted and actual frailty may not have been observed due to the multifactorial nature of frailty, which is also influenced by factors such as age and gender. Better characterization of frailty by expanding the criterion as well as accounting for proteins that were not analyzed in the SomaScan array will help improve the concurrent validity of our biological frailty prediction model with physical frailty in the future. Predicted frailty (proteomic) was more strongly correlated with chronological age than actual frailty (FI) in the validation cohort. Hence, proteomic or biological models might become better predictors for frailty and chronological age. Further studies are warranted in this direction.
The current study has many strengths. We examined over 4000 proteins, making it one of the largest studies of proteomics of frailty to date, based on proteome and cohort sizes. Another strength of the study is the well-characterized LonGenity cohort, which undergoes systematic clinical assessments and includes a validated and reliable cumulative deficit FI for capturing the multidimensional aspects of the frailty phenotype (Lehallier et al., 2019;Sathyan et al., 2018). In conclusion, this study identified novel associations of proteins as well as pathways and frailty using the SomaScan array. This study also suggested the possibility of developing a better biological signature for frailty that can be defined by biomarkers. Future studies will need to investigate whether this proteomic signature can accurately identify and predict frailty in diverse populations. Further examination of the frailty-associated proteins identified in this study may help develop potential interventions to mitigate frailty and to maintain functional independence in older adults.

| Frailty
The two most common approaches adopted to define frailty is either as a clinical syndrome (Fried et al., 2001) or as a cumulative deficit index (Rockwood & Mitnitski, 2007). In the present study, we used the cumulative deficit index proposed by Rockwood et al. (Searle et al., 2008) as it assesses a broader spectrum of disorders and conditions compared to the syndromic frailty definition. The cumulative deficit index also provides a continuous variable with meaningful quantification of frailty status independent of functional status or age (Kulminski et al., 2008). Phenotypic frailty proposed by Fried as clinical syndrome considered frailty as categorical variable with meaningful results restricted to non-disabled older person only (Fried et al., 2001). The variables selected for the FI construction were based on standardized criteria that includes the following: association with health status, accumulates with age, biologically relevant, and must represent multiple organ systems (Searle et al., 2008).
Further, variables should not saturate early with age like presbyopia, which are quite common by age 55 and are therefore excluded.
A minimum of 30 variables is recommended for developing the FI (Rockwood & Mitnitski, 2007), which has been shown to predict deteriorating health status, institutionalization, and death (Rockwood & Mitnitski, 2007). Based on the recommended approach, 41 variables were included in the construction of the FI (Rockwood & Mitnitski, 2007). In case of binary variables, 0 represents no deficit and 1 represents a deficit. Continuous or rank variables were graded from 0 (no deficit) to 1 (maximum deficits). The variables and cutoffs used for construction of the FI are shown in Table S9. The FI was calculated by adding the number of deficits (value = 1) and dividing the sum by the total number of variables per participant, which resulted in a range of scores from 0 to 1 for each individual (Rockwood & Mitnitski, 2007). The FI showed a similar distribution to that obtained in earlier studies (Searle et al., 2008).

| Proteomic assessment
Proteomics assessment was carried out using SomaScan assay from human plasma collected at baseline in LonGenity. Plasma samples were stored at −80°C, and 150 µl aliquots of plasma were sent to SomaLogic on dry ice. 5.0k SomaScan Assay includes 5284 SOMAmer reagents consisting of 5209 SOMAmer reagents that recognize human proteins with the remaining including 7 deprecated proteins, 12 hybridization control elution, 10 non-biotin, 4 noncleavable, 22 non-human proteins, and 20 spuriomers. SomaScan data standardization was carried out at SomaLogic, Inc., as previously described (Candia et al., 2017;Lehallier et al., 2019

| Statistical analysis
Baseline characteristics of participants were compared using descriptive statistics (Table 1). Relative fluorescence unit (RFU) values observed after data normalization procedures for each SOMAmer reagent were natural log transformed. Outliers were removed using median absolute deviation method. The primary objective of this study was to identify the association between proteins and frailty using a linear regression analysis. Analyses were adjusted for age, sex, and cohort status (OPUS-OPEL). There are reports of higher frailty prevalence in females compared to males with age (Fried et al., 2001); therefore, we carried out a sex stratified analysis adjusting for age and cohort status. Multiple testing correction was carried out, and Bonferroni corrected p-values of less than 1.17 × 10 −5 (0.05/4265) were considered statistically significant.

| Pathway analysis
Pathway analysis was conducted using frailty-associated proteins to discover the biological pathways related to frailty. This was carried out using QIAGEN's Ingenuity ® Pathway Analysis (IPA ® , QIAGEN Redwood City, www.qiagen.com/ingen uity) (Krämer, Green, Pollard, & Tugendreich, 2013). In this analysis, we included 143 proteins that were significantly associated with frailty in our initial analysis. IPA network analysis output consisted of a list of biological functions and sets of proteins, as well as scores (Score = − log10 (p-value)) according to the fit of the set of proteins.
We selected a biological (cholesterol) and clinical marker (BMI) linked to lipid metabolism to explore the relevance of our results.

| Frailty prediction using proteomic markers
Proteomic frailty predictor was constructed by utilizing a penalized regression model using the glmnet R package (Friedman, Hastie, & Tibshirani, 2010). Participants in the training set were selected using stratified random sampling method. Participants were selected from each of the 0.03 frailty score strata ( analysis, and interpretation of the data; and preparation, review, or approval of the manuscript.

CO N FLI C T O F I NTE R E S T
None declared.

AUTH O R CO NTR I B UTI O N S
SS, NB, and JV contributed to the design of the study and interpretation of the data. SS, NB, SM, TG, and EA contributed to the acquisition of data and writing of the manuscript. EA and TG contributed to the analysis of the data. SS, EA, TG, SM, NB and JV contributed to the critical revisions of the manuscript. All the authors approved the final version of the manuscript and agree to be accountable for all aspects of the work.

DATA AVA I L A B I L I T Y S TAT E M E N T
Proteomic data used in this study are available upon request. Please contact the corresponding author for further information.