Differential proteomic profiles between cognitive frail and robust older adults from the MELoR cohort

Abstract

The present study explored for the first time the blood-based proteomic signature that could potentially distinguish older adults with and without cognitive frailty (CF). The participants were recruited under the Malaysian Elders Longitudinal Research (MELoR) study. Cognition and physical frailty were determined using the Montreal Cognitive Assessment (MoCA) and Fried’s criteria, respectively. The differential protein expression in the blood samples (38 CF vs 40 robust) were then determined using the Sequential Window Acquisition of All Theoretical Mass Spectra (SWATH) analysis. A total of 294 proteins were found to be differentially expressed in the CF group as opposed to the robust group. Considering proteins with fold change (FC) ≥  ± 2 and p-values < 0.05, 13 proteins were significantly upregulated and nine proteins significantly downregulated in the CF group when compared to the robust group. Subsequent correlation analysis identified nine dysregulated proteins, namely APOA1, APOA2, APOA4, APOC1, APOE, GPX3, RBP4, SERPINC1 and TTR, to exhibit significantly and moderately strong correlations with parameters of cognitive and/or frailty assessments. These proteins could potentially serve as useful proteomic signature of CF given their sensitivity > 78%, specificity > 75%, accuracy > 80% and area under the curve (AUC) > 0.8. The major biological pathways that could be potentially dysregulated by the nine proteins were associated with lipid metabolism and the retinoid system. The present findings warrant further validation in future studies that involve a larger cohort.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11357-024-01462-z.

Introduction

The rapidly ageing global population is attributable to the success in extension of life expectancy since the second half of the twentieth century. This has ironically led to concerns with regard to the increased burden of conditions which are more common in later life, including dementia and frailty. Age-related conditions are associated with increased risk of functional dependency, hospitalisation and mortality, which could in turn lead to increased health expenditure and a reduction in quality of life [1].

Cognitive frailty (CF), which is characterised by the co-occurrence of mild cognitive impairment (MCI) and physical frailty, is deemed a potential precursor to both dementia and adverse physical outcomes [2]. Since the establishment of the consensus on the criteria that define CF in 2013 [2], the emerging interest in CF has resulted in a growing number of published studies. Nevertheless, this has also led to the identification of potential areas of confusion, suggesting a need to refine the current CF construct [3]. Prior to the development of CF, cognitive impairment and physical frailty are often independently assessed despite their common aetiologies [4]. In fact, conventional cognitive performance assessment tools do not account for motor capacity whilst conventional physical frailty assessment tools do not account for cognitive performance [5].

Although the latest classification has broadened the CF spectrum into reversible CF (RCF) and potentially RCF, the determinants of RCF or asymptomatic preclinical CF remain to be fully elucidated before effective strategies can be identified to delay the onset or progression of cognitive impairment and physical frailty [6]. The need for early diagnosis and early intervention that can delay or reverse CF and thereby promote healthy ageing have led to the search for potential biomarkers. In this regard, several attempts have been made on blood-based biomarker discovery for CF, but they were predominantly based on targeted approach [7–9] except for Kameda, Teruya [10], who identified potential frailty metabolite markers involved in antioxidation, cognition and mobility through untargeted, comprehensive liquid chromatography–mass spectrometry (LC–MS) metabolomic analysis. There are also reports on other potential non-invasive biomarkers, including digital gait metrics [5] and neuroimaging [11], which had been explored within the separate constructs of either cognitive impairment or physical frailty. Other efforts of CF biomarker discovery were based on in silico [12] and machine learning [4].

The proteome, which represents the set of proteins in a given time and space with varied composition in different cells or tissues, is increasingly recognised as a reliable source of diagnostic biomarkers for various diseases [13]. The use of proteomic approach in understanding frailty in older adults (general or those presented with specific diseases) has been previously described but often limited only to its physical aspect [14–17]. Given the clinical relevance and reliability of the proteomic approach as well as the lack of untargeted proteomic studies amongst those with CF, the present study explored untargeted potential blood-based proteomic profiles of Malaysian adults aged 55 years and over, with and without CF, through the sequential window acquisition of all theoretical mass spectra (SWATH) analysis.

Methods

Ethics approval and study design

The present study enrolled participants from the Malaysian Elders Longitudinal Research (MELoR) study, which was designed to investigate various dimensions of ageing amongst urban-dwelling older adults in Malaysia. A details of this cohort has been previously described by Alex, Khor [18]. The MELoR study was approved by the University of Malaya Medical Centre Medical Ethics Committee (ethics approval ID: 925.4). Written informed consent was acquired from each participant.

Recruitment of participants and CF assessment

Cognition was assessed using the Montreal Cognitive Assessment (MoCA). Physical frailty was determined using the Fried’s criteria (i.e. weakness (hand grip strength measurement), slowness (gait speed measurement), low physical activity (low physical activity if the International Physical Activity Questionnaire (IPAQ) scores were to fall within the lowest quintile for the study population) and shrinking (body mass index (BMI) < 18.5 or self-reported weight loss of 10 pounds or more in the past year) [19]. Upper limb strength was determined by handgrip strength while lower limb strength and balance were determined by the timed-up and go (TUG) test [20]. The present study has adopted distinct criteria that clearly defined robust and CF individuals. Basically, individuals with MoCA scores ≥ 23 or above and did not fulfill any of the Fried’s criteria were categorised as “robust”, whereas those with MoCA scores < 23 and fulfilled at least three of the Fried’s criteria were classified as “CF”. Individuals who fulfilled three or more of the Fried’s criteria were classified as frail [21]. Psychosocial factors were determined using the 21-item Depression, Anxiety and Stress Scale (DASS-21) for anxiety and depression, 6-item Lubben’s Social Network Scale (LSNS-6) for social network, Katz scale for basic activities of daily living (ADL) and Lawton scale for instrumental ADL.

Sample preparation

To improve the identification and quantification of low-abundant proteins, the plasma samples of robust (n = 40) and CF (n = 38) participants (Fig. 1) were depleted for albumin using the single-use spin columns of the ITSIPrep Albumin Removal Kits (ITSI Biosciences, PA, USA). Samples were lysed with sodium deoxycholate (1%) in 100 mM triethylammonium bicarbonate and proteins assayed using commercial kits. Briefly, samples were reduced and alkylated in 1 M dithiothreitol (DTT) and 0.5 M iodoacetamide (IAA), respectively. Proteins were trypsin-digested and then cleaned using styrenedivinylbenzene reverse phase sulfonate (SDB-RPS) StageTips (Fischer Scientific, MA, USA). The samples were then loaded and the StageTips were washed twice with 200 μL 0.2% trifluoroacetic acid. Elution from the StageTips was performed using 100 μL 80% acetonitrile and 5% ammonium hydroxide in water. Eluent was dried through vacuum centrifugation and reconstituted in 90 μL 0.1% formic acid in water.

Fig. 1.

Flowchart of participant selection for the present study

High pH reverse-phase high-performance liquid chromatography

The samples (2 μg each) were pooled for ion library generation through the HpH-RP-HPLC as outlined in a previous study [22]. High pH fractionation was performed and the eluent collected at 2-min intervals at the start of the gradient and at 1-min intervals for the remainder of the gradient. A total of 17 fractions were concatenated, subjected to drying, and reconstituted in 30 μL loading buffer. Subsequently, 10 μL from each fraction was analysed using the two-dimensional information-dependent acquisition (2D-IDA).

Data acquisition

The LC–MS data acquisition was performed using the Triple TOF 6600 (Sciex, Toronto, Canada) with a nano LC system Eksigent Ultra nanoLC system (Eksigent, CA, USA) as described elsewhere [22]. Briefly, each sample (10 μL) was introduced onto a C18 peptide trap (Thermo Fisher Scientific, MA, USA) for pre-concentration and desalted with loading buffer (0.1% formic acid). The peptide trap transitioned in line with the analytical column, and peptides were eluted. The LC eluent was used in the positive ion nano-flow electrospray analysis. In terms of information dependent acquisition (IDA) mode, a time-of-flight mass spectrometry (TOF–MS) survey scan (m/z 350–1500, 0.25 s) was conducted with subsequent MS/MS analysis for the 20 most intense ions (2⁺–5⁺; exceeding 200 counts per s). In terms of data independent acquisition SWATH mode, analysis involved TOF–MS survey scans (m/z 350–1500, 50 ms), followed by MS/MS analysis of 100 predefined m/z ranges were performed. The MS/MS spectra were accumulated for 30 ms in the mass range of 350–1500 m/z.

Data processing and analysis

Protein identification was conducted using the ProteinPilot (v5.0) whilst the Paragon™ algorithm was performed as described elsewhere [22], but with some modifications. UniProt, subset for Homo sapiens, encompassing 26,615 proteins, served as the reference database. The local ion library (353 proteins) and SWATH data were processed using PeakView (v2.2). Parameters included extraction of the top six intense fragments per peptide, a 75 ppm mass tolerance, and a 5-min retention time window. Peptides (≤ 100 per protein) with confidence ≥ 99% and false discovery rate (FDR) ≤ 1% were used for quantitation. SWATH protein peak areas were normalised and subjected to pairwise-relative abundance comparison. Proteins with p < 0.05 and fold change (FC) ≥  ± 2 were deemed as differentially expressed.

Statistical and bioinformatic analyses

The statistical analyses were performed using the GraphPad Prism™ version 9.0.0 (GraphPad Software Inc., CA, USA). The normality of data was determined using the Shapiro–Wilk test. For continuous variables, the significance of differences between demographic parameters of the CF and robust groups was assessed using the independent t-test and Mann–Whitney test for normally and non-normally distributed data, respectively. Categorical variables were compared using the two-sided Fisher’s exact test.

The statistical analyses for correlation, biomarkers and multivariate analysis of variance (MANOVA) were performed using the R programming language version 4.4.0 (R Development Core Team, Vienna, Austria). To understand the relationship between protein expression levels and physical assessments with significant differences, the Spearman correlation analysis was performed for each participant using the Hmisc package of R version 4.4.0. A pairwise correlation coefficient of either a positive or negative correlation was interpreted as follows: < 0.3 (poor), 0.3–0.5 (fair), 0.6–0.8 (moderate) and > 0.8 (very strong) [23]. On another note, the biomarker analysis was performed by generating the receiver operating characteristic (ROC), from which the area under the ROC curve (AUC) was derived. The cut-off value of protein abundance was determined using the Youden index [24]. Subsequently, sensitivity, specificity and accuracy were calculated based on the confusion matrix. The MANOVA was performed to assess the impact of age, height, gender, BMI, diabetes, hypertension and CF or robust status on protein expression levels. It was also used to examine the discrimination between CF and robust status on the nine dysregulated protein expression levels after adjusting for gender, height and BMI using linear regression.

The differentially expressed proteins and genes were retrieved from the UniProt database (https://www.uniprot.org/). The SPRING version 12.0 (https://string-db.org/) was utilised to discern protein–protein interactions, whereas the Reactome pathway browser (https://reactome.org/) was used to identify pathways that are associated with each individual protein and protein subset. The principal component analysis (PCA) and hierarchical clustering of the participants involved were performed using the MetaboAnalyst version 6.0 (https://www.metaboanalyst.ca/).

Results

Demographic characteristics of participants

CF participants were significantly (p < 0.05) older and of lower height compared to robust participants. They had significantly (p < 0.05) lower haemoglobin and packed cell volume (PCV) compared to their robust counterparts. As expected, the CF participants recorded significantly (p < 0.05) higher total frailty scores (i.e. slower gait speed, lower grip strength, longer TUG test scores, higher dependency, lower MoCA test scores and were more likely to be at risk of social isolation relative to the robust participants) (Table 1).

Table 1.

Demographic characteristics of participants

	CD (n = 38)	Robust (n = 40)	p-value
Age (years)	72.58 ± 9.34	63.83 ± 5.67	< 0.0001
Gender			< 0.9999
Male	19 (50.0%)	20 (50.0%)
Female	19 (50.0%)	20 (50.0%)
Height (cm)	154.21 ± 7.414	161.70 ± 8.20	< 0.0001
Weight (cm)	62.16 ± 13.60	67.65 ± 11.74	0.0641
BMI (kg/m2)	26.11 ± 5.37	25.84 ± 3.94	0.7941
Blood and urine test
Haemoglobin (g/L)	131.62 ± 17.62*	141.33 ± 14.64	0.0067
Glucose (mmol/L)	6.48 ± 2.01	5.96 ± 1.54	0.1469
Cholesterol (mmol/L)	4.91 ± 1.23	5.07 ± 0.98	0.3876
Triglycerides (mmol/L)	1.47 ± 0.47	1.45 ± 0.80	0.3354
Total CH/HDL ratio	3.94 ± 0.90	3.69 ± 0.83	0.2240
Albumin (g/L)	43.76 ± 2.23*	44.80 ± 3.05	0.0939
Vitamin_B12 (pmol/L)	459.21 ± 256.48	422.18 ± 169.99	0.7242
PCV (L/L)	0.40 ± 0.05	0.43 ± 0.04	0.0021
MCV (fL)	87.16 ± 7.60	86.88 ± 7.05	0.8555
MCH (pg)	29.13 ± 2.18	28.68 ± 2.65	0.5184
MCHC (g/L)	330.29 ± 10.67	328.28 ± 11.96	0.4403
RDW (%)	14.12 ± 1.40	13.59 ± 0.85	0.1112
Incidence of diseases
Eye disease	18 (50.0%)#	14 (37.84%)^	0.3497
Hearing problem	12 (31.58%)	10 (25.0%)	0.6173
Diabetes	23 (60.53%)	11 (27.50%)	0.0058
Hypertension	27 (71.05%)	15 (37.50%)	0.0035
High cholesterol	23 (60.53%)	20 (50%)	0.3724
Respiratory disease	5 (13.16%)	4 (10.00%)	0.7337
Cerebrovascular disease	3 (7.89%)	1 (2.50%)	0.3524
Malignancy	1 (2.63%)	4 (10.0%)	0.3593
Cognitive frailty assessment
Frailty score	3.32 ± 0.46	0	< 0.0001
Slowness	34 (89.47%)	0 (0%)
Weakness	30 (78.95%)	0 (0%)
Low Activity	27 (71.05%)	0 (0%)
Shrinking	3 (7.89%)	0 (0%)
Exhaustion	32 (84.21%)	0 (0%)
Unintentional weight loss	4 (12.5%)$	2 (5.41%)^	0.4052
Grip strength (kg)	19.39 ± 6.91	29.47 ± 7.51	< 0.0001
TUG (s)	16.32 ± 4.57	11.03 ± 1.92	< 0.0001
Gait speed (m/s)	0.57 ± 0.15	0.91 ± 0.14	< 0.0001
Basic ADL	5.26 ± 0.50	5.18 ± 0.38	0.3579
Instrumental ADL	6.08 ± 1.22	6.73 ± 0.59	0.0034
MoCA	14.79 ± 3.84	26.50 ± 2.11	< 0.0001
Depression	2 (5.26%)	1 (2.5%)	0.6104
Anxiety	2 (5.26%)	0 (0%)	0.2341
Social network	10.89 ± 5.71	13.98 ± 3.65	0.0198
Medications
Alimentary tract	20 (55.56%)#	24 (60%)	0.8167
Blood forming	11 (30.56%)#	12 (30%)	> 0.9999
Cardiovascular system	25 (69.44)#	26 (65%)	0.8078
Lipid modifying agents	18 (62.07%)+	21 (63.64%)@	> 0.9999
Statins	18 (62.07%)+	18 (54.55%)@	0.6120
Fibrates	0 (0%)+	1 (3.03%)@	> 0.9999
Omega	0 (0%)+	4 (12.12%)@	0.1159
Dermatologicals	2 (5.56%)#	0 (0%)	0.2211
Genito-urinary	1 (2.78%)#	2 (5%)	> 0.9999
Hormonal medication	0 (0%)#	3 (7.50%)	0.2421
Antineoplastic and immunomodulating agents	0 (0%)#	2 (5%)	0.4947
Musculoskeletal system	4 (11.11%)#	7 (17.50)	0.5236
Nervous system	4 (11.11%)#	0 (0%)	0.0459
Respiratory system	1 (2.78%)#	3 (7.50%)	0.6170
Sensory organ	0 (0%)#	1 (2.50%)	> 0.9999
Health supplements	1 (2.78%)#	10 (25%)	0.0076

Each data represents either mean ± standard deviation (SD) or n (%)

Health supplements including “transfer factor, shiitake mushroom, beta glucan, soya bean, olive leaf, zinc, inositol phosphate 6, maitake mushroom, agaricus, cordycep, alpha mannans (aloe vera)”, “tomato powder, psyllium husk, spirulina, seaweed, alfalfa, maltooligosaccharide”, stem cell, saw palmetto, arabic gum, cordyceps, garlic, honey, bee pollen, evening primrose oil, transfer factor, garlic oil, chlorella

BMI, body mass index; CF, cognitive frailty; CH, cholesterol; Frailty15ft, walking speed for 15 feet distance; HDL, high-density lipoprotein; ADL, activities of daily living; MCV, mean corpuscular volume; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular haemoglobin concentration; MoCA, Montreal Cognitive Assessment; PCV, packed cell volume; RDW, red cell distribution width; TUG, timed-up and go

^*Information of one participant (n = 1) was unattainable

^#Information of two participants (n = 2) was unattainable

^Information of three participants (n = 3) was unattainable

^$Information of six participants (n = 6) was unattainable

^@Information of seven participants (n = 7) was unattainable

⁺Information of nine participants (n = 9) was unattainable

Differential protein expressions between participants of the CF and robust groups

The SWATH analysis has identified a total of 294 proteins that were differentially expressed in the CF group as opposed to the robust group. Considering proteins with FC ≥  ± 2 and p-values < 0.05, 13 proteins were found to be significantly upregulated whereas nine proteins were significantly downregulated in the CF group compared to the robust group (Fig. 2). Table 2 lists both the upregulated and downregulated proteins in the CF group with FC ≥  ± 2 and p-values < 0.05 compared to the robust group (see Table S1 for the list of upregulated and downregulated proteins in the CF group with ± 1.5 ≤ FC <  ± 2.0 compared to the robust group). Of the 13 differentially upregulated proteins in the CF group, immunoglobulin kappa variable 2–30 accounted for the highest FC (+ 8.7-fold) compared to the robust group. Of the nine differentially downregulated proteins in the CF group, dynein axonemal heavy chain 1 accounted for the highest FC (− 5.3-fold) compared to the robust group. The MANOVA revealed that gender also significantly influenced the protein expression levels (Table S2). However, the number of men and women in CF and robust groups was equally matched (Table 1).

Fig. 2.

Volcano plot of the differentially expressed proteins in the CF group as opposed to the robust group. Proteins were deemed differentially expressed between the CF and robust groups when FC ≥  ± 2 and p < 0.05

Table 2.

List of upregulated and downregulated proteins in the CF group with FC ≥  ± 2 and p < 0.05 when compared to the robust group

Gene	Uniprot ID	Protein name	FC	p-value
IGKV2-30	P06310	Immunoglobulin kappa variable 2–30	8.7	1.48E-04
IGLC2	P0DOY2	Immunoglobulin lambda constant 2	3.0	8.64E-08
IGHG4	P01861	Immunoglobulin heavy constant gamma 4	2.9	2.58E-03
IGHV5-10–1	A0A0J9YXX1	Immunoglobulin heavy variable 5–10-1	2.6	5.95E-04
IGLV2-11	P01706	Immunoglobulin lambda variable 2–11	2.5	1.26E-04
TTR	P02766	Transthyretin	2.4	4.58E-10
GPX3	P22352	Glutathione peroxidase 3	2.4	2.05E-11
IGHV5-51	A0A0C4DH38	Immunoglobulin heavy variable 5–51	2.3	1.12E-04
SERPINC1	P01008	Antithrombin-III	2.2	9.00E-11
KRT10	P13645	Keratin, type I cytoskeletal 10	2.1	9.88E-04
HPR	P00739	Haptoglobin-related protein	2.1	1.96E-07
ITIH4	Q14624	Inter-alpha-trypsin inhibitor heavy chain H4	2.1	2.97E-08
RBP4	P02753	Retinol-binding protein 4	2.0	1.18E-08
APOA2	P02652	Apolipoprotein A-II	− 2.0	3.16E-09
APOC2	P02655	Apolipoprotein C-II	− 2.1	8.45E-06
HBB	P68871	Haemoglobin subunit beta	− 2.2	7.41E-04
APOE	P02649	Apolipoprotein E	− 2.9	5.47E-12
APOA4	P06727	Apolipoprotein A-IV	− 3.0	2.19E-09
EEF1DP3	Q658K8	Putative elongation factor 1-delta-like protein	− 3.6	5.03E-06
APOC1	P02654	Apolipoprotein C-I	− 4.0	3.32E-10
APOA1	P02647	Apolipoprotein A-I	− 4.2	2.11E-10
DNAH1	Q9P2D7	Dynein axonemal heavy chain 1	− 5.3	1.20E-07

FC, fold change

Correlations between differentially expressed proteins and CF parameters

(Fig. 3A depicts the heatmap of differentially expressed proteins and parameters of cognitive and frailty assessments. Considering coefficient correlations ≥ 0.6 and p-values < 0.05, nine proteins exhibited significantly and moderately strong correlations with parameters of cognitive and/or frailty assessments (Fig. 3B). APOA1, APOA4, APOC1, APOE, GPX3 and TTR exhibited moderately strong correlations with parameters of both cognitive and frailty assessments. APOE, GPX3 and TTR, in particular, were correlated to MoCA, frailty score and gait speed. APOA2, RBP4 and SERPINC1 elicited moderately strong correlations with parameters of frailty (i.e. frailty score) but not cognitive assessments.

Fig. 3.

Correlations between differentially expressed proteins (FC ≥  ± 2, p < 0.05) and parameters of cognitive and frailty assessments (significantly different, p < 0.05). A Heatmap of correlation analysis. B List of proteins that exhibited moderately strong correlation (correlation coefficient ≥ 0.60, p < 0.05) with parameters of cognitive and/or frailty assessments. *The coefficient value ≥ 0.60. Abbreviations: PCV, packed cell volume; MoCA, Montreal Cognitive Assessment; TUG, timed-up and go

Discriminatory strength and biomarker potential of the nine dysregulated proteins that exhibited moderately strong correlation with CF parameters

Based on the nine dysregulated proteins that exhibited moderately strong correlation with CF parameters, the PCA plot (Fig. 4A) and hierarchical clustering heatmap (Fig. 4B) indicated a distinct separation between most of the participants in the CF and robust groups.

Fig. 4.

Separation between the CF and robust participants based on the nine dysregulated proteins that exhibited moderately strong correlation with parameters of cognitive and/ or frailty assessments. A PCA plot of the participants. B Hierarchical clustering heatmap that depict the expression trend of the nine dysregulated proteins in each participant. See Fig. S1 for the expression trend of the nine dysregulated proteins in each participant based on diabetes and hypertension incidences.

Subsequent analyses of sensitivity, specificity and accuracy as well as AUC (Table 3 and Fig. S2) revealed that all nine dysregulated proteins yielded sensitivity, specificity and accuracy ranging 78.05–97.22%, 75.51–92.31% and 80.77–92.31%, respectively. Except for APOC1 and RBP4, the proteins yielded both sensitivity and specificity of > 80%. APOA1, in particular, yielded both sensitivity and specificity of ≥ 90%. The ROC analysis indicated that the AUC of the nine dysregulated proteins ranged between 0.88 and 0.95.

Table 3.

Sensitivity, specificity and accuracy of the nine dysregulated proteins (FC ≥ 2, p < 0.05) that exhibited moderately strong correlation (correlation coefficient ≥ 0.60, p < 0.05) with parameters of cognitive and/or frailty assessments. See Table S3 for false positive test on diseases using single and combination of dysregulated proteins (FC ≥ 2, p < 0.05) that exhibited moderately strong correlation (correlation coefficient ≥ 0.60, p < 0.05) with parameters of cognitive and/ or frailty assessments

Protein	Optimal abundance value cutoff	Sensitivity (%)	Specificity (%)	Accuracy (%)	AUC (95% CI)
APOA1	195,731,096.6	94.74	90.00	92.31	0.92 (0.84, 1.00)
APOA2	11,892,998.6	87.50	86.84	87.18	0.89 (0.81, 0.97)
APOA4	24,091,052.3	80.95	83.33	82.05	0.89 (0.82, 0.96)
APOC1	3,561,394.5	96.55	75.51	83.33	0.91 (0.86, 0.97)
APOE	6,236,817.4	97.22	88.10	92.31	0.95 (0.91, 1.00)
GPX3	678,282.5	89.74	92.31	91.03	0.94 (0.88, 1.00)
RBP4	11,117,975.4	78.05	83.78	80.77	0.88 (0.80, 0.95)
SERPINC1	31,070,160.2	90.91	82.22	85.90	0.93 (0.88, 0.98)
TTR	23,364,280.4	87.18	89.74	88.46	0.91 (0.84, 0.98)
3-protein combination#		94.74	95.00	94.87	0.97 (0.94, 1.00)
6-protein combination^		100	93.02	96.15	0.98 (0.94, 1.00)
9-protein combination$		100	93.02	96.15	0.99 (0.97, 1.00)

AUC, area under ROC curve; CI, confidence interval; FC, fold change

^#3-protein combination include APOA1, APOE and GPX3; these three proteins exhibited accuracy ≥ 90% in individual biomarker analysis

^{^}6-protein combination include APOA1, APOA4, APOC1, APOE, GPX3 and TTR; these six proteins were correlated (correlation coefficient ≥ 0.6) with both cognitive and physical assessments

^$9-protein combination include APOA1, APOA2, APOA4, APOC1, APOE, GPX3, RBP4, SERPINC1 and TTR; these nine proteins were correlated (correlation coefficient ≥ 0.6) with parameters of cognitive and/or frailty assessments

Subsequent MANOVA of the nine dysregulated proteins indicated that, apart from CF or robust status, both gender and height of the participants also significantly affected protein expression levels (Table S4). However, after adjusting for potential confounders (i.e. gender, height and BMI), CF or robust status remained as the significant factor influencing the expression levels of the nine dysregulated proteins (Table S5 and Fig. S3).

Protein–protein interaction network and pathway enrichment of the nine dysregulated proteins that exhibited moderately strong correlation with CF parameters

(Fig. 5A–C illustrates the protein–protein interaction network of the nine dysregulated proteins that exhibited moderately strong correlation with CF parameters and with the rest of the differentially expressed proteins (FC ≥  ± 2). Considering only the upregulated proteins (Fig. 5A), five differentially expressed proteins showed interaction in the matched PPI networks. Except for GPX3, three (i.e. RBP4, SERPINC1 and TTR) of the nine dysregulated proteins were found to be part of this protein–protein interaction network. SERPINC1 and TTR, in particular, were presented with the highest degree of connectivity. Considering only the downregulated proteins (Fig. 5B), seven differentially expressed proteins showed interaction in the matched PPI networks. Five (i.e. APOA1, APOA2, APOA4, APOC1 and APOE) of the nine dysregulated proteins were all part of the protein–protein interaction network. APOA4, APOC1 and APOE, in particular, were presented with the highest degree of connectivity. Considering both upregulated and downregulated proteins (Fig. 5C), all nine dysregulated proteins were part of the protein–protein interaction network.

Fig. 5.

STRING network of protein–protein interactions (PPI) and potential biological pathways significantly regulated of the identified nine proteins. A PPI of upregulated proteins. B Downregulated proteins. C Combined upregulated and downregulated proteins. Remarks: The proteins highlighted in the coloured box are the nine dysregulated proteins that exhibited moderately strong correlation with frailty (boxes with blue outlines) or both cognition and frailty parameters (boxes with yellow outlines). D Significant pathways involving upregulated proteins. E Significant pathways involving downregulated proteins

(Fig. 5D–E depicts the significantly dysregulated biological pathway. Considering only the upregulated dysregulated proteins (i.e. GPX3, RBP4, SERPINC1 and TTR) (Fig. 5D), a total of 18 biological pathways were significantly affected, of which six were associated with the retinoid system. Other pathways included those related to the neuronal system, vitamin metabolism, blood clotting, oxidative stress, protein transportation and amyloid fiber formation. As for the downregulated proteins (i.e. APOA1, APOA2, APOA4, APOC1 and APOE) (Fig. 5E), of the top 25 biological pathways that were significantly affected, 14 were involved in lipid metabolism. Other pathways included those associated with the retinoid system, vitamin metabolism, protein transportation and amyloid fiber formation.

Discussion

The present study reported, for the first time, the dysregulation of plasma APOA1, APOA2, APOA4, APOC1, APOE, GPX3, RBP4, SERPINC1 and TTR in urban-dwelling Malaysians aged 55 years and over with CF using the SWATH analysis. Six of the nine dysregulated proteins, which included the downregulated APOA1, APOA4, APOC1 and APOE as well as the upregulated GPX3 and TTR in participants with CF, exhibited moderately strong correlations with parameters of both cognitive and frailty assessments. More importantly, the present study found all nine dysregulated proteins to be a potentially useful signature of CF (sensitivity > 78%, specificity > 75%, accuracy > 80% and AUC > 0.8). Six of the nine dysregulated proteins, namely APOA1, APOC1, APOE, GPX3, TTR and SERPINC1, yielded AUC > 0.9, an indication of excellent clinical usability [25].

Interestingly, seven of the nine dysregulated proteins (i.e. APOA1, APOA2, APOA4, APOC1, GPX3, SERPINC1 and TTR) have been reported in previous proteomic profiling studies despite the use of different platforms. Nevertheless, these proteins were previously identified in older adults (different geographical regions) with either MCI or physical frailty but not both conditions. Liu, Austin [17], for instance, reported the association of APOA1 with lower odds of prefrailty in older American adults using the SomaScan platform. Mehta, Mohebbi [26], another example, reported the correlation of APOA1, APOC1 and GPX3 with cognitive function and significant interaction with mood disorders or bone health-related variables amongst cognitively unimpaired older Australian men using the targeted mass spectrometry-based proteomic assay. Song, Poljak [27], yet another example, reported the downregulation of plasma APOA2 and APOA4 but upregulation of plasma TTR in older Australian adults with MCI using the iTRAQ quantitative proteomics. Elsewhere, Lin, Liao [28], found serum SERPINC1 (i.e. antithrombin III), which was upregulated in frail older Chinese adults based on the ultra-high-performance liquid chromatography-tandem mass spectrometry, to be highly correlated with grip strength. On another note, several previous proteomic studies of either cognitively impaired or frail subjects from other regions (i.e. Australia, Italy, Sweden, Taiwan and the USA) had also identified proteins that were undetected by the present SWATH analysis [14, 16, 17, 26–30]. These variations could be attributed to the different proteomic platforms employed in these previous studies (i.e. LC–MS, proximity extension assay (PEA), SomaScan and nanoflow ultra-high-performance liquid chromatography-tandem mass spectrometry (nUPLC-MS/MS)), which certainly vary in terms of their underlying principles, measurement techniques and the nature of the results produced [31, 32].

Based on their protein–protein interaction network and pathway enrichment, the present study proposed the potential crosstalk between the nine dysregulated proteins during the pathogenesis of CF via the brain-skeletal muscle axis (Fig. 6). In this regard, the present study has identified lipid metabolism as the key dysregulated biological pathways in older adults with CF given that the said pathway involved five of the nine dysregulated proteins (i.e. the downregulated APOA1, APOA2, APOA4, APOC1 and APOE). This confirmed the previous postulation from a population predictive model developed by Sargent, Nalls [4], whereby CF could be potentially driven by dysregulation across multiple cellular processes, including lipid metabolism. This is also in line with the association between cognitive decline and impaired cholesterol metabolism [33, 34].

Fig. 6.

Proposed pathways that may involve the potential crosstalk between the nine dysregulated proteins during the pathogenesis of CF via the brain-skeletal muscle axis. Remark: Part of this image was created using BioRender (BioRender.com)

In the context of cognitive impairment, the present findings on the downregulated apolipoproteins in older adults with CF were, in general, consistent with previous findings except for APOA2 and APOE, both of which reportedly yielded mixed results. A comparison of older Han Chinese men with normal, mildly impaired and impaired cognitive function, for example, implied that those with a higher level of APOA1 may have a lower risk of cognitive impairment (i.e. APOA1 as a protective factor) whereas those with a higher level of APOA2 may have a higher risk of cognitive impairment (i.e. APOA2 as an adverse factor) [35]. Elsewhere, a community based study in Australia found lower plasma levels of APOA1 and APOA2 but higher levels of APOE in individuals with MCI. Interestingly, lower APOA1 was the most significant predictor of cognitive decline over 2 years in cognitively normal individuals [36]. In fact, a lower level of serum and/ or plasma APOA1 was found to be associated with Alzheimer’s disease (AD) [37, 38]. On the other hand, plasma APOA2 and APOA4 were reportedly downregulated in older Australian adults with MCI [27]. Low levels of serum APOA4 were also found to be associated with the risk of AD [37]. Besides, the causative relationships of apolipoproteins and the central nervous system (CNS) had also been demonstrated in preclinical studies. APOC1 knocked out mice, for instance, were presented with impaired memory function in object recognition [39]. In fact, both the absence [39] and overexpression [40] of APOC1 in mice have been found to impair memory functions, suggestive of an important, bell-shaped gene-dose–dependent role for APOC1 in appropriate brain functioning. On another note, APOE was found to be associated with lifespan and cognitive function in exceptionally long-lived older Australian adults. More specifically, APOE exhibited significant negative associations with global cognition scores in the 56 to 86 age group, whereas positive associations were demonstrated in the 95 to 105 age group [41]. The downregulation of APOE as observed in the present study may be due to ATP-binding cassette A1 (ABCA1) deficiency. It has been reported that APOE secretion is compromised in ABCA1-/- astrocytes and microglia that are characterised by lipid accumulation [42]. Overall, the downregulated apolipoproteins as observed in the present study may have dysregulated lipid metabolism (Fig. 6A), which resulted in excess free cholesterol. The increased free cholesterol could be possibly converted to cholesteryl that would in turn enhance the release of β-amyloid (Aβ) (Fig. 6B) [43]. Furthermore, metabolic dysregulation of cholesterol could also cause neurodegeneration (Fig. 6C) [44].

In the context of physical frailty, the present findings of the downregulated apolipoproteins in participants with CF were also in agreement with previous findings. Older Chinese adults with early sarcopenia, for instance, have reduced level of APOA2 [45]. Basically, apolipoproteins are known to play important roles in lipid metabolism and transportation [46]. APOA1, in particular, would markedly increase the levels of ABCA1 protein whereas ABCA1 would, in turn, stabilise APOA1, thereby activating signalling molecules that modulate posttranslational ABCA1 activity or lipid transport activity (Fig. 6D) [47]. More specifically, ABCA1 is involved in the biogenesis of high-density lipoprotein (HDL), whereby it would mediate the efflux of cholesterol and phospholipids to APOA1. Deficiency of ABCA1 would, however, result in the lack of circulating HDL, which would then greatly reduce APOA1 [42]. This would have, in turn, decreased ABCA1 activity, leading to cholesterol accumulation in the skeletal muscles [48] and consequently, skeletal muscle damage (Fig. 6E) [49]. On another note, APOA1 and APOA4 are also known to aid glucose uptake in skeletal muscle [50, 51]. The downregulation of the apolipoproteins in older adults with CF would have decreased ABCA1 activity, hence impairing insulin signalling [48]. Basically, the lower peripheral glucose uptake by skeletal muscle (Fig. 6F) would cause hyperinsulinaemia and insulin resistance (Fig. 6G), thereby increasing frailty incidence [52]. Furthermore, the downregulation of apolipoproteins, especially APOA1, in older adults with CF may also lead to other complications outside the CNS given its roles in preserving mitochondria function (Fig. 6H) [53] and suppressing inflammation (Fig. 6I) [54].

Additionally, the present study has also identified the retinoid (vitamin A) system as the second key dysregulated biological pathways in older adults with CF given that it involved three of the nine dysregulated proteins (i.e. the upregulated GPX3, RBP4 and TTR). The upregulated TTR, in particular, is in line with previous finding in Australian and Taiwanese MCI subjects [27, 55]. In the CNS, vitamin A is crucial for neuroplasticity and aspects of brain function necessary for cognition [56]. Deficiency of vitamin A was, however, found to be linked to cognitive impairment (Fig. 6J) [57]. It was reported that low serum retinol in older Mexican adults was strongly associated with MCI [58]. Low serum vitamin A has also been found to be significantly correlated with frailty (US participants) [59]. Elsewhere, a preclinical study reported that a consistently inadequate intake of vitamin A would give rise to negative effect on skeletal muscle function [60]. Deficiency of vitamin A could result in oxidative stress (Fig. 6K) and impaired mitochondrial function (Fig. 6L) [61]. Yet another preclinical study of psoriasis mice model implied that a deficiency in vitamin A would have increased circulating RBP4 to mobilise retinol from the liver to target tissues [62]. As such, the present finding of the upregulated RBP4, which is consistent with the higher level of RBP4 found in HIV-infected frail subjects [63], may be a positive feedback mechanism in older adults with CF and vitamin A deficiency (Fig. 6M). This may also explain the present finding of upregulated TTR, which binds to RBP [64] in transporting vitamin A (Fig. 6N) [65].

Apart from its role in transportation of vitamin A, elevation of RBP4 would stimulate pro-inflammatory response (Fig. 6O) and subsequently induce insulin resistance (Fig. 6P) [66]. This would certainly compromise skeletal muscle function which is associated with insulin sensitivity [67, 68]. It was reported that elevation of RBP4 is associated with increased serine 307 phosphorylation of insulin receptor substrate-1 (IRS-1), which would reduce its affinity for PI3K and suppress insulin signalling in the muscle [69].

On the other hand, the present finding of upregulated GPX3 may be explained by the previous preclinical study which detected overexpression of GPX3 a protective mechanism against hydrogen peroxide (H₂O₂)-induced oxidative stress in tendons (Fig. 6Q) [70]. Elsewhere, an in vitro study using myoblasts indicated the beneficial effects of GPX3 overexpression on cellular signalling pathways that are typically impaired by ROS [71]. It is also possible that the upregulated GPX3 may coincide with the mobilisation of retinol from the liver to target tissues in older adults with CF and vitamin A deficiency. A preclinical study demonstrated that pretreatment of retinoic acid, a precursor of vitamin A, reduced H₂O₂-induced myoblast death by increasing GPX3 activity in myoblasts, suggesting GPX3 as a retinoic acid target gene [72]. The present finding of upregulated SERPINC1 (i.e. Antithrombin III) is in agreement with a previous report of frail individuals in Taiwan [28]. SERPINC1 would inhibit the activity of thrombin (Fig. 6R) [73] which functions in a signal transduction cascade in skeletal muscle (Fig. 6S) [74]. In the CNS, the inhibition of thrombin (Fig. 6T) would in turn prevent the conversion of fibrinogen into fibrin (Fig. 6U), which may explain the higher level of fibrinogen in MCI groups than non-MCI group [75]. Increased fibrinogen level was associated with an elevated risk of vascular dementia and AD [76].

Strengths and limitations of the study

The present study had adopted SWATH-mass spectrometry (MS), which is a powerful and advanced proteomic technology that permits more precise identification of disease-specific changes in large protein pools [77]. It is suited for biomarker studies, clinical drug/perturbation studies or exploratory basic research [78]. Basically, it consists of data-independent acquisition and a targeted data analysis strategy that aims at maintaining the favourable quantitative characteristics (accuracy, sensitivity and selectivity) of targeted proteomics at large scale. In other words, SWATH-MS has the combined advantages of high reproducibility and sensitivity of targeted methods with the increased proteome depth. SWATH-MS is versatile and has been used in diverse applications (i.e. model organisms, diseases states and bacteria). SWATH-MS has also been useful in characterising of low abundance sub-proteomes including post-translational modifications [79]. More importantly, SWATH-MS studies have shown high intra- and inter-lab reproducibility [80].

A current drawback of SWATH-MS compared to the classical targeted proteomic approaches is that peptide quantification with SWATH-MS is still three- to tenfold less sensitive. A further drawback of SWATH-MS in comparison with data-dependent acquisition-based methods is the required upfront effort on experimental or in silico spectral library and peptide query parameters generation and optimisation [78]. Besides, each sample from the comparison group is run individually in a mass spectrometer for quantitative analysis using SWATH, resulting in inter-run variation, which may influence relative quantification and identification of biomarkers. As such, normalisation of data to diminish this variation becomes an essential step in SWATH data processing [81]. On another note, the method used in the present study indicated abundance of the protein but not the absolute concentration of the protein. Hence, a targeted approach that quantifies the protein level will be useful for future studies.

The present study acknowledged the limitation of a small sample size. There are concerns that smaller samples yield progressively smaller coverage of the expressed proteome [82] which may be influenced by random variation [83]. Besides, small cohorts of complex human samples are associated with inter- and intraindividual variations and systematic effects that may obscure a differential analysis, leading to high false discovery rates and irrelevant results when an improper study design is applied [84]. The uncertainty in the sample variability estimates due to small sample size may give rise to proteins exhibiting a large fold change that are often declared non-significant because of a large sample variance, while at the same time small observed fold changes may be declared statistically significant, because of a small sample variance [85]. Therefore, findings from the study with small samples may not be generalisable to broader population [86]. This warrants validation in a longitudinal study that will involve an independent cohort of a larger sample size to assure high accuracy and reproducibility of the proteomic signature. Besides, proteomic detection of protein expression from small samples can be enriched by pathway analysis followed by targeted proteomics (82).

Conclusion

The present study revealed nine dysregulated proteins (i.e. APOA1, APOA2, APOA4, APOC1, APOE, GPX3, RBP4, SERPINC1 and TTR) which could potentially serve as the proteomic signature of older adults with CF. These interesting findings certainly provide important insights into the pathogenesis of CF that could be associated with lipid metabolism and the retinoid system.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

2D-IDA

Two-dimensional information-dependent acquisition

Aβ

β-Amyloid

ABCA1

ATP-binding cassette A1

AD

Alzheimer’s disease

ADL

Activities of daily living

AUC

Area under the curve

BMI

Body mass index

CF

Cognitive frailty

CNS

Central nervous system

DASS-21

Depression, Anxiety and Stress Scale

DTT

Dithiothreitol

FC

Fold change

FDR

False discovery rate

HDL

High-density lipoprotein

HpH-RP-HPLC

High pH reverse-phase high-performance liquid chromatography

IAA

Iodoacetamide

IDA

Information-dependent acquisition

IPAQ

International Physical Activity Questionnaire

IRS-1

Insulin receptor substrate-1

LC-MS

Liquid chromatography–mass spectrometry

LSNS-6

Lubben’s Social Network Scale

MANOVA

Multivariate analysis of variance

MCI

Mild cognitive impairment

MELoR

Malaysian Elders Longitudinal Research

MoCA

Montreal Cognitive Assessment

nUPLC-MS/MS

Nanoflow ultra-high-performance liquid chromatography-tandem mass spectrometry

PCA

Principal component analysis

PCV

Packed cell volume

PEA

Proximity extension assay

RCF

Reversible cognitive frailty

ROC

Receiver operating characteristic

SDB-RPS

Styrenedivinylbenzene reverse phase sulfonate

SWATH

Sequential window acquisition of all theoretical mass spectra

TOF-MS

Time-of-flight mass spectrometry

TUG

Timed-up and go

Funding

This Transforming Cognitive Frailty to Later-Life Self-sufficiency (AGELESS) study is funded by the Ministry of Higher Education Malaysia under the Long-Term Research Grant Scheme (LRGS/1/2019/UM/01/1/3).

Data availability

The data that supports the findings of this study is available upon request from the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.