Abstract
Using a non-invasive biofluid (saliva), we apply a powerful metabolomics workflow for unbiased biomarker discovery in Alzheimer’s disease (AD). We profile and differentiate Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and AD groups. The workflow involves differential chemical isotope labeling liquid chromatography mass spectrometry using dansylation derivatization for in-depth profiling of the amine/phenol submetabolome. The total sample (N = 109) was divided in to the Discovery Phase (DP) (n = 82; 35 CN, 25 MCI, 22 AD) and a provisional Validation Phase (VP) (n = 27; 10 CN, 10 MCI, 7 AD). In DP we detected 6,230 metabolites. Pairwise analyses confirmed biomarkers for AD versus CN (63), AD versus MCI (47), and MCI versus CN (2). We then determined the top discriminating biomarkers and diagnostic panels. A 3-metabolite panel distinguished AD from CN and MCI (DP and VP: Area Under the Curve [AUC] = 1.000). The MCI and CN groups were best discriminated with a 2-metabolite panel (DP: AUC = 0.779; VP: AUC = 0.889). In addition, using positively confirmed metabolites, we were able to distinguish AD from CN and MCI with good diagnostic performance (AUC > 0.8). Saliva is a promising biofluid for both unbiased and targeted AD biomarker discovery and mechanism detection. Given its wide availability and convenient accessibility, saliva is a biofluid that can promote diversification of global AD biomarker research.
Keywords
INTRODUCTION
The global impact of dementia is both staggering and expanding. Current worldwide future estimates include nearly 10 million new cases, with a total of about 47 million people affected, with costs of multiple $100B (US) [1, 2]. Notably, much projected growth in numbers will occur also in middle (and lower) income countries. Alzheimer’s disease (AD) accounts for over 60% of dementia cases. Despite clarifying advances in AD diagnosis, expansion of the evidentiary base of the neuropathological hallmarks of AD, and the completion of multiple well-designed trials, no effective disease-modifying treatments have appeared [3, 4]. Disease-modifying or delaying interventions could be targeted earlier in the course of AD, during the long pre-AD period of early pathophysiological perturbation referred to as preclinical AD or mild cognitive impairment (MCI) [5–7]. Given the lack of diagnostic and prognostic precision in detecting and confirming early signals of AD in the pre-diagnostic period, much effort has been devoted to discovering and validating early-appearing biomarkers [8]. Prominent among the technologies and approaches employed to date are those targeting AD hallmark markers in cerebral spinal fluid assays (e.g., amyloid-β 42, phosphorylated tau), multiple candidate blood-based markers, neuroimaging techniques (e.g., PET amyloid, volumetric MRI), multi-modal data-driven approaches, and big-data network approaches [7, 9–11]. Although powerful and promising, some of these procedures may involve 1) relatively invasive technologies, 2) limitations to world-wide and diversity distribution, and 3) a focus on specific candidate markers of the pathophysiology of established AD.
We report a complementary approach to detection of early biomarkers of AD and other neurodegenerative diseases. Our approach combines two main elements. First, we apply an unbiased and rapid throughput assay of metabolic perturbations representing a range of potential biochemical pathways leading to clinically diagnosable AD. With metabolomics analyses we are able to examine the metabolic or physiologic state of the organism; this metabolome represents the end and transitional products of the interaction between gene and environment [12, 13]. The result of a metabolomics analysis is an empirically and quantitatively derived set of metabolites (intermediate and end products of metabolism) that form a set of biomarkers that discriminate between two clinical groups and provide targets for further analyses of mechanisms, staging, and intervention endpoints. In metabolomics investigations of AD and MCI biomarkers, CSF, blood, and tissue analyses have appeared (e.g., [11, 14–19]. For example, Trushina and colleagues [19] examined both CSF and plasma from small groups (n = 15) of Cognitively Normal (CN), MCI, and AD patients. Although relatively few metabolites were identified (65–74), key group differences suggested reasonable hypotheses about specific metabolites and pathways. With metabolic profiling of brain tissue, Snowden et al. [17] reported profile differences across small groups AD (n = 14), asymptomatic AD (n = 15) and controls (n = 14) and associations with cognitive performance.
The second main element is that we use saliva, a non-invasive and easily collected and stored biofluid. Saliva, often characterized as “the mirror of the body”, is secreted from three pairs of major salivary glands and many salivary glands lying beneath the oral mucosa [20]. Salivary metabolomics has been established as a viable approach to screen for potential diagnostic and prognostic biomarkers to distinguish different states of diseases [21]. For example, saliva metabolomic profiles of healthy controls and of patients with oral, breast, or pancreatic cancer have been used to differentiate clinically diagnosed groups [22]. Although saliva is a biofluid often used for DNA extraction and genetic risk analyses [23, 24], a broader value for disease-related biomarker development has been recognized [25–27]. For AD, its value includes the complementary facts that it is non-invasive and relatively inexpensive and could rapidly expand the AD biomarker data-base to include a diversity of global populations, thereby extending the range of application and translation possibilities. As a secretory fluid, saliva includes a range of molecules that represent current systemic disease, including neurobiological health [27]. In our first exploration of salivary metabolomics [21], we used new techniques to detect 1052 metabolites and subsequently identify a set of 18 that validly discriminated age-and-sex-matched small groups of CN and MCI participants in Canada. Recently, in China, Liang and colleagues [28] used larger groups to detect 6 tentatively identified metabolites for early AD diagnosis and develop a 3-metabolite panel that produced substantial AUC values in discriminating AD patients from normal controls. Other efforts are appearing using a variety of technologies and clinically diagnosed groups, including useful preliminary reports [29, 30] and larger investigations based on extant clinical cohorts [31].
In the present study, we apply advanced metabolomics technology to salivary samples from three well-characterized and objectively classified groups (CN, MCI, and AD). We thus examine salivary metabolomic profiles across three points in the continuum of known clinical impairment in AD. Our current work proceeds in two phases, including both discovery and validation groups for each of the groups, thus providing a test of the validity of the discovered biomarkers. Like one previous investigation [31], we select leveraged biosamples and clinical data from larger-scale longitudinal databases in brain aging and dementia.
An unbiased and data-driven search for new and potentially early biomarkers of MCI and AD requires an analytical technique that allows accurate quantification and high metabolome coverage. To this end, we have recently reported a quantitative technique based on 13C-/12C-isotope dansylation labeling combined with liquid chromatography Fourier-transform ion cyclotron resonance (FTICR) mass spectrometry (MS) for in-depth profiling of the amine/phenol submetabolome in human saliva biosamples [21]. In the present study, we describe a workflow with expansive submetabolome coverage for saliva metabolomics and report the pairwise comparison results pertaining to new metabolite biomarkers differentiating CN, MCI, and AD. Our new technology produces strong biomarker discovery results using three moderate-sized groups, as validated by smaller independent groups and analyses. We examine three major research goals. First, we apply an unbiased metabolomics workflow to examine the salivary submetabolome of CN, MCI, and AD groups. Second, we determine the number of important biomarkers for discriminating each of the clinical group pairs in both the discovery and provisional validation phases. Third, we develop diagnostic metabolite panels that maximize the discrimination of each pairwise clinical comparison (CN versus MCI, CN versus AD, MCI versus AD).
MATERIALS AND METHODS
Participants
Participants were community-dwelling older adult volunteers from the Victoria Longitudinal Study (VLS), an ongoing multi-cohort investigation of biomedical, genetic, neurocognitive, and other aspects of aging, impairment, and dementia. All participants provided written informed consent and all data collection procedures were in full and certified compliance with human research ethics (University of Alberta). Detailed information on recruitment, research design, and participant characteristics are available elsewhere [23, 32]. For the present study, the CN and MCI participants were drawn from a subset of the main cohort that participated in the VLS biomarker and genetics initiative (2009–2012). The AD patients were recruited independently as part of this initiative and are described below. All participants received a small honorarium for their contributions. The present research proceeded in two phases with two independent cohorts from a common population. First a Discovery Phase (DP; n = 82) included adults classified as CN (n = 35; age 64–75 years; 62.9% female), MCI (n = 25; age 64–75 years; 60% female), and newly diagnosed AD (n = 22; age 52–91; 72.7% female). Second, a small and provisional Validation Phase (VP, n = 27), included an additional group of CN (n = 10; age 68–75 years; 50% female), MCI (n = 10; age 67–75 years; 50% female), and newly diagnosed AD (n = 7; age 53–91 years; 71.4% female) patients. Participant demographic characteristics and clinical information are presented in Table 1.
Clinical characteristics of the discovery and validation samples
Classification and diagnosis
CN and MCI classification
For both sets of CN and MCI participants, we performed an extensive and fully objective (non-clinical) classification. Exclusionary criteria included no diagnosed dementia, cardiovascular disease, stroke history, or psychiatric illness, with minimum Mini-Mental State Examination (MMSE = 24) [33]. Inclusionary criteria included two waves (4.5 years) of longitudinal data, and complete data on the VLS cognitive status reference battery. With this battery we implemented an established cognitive classification procedure that requires strict adherence to specific assessment and objective selection rules [34–36]. At each of two waves all eligible participants completed performance tests of five fundamental cognitive domains: perceptual speed, inductive reasoning, episodic memory, verbal fluency, and semantic memory. The four-step classification procedure was as follows: Source participants were 1) stratified into two age (64–73 and 74–95) and education (0–12 years and 13+ years) groups, 2) placed in appropriate age×education subgroups, 3) analyzed for mean cognitive scores on all tests, and 4) evaluated by score within respective age×education subgroups. We applied a standard criterion to establish higher or lower (“impaired”) group based on one SD below the subgroup mean for any cognitive test. The remainder were classified as CN. We repeated this procedure independently at wave 2. Because the pool of CN was larger, we then identified MCI participants (n = 25 for DP) who met the additional criterion of being selected at two consecutive waves. For the DP phase this resulted in n = 25, who we then matched (age, sex) with CN adults and supplemented with randomly selected additional participants (n = 35). One purpose of the meticulous (two-wave) classification procedure was to ensure that both the CN and MCI groups were relatively homogeneous in status and not immediately subject to clinical transitions.
AD diagnosis
Dementia patients with AD were recruited from the Geriatric and Cognitive Neurology clinics at the Glenrose Rehabilitation Hospital in Edmonton, Alberta. Patients were required to have a diagnosis of AD based on DSM-IV criteria for Dementia of the Alzheimer Type [37]. Clinical assessments were performed as part of routine clinical evaluation, which included caregiver report of cognitive decline and impaired functional status, mental status evaluation of the patient (including the MMSE [33] and Montreal Cognitive Assessment [38]), and a physical and neurological examination. All patients had a medication review and had routine laboratory assessment for causes of dementia, including blood work and brain imaging according to Canadian Consensus Guidelines. Patients could not have vascular dementia based on imaging and a modified ischemic score >4. Medical comorbidity was recorded using the modified Cumulative Illness Rating Scale. The protocol was approved by the Human Ethics Committee of the University of Alberta and participants signed an informed consent allowing access to clinical data performed as part of clinical assessment.
Saliva sample collection
We used the Oragene® •DNA Self-Collection Kit OG-500 (DNA Genotek, Inc., Ottawa, Ontario, Canada). Whole saliva was collected, placed inside the kit, and shaken. The kit contained an Oragene DNA-preserving solution. The ingredients of Oragene solution include ethyl alcohol (<24 %) and Tris-HCl buffer (pH 8). Salivary samples were collected and prepared according to the manufacturer’s protocol. The protocol included the provisions that no food had been consumed in the previous hour and light washing prior to saliva collection was administered. This procedure was relatively representative, methodologically replicable, and clinically manageable. Remnants of any prior food intake were minimized in our analyses by examining only commonly detected metabolites (existing in at least 50% of the samples). Such a selection virtually assures that further analyzed metabolites were endogenous. As provided by established procedures, samples were stored at room temperature before the dansylation labeling procedures and were preserved in –80°C after labeling for long-term storage and follow-up studies.
Chemicals and reagents
13C-dansyl chloride (DnsCl) was synthesized in-house [39]. 12C-dansyl chloride was purchased from Sigma-Aldrich (Milwaukee, WI). All reagents were of ACS grade or higher with water and organic solvents being of LC-MS grade.
Metabolite extraction and isotope labelling
We adapted and extended the workflow from earlier pilot work on saliva metabolome profiling [21]. Specifically, an aliquot of 5 μL of saliva sample was dissolved in 20 μL acetonitrile (ACN)/ H2O (50/50) in a screw cap vial. 12.5 μL of NaHCO3/NaH2CO3 buffer solution (250 mM, 1:1, v/v) was mixed in the solution and the vial was vortexed and then spun down. 36.6 μL of freshly prepared 12C-DnsCl or 13C-DnsCl in ACN (12 mg/mL) was added into the vial. The solution was vortexed, spun down again, and then let to react for 60 min in an oven at 60°C. After the reaction, 5 μL of NaOH (250 mM) was added to quench the excess DnsCl by another 10-min incubation in the 60°C oven. Finally, 25 μL of formic acid in ACN/ H2O (425 mM, 1:1, v/v) was added to neutralize the solution. Each individual saliva sample was directly labeled with 12C-DnsCl. A pooled saliva sample was prepared by pooling 10 μL of each from all 109 saliva samples together. This pooled sample was labeled with 13C-DnsCl under the exact same reaction condition. Both the DP and VP sample sets used the same 13C-labeled pooled sample to mix with their 12C-labeled individual saliva samples.
Ultra performance liquid chromatography with UV detection
After being labeled with 12C- or 13C-DnsCl, the total concentration of the labeled submetabolome in a sample was quantified by a step-gradient Ultra Performance liquid chromatography (UPLC) with UV detection at 338 nm [40]. An ACQUITY UPLC system (Waters Corporation, Milford, MA) equipped with photo diode array (PDA) detector, and a Waters ACQUITY UPLC BEH (Ethylene Bridged Hybrid) C18 column (2.1 mm×50 mm, 1.7 μm particle size, 130 Å pore size) were used for LC-UV analysis. LC solvent A was 0.1% (v/v) formic acid in 5% (v/v) ACN/H2O, and solvent B was 0.1% (v/v) formic acid in ACN. The gradient elution profile was as follows: t = 0 min, 0% B; t = 1.00 min, 0% B; t = 1.01 min, 95% B; t = 2.50 min, 95% B; t = 3.00 min, 0% B; t = 6.00 min, 0% B. The flow rate was 450 μL/min, and the sample injection volume was 2 μL.
Liquid chromatography mass spectrometry (LC-MS)
Metabolomics analyses were performed using an Agilent 1100 series capillary high performance liquid chromatography (HPLC) system (Agilent, Palo Alto, CA, USA) connected to a Bruker 9.4 T Apex-Qe fourier transform ion cyclotron resonance (FTICR) mass spectrometer (Bruker, Billerica, MA, USA) equipped with an electrospray ionization (ESI) interface operating in positive ion mode. Reversed phase (RP) chromatographic separation was carried out on an Eclipse C18 column (2.1 mm×100 mm, 1.8 μm particle size, 95 Å pore size), with solvent A being water with 0.1% (v/v) formic acid and 5% acetonitrile (ACN) (v/v), and solvent B being ACN with 0.1% (v/v) formic acid. The LC flow rate was 180 μL/min and running time was 26.50 min. The gradient was: t = 0 min, 20% B; t = 3.50 min, 35% B; t = 18.00 min, 65% B; t = 21.00 min, 95% B; t = 21.50 min, 95% B; t = 23.00 min, 98% B; t = 24.00 min, 98% B; t = 26.50 min, 99% B. The sample injection volume was 6 μL and the flow was split 1:2 and 60 μL/min of the LC eluate entered the MS system.
Data processing and statistical analyses
The 12C-/13C-ion pairs were extracted from raw LC-MS data by a peak pair picking program, IsoMS [41]. The peak pairs of all individual samples were then aligned by retention time and accurate mass to produce a metabolite-intensity table. The alignment parameters were set as retention time tolerance of 30 seconds and accurate mass tolerance of 8 ppm. A Zero-fill program [42] was then applied to all the peak pairs in the table to retrieve the missing values from the raw LC-MS data. Finally, a peak reconstruction program [43] was used to reconstruct the 12C-/13C-MS intensity ratios using the LC-MS chromatographic peak areas for better quantification accuracy.
Multivariate statistical analysis of the LC-MS data was carried out using SIMCA-P+ 12.0 (Umetrics, Umeå, Sweden). Principal component analysis (PCA) and orthogonal partial least squares-discriminant analysis (OPLS-DA) were used to analyze the data. Receiver operating characteristic (ROC) analysis and linear SVM model was performed using MetaboAnalyst (http://www.metaboanalyst.ca/). Mean center and autoscaling were used to normalize all the peak ratio values prior to the statistical analysis.
Metabolite identification
For positive or definitive metabolite identification, the peak pairs were matched against a Dns-standards library by retention time and accurate mass [44]. In addition, putative metabolite identification was performed based on accurate mass match of the peak pairs to those of the metabolites in the Human Metabolome Database (HMDB) [45] and the Evidence-based Metabolome Library (EML) using MyCompoundID [46], with a mass tolerance of 5 ppm.
RESULTS
Workflow
The workflow required three successive two-group metabolomics analyses (discriminating AD versus CN, AD versus MCI, and MCI versus CN; see Fig. 1). As shown in the figure, a 5-μL saliva sample was aliquoted out from each individual sample and labeled with 12C-DnsCl. A pooled sample, prepared by mixing small aliquots of individual samples, was labeled with 13C-DnsCl. The 12C-labeled individual sample was then mixed with 13C-labeled pooled sample in a 1:1 mole amount ratio after the total concentration of the labeled metabolites was determined by LC-UV. After the LC-MS analysis of the 13C-/12C-mixtures, automatic data preprocessing was performed to extract the peak pairs belonging to the labeled amine/phenol submetabolome. To discover metabolites best contributing to the clinical group discriminations, we performed pair-wise statistical comparisons using OPLS-DA and volcano plot analyses. The discrimination power of the common metabolites that were highly ranked by both statistical tools was then evaluated using ROC analysis. The top-ranked metabolites were identified and externally validated by an independent set of saliva samples from equivalent groups. Independent and panel metabolite biomarkers were evaluated for application to detection of neurobiological perturbations associated with AD-related neurodegeneration. Detailed experimental work on workflow optimization can be found in Supplementary Material 1 and Supplementary Figure 1.
Schematic workflow for the salivary biomarker discovery.
Characterizing the saliva metabolome
Our procedures led to an overall metabolome coverage (with consistent peak ratio values) that was higher than any known salivary metabolomics work in neurodegeneration [21, 28] and higher than most blood-based metabolomics research in the field [19, 48]. Specifically, in the DP analyses we profiled the 82 saliva samples from the three groups. Relative standard deviations (RSD) for peak pair ratios ranged from 0.1% to 15%, with an average of 2% for the triplicate experiments. A set of data that fell within the range of 10–15% was subjected to Grubbs test at 99% confidence level to detect any statistical outlier. A total of 6,230 unique pairs or metabolites (defined by molecular ion m/z coupled with its retention time) were obtained from the LC-FTICR-MS analysis with an average of 3,669 peak pairs detected from each sample. Among them, 3801 peak pairs were commonly detected in more than 50% of the samples. By searching these peak pairs against the Dns-library composed of 273 labeled standards, using mass tolerance of 5 ppm and RT tolerance of 15 s, we positively identified 79 metabolites based on mass and RT matches (see Supplementary Table 1 for the list). Using MyCompoundID MS, a search based on accurate mass of peak pairs with mass tolerance of 5 ppm, 616 (9.9%) metabolites were putatively identified using the HMDB library (see Supplementary Table 2) and 2972 (47.7%) were identified using the predicted human metabolite library with one reaction (see Supplementary Table 3). These results demonstrate a very high coverage of the amine/phenol submetabolome in saliva using the optimized dansylation LC-MS workflow and also show the complexity and great diversity of the salivary metabolites present in a sample.
Principal component analyses for CN, MCI, and AD groups
Principal component analysis (PCA) was used to obtain an overview of the CN, MCI and AD salivary metabolomic data. As can be seen in Fig. 2A, the PCA score plot displays some initial separation among the groups (e.g., the AD patients are located in the top right corner marker in green). However, the separation between CN and MCI samples is not very obvious, indicating that the metabolomic differences between MCI and CN samples are not as large as those of AD and CN.
The results of multivariate statistical analysis in PCA score plot (A), OPLS-DA score plot (B), and OPLS-DA 3D score plot (C).
We next applied OPLS-DA to examine the metabolomic differences in CN, MCI, and AD (see Fig. 2B). The quality of the OPLS-DA model was evaluated via an internal validation method using a seven-fold cross-validation step, from which the values of Q2Y (predictive ability of the model) and R2Y (goodness of fit parameter) were calculated. The score plot in Fig. 2B shows a very clear separation among the three groups with high validation parameters (R2Y = 0.93 and Q2Y = 0.79), demonstrating the robustness of the model. To view the progression of saliva metabolomic variations from CN to MCI, and then to AD, a 3D OPLS-DA score plot is shown in Fig. 2C. There were clear differences of metabolic changes across the three clinical conditions. Specifically, the AD cluster was further away from the CN cluster, with the MCI cluster at an intermediate location. A misclassification analysis revealed that only one sample was misclassified in the OPLS-DA model and that the overall p value (Fisher probability: 3.9×10-17) was extremely small (data not shown). This result validates the group separation observed in these figures.
AD biomarker discovery results
To determine the most significant metabolites that differentiate paired groups (i.e., AD versus CN, AD versus MCI, and MCI versus CN), both multivariate (OPLS-DA) and univariate (Volcano plot) statistical tools were applied in a cross-selection procedure. After the common metabolites were extracted, we applied ROC analyses to evaluate their diagnostic performance. Metabolites with high diagnostic performance were considered as the potential discovered biomarkers (and subsequently externally verified with the validation samples). Our goal was to optimize the discovery biomarkers by using as few identified metabolites as possible (e.g., the top one to three ranked metabolites) to build a diagnostic model. We reasoned that larger panels would gain little in diagnostic precision but be less practical or cost-efficient in clinical applications.
Results of this OPLS-DA analysis for pair-wise comparisons are shown in Supplementary Figure 2. All OPLS-DA models demonstrate clear group separation with high validation metrics, confirming the goodness of fit and good predictive capabilities of the models. From OPLS-DA analysis, metabolites with Variable Importance for Projection (VIP) scores larger than 1.5 in all three comparisons were retained (see Supplementary Table 4). Next, volcano plots analysis was performed to find metabolites with high fold-change and low p values; these results are shown in Supplementary Figure 3. Thresholds of p-value of 0.01 and fold-change of 1.2 were used to discriminate between the significantly up- and downregulated metabolites (see Supplementary Table 5).
We then compared the results of OPLS-DA and volcano plot analyses in order to select only the metabolites shown to be significant in both analyses. ROC analyses were then carried out on these selected and common metabolites, imposing an AUC cut-off value of 0.75 or above to generate the final pool of significant metabolites. Notably, this process resulted in observing 175 significant metabolites in the AD versus CN comparison, 142 metabolites in the AD versus MCI comparison and 59 metabolites in the MCI versus CN comparison (see Supplementary Table 6).
AD biomarker validation results
To provisionally validate selected significant metabolites detected in the discovery phase, we applied the identical procedures and analyses to the independent VP samples representing the same clinical groups. Each individual sample was labeled by 12C-DnsCl and experimental triplicate was performed. The same 13C-labeled pooled sample was used and mixed with each 12C-labeled individual sample at a 1:1 mole sample amount, followed by LC-MS analysis. We detected an average of 2,981 peak pairs or metabolites per sample with a total of 4,157 peak pairs detected in the 27 samples across the three groups. Among them, 3,184 peak pairs were commonly found in more than 50% of the samples. These peak pair numbers are smaller than those observed in the DP, due to the fact that using the zero-fill program the low abundance peak pairs are not recovered as efficiently in smaller samples.
As the purpose of VP is to corroborate the discovered biomarkers and diagnostic models, only those significant DP metabolites were studied in the validation dataset. Based on the criteria of AUC≥0.75 in both the DP and VP, we were able to identify numerous metabolites that produced consistently good ROC performance for AD versus CN (n = 63), AD versus MCI (n = 48), and MCI versus CN (n = 2) comparisons, respectively. These metabolites are listed in Supplementary Table 7. Among them, 4 metabolites (phenylalanyl-proline, urocanic acid, phenylalanyl-phenylalanine, and tryptophyl-tyrosine) were definitively identified for differentiating AD and CN. These 3 metabolites (phenylalanyl-proline, alanyl-phenylalanine, phenylalanyl-glycine) were definitively identified for differentiating AD and MCI.
Development and validation of a diagnostic model
A single metabolite may have an independently good prediction power for disease diagnosis. However, a well-developed diagnostic model or panel using multiple biomarkers may produce superior performance [24, 49]. We used a linear support vector machine tool (linear SVM) in MetaboAnalyst [50] to develop a diagnostic model for each of the three pair-wise comparisons using the DP. Its diagnostic performance was further evaluated using the VP data set.
Table 2 presents the summary of the three linear SVM-based tools applied to the diagnostic models. Using the top 3 metabolites (#6112 putatively identified as methylguanosine, #7628 putatively identified as histidinyl-phenylalanine, and #4489 putatively identified as choline-cytidine), we distinguished AD from CN with AUC = 1.000 (0.993–1.000 at 95% CI) in the discovery set. This result was validated in the validation set with AUC = 1.000. The diagnostic sensitivity was 100% and the specificity was 100%. Similarly, using the corresponding three-metabolite panel (#1429 putatively identified as amino-dihydroxybenzene, #3731 putatively identified as glucosylgalactosyl hydroxylysine - H2O, and #943 putatively identified as aminobutyric acid + H2), we separate AD from MCI with AUC = 1.000 (1.000–1.000 at 95% CI) in the discovery set and AUC = 1.000 in the validation set. The sensitivity was 100% and the specificity was 100%. In the case of MCI versus CN, using two metabolites (#3731 putatively identified as glucosylgalactosyl hydroxylysine- H2O and #7500 putatively identified as glutamine-carnitines), we differentiated MCI and CN with AUC = 0.779 (0.625–0.917 at 95% CI) in discovery set and AUC = 0.889 in the validation set. The sensitivity was 100% and the specificity was 70.0%. The discrimination performance observed for two of the panels in the pairwise comparisons (CN versus AD and MCI versus AD) was substantially better than the third (MCI versus CN). For discriminating MCI and CN, we attempted to use a larger panel by including more metabolites for comparison. However, the discrimination performance for discriminating MCI and CN were found not to be better than the 2-metabolite panel (#3731 and #7500).
Metabolic diagnostic model for pair-wise comparison using top-ranked metabolites
The biomarkers listed in Table 2 were putatively identified, but their structures have not yet been positively confirmed. We next built diagnostic models using only positively identified metabolites. The results of these biomarker panels are shown in Table 4. For separating AD and CN, a biomarker panel using three identified metabolites, phenylalanyl-proline, phenylalanyl-phenylalanine and urocanic acid, produced an AUC of 0.820 (0.711–0.929 at 95% CI) in the DP and 0.814 in the VP with 71.4% sensitivity and 90.0% specificity. For separating AD and MCI, a two-biomarker panel comprised of alanyl-phenylalanine and phenylalanyl-proline produced an AUC of 0.881 (0.743–0.986 at 95% CI) in DP and 0.786 in VP with 71.4% sensitivity and 80.0% specificity. Although the biomarker panels achieved somewhat lower AUCs than the top 3 individual biomarkers, the values are strong (>0.80) and the metabolites in these panels were positively identified. Therefore, these panels could be readily tested for translation clinical application, whereas the putatively identified top-ranked individual metabolites require further study for chemical structure confirmation.
Because the sample size of the validation data set is relatively small, we have referred to it as “provisional”. To check the results of the VP data, we performed another set of ROC analyses on the combined data sets (i.e., DP and VP together). Analyzing the two data sets together provides an indirect means of cross-validating the performance of the biomarker panels reported above. For the overall analyses, the AUC, sensitivity and specificity results are shown in Tables 2 and 3 . The ROC curves and permutation tests are presented in Supplementary Figures 4 and 5. The results were encouraging to the provisional conclusion of validation. For differentiating AD versus CN using three putatively identified metabolites as listed in Table 2, the overall AUC is 0.997 with sensitivity of 98.52% and specificity of 96.55%. The permutation results (100 repeats) show a p-value of less than 0.01, indicating the significance of the model (Supplementary Figure 4). These ROC parameters are only slightly less than those obtained in original validation data set (AUC = 1 with 100% sensitivity and 100% specificity). In the case of AD versus MCI (see Table 2), the comparison results are similar to those of AD versus CN. For differentiating MCI versus CN (Table 2), the overall AUC from the combined data sets is 0.864, which is comparable to 0.889 in the original validation data set. The overall sensitivity and specificity are 80.0% and 91.80%, respectively, compared to 100% and 70% in validation set. Using three positively identified metabolites as the biomarker panel to differentiate AD and CN, the overall AUC is 0.831 (see Table 3 and Supplementary Figure 5) with 82.22% sensitivity and 73.56 specificity; these results are comparable to those observed in the validation sample (i.e., AUC of 0.814 with 71.4% sensitivity and 90.0% specificity). Similarly, using the two identified metabolites as the biomarker panel to differentiate AD versus MCI, the overall AUC is 0.843 (see Table 3) with 81.90% sensitivity and 72.41% specificity; these values are comparable to AUC of 0.786 with 71.4% sensitivity and 80.0% specificity as determined in the validation data. In all cases, the permutation (100) test results are significant (p < 0.01). This full-sample check of the original validation data indicates a robust performance of the biomarker panels shown in Tables 2 and 3. These panels effectively differentiate the clinical groups in both the combined and separate data sets.
Metabolic diagnostic model for pair-wise comparison using positively identified metabolites
Molecular pathways
As discussed above, six dipeptides were significantly dysregulated in the comparison of AD versus CN and AD versus MCI. Protein malfunction and dysregulation is one of the hallmarks in AD and other neurodegenerative diseases [51]. Protein dysregulation leads to the dramatic alteration of peptide levels in the CSF of AD patients that could serve as diagnostic biomarkers [52]. It has been shown that many of these peptides are degraded products of amyloid or tau precursor proteins which are the biomarkers detectable primarily through CSF sample analysis for indicating AD progression [51]. In the present study, we discovered specific affected dipeptides with significantly upregulated concentrations in saliva that may serve as diagnostic biomarkers for distinguishing AD from CN, and even AD from MCI. These dipeptides may be targeted for validation in future work of using salivary samples from other independent and larger cohorts. We note that, compared to CSF, saliva samples can be noninvasively obtained and thus analyzing potential dipeptide biomarkers and panels in saliva could be efficient and effective for early detection, diagnosis staging of AD.
Urocanic acid was also detected with consistent alteration in the AD versus CN comparison in both discovery and validation groups. Urocanic acid is involved in the histidine degradation pathway and is formed from L-histidine through the action of histidine ammonialyase by elimination of ammonium. The disruption of the histidine degradation pathway affects the biosynthesis of histamine, which is associated with increased production of pro-inflammatory cytokines as an inflammatory response against pathogens and has also been implicated as involved in the inflammatory response against pathogens and neurotransmission failures in AD [53, 54]. A previous study reported a decreased level of urocanic acid in the serum of transgenic AD mouse model [55]. Our study also found the dysregulated urocanic acid level in the saliva of AD patients, but with a 3-fold higher concentration in the AD group compared with the CN group. The difference may be due to species specificity in urocanic acid metabolism. Nevertheless, the dysregulation of urocanic acid in saliva supports the potential biological failures of the histaminergic system in AD patients and indicates the intriguing importance of immune regulation and inflammatory processes during AD progression.
DISCUSSION
The overall purpose of this study was to apply a new and powerful unbiased and rapid throughput metabolomics assay to salivary samples in order to detect AD-related perturbations leading to biomarker discovery and validation. Salivary samples, a non-invasive and widely available biofluid, are of considerable promise to diversify and clarify research in neurodegenerative diseases. We used objectively classified CN and MCI participants and well-diagnosed AD participants. The workflow (Fig. 1) was conducted in both discovery and validation phases (DP, VP) and performed successfully in all two-group comparisons. Not only was an unprecedented number of metabolites detected, but each clinical group comparison resulted in discriminative biomarkers with meaningful sensitivities and specificities. Given the favorable results in both phases, we developed diagnostic panels of the top biomarkers to be used in further research and application. We review our main findings.
Leading edge advances in dansylation isotope labelling LC-FTICR-MS method was developed for metabolite biomarker discovery using human salivary samples. With only a very small amount of starting material (5 μL of individual saliva) needed for the analyses, our new technology detected a total of 6230 metabolites in the amine/phenol submetabolome in the DP. In the DP, we tested the top 3 metabolites commonly found by both OPLS-DA and volcano plot analyses. Excellent sensitivity (100%) and specificity (100%) was achieved for differentiating AD from CN and AD from MCI. In addition, good sensitivity (100%) and specificity (70.0%) was obtained for separating MCI from CN. These results were provisionally validated using a smaller set of salivary samples in the VP. It is important to note that for the purpose of developing clinically robust biomarkers, the saliva samples were collected after at least one hour of fasting. To reduce the influence of earlier food intake in the biomarker discovery phase, we filtered metabolomics data by examining only those metabolites that were present in at least 50% of the samples. This procedure virtually assured that the commonly observed metabolites in different individuals were generated from endogenous sources.
By applying our technology to the present salivary samples, we were able to not only detect and quantify key metabolic changes for the AD versus CN, AD versus MCI, and (moderately) MCI versus CN comparisons but also to discover a number of significant metabolites, which could be used to discriminate these groups. Although some of the metabolites were positively identified, the identities of other significant metabolites remain as important challenges for future research. The current dansyl standard library did provide us with positive identities of 6 dipeptides that could be targeted in future neurobiological and clinical research. Among the significant metabolites we detected and identified, the biomarker panel discriminating AD from CN included Methylguanosine, Histidinyl-Phenylalanine, and Choline-cytidine, AD from MCI included Amino-dihydroxybenzene, Glucosylgalactosyl hydroxylysine - H2O, and Aminobutyric acid + H2, and MCI from CN included Glucosylgalactosyl hydroxylysine- H2O and Glutamine-carinitine. Following further validation in larger samples, we will carry out detailed structural identification work based on metabolite fractionation, MS/MS and perhaps nuclear magnetic resonance (NMR) spectroscopy, and synthesis of standard(s).
The application of metabolomics analyses to brain aging, impairment and dementia is in its early phases, with multiple approaches and sample sources being explored. In addition to early studies there is very little published work in salivary metabolomics of aging and AD to which to compare the present results. The available studies have used different technologies, biofluid samples, and national populations [28–31]. One recent study reported the efficacy of a saliva-based diagnostic metabolite panel for discriminating AD and controls [28]. Their panel, comprised of sphinganine-1-phosphate, ornithine, and phenyllactic acid discriminated these two clinically distant groups at a high level (AUC > 0.8). Phenyllactic acid is the metabolic product of phenylalanine. Our biomarker panels include dipeptides containing phenylalanine. Both of these studies implicate the possible dysregulation of phenylalanine-associated pathways in saliva from AD patients. Another recent study discussed the possibility of using NMR-based metabolomics to discover AD biomarkers in saliva samples [31]. In their diagnostic panel, histamine was selected due to its significantly increased concentration in the saliva of AD or vascular dementia patients compared with normal controls. It is worth noting that both histamine and urocanic acid, the metabolic biomarker for AD diagnosis found in our study, belong to the histamine biosynthesis pathway. Further research is needed to clarify the extent to which such panels can be used as non-invasive diagnostic biomarkers. On the other hand, the unique saliva-based metabolic biomarkers discovered in these complementary analytical platforms could be tested in combinations to develop a refined biomarker panel with enhanced sensitivity and specificity.
Compared to metabolite biomarkers of other biofluids, we note a recent plasma and CSF-based metabolomics study [19] where a small number of samples of the same groups (AD, MCI, and CN) were examined. Their results indicated that lysine metabolism in plasma and the Krebs cycle in CSF were among the strongest differences between MCI and CN whereas for AD versus CN, cholesterol and sphingolipids transport were significantly altered in both CSF and plasma. Direct comparison of our panel of metabolites with these reported is difficult, as we used the dansylation LC-MS method to target the analysis of the amine/phenol submetabolome and thus lipids were not analyzed. In the present study, we did not observe significant changes in metabolites related to lysine metabolism and the Krebs cycle. Nevertheless, we validated our metabolite panel to differentiate all three pairwise clinical groups with greater than 70% specificity and sensitivity and AUC > 0.8 in an independent sample. Conceivably, given the thousands of metabolites representing potential pathways of MCI and AD development— and the sensitivity of these biomarkers— different biomarkers (or sets of biomarkers) could be prominent across comparisons, depending on such factors as biofluid type, diagnosis, severity, technology, and group compositions. Future research comparing and contrasting results from multiple biofluids simultaneously will contribute to further advancements and applications [11, 56].
Among the limitations of the present study are the relatively small sample sizes, especially in the VP. For background, we note that other published non-salivary metabolomics research have both smaller [19, 57] and larger [28, 59] participant groups. In the recently emerging collection of salivary metabolomics work, other sample sizes are also both smaller [29, 30] and comparable or larger [28, 31]. We have several comments about this important limitation and our efforts to address it. First, not all previous studies included a validation analysis. For our study, the VP results were coordinated with both the larger DP and overall (combined sample) results. We view these observations as encouraging (but not definitive) to the validation interpretation. Second, a specific limitation for small sample sizes is possible over-fitting of models. We applied a strategy to mitigate this concern: We used univariate (volcano plot) and then multivariate (OPLS-DA) analyses to reduce the number of extracted biomarkers; the VP analyses were conducted with these data. Third, the present promising results will be useful in guiding our own and other future research. Specifically, our recent analytical development, universal metabolome standard (UMS) strategy [60], can be used to directly merge and compare the current study with future data sets generated with our CIL LC-MS analyses. The underlying rationale is that in CIL LC-MS-based metabolomics studies, the peak ratio of an individual metabolite in a given sample can be referenced in larger pooled samples. By finding out the concentration relation between the old and new pooled samples using UMS strategy, we can merge all the peak ratios of the two sample sets for combined analyses.
An additional limitation is that biomarker panel for discriminating the MCI and CN groups was moderately successful. One possible explanation for the modest effect for these two groups derives from our cohort sampling and classification procedures. In order to verify the MCI classification, we applied a rule that required all baseline MCI participants to be independently and objectively classified in status both at baseline and a 4-year longitudinal follow-up. This assured us that this group was comprised of relatively chronic (but not subject to either post-baseline reversion or precipitous decline), but it may also have ensured that this group is closer phenotypically to cognitively older adults. The latter CN group was also required to be stable in status at two consecutive waves. However, we also note that the sensitivity and specificity values for this discrimination were as good as similar values obtained with other biofluids. Another notable limitation is that although our three clinical groups were well-characterized and objectively classified, the cognitively impaired group was not divided into subgroups (e.g., amnestic MCI). In addition, the criterion was strictly applied but relatively low (<1 SD below the reference group mean for any domain). This group was, however, confirmed in status longitudinally, providing validation of classification. Further research with additional MCI groups and with subgroups converting to dementia (see [61]) would be useful. We note that we were not powered to compare metabolic differences between male and female participants. A growing number of studies have identified sex differences in AD vulnerability and mechanisms [62, 63]. We recommend further investigation of these issues in future metabolomics of neurodegeneration work. Finally, we recommend that all metabolomics studies examine high-performance biomarkers, as discovered in other investigations, so that these may be identified and validated in future and large-scale research (e.g., [11]).
In sum, an important goal of current biomarker research in brain aging and neurodegenerative disease is to discover new biomarkers that can be useful for 1) detecting early disease-related perturbations, 2) implicating pre-clinical mechanisms and pathways, and 3) functioning as sensitive outcome measures in pre-AD-related trials. We used saliva, a non-invasive and widely available biofluid, increasingly used to discover diagnostic and prognostic biomarkers of disease, including AD. Salivary samples from three groups were subjected to unbiased pairwise metabolomics analyses in both discovery and validation phases. Overall, our procedures produced excellent metabolome coverage, with a total of over 6,000 unique pairs or metabolites and an average of 3,669 peak pairs from each salivary sample. From this pool our statistical analyses converged on subsets of individual biomarkers selected for their ability to discriminate among the three clinical groups. Finally, using machine-learning statistical techniques we developed small (2- and 3-metabolite) biomarker panels that optimized the discrimination of the groups. Future research will include replication with new groups, application of both targeted and unbiased approaches, and integration across biomarker modalities. We expect that advances in salivary biomarker research in AD and other neurodegenerative diseases will provide opportunities for discovery and application research in diverse, if not worldwide, populations.
Footnotes
ACKNOWLEDGMENTS
This work was funded by grants from the (a) Canadian Institutes of Health Research (CIHR; to LL and RAD), (b) the Canada Research Chairs program (to LL and RAD), (c) Natural Sciences and Engineering Research Council of Canada, Genome Canada and Alberta Innovates (to LL), (d) the National Institutes of Health (National Institute on Aging, R01 AG 008235; to RAD), and (e) the Canadian Consortium on Neurodegeneration in Aging (with funding from CIHR and partners, including SANOFI-AVENTIS R&D; to RAD).
All research has been approved continuously by relevant institutional review boards. Certificates are available and on file in the University of Alberta Research Services Office and the US National Institutes of Health. All participants have completed and signed informed consent forms.
