|Home | About | Journals | Submit | Contact Us | Français|
Conceived and designed the experiments: RJP RC-S DMH JCM AMF RRT. Performed the experiments: RJP RC-S JPM AED PG RRT ARS. Analyzed the data: RJP RC-S JPM AED PG RRT ARS CMR DMH AMF. Contributed reagents/materials/analysis tools: RRT ERP GL DRG CMC JFQ JAK DMH JCM AMF. Wrote the paper: RJP RC-S JPM RRT. All authors revised the manuscript for important intellectual content and gave final approval of the version to be published.
Ideally, disease modifying therapies for Alzheimer disease (AD) will be applied during the ‘preclinical’ stage (pathology present with cognition intact) before severe neuronal damage occurs, or upon recognizing very mild cognitive impairment. Developing and judiciously administering such therapies will require biomarker panels to identify early AD pathology, classify disease stage, monitor pathological progression, and predict cognitive decline. To discover such biomarkers, we measured AD-associated changes in the cerebrospinal fluid (CSF) proteome.
CSF samples from individuals with mild AD (Clinical Dementia Rating [CDR] 1) (n=24) and cognitively normal controls (CDR 0) (n=24) were subjected to two-dimensional difference-in-gel electrophoresis. Within 119 differentially-abundant gel features, mass spectrometry (LC-MS/MS) identified 47 proteins. For validation, eleven proteins were re-evaluated by enzyme-linked immunosorbent assays (ELISA). Six of these assays (NrCAM, YKL-40, chromogranin A, carnosinase I, transthyretin, cystatin C) distinguished CDR 1 and CDR 0 groups and were subsequently applied (with tau, p-tau181 and Aβ42 ELISAs) to a larger independent cohort (n=292) that included individuals with very mild dementia (CDR 0.5). Receiver-operating characteristic curve analyses using stepwise logistic regression yielded optimal biomarker combinations to distinguish CDR 0 from CDR>0 (tau, YKL-40, NrCAM) and CDR 1 from CDR<1 (tau, chromogranin A, carnosinase I) with areas under the curve of 0.90 (0.85–0.94 95% confidence interval [CI]) and 0.88 (0.81–0.94 CI), respectively.
Four novel CSF biomarkers for AD (NrCAM, YKL-40, chromogranin A, carnosinase I) can improve the diagnostic accuracy of Aβ42 and tau. Together, these six markers describe six clinicopathological stages from cognitive normalcy to mild dementia, including stages defined by increased risk of cognitive decline. Such a panel might improve clinical trial efficiency by guiding subject enrollment and monitoring disease progression. Further studies will be required to validate this panel and evaluate its potential for distinguishing AD from other dementing conditions.
Clinicopathological studies suggest that Alzheimer's disease (AD) pathology (amyloid plaque formation, followed by gliosis and neurofibrillary tangle formation) begins 10–15 years before the onset of very mild dementia , . This period of ‘preclinical AD’ could provide an opportunity for disease modifying therapies to prevent or forestall the synaptic and neuronal losses associated with cognitive impairment –. However, before such interventions can be developed and judiciously administered, accurate tools must be in place to diagnose and monitor the pathophysiological condition of individuals with preclinical AD and very early stage AD dementia. Clinical examination cannot detect preclinical disease or measure cellular and molecular changes within the brain, and, in general, has limited accuracy when diagnosing the very earliest symptomatic stages of AD. Therefore, there is an urgent need to identify biomarkers that can do so. Because its composition is rapidly and directly influenced by the brain, the cerebrospinal fluid (CSF) proteome represents an appealing source for such biomarkers.
Indeed, a few CSF proteins have already shown promise as diagnostic biomarkers for clinical AD (dementia of the Alzheimer type [DAT]) and even preclinical AD. Lower mean levels of CSF Aβ42 and higher mean levels of tau and phosphorylated tau can distinguish groups with DAT from cognitively normal controls , . Unfortunately, value ranges for each biomarker show substantial overlap between groups.
Recently, using positron-emission tomography PET imaging with Pittsburgh Compound B (PIB) to measure brain amyloid in vivo, we and others have demonstrated that low CSF Aβ42 can serve as an indicator of amyloid deposition –, and that CSF tau levels correlate positively with in vivo brain amyloid load , . Importantly, both of these associations are independent of clinical diagnosis –, though CSF tau does correlate with more sensitive measures of cognition . These findings suggest that the overlap of biomarker values between clinical groups may, in part, reflect “contamination” of control groups by cognitively normal individuals exhibiting amyloid plaques and early neurodegeneration (preclinical AD), low CSF Aβ42 and elevated CSF tau. Supporting this notion, elevated ratios of tau/Aβ42 and p-tau181/Aβ42 (consistent with the presence of amyloid plaques and neurodegeneration) have been associated with increased risk of converting from cognitive normalcy to mild cognitive impairment or dementia , , and with increased rate of cognitive decline among those with very mild dementia . Together, these findings suggest that CSF biomarkers can describe neuropathological state and trajectory. They also suggest that a pathological staging system based on biomarkers might be a favorable alternative or adjunct to clinical staging for guiding treatment decisions or designing clinical trials.
Beyond amyloid plaque formation, other features of AD pathophysiology might also be exploited as therapeutic targets, sources of diagnostic biomarkers, or measures of disease progression. In addition to Aβ42 and tau, many other candidate AD biomarkers have been identified by either targeted or unbiased proteomics screens –. Only a few of these studies have tested large, well-characterized cohorts, however. Even fewer have evaluated biomarkers for their ability to distinguish the very early stages of AD pathophysiology. Thus, there remains a critical need for validated AD biomarkers that can properly categorize individuals by early pathological stage; such markers may have potential for monitoring neuropathological decline and, thereby, for evaluating response to disease-modifying therapies.
The goal of this study, therefore, is to identify such CSF protein biomarkers for AD using the unbiased proteomic technique of two-dimensional difference-in-gel electrophoresis (2D-DIGE) coupled with liquid chromatography and tandem mass spectrometry (LC-MS/MS), and to evaluate them further in a larger independent cohort using quantitative enzyme-linked immunosorbent assays (ELISA). Our findings suggest that a small ensemble of novel biomarkers may be able to distinguish several stages of cognitive decline in early AD, and improve the ability of current leading biomarkers tau and Aβ42 to discriminate early symptomatic AD from cognitive normalcy.
The study protocols were approved by the institutional review boards of the University of Washington, the Oregon Health and Science University, the University of Pennsylvania, the University of California San Diego, and Washington University. Written informed consent was obtained from all participants at enrollment. All aspects of this study were conducted according to the principles expressed in the Declaration of Helsinki.
Participants (n=48), community-dwelling volunteers from University of Washington [n=18], Oregon Health and Science University [n=11], University of Pennsylvania [n=11], and University of California San Diego [n=8], were 51–87 years of age and in good general health, having no other neurological, psychiatric, or major medical diagnoses that could contribute to dementia, nor use of exclusionary medications (e.g. anticoagulants) within 1–3 months of lumbar puncture (LP). Cognitive status was evaluated based on criteria from the National Institute of Neurological and Communicative Diseases and Stroke-Alzheimer's Disease and Related Disorders Association . In the morning after overnight fasting, CSF was obtained by LP, collected and aliquoted in polypropylene tubes, and immediately frozen at −80°C. Participants who were cognitively normal (Clinical Dementia Rating [CDR] of 0 [n=24]) , or had mild “probable AD” (CDR 1) (n=24), were selected from a larger group of 120 individuals on the basis of CSF Aβ42 (relatively high and low values, respectively), and, when possible, CSF tau (relatively low and high values, respectively) to increase the likelihood of CDR 1 participants having and CDR 0 participants not having AD pathology. CSF Aβ42 and tau levels for the discovery cohort were all measured in a single laboratory using well-established ELISA assays ( and Innotest, Innogenetics, Ghent, Belgium). Although quantitative thresholds were not defined prior to sample selection, the lowest CDR 0 value and the highest CDR 1 value for CSF Aβ42 in this ‘discovery cohort’ were 609 and 361 pg/mL, respectively; ranges for CSF tau were 141–461 pg/mL for CDR 0 and 215–1965 pg/mL for CDR 1.
Participants (n=292), community-dwelling volunteers enrolled at the Knight Alzheimer Disease Research Center at Washington University (WU-ADRC), were ≥60 years of age and met the same exclusion criteria as the discovery cohort. The study protocol was approved by the Human Studies Committee at Washington University, and written and verbal informed consent was obtained from participants at enrollment. Cognitive status was determined as with the discovery cohort. Participants who were cognitively normal (CDR 0, n=198), very mildly demented (CDR 0.5, n=65) or mildly demented (CDR 1, n=29) at the time of LP were selected without regard to previously measured biomarkers. Some CDR 0.5 participants met criteria for mild cognitive impairment (MCI) and some showed even milder impairment, and could be considered “pre-MCI” . All CDR 1 individuals had received a diagnosis of DAT (See Table 1 for demographic characteristics). Apolipoprotein E (APOE) genotypes were determined by the WU-ADRC Genetics Core. Fasted CSF (20–30 mL) was collected, gently mixed, centrifuged, aliquoted and frozen at −80°C in polypropylene tubes .
A pooled CSF sample, containing an equivalent volume from every ‘discovery’ cohort sample, was prepared as an internal standard for 2D-DIGE to facilitate the matching of gel features, and to allow normalization of the intensity of each gel feature among different gels. To enrich for proteins of low-abundance prior to 2D-DIGE, each CSF sample was depleted of six highly-abundant proteins (albumin, IgG, IgA, haptoglobin, transferrin, and α-1-antitrypsin) by immunoaffinity chromatography (Agilent Technologies, Palo Alto, CA) according to the manufacturer's instructions and as described previously . Depleted samples were then concentrated using 10 kDa exclusion filters to retain larger molecules. As a ‘benchmark’ of immunodepletion column performance, an aliquot of reference CSF was depleted after every group of seven experimental chromatographic depletions. Non-depleted reference CSF, depleted CSF and the proteins that were retained by the column were analyzed by 2D-DIGE as previously described , ; gel images obtained from all reference CSF depletion analyses were similar (data not shown), indicating consistent column performance over time.
2D-DIGE was performed as described previously , . Briefly, CDR 0 and CDR 1 samples were randomly paired. 50 micrograms of protein from each paired sample and from an aliquot of the pooled CSF sample were labeled with one of three N-hydroxysuccinimide cyanine dyes. The labeled proteins and 100 micrograms of unlabeled protein from each sample were mixed and equilibrated with an immobilized pH gradient strip for isoelectric focusing (first dimension), after which the strip was treated with reducing and alkylating solutions prior to SDS-PAGE (second dimension). Cy2, Cy3 and Cy5-labeled images were acquired on a Typhoon 9400 scanner (GE Healthcare, United Kingdom) at excitation/emission wavelengths of 488/520, 532/580, and 633/670 nm, respectively.
The comparative two-dimensional gel analysis was performed using an established experimental design  in which the high variation between gels is minimized by including a common, labeled pooled sample in all gels. Intra-gel feature detection, quantification and inter-gel matching and quantification were performed using the Differential In-Gel Analysis (DIA) and Biological Variation Analysis (BVA) modules of DeCyder software v 6.5 (GE Healthcare), respectively, as described previously . This process (DIA analysis) resulted in approximately 5,000 gel features per gel image. In five gels, one sample contained significant amounts of hemoglobin indicating possible blood contamination. Therefore, all images from gels with these hemoglobin-containing samples were removed from further analysis. Remaining gel images were separated into three sets: standard (pool of all samples), CDR 0 and CDR 1. The pooled sample image with the largest number of well-resolved gel features was chosen as a master image. Gel features in each remaining pooled sample image were hand matched to gel features in the master image. For each gel feature that was matched across >50% of the gels (n=764), a Student's t-test (α=0.05) was performed to determine the statistical significance of CDR 0/CDR 1 ratios, using the DeCyder EDA (Extended Data Analysis) module. To maximize discovery rate and minimize type II error, no multiple test correction was applied. The image intensity data for the statistically significant gel features (n=119) were then subjected to unsupervised hierarchical clustering (DeCyder EDA module).
Gel features with significant intensity differences were targeted by a robotic gel sampling system (ProPic; Genomics Solutions, Ann Arbor, MI) and transferred into 96 well plates for in-gel digestion with trypsin using a modification of a method  described previously . Aliquots of these digests were processed for and analyzed by LC-MS/MS using a capillary LC (Eksigent, Livermore CA) interfaced to a nano-LC-linear quadrupole ion trap Fourier transform ion cyclotron resonance mass spectrometer (nano-LC-FTMS)  QStar  or LTQ . The tandem spectra were searched against the National Center for Biotechnology Information non-redundant protein database NR (downloaded on 02-18-2007) using MASCOT, version 2.2.04 (Matrix Sciences, London). The database searches were constrained by allowing for trypsin cleavage (with up to two missed cleavage sites), fixed modifications (carbamidomethylation of Cys residues) and variable modifications (oxidation of Met residues and N-terminal pyroglutamate formation). Protein identifications were considered genuine if at least two peptides were matched with individual MASCOT ion scores ≥40.
Using nano-LC-MS/MS, multiple proteins were identified in the majority of individual gel features. The frequent observation of multiple proteins in single gel features was attributed to the sensitivity and greater peptide coverage that can be achieved with nano-LC-MS methods as compared to, for example, MALDI-MS analysis of peptides from gel features. Assignment of the major protein(s) from each gel feature was achieved using quantitative proteomics from spectra counting . The detection of multiple proteins within single gel features could also be attributed to artifacts and technical issues associated with 2D gel electrophoresis: 1) incomplete resolution of proteins by gel electrophoresis (due to similar charge and size characteristics, excessive abundance of neighboring proteins, or artifactual streaking); 2) changes in molecular weight associated with cyanine dye labeling, particularly for lower molecular weight proteins; and 3) sample ‘carryover’ during robotic gel sampling or during nano-LC-MS/MS.
All relevant proteomics data are detailed in Table S1.
CSF samples were analyzed by ELISA in duplicate for Aβ42, total tau, and phospho-tau181 (Innotest, Innogenetics, Ghent, Belgium) after one freeze-thaw cycle, and in triplicate for all other ELISAs after two freeze-thaw cycles. Samples were evaluated using commercially available ELISAs for NrCAM (R&D Systems Inc., Minneapolis, MN), YKL-40 (Quidel Corporation, San Diego, CA), apolipoprotein E (Medical and Biological Laboratories Company, Ltd., Nagoya, Japan), clusterin/apolipoprotein J (ALPCO Diagnostics, Salem, NH), pigment epithelium-derived factor (PEDF)/serpin-F1 (Chemicon International Inc./ Millipore Corporation, Billerica, MA), beta-2 microglobulin (ALPCO Diagnostics), ceruloplasmin (Assaypro, St. Charles, MO), chromogranin A (ALPCO Diagnostics, low binding capacity manufacturing protocol), transthyretin (Assaypro), and cystatin C (US Biological, Swampscott, MA), according to manufacturer's instructions, with adjustments for the analysis of CSF. A sandwich ELISA was developed for carnosinase I using goat anti-human carnosinase I antibody (2 µg/mL, R&D Systems Inc.) for capture, rabbit anti-human carnosinase I antibody (1 µg/mL, Sigma-Aldrich Corporation, St. Louis, MO) for detection, goat anti-rabbit:horseradish peroxidase (15000, Upstate Biologicals Inc./Millipore Corporation) for reporting, and TMB (3,3′,5,5′-tetramethylbenzidine) Super Slow (Sigma-Aldrich Corporation) for color development; recombinant carnosinase I (R&D Systems Inc.) was used as standard.
Statistical analyses were performed using commercially available software: SAS 9.2 (SAS Institute Inc., Cary, NC) for Receiver Operating Characteristic (ROC)/area under curve (AUC) calculations and logistic regression analyses, and SPSS 18 (SPSS Inc., Chicago, IL) for all other analyses.
Comparisons between CDR 0 and CDR 1 groups of the ‘discovery’ cohort (one sample was unavailable for re-evaluation, n=47) were performed using unpaired t-test. For the ‘validation’ cohort (n=292), correlations with age and gender were evaluated using the Spearman rho correlation coefficient (α=0.05). Chi-square analyses were performed to evaluate need for adjustment for observed correlations. Comparisons between the three CDR groups were performed using one-way analysis of variance (ANOVA), with Bonferroni and LSD post-hoc tests for pair-wise group comparisons, with the following exceptions: one-way ANOVA with Welch's correction was applied for markers (transthyretin) demonstrating unequal variances (Levene <.05); markers correlating with age (tau, p-tau181, Aβ42, YKL-40) were evaluated by analysis of covariance (ANCOVA) adjusting for age, followed by Bonferroni and LSD post-hoc tests. Multiple post-hoc tests were applied in recognition of their different levels of stringency (Bonferroni > LSD), and their non-uniform popularity among statisticians. For CDR 0 vs >0 comparisons and CDR 1 vs <1 comparisons, unpaired t-test was used; Welch's correction for unequal variances was applied for YKL-40, p-tau181, tau, and Aβ42. For each biomarker measured in the larger ‘validation’ cohort, the ROC curve and the AUC were calculated for predicting CDR 0 versus CDR>0. A stepwise logistic regression analysis was used to identify an optimal combination of these biomarkers for this data set. These analyses were repeated for CDR 1 vs CDR<1.
To identify new candidate biomarkers for AD, we utilized an unbiased proteomics approach, 2D-DIGE LC-MS/MS , , to compare the relative concentrations of CSF proteins in individuals with mild “probable AD” (CDR 1, n=24) to those in individuals with normal cognition (CDR 0, n=24). The two clinical groups were selected on the basis of relative biomarker values for CSF Aβ42 and tau (see Methods), and differed somewhat with respect to age at LP and gender (CDR 0: 64.8±8.8 yrs, 38% female; CDR 1: 72.8 yrs ±7.9 yrs, 54% female). Five samples showed evidence of blood contamination by 2D-DIGE; the five gels containing these samples were excluded from subsequent image analyses. The remaining individual sample images (n=38, from 19 gels) were aligned using the BVA module (described under Methods).
Among the 764 gel features that were present in >50% of the gels, 119 were found to have significant intensity differences between CDR 0 and CDR 1 groups (Student's t-test [α=0.05]) (Figure 1). The image intensity data for these 119 gel features were subjected to unsupervised hierarchical clustering (EDA module, DeCyder software) and the gel features themselves were analyzed for protein composition.
LC-MS/MS identified single dominant proteins in 78 of the 119 gel features (Table 2). In 29 gel features, our analyses identified two or more co-dominant proteins. The 12 remaining gel features were not annotated from the nano-LC-MS/MS data. Among the characterized gel features, there was considerable redundancy in protein identifications, with some proteins appearing in multiple gel features. Such ‘redundant’ gel features, likely representing a modified form or variant of the same ‘parent’ protein, generally migrated with some proximity on 2D-gel electrophoresis (Figure 1). Forty-seven unique proteins were identified (Table 2). Thirteen of these unique proteins had been identified in our previous studies ,  (including chromogranin B, cystatin C, prostaglandin H2 D-isomerase/beta trace, neuronal pentraxin receptor, gelsolin, beta-2 microglobulin, carnosinase I, angiotensinogen, apolipoprotein H, secretogranin III, alpha-1-antichymotrypsin, chitinase 3-like 1/YKL-40, and kininogen I) and others have been reported by other groups , , , , , . These previous reports provide supporting evidence that this list of proteins may contain viable candidate biomarkers for AD that are worthy of pursuit in validation experiments.
The intensity data from the 119 gel features of interest were subjected to an unsupervised clustering analysis to evaluate their ability to segregate the CDR 0 and CDR 1 samples, and to assess their collective potential as a diagnostic biomarker panel (Figure 2). The ‘heatmap’ generated from this analysis appeared to segregate CDR 0 and CDR 1 individuals (indicated by green and red ovals, respectively) almost completely, with only four participants ‘misclassified.’ However, closer examination revealed an additional layer of segregation on the basis of APOE genotype (indicated by ‘ApoE 4+ Cluster’ and ‘ApoE 4 – Cluster’) which showed perfect sample segregation. Given that the APOE-ε4 allele is a dominant genetic risk factor for AD, some clustering of individuals by APOE genotype might be expected simply from successful segregation of CDR 0 and CDR 1 individuals. However, we hypothesize that the apoE protein exerts a dominant clustering influence through the markedly different electrophoretic profiles of its different isoforms derived from APOE-ε2, APOE-ε3 and APOE- ε4 alleles (illustrated in Figure S1). ApoE was present in 24 of the 119 gel features found to differ in intensity between the CDR groups, and was found to be the primary protein in 12 of these gel features. This heterogeneous electrophoretic mobility of apoE results from the inherent charge differences of the three major apoE isoforms (-E2, -E3, -E4) and the appearance of each isoform as an array of multiple distinct gel features caused by post-translational modifications. These isoform-specific differences are reflected in the prominent red and green clusters, located within the lower third of Figure 2 (corresponding to gel features 83–90, 107–117, and 119), that correlate very closely with participant APOE genotypes. Recognizing this correlation, we hypothesized that APOE genotypes were in large part driving the clustering of participant samples in Figure 2. To test this hypothesis, we performed a second unsupervised clustering analysis, including only those gel features from the initial analysis that did not contain apoE protein (Figure 3). Although this ‘apoE-free’ analysis segregated CDR 1 and CDR 0 groups less completely, it appropriately re-clustered (by CDR status) several samples (#12, 36, 37) that were aberrantly segregated in Figure 2, potentially due to their APOE genotypes. Moreover, clustering of participant samples into APOE genotype subgroups in Figure 3 appears negligible. The underlying benefit of this ‘apoE-free’ analysis is that it reveals the sample-clustering potential of other gel features, which was previously obscured by the inclusion of apoE-containing gel features. As can now be better visualized in Figure 3, gel features appearing within the upper three-fourths of the heatmap appear to show greater intensity in CDR 1 samples; the converse is true of gel features within the lower fourth. It is important to note that measurements of Aβ42 and tau (two proteins measured by ELISA and not detected by 2D-DIGE) were not included in these clustering analyses; because these ‘discovery’ samples were selected for this study on the basis of CSF Aβ42 and tau levels, such inclusion would presumably yield perfect or near-perfect segregation by CDR status in this ‘discovery’ cohort. Therefore, this analysis reflects the potential of these candidate biomarkers to segregate CDR 0 and CDR 1 individuals independent of any contribution from current leading CSF biomarkers Aβ42 and tau. It does not address whether these biomarker candidates might improve upon the utility of Aβ42 and tau, however.
Before evaluating a subset of these candidate biomarkers in a larger independent sample set, we first assessed the capacity of protein-specific quantitative ELISAs to detect significant differences between the CDR 0 and CDR 1 groups of the original ‘discovery’ cohort. When possible, to facilitate future reproduction of our findings by other groups and potential translation to clinical use, we applied commercially available ELISA kits.
Of the eleven ELISAs applied to the ‘discovery’ cohort (n=47, one sample was unavailable for validation), six (NrCAM, YKL-40, chromogranin A, carnosinase I, transthyretin, cystatin C) showed statistically significant or near-significant differences between CDR 0 and CDR 1 groups (Figure 4); five others (PEDF, beta-2 microglobulin, clusterin/apoJ, ceruloplasmin, apoE) did not.
The six ELISAs that measured differences between the CDR 0 and CDR 1 CSF samples of the ‘discovery’ cohort were subsequently applied to a larger, independent set of CSF samples (n=292) collected from volunteer participants studied by the WU-ADRC. This ‘validation’ cohort included a CDR 0.5 group in addition to CDR 0 and CDR 1 groups, allowing for biomarker assessment in the very early clinical stage of AD. Demographic, clinical, and genetic characteristics of these individuals at time of sample collection are presented in Table 1. Unlike the ‘discovery’ cohort, this ‘validation’ cohort was not preselected on the basis of prior biomarker values (CSF Aβ42 and tau), although assays for CSF Aβ42, tau and p-tau181 were performed.
Because the age and gender compositions differed among the clinical groups of the ‘validation cohort,’ we evaluated each of these 9 biomarkers (six novel candidates, Aβ42, tau, and p-tau181) for age and gender correlations in order to apply covariate analyses appropriately. Correlating with age were tau (r=0.318, p<0.0001), p-tau181 (r=0.2216, p<0.001), Aβ42 (r=−0.2334, p<0.0001) and YKL-40 (r=0.4001, p<0.001); no biomarkers correlated with gender (p>0.05).
As shown in Figure 5, statistically significant differences between clinically defined groups were measured for Aβ42, tau, p-tau181, NrCAM, YKL-40, chromogranin A, and carnosinase I; for transthyretin and cystatin C, non-significant trends were measured. These differences appeared in three patterns: Aβ42 showed a pronounced decrease from CDR 0 to CDR 0.5 and a lesser reduction from CDR 0.5 to CDR 1; tau, p-tau181, and YKL-40 showed increases that were equivalent in CDR 0.5 and CDR 1 relative to CDR 0; NrCAM, chromogranin A, and carnosinase I showed decreases relative to CDR 0 only in CDR 1, and not in CDR 0.5.
To evaluate and compare the potential of the validated candidate biomarkers and Aβ42, tau, and p-tau181 for identifying either very mild to mild dementia (combined CDR 0.5 and CDR 1) or mild dementia (CDR 1), ROC curves and AUCs were calculated for each biomarker using data from the ‘validation’ cohort (Figure 6A, B, Tables 3, ,4).4). Stepwise logistic regression analyses indicated that, among the nine biomarkers under consideration, YKL-40, NrCAM and tau yielded the highest AUC (0.896) in discriminating cognitive normalcy (CDR 0) from very mild to mild dementia (CDR>0) (Figure 6C, Table 3); for discriminating mild dementia (CDR 1) from CDR<1, carnosinase I, chromogranin A and tau yielded the highest AUC (0.876) (Figure 6D, Table 4).
Using an unbiased proteomics approach (2D-DIGE LC-MS/MS), this study identified 47 novel candidate CSF protein biomarkers for early AD. Subsequently, by evaluating a subset of these candidate biomarkers by ELISA, this study validated the utility of four candidate biomarkers for distinguishing groups with mild, very mild, or no dementia (CDR 1, 0.5, 0, respectively). Further statistical analyses demonstrated that these biomarkers could improve the accuracy of ‘established’ biomarkers Aβ42 and tau for the diagnosis of early AD.
The results from the 2D-DIGE LC-MS/MS portion of this study suggest that many of the recognized neuropathological changes of AD are represented by changes in the CSF proteome. Most of the 47 candidate biomarker proteins identified in this study can be placed into structural and/or functional categories (e.g. synaptic adhesion, synaptic function, dense core synaptic vesicle proteins, inflammation/complement, protease activity/inhibition, apolipoproteins, etc.) associated with accepted neuropathophysiological changes in AD (Table 5). Unsupervised clustering analyses of these 2D-DIGE data, performed without the influence of CSF Aβ42, tau, p-tau181 and APOE genotype, additionally suggest that these biomarker candidates collectively show utility for discriminating groups with and without mild DAT (Figure 3).
In the second phase of this study, designed to measure a subset of candidate biomarker proteins in two independent sample sets by ELISA, four of the eleven candidate biomarkers that were tested showed capacity to distinguish clinical groups. However, seven candidate biomarkers did not show statistically significant differences between clinical groups in either the smaller ‘discovery’ cohort or the larger ‘validation’ cohort. Superficially, this ‘failure rate’ might cast doubt on the list of candidate biomarkers identified through 2D-DIGE. However, it is important to note that 2D-DIGE is sensitive to changes in concentrations of minor protein isoforms and post-translational modifications that may not significantly alter the global concentrations of a ‘parent’ protein, which would be measured by ELISA. Therefore, it is not surprising that some of the candidate biomarker ELISAs did not replicate the findings from 2D-DIGE. Transthyretin provides a prime example: all of the significant gel-features ascribed to transthyretin (gel features # 20, 52, 57, 58, 60, 77, 78, 79, 84, 87, 110, 115; Table 2) showed unusual electrophoretic patterns and were dwarfed by the canonical transthyretin gel features that did not individually show statistical differences (Figure 1). In fact, whereas most of the significant transthyretin 2D-DIGE gel features were decreased in AD, the global transthyretin levels measured by ELISA in the ‘discovery’ and ‘validation’ cohorts were actually mildly increased in groups with cognitive impairment (CDR>0) relative to those without (CDR 0) (Figures 4 and and5).5). To measure the sub-species of transthyretin that were identified by 2D-DIGE as decreasing in AD will require assays that specifically target relevant post-translational modifications and exclude other forms of transthyretin. Similarly, other 2D-DIGE biomarker candidates may also require specifically tailored assays for accurate, high-throughput measurement.
Nevertheless, four candidate biomarkers were successfully validated in both cohorts, and two others showed non-significant trends by ELISA in the larger ‘validation’ cohort (Figure 5). This larger cohort represented three different cognitive stages: normalcy, very mild dementia, and mild dementia (CDR 0, CDR 0.5, CDR 1, respectively), and revealed different patterns of CSF biomarker levels, vis-a-vis cognitive status. The CSF concentration of YKL-40, an astrocytic marker of plaque-associated neuroinflammation –, is increased by the very earliest stage of clinical disease (CDR 0.5). Transthyretin , , , , – and cystatin C , , –, two proteins with neuroprotective qualities that are implicated in preventing amyloidogenesis of Aβ peptide, show a similar pattern. In contrast, the concentrations of NrCAM, a synaptic adhesion molecule , –, chromogranin A, a dense core synaptic vesicle protein , , , –, and carnosinase I, a neuronal dipeptidase responsible for degradation of the anti-oxidant and metal-chelating dipeptide carnosine , – do not decline until mild dementia ensues (CDR 1).
Like the current leading CSF biomarkers for AD (Aβ42, tau and p-tau181), all of these biomarker candidates show ranges with substantial overlap between clinically defined groups. This issue of overlapping values, common among candidate AD CSF biomarkers reported to date, suggests that any one biomarker will be insufficient to accurately identify early AD, and that an ensemble of complementary biomarkers will be required to provide adequate sensitivity and specificity. Therefore, to identify an optimal combination of these biomarkers that can distinguish the early clinical stages of AD from cognitive normalcy, we applied stepwise logistic regression analyses to the ELISA data from our ‘validation’ cohort (Figure 6, Tables 3 and and4).4). These analyses suggest that four candidate AD biomarkers (YKL-40, NrCAM, chromogranin A, carnosinase I) can improve the ability of tau to classify individuals into CDR 0, CDR 0.5 and CDR 1 groups with appreciable accuracy.
It may appear counter-intuitive that Aβ42 and p-tau181, which individually discriminate very mild AD and mild AD from cognitively normal groups quite well, were not incorporated into either ‘optimal’ biomarker panel by the stepwise logistic regression analyses. Likewise, NrCAM was included in the optimal CDR 0 vs CDR>0 biomarker panel (AUC 0.896) even though its mean levels did not independently show a statistical difference between CDR 0 and CDR>0 groups. In considering this outcome, it may be worth noting that if NrCAM, transthyretin, chromogranin and cystatin C are removed from consideration, the stepwise logistic regression model for the CDR 0 vs CDR>0 comparison yields an ‘optimal’ biomarker panel that includes only tau, Aβ42 and carnosinase I, with an AUC of 0.849 (not shown). In this restricted analysis, the paired contribution of Aβ42 and carnosinase I to tau is apparently greater than that of YKL-40. These analyses illustrate how ‘unpredictable’ and context-dependent optimal biomarker combinations can be, and suggest that biomarker complementarity may be more important to consider than each biomarker's independent performance, when choosing a biomarker panel. Of course, it will be necessary to replicate these findings in additional independent cohorts. It will also be essential to evaluate a greater number of candidate biomarkers in similar fashion, in order to construct a biomarker panel with even greater accuracy.
Another worthwhile feature to consider when evaluating and selecting CSF biomarkers is relative concentration in the blood (plasma, serum), because biomarker measurements in CSF can be artifactually influenced by subtle blood contamination at the time of lumbar puncture; from this perspective, ideal CSF biomarkers show CSF concentrations that are equal to or greater than those in blood. An additional reason to assess plasma/serum concentrations of candidate CSF biomarkers is to determine if venipuncture, which is more easily performed than lumbar puncture, might yield equivalent information. Among the six CSF biomarkers identified by stepwise logistic regression analysis in the current study, Aβ42 and tau –, YKL-40 , and chromogranin A  show higher levels in CSF than in plasma; carnosinase I levels appear similar in CSF and serum ; NrCAM levels appear higher in serum than in CSF, although the forms of NrCAM present in these fluids may differ . Concerning independent utility as biomarkers for AD, only plasma YKL-40 and serum NrCAM have shown promise , , albeit inferior to that of CSF YKL-40 and NrCAM demonstrated here. Plasma tau concentrations in AD and controls are below the level of detection of the most commonly used tau assays, and plasma Aβ42 – and plasma chromogranin A (R.Perrin et al., unpublished data) concentrations show no significant differences among CDR groups. Serum carnosinase activity likewise has not shown significant differences between AD and controls in one small study , though a difference between AD and mixed dementia (including vascular dementia) has been reported . To our knowledge, an evaluation of plasma or serum carnosinase I concentrations in the context of AD has not yet been performed or reported. Further assessment of the potential of these and other proteins as candidate AD biomarkers in plasma or serum, complete with evaluation of their performance as ensembles, remains an important task for future studies. Currently, however, this panel of six biomarkers appears likely to show much greater promise in its application to CSF.
Indeed, by providing proof of concept, this study outlines a scheme to categorize the early stages of AD using CSF protein biomarkers that reflect established features of the pathophysiological evolution of the disease (Figure 7). Building upon previous findings that low CSF Aβ42 can identify cognitively normal individuals with plaques (preclinical AD) , , and that tau/Aβ42 and YKL-40/Aβ42 ratios can predict risk of developing cognitive impairment , , , this minimal panel of six CSF biomarkers (YKL-40, NrCAM, chromogranin A, carnosinase I, tau and Aβ42) begins to segregate individuals into six clinicopathological categories: normal cognition without amyloid plaques, normal cognition with amyloid plaques (preclinical AD), normal cognition at increased risk to develop dementia (converters), very mild dementia (CDR 0.5), very mild dementia at increased risk for progression, and mild dementia (CDR 1) (Figure 7).
We acknowledge that this minimal panel of biomarkers currently has insufficient sensitivity and specificity for clinical application, particularly because it has not been fully evaluated for its ability to discriminate AD from non-AD causes of dementia (although Aβ42, p-tau181, tau, and specific fragments of chromogranin A and cystatin C have shown some ability to distinguish AD from frontotemporal lobar degeneration [FTLD]) , , . The incorporation of additional biomarkers that are likely to discriminate early AD from cognitive normalcy, such as those identified in the first phase of this study, or other biomarkers that have already shown promise for distinguishing AD from other leading causes of dementia (e.g. agouti related peptide, eotaxin-3, and hepatocyte growth factor , complement C3a des-arg and integral membrane protein 2B CT , for FTLDs; and alpha-synuclein , apoH and vitamin D binding protein  for Lewy body disorders), would likely improve the panel's diagnostic utility. However, even in its current form, this initial panel might show value if applied in the context of clinical trial design, wherein simple enrichment of study populations for characteristics of interest would increase efficiency and power and reduce duration and cost. A biomarker panel like this one might also allow clinical trials to evaluate stage-specific responses to treatment, which may differ. Finally, because most of these biomarkers reflect underlying pathological changes in real time, it is appealing to speculate that these biomarkers may have additional utility for evaluating clinically imperceptible treatment responses (as in ) and for monitoring neuropathological – rather than cognitive – decline.
ApoE protein isoforms appear in different gel features on 2D-DIGE. Overlays of fluorescent 2D-DIGE images from gels representing CSF from two individuals with homozygosity for APOE-ε2 (green) or APOE-ε3 (red) (panel A) and for APOE-ε3 (green) or APOE-ε4 (red) (panel B) illustrate the heterogeneity of signal distribution by isoelectric point and molecular weight among apoE protein isoforms derived from different alleles. In panels C, D, E, F, G, H, signal intensities of individual CSF samples, grouped by genotype (2/2, 3/3 and 4/4 represent homozygotes; 2/3, 3/4 represent heterozygotes) are indicated for six apoE gel features (labeled C, D, E, F, G, H in panels A and B), illustrating that gel features C and D represent apoE2; gel feature E represents multiple forms; gel feature F represents apoE3; and gel features G and H, apoE4.
Mass spectrometry and protein identification data for 2D-DIGE gel features that differ in AD CSF. Results are ordered sequentially by “heat map #” [column A], corresponding to the ‘heat map’ row numbers in Figure 2. “Spot” [column B] refers to BVA number (see Methods). “(Accession) primary protein name” [column C] provides the gi number and protein name from the NCBI database. “Protein molecular weight” [column D] is the gene product molecular weight in Daltons. “Protein score” [column E] is the MASCOT-generated protein score. “Protein ID probability” [column F] indicates Scaffold's percent probability that the protein identification is correct. “Spectral count” [column G] is the number of spectra assigned to the protein by Scaffold. “Unique proteins” [column H] refers to the number of recognized tryptic peptides attributed to the protein by MASCOT. “Peptide sequence” [column I] indicates the amino acid sequence of the tryptic peptide predicted by MASCOT. “MASCOT ion score” [column J] is the MASCOT quality assessment of the peptide sequence assignment. “M/Z (observed)” [column K] is mass/charge ratio. “Mass (observed)” [column L] of peptide is indicated in Daltons. “Mass (theoretical)” [column M] is idealized tryptic peptide mass as predicted by NCBI. “Mass error (ppm)” [column N] is the error in parts per million determined through comparison of theoretical peptide mass to data generated by mass spectrometry. “MS source” [column O] reflects the mass spectrometer that produced the observed data (Q-STAR or LTQ-FT). “Modifications” [column P] lists variable post-translational modifications identified by mass spectrometry peptide sequence analysis.
The authors would like to express our appreciation to the Biomarker Core, Clinical Core, Data Management and Statistics Core, Genetics Core, lumbar puncture physicians, and volunteer participants of the Knight ADRC of Washington University in St. Louis, and to the volunteer participants of the University of Washington, the Oregon Health and Science University, the University of Pennsylvania, and the University of California San Diego.
Competing Interests: D.M.H. co-founded C2N Diagnostics, is on the scientific advisory board of C2N Diagnostics, En Vivo, and Satori, consulted for Pfizer, and receives grants that did not support this work from Eli Lilly, Pfizer, and Astra-Zeneca. J.C.M. has consulted for Astra Zeneca, Genentech, and Merck, has received honoraria from ANA Soriano lecture, payment from Journal Watch for preparation of other manuscripts, royalties from Blackwell Publishers and Taylor and Francis, and travel funding for ANA meeting Baltimore. C.M.C. became emeritus at University of Pennsylvania on January 1, 2010, and now works for Avid Radiopharmaceuticals, which did not provide any funding for this study, and did not have any involvement or influence in data production, data analysis, decision to publish, or manuscript preparation. None of the above stated competing interests alter our adherence to all the PLoS ONE policies on sharing data and materials.
Funding: National Alzheimer's Coordinating Center (NACC) Grant U01 AG16976 (A.M.F.); P50 AG05681 (C.M.R., E.A.G.); P01 AG026276 (C.M.R., E.A.G.); P50 AG05136; P01 AG03991(J.C.M.); P30 NS057105 (D.M.H.); UL1 RR024992 (R.R.T.; C.M.R.); W.M. Keck Foundation (R.R.T.); P41 RR00954 (R.R.T); Dept. of Veterans Affairs (E.R.P.); AG020020 (G.L.); AG23185 (D.R.G.); AG08017 (J.F.Q.); AG08017 (J.A.K.); AG10124 (C.M.C); T32 NS007205 (R.J.P.); and the Charles and Joanne Knight Alzheimer Research Initiative (A.M.F.); Avid Radiopharmaceuticals (C.M.C., employment unrelated to this study). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.