Alzheimer’s disease (AD) is the most common form of dementia and the 6^th leading cause of death in America [1]. The pathological hallmarks of AD are extracellular plaques composed of amyloid beta (Aβ) and neurofibrillary tangles composed of hyperphosphorylated tau (p-tau). Reductions in cerebrospinal fluid (CSF) Aβ₄₂ and increases of CSF tau and p-tau₁₈₁ are observed in AD patients in comparison with aged, cognitively-normal individuals, and these changes in CSF correlate with the pathological hallmarks of AD [2, 3]. The change in levels of Aβ₄₂ and tau begin 1–2 decades prior to onset of symptoms, with the change in Aβ₄₂ preceding that of tau [4, 5].

CSF Aβ₄₂, tau and p-tau₁₈₁ are routinely used in AD research and drug development. However, while the rank order performance (relative quantitation) of these biomarkers has been consistent, the inter-laboratory variability in the absolute value has been high [6–8]. In addition to variable performance, there is significant overlap between levels in controls and AD patients. Further, none of the three established CSF biomarkers have been directly linked to efficacy in a clinical trial. For example, while CSF levels of tau and p-tau₁₈₁ levels did trend toward decline after treatment with bapineuzumab, no correlation with cognitive benefit was observed (reviewed in [9]).

There is a need to develop additional biomarkers that can aid in the diagnosis and treatment of AD. Many proteomic studies have identified potential AD biomarkers (reviewed by [10, 11]), generating a substantial list of candidates that need to be evaluated for their utility. However, follow-up studies focusing on individual candidates are limited and usually focused on markers for which existing immunoassays are available. The development time for a new immunoassay can be significant and relies on successful identification of specific and high affinity antibodies. Targeted-proteomics is an emerging tool to accelerate hypothesis driven biomarker assay development by enabling the quantitative assessment of potentially hundreds of biomarkers simultaneously (reviewed in [12–14]).

Candidate AD biomarkers in CSF were selected based upon the consistency of their identification in proteomic discovery experiments [15–23]. These experiments used many different proteomic approaches, including differences in sample preparation (in-gel vs. in-solution digestion), the mass spectrometer (MS) used for analysis (MALD-ToF vs. FT-ICR), and the quantitation technique (label-free vs. iTRAQ). The differences in the proteomic techniques may account for some discrepancies in the trend of change observed for certain markers, but more importantly highlight the need for targeted-follow-up of candidate biomarkers within a single sample set. Independent of the proteomic experiments, additional candidate biomarkers were selected for evaluation based on their known relationship with AD (e.g. apolipoprotein E4) [24], or their emerging promise as disease progression markers (ex. Visinin-like-protein 1) [25, 26].

A multiplexed, absolute quantitative LC-MS/MS multiple reaction monitoring (MRM) assay was developed to quantitate candidate biomarker-specific peptides in CSF samples from cognitively-normal, aged control, mildly-cognitively impaired (MCI) and AD patients. We evaluated the diagnostic utility of all the protein-specific peptides in comparison with the published proteomic literature and also compared with the classic AD CSF biomarkers Aβ₄₂, tau and p-tau. The longitudinal performance of each biomarker (sampled serially from the same patient) was established, with the majority of peptides demonstrating stability over the course of a year. Four candidates decreased over time in AD patients. Results from this targeted-proteomic assay narrow the list of candidates for more rigorous follow-up and increase the rate of development of novel clinical biomarkers for AD.

Using a multiplexed, targeted-proteomics approach we evaluated candidate biomarkers for AD that were previously reported in discovery proteomic experiments to change with disease. At least one surrogate peptide was selected for each protein, and 4 specific transitions were selected for each peptide. Stable-isotope labeled peptides (heavy) were used as internal standards for every endogenous peptide (light), and the ratio of the peptide pair (light to heavy) was used for absolute quantitation. Using CH3L1 as an example (Figure 1D), absolute quantitation by this approach yields results similar to that of traditional immunoassays [22, 23, 44, 65].

One of the major benefits of our approach, especially when working in a less-complex matrix like CSF, is that sample preparation does not require sample enrichment approaches. For example, sample enrichment using antibodies, even in the case of depleting the most abundant proteins, such as albumin, can add bias due to co-depletion of proteins of interest [66]. However, enrichment strategies are usually needed when additional sensitivity is required. In this assay, the concentration range in CSF for the 28 proteins detected was 28 ng/mL to 13.8 μg/mL, which is consistent with the expected dynamic range of current instruments. Visinin like protein 1 (VILIP-1) was below the level of detection for all three peptides that were monitored; targeted-enrichment strategies or development of a sensitive immunoassay will be required to evaluate this candidate biomarker.

While not a primary focus of our study due to our small group sizes, we did evaluate the diagnostic performance of all the candidate biomarkers in comparison with the classic AD CSF biomarkers. The directionality and fold-difference observed between diagnostic groups was as expected for Aβ₄₂, tau and p-tau in our sample set (Figure 1, Table 3); this result gave us confidence that these samples are representative of their respective diagnostic groups and that they could be used to evaluate the performance of the candidate biomarkers. After correction for multiple comparisons testing, a number of peptides reached significance in differentiating the diagnostic groups (Table 3). However, due to the small group size, replication with more individuals needs to be performed, especially for the MCI group comparisons. The observed modest differences between groups (0.6-2 fold-differences) are comparable with what others have reported; proteomic studies of larger sample sets have demonstrated that the majority of candidates change by less than 2-fold in AD [16, 22]. Extending from sporadic AD to familial AD, Ringman et al. also reported that complement component C3 (CO3) increased in disease by 1.3 [34], in comparison with our observed changes of 1.4. Also consistent with previous reports, the traditionally used biomarkers Aβ₄₂, tau and p-tau₁₈₁ were the most significant and specific markers differentiating AD patients from control. The other significant biomarkers detected lack specificity; they have also been reported to change in other neurodegenerative diseases including Parkinson’s disease and multiple sclerosis [15, 16, 67]. The biomarkers evaluated in this study may have greater utility for disease-staging rather than diagnostics [22] or for monitoring therapeutic response, especially if used in combination; if multiple molecules from the same pathway change in response to treatment, it would give one higher confidence of a therapeutic effect.

It is important to characterize the longitudinal stability of a candidate biomarker prior to use in the clinic. Multiple studies have demonstrated that CSF Aβ₄₂ levels stabilize and plateau in established AD [59–63]. Similarly CSF tau and p-tau appear to demonstrate longitudinal stability [68], though some studies have noted a slight continued longitudinal increase in tau [60, 63]. In our longitudinal assessment of the classic AD biomarkers, Aβ₄₂ and tau levels were stable over the course of a year in our AD patients (Figure 2), consistent with previous reports. Similar to Aβ₄₂ and tau, the majority of peptides were stable over time (11–16 months) in AD patients (Figure 3). There were four peptides (A4_117, CMGA_322, NPTXR_22, and NRCAM_806) that declined significantly by approximately 10% per year in the AD patients but not in aged control or MCI patients (Figure 4). P-tau trended toward decline over time in 14 of 24 patients (Figure 2). A decline in p-tau (2 pg/mL/year) has been observed previously in sporadic AD patients [69], and more recently in autosomal-dominant AD patients [70]. Fagan et al. also showed that levels of VILIP-1 decreased after the onset of disease [70], suggesting that at later stages of disease a decline in CSF biomarkers, like VILIP-1, are reflective of neuronal loss. The four markers identified in this study are also neuronal markers, and are potential progression markers in Alzheimer’s patients that will be further evaluated in additional longitudinal sample sets.

Interestingly, the potential progression markers CMGA, NrCAM and NPTXR were also highly correlated (Figure 5). The fourth candidate progression marker amyloid precursor protein (APP) peptide A4_117 was also correlated with the three other peptides, however, the Spearman rank correlation coefficient was below our |R > 0.8| cut-off (A4_117 correlations: NrCAM R = 0.65, CMGA R = 0.62, NPTXR R = 0.6). Another APP peptide quantitated in this study, A4_688, did not demonstrate significant longitudinal change. The lack of strong correlation between the two APP peptides, A4_117 and A4_688 (Spearman rank R = 0.42), may be a consequence of differential CSF kinetics of APP cleavage fragments (Figure 6). Based on the full-length sequence of APP A4770, A4_117 corresponds to amino acids (a.a.) 117–132, which is present in soluble APP (sAPP) α (a.a.18-687) and sAPPβ (a.a.18-671) as well as N-APP (a.a.18-286). In contrast, A4_688 corresponds to a.a. 688–699 from full-length APP as well as a.a. 17–28 within Aβ [71]. The differential behavior of these peptides demonstrates the importance of distinguishing peptide-level quantitation from total protein quantitation, and caution against algorithms that automatically combine peptide quantitation into a single protein quantitation.

The correlations were remarkably high for some of these biomarkers and possible analytical factors were evaluated. The majority of potential analytical artifacts were ruled out; for example, peptide properties including retention time and m/z were different in each case. To support the hypothesis that the correlations are not an artifact of the detection method, NrCAM and CMGA levels were analyzed using ELISAs. A similar rank order was observed in the MS assay and ELISA (Figure 5D). Interestingly, the absolute levels measured by the NrCAM ELISA were significantly lower than that measured by the MS assay. The difference in absolute value between the NrCAM ELISA and MS result may be due to the fact that NrCAM exists as multiple isoforms in CSF. The antibody used in the ELISA may not detect all the isoforms; however, the peptide used for the MS assay is present in all 5 NrCAM isoforms, which would account for the higher value observed by MS. Although the absolute values were different from the two methods used, the fact that the rank orders are maintained support the hypothesis that the high correlation for some biomarkers reflects a common biological process. CMGA, NPTXR, NrCAM, and A4_117 and A4_688 are all produced in neurons, and changes in these peptides may be reflective of neurodegeneration. NrCAM is an adhesion molecule present on synapses [45, 72]. CMGA is a prohormone that is packaged in dense-core synaptic vesicles [46, 73, 74]. NPTXR and APP are both integral-membrane proteins. NPTXR has been implicated in synaptic plasticity [75, 76], and APP is known to be critical to synapse formation and function [77]. Given the diverse biological roles of these proteins, the observed correlations suggest that coordinated changes in CSF may reflect pathologies beyond Aβ and tau (e.g. synaptic function or secretion mechanisms). Interestingly, CMGA, NPTXR, and NrCAM are also positively correlated with tau within our AD samples (R = 0.69, 0.71, 0.74, respectively, Additional file 1: Table S2), although not as strongly as they are correlated with each other. VILIP-1, a potential marker of neurodegeneration, is also positively correlated with tau [26]. Future studies are needed to increase our understanding of the underlying cause of these correlations, the spectrum of other CSF proteins that are correlated with this initial cluster, and the longitudinal performance and utility of the correlations.

It is now clear that Alzheimer’s disease pathology begins years before the onset of clinical symptoms, and therapeutic trials are beginning to focus on treating patients prior to the onset of dementia. Thus, the need for biomarkers that can increase the accuracy of early diagnosis and reflect the success of treatment is critical for AD researchers, clinicians, and patients. Ideally, biomarkers will be discovered that reflect multiple processes such as neuroinflammation, neuronal stress and neuronal dysfunction, which will enable intervention prior to significant neuronal loss. In addition, biomarkers are needed to monitor progression of disease in established patients. To accelerate the assessment of candidate biomarkers, we developed an absolute quantitative assay for 30 candidate protein biomarkers (Table 1). Biomarkers that were significantly different between diagnostic groups at baseline (fold differences ranging from 0.6-1.6) (Table 3), reproduced published discovery proteomic results. Four of the markers emerged as potential progression markers in AD (Figure 4). Multiplexing enabled the exploration of relationships between CSF biomarkers; we discovered highly correlated proteins that may serve as markers of neurodegeneration in established AD patients (Figure 5). Multiplexed, targeted-proteomics proved to be a robust approach with relatively rapid assay development time enabling the assessment and prioritization of biomarker discovery candidates. This approach promises to fill a much needed gap in clinical biomarker development.