A biomarker signature that identifies people on the brink of Alzheimer’s disease would help clinicians recruit the most appropriate volunteers for prevention trials. Researchers have mostly focused on select markers known to be associated with AD—from imaging, cerebrospinal fluid (CSF), or plasma—but nothing has emerged that predicts onset with high sensitivity and specificity. In an effort to find a more robust signal, scientists led by Markus Britschgi, F. Hoffmann–La Roche Ltd., Basel, Switzerland, took an unbiased sampling of more than 200 markers in the CSF, blood, and brains of participants in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) who had mild cognitive impairment (MCI) to see if a combination of marker types could better predict AD. “We took a holistic approach, asking if a combination of different clinical readouts together with imaging and fluid markers could predict progression better than individual or small sets of markers,” Britschgi told Alzforum. Trumping the stalwarts, Aβ and tau, his group found a combination of six factors from the plasma and CSF that predicted who would progress to Alzheimer’s disease (AD) in three years with 80 percent accuracy. 

“The identification of predictive markers for progression from MCI to AD is important,” said Sid O’Bryant, University of North Texas Health Science Center, Fort Worth, Texas, who was not involved in the research. “This is an interesting article that leverages a tremendous dataset in ADNI and this preliminary work certainly warrants further investigation.” Henrik Zetterberg, University of Gothenburg, Sweden, also praised the effort. “We need tau- and Aβ-independent markers of AD,” he wrote to Alzforum. “They would be useful to evaluate treatment effects in tau- and Aβ-targeted treatment trials alongside the standard markers.” However, they and other researchers contacted for this article cautioned that the analysis was not repeated in an independent cohort, causing them to question whether the results could be replicated.

ADNI provides the biggest publicly available sample of people with MCI. Previously, researchers have used this data to look for a biomarker signature that would predict progression to disease within three years. Scientists have tried using imaging measures, CSF proteins, and their combination with limited success, but nothing has proved reliable (Young et al., 2013Trzepacz et al,. 2014; Eskildsen et al., 2013). Britschgi and colleagues hoped to find a marker signature that predicted conversion with higher specificity and sensitivity.

First author Benoit Lehallier, Stanford University School of Medicine, California, included data from 928 people in ADNI who had MCI at enrollment. He and colleagues considered 249 baseline variables that might predict disease onset, including clinical and demographic characteristics, cognitive scores, MRI data, genetic status, and fluid markers. Because complete data sets were available for only a few volunteers, the researchers selected a subgroup to maximize both sample size and the number of markers. They settled on 94 people for whom data on 224 variables were available. The researchers analyzed, year by year, which markers best separated those who had stayed cognitively stable and those who had progressed to AD over a six-year period. The sample size shrank each year, due to missing data. Sixteen subjects were included in the final year-six analysis.

Without a data set comparable to ADNI with which to replicate this analysis, the researchers used a statistical method that cross-validates within the sample set. In a nutshell, they tested the sample 1,000 times, each time dividing the participants into learning and test groups a different way. For each iteration, the algorithm—called elastic net—selected the markers most highly associated with progression. Elastic net is a standard methodology used when the number of predictors is much bigger than the number of observations. In the end, the top-performing biomarkers were those that cropped up most often using this algorithm. These markers were then used in a forward classification strategy to find out how well they predicted, at baseline, progression from MCI to AD. The authors found that marker levels two and three years before diagnosis gave the best sensitivity and specificity. This time window is the most relevant for clinical trials.

They found that the typical Alzheimer’s markers, such as Aβ42, total tau, and phosphorylated tau, made unreliable predictors in this data set. Instead, a combination of six different factors—two in plasma and four in CSF—predicted progression within three years with 88 percent sensitivity and 70 percent specificity. Those who converted to AD had high apolipoprotein A-II and cortisol levels in the blood, as well as high fibroblast growth factor 4 (FGF-4) and low tumor necrosis factor–related apoptosis-inducing ligand receptor 3 (TRAIL-R3) in the CSF. They also had slightly higher calcitonin and heart-type fatty acid binding protein (FABP-heart) in the CSF. The last two factors, while not significant on their own, added to the model in combination with the other factors.

While Britschgi said he is unsure what cortisol and these five proteins say about AD or neurodegeneration in general, they point to biological pathways that previous studies have tied to Alzheimer’s. Scientists have found that levels of CSF FABP-heart and TRAIL-R3 levels, as well as plasma cortisol and ApoA-II, vary in AD (Olsson et al., 2013Craig-Schapiro et al., 2011Csernansky et al., 2006Song et al., 2012). In one study, Britschgi and colleagues found that CSF levels of FABP-heart were associated with tau and p-tau levels in AD patients and controls (Britschgi et al., 2011). 

The authors wrote that the lack of replication in an independent data set and the relatively small sample size limit the study, but they contend that their statistical methods make the best use of the limited data available. However, some scientists commented privately to Alzforum that this kind of analysis run on a small sample still runs great risk of detecting error or noise in the data instead of finding true progression markers, and they predicted replication would prove difficult.

Aside from a paucity of samples, William Hu, Emory University, Atlanta, pointed out that another challenge faced by the field is that results from many protein assays are highly variable in what and how much they detect, and how they perform in different labs. Many scientists from different institutions collect data for ADNI, so the data are prone to variability, and this makes results from panel analyses hard to replicate, he said. “The multicentered nature of ADNI makes it a better validation set than a discovery set,” he told Alzforum. In his experience, fewer than a third of CSF and plasma analytes found to be associated with AD in ADNI are validated. He said that people often look past how the assays perform if these tests identify markers previously associated with AD. However, he noted that the assays used in ADNI were designed to pick up proteins tied to AD in previous reports, so those associations are already likely (see full comment below).

“If this approach is independently cross-validated in AIBL [the Australian Imaging, Biomarker & Lifestyle Flagship Study of Ageing] or some other cohort, that would be tremendous and would set the stage for a prospective clinical study of newly diagnosed MCI cases,” O’Bryant wrote to Alzforum. “Validation is the key next step for this sort of work.” Britschgi agreed that these unvalidated markers are not ready for prime time, but said that other researchers might now explore whether the proteins help stratify their own patient samples. He also suggested other researchers analyze large data sets in a similar, integrative way using this type of algorithm.—Gwyneth Dickey Zakaib


  1. This is an interesting methodology paper that integrates different types of data (clinical, genetic, MRI features, and CSF and plasma proteins). Even though ADNI is a convenient sample that has been used for discovery purposes, the multi-centered nature of ADNI makes it a better validation set than a discovery set because it is more prone to variable signal-to-noise ratios from the different centers.

    We have some experience validating discovery findings from ADNI. Among CSF and plasma analytes identified as associated with AD in ADNI, fewer than one-third of the associations have had some type of validation. A major challenge so far has been the technical validity of assays. While we and others use these assays to measure proteins in biological fluids, all the assays are most reproducible in buffer. There are many proteins in the fluids of interest (CSF or plasma) that can interfere with antibody-antigen binding, either by enhancing it or reducing it. Some of these processes can be stochastic (or are assumed to be), which makes precise (reproducible) measurements sometimes difficult, especially if the solute levels are low (as is often the case in CSF). Second, in ADNI the Myriad RBM (Luminex) assays are performed by third-party vendors, and we do not get control data of sufficient quality. This contrasts with in-house experiments that generate quite a bit of QC data that allow us to determine whether a particular plate/batch is good or bad. Third, the analytical approaches chosen (traditional statistics, modeling, network analysis, etc.) do not generate consistent outcomes because of intrinsic differences in the data structure of these biofluid proteins. Without independent discovery and validation sets, it is very easy for the combination of these factors to grossly overfit the data, which is possibly the most common cause for non-replication (even with cross-validation, which this paper did). 

    In these studies, people are likelier to believe a particular analyte is truly associated with Alzheimer's if there are reports to that effect, and they sometimes overlook the issue of how the assays perform. The likelihood of an association in ADNI is very high, because the 190 analytes measured in ADNI were assembled after a literature review of proteins previously associated with AD. In keeping with the increasing emphasis on rigor and reproducibility in the broader scientific community, we all share the responsibility of redirecting some of the precision-associated enthusiasm towards technical validation. There is now an NIH initiative that solicits journals to encourage reproducible, robust, and transparent data analysis and reporting (see NIH: Research and Reproducibility), but there is quite a bit of catching up to do in the field.


Make a Comment

To make a comment you must login or register.


Paper Citations

  1. . Accurate multimodal probabilistic prediction of conversion to Alzheimer's disease in patients with mild cognitive impairment. Neuroimage Clin. 2013;2:735-45. PubMed.
  2. . Comparison of neuroimaging modalities for the prediction of conversion from mild cognitive impairment to Alzheimer's dementia. Neurobiol Aging. 2014 Jan;35(1):143-51. PubMed.
  3. . Prediction of Alzheimer's disease in subjects with mild cognitive impairment from the ADNI cohort using patterns of cortical thinning. Neuroimage. 2013 Jan 15;65C:511-521. PubMed.
  4. . Cerebrospinal fluid levels of heart fatty acid binding protein are elevated prodromally in Alzheimer's disease and vascular dementia. J Alzheimers Dis. 2013;34(3):673-9. PubMed.
  5. . Multiplexed immunoassay panel identifies novel CSF biomarkers for Alzheimer's disease diagnosis and prognosis. PLoS One. 2011;6(4):e18850. PubMed.
  6. . Plasma cortisol and progression of dementia in subjects with Alzheimer-type dementia. Am J Psychiatry. 2006 Dec;163(12):2164-9. PubMed.
  7. . Plasma apolipoprotein levels are associated with cognitive status and decline in a community cohort of older individuals. PLoS One. 2012;7(6):e34078. PubMed.
  8. . Modeling of pathological traits in Alzheimer's disease based on systemic extracellular signaling proteome. Mol Cell Proteomics. 2011 Oct;10(10):M111.008862. PubMed.

Further Reading


  1. . Longitudinal Protein Changes in Blood Plasma Associated with the Rate of Cognitive Decline in Alzheimer's Disease. J Alzheimers Dis. 2015;49(4):1105-14. PubMed.
  2. . Association of Elevated Amyloid Levels With Cognition and Biomarkers in Cognitively Normal People From the Community. JAMA Neurol. 2016 Jan;73(1):85-92. PubMed.
  3. . Visual Versus Fully Automated Analyses of 18F-FDG and Amyloid PET for Prediction of Dementia Due to Alzheimer Disease in Mild Cognitive Impairment. J Nucl Med. 2016 Feb;57(2):204-7. Epub 2015 Nov 19 PubMed.
  4. . Integrating Biomarkers for Underlying Alzheimer's Disease in Mild Cognitive Impairment in Daily Practice: Comparison of a Clinical Decision Support System with Individual Biomarkers. J Alzheimers Dis. 2015;50(1):261-70. PubMed.

Primary Papers

  1. . Combined Plasma and Cerebrospinal Fluid Signature for the Prediction of Midterm Progression From Mild Cognitive Impairment to Alzheimer Disease. JAMA Neurol. 2015 Dec 14;:1-10. PubMed.