A broad screen of CSF samples using 11 Olink Target 96 panels analyzed patients with Alzheimer’s Disease with dementia, patients with mild cognitive impairment and who are positive for β-amyloid (MCI Aβ+), and patients with dementia but are negative for β-amyloid (Aβ-). Over a hundred differentially expressed proteins were shown to be dysregulated, and from this data an 8-marker and a 9-marker Olink Focus Panel were designed to discriminate between these three groups with high accuracy (AUC of 0.8-0.99) in an external independent cohort. These marker panels have real diagnostic potential as well as utility in design of clinical trials of drugs targeting various pathological mechanisms of Alzheimer’s Disease.
The multifactorial nature of Alzheimer’s Disease: a spectrum of diverse pathological pathways
Alzheimer’s Disease (AD) and related dementias are the cause of 6.5 million deaths in the United States every year, and the National Institutes of Health estimates this number will double in the next four decades.1 Another study has found individuals in their 70’s have an almost one in three chance of getting Alzheimer’s Disease, and these were individuals who were born in the 1920’s. Those born after 1920 naturally will be living longer, and thus are at an even higher risk.2
Alzheimer’s Disease is the most common age-related dementia, accounting for 60% to 80% of patients with dementia. However, the symptoms and pathology of AD overlap the other two causes of dementia, called dementia with Lewy Bodies (DLB) and frontotemporal dementia (FTD); it is estimated that misdiagnosis of AD is up to 30%. This misdiagnosis challenges the effectiveness of any clinical trial, and a better understanding of the molecular pathology of AD is critically needed.
Current markers of AD include presence of the protein Aβ42 or a ratio of proteins Aβ42/40 reflecting a later stage of AD, and hyperphosphorylated tau (p-tau) or total tau (tTau) for early diagnosis of AD. A weakness of these two earlier markers of AD is their lack of discriminating power for non-AD dementias, as well as its inability to characterize the multifactorial nature of AD pathology. New markers are needed, not only for characterizing the biological definition of AD, but also to reveal new potential drug targets, as well as provide surrogate endpoints for clinical trials.
Two goals of this work: To define new CSF proteomic changes underlying AD, and to identify and validate biomarkers to aid in early and specific diagnosis of AD
The researchers used eleven Olink® Target 96 panels (a total of 979 proteins) to assess 797 individuals in their discovery cohort (numbers listed below in Figure 1). After identification of over 100 dysregulated proteins, a classification model separated clinical groups with very high accuracy (Figure 2). The Receiver Operator Characteristic (ROC) curves were determined to have an area under the curve (AUC) ranging from 0.85 to 0.99.
The numbers of CSF samples and their sample type in the Discovery Cohort, adapted from reference 3. MCI(Aβ+) = Mild cognitive impairment patients with positive amyloid-β; AD = Alzheimer’s Disease
The number of significantly change markers by CSF sample type, adapted from reference 3.
Four clinical groups with both unique and overlapping proteins
The four clinical groups – a group with mild cognitive impairment and positive for β-amyloid [called MCI(Aβ+)], a group with AD, a group with dementia but not AD, and a control group – had proteins unique to themselves, as well as other proteins in common among them. 281 proteins were discovered in the AD group; 189 of these 281 were abnormal only in the AD dementia stage, and of these 63 were also dysregulated in the MCI(Aβ+) group.
Thus 63 proteins were dysregulated in the early stages of AD and persist through AD dementia; another 41 proteins were dysregulated in the MCI(Aβ+) samples but not observed in AD dementia. This demonstrates the uniqueness of each stage of AD, and the presence of markers that are stage-specific.
Broadly, proteins uniquely dysregulated in the early AD group called MCI(Aβ+) related to protein breakdown (called catabolism), energy metabolism and oxidation, and proteins dysregulated only in AD dementia related to cell remodeling, vascular function and immune response.
A classification challenge: what is the minimum number of markers with the maximum discriminating power?
A data-driven model determined the smallest protein signature that gave the highest discriminating power. They used a “penalized generalized linear modeling with an elastic net penalty (a linear combination of lasso and ridge penalties)”3 and the performance of each selection was tested though a fivefold cross-validation repeated 1,000 times.
By modeling multiple biomarkers simultaneously, those with the highest power to reproducibly discriminate between classes were then selected. This is in stark contrast to single-marker analysis, which does not take into account the biological relationships between the markers in the pathobiology of disease.
The authors point out that individual markers have a limited magnitude of change, and are therefore of marginal utility as a biomarker of disease. with a large sample size (which this study fulfills) and “robust high-throughput, reproduceable and translatable technologies”3 they are able to achieve discrimination between the different stages of AD.
Details around a custom 15 biomarker Olink Focus panel
Having identified fifteen proteins that can distinguish between the three affected groups [that is, MCI(Aβ+), AD, and Non-AD Dementia to healthy controls], a custom Olink Focus panel was constructed. These fifteen proteins were divided into two groups, with two proteins (ABL1 and ITBG2) in common among the two.
|Panel to distinguish between MCI(Aβ+) with AD and healthy normal
|ABL1, SDC4, CLEC5A, MMP-10, ITGB2, TREM1, THBD, SPON2
|Receiver Operator Characteristic Area Under Curve: 0.96
|Panel to distinguish between dementia patients with and without AD
|ABL1, THOP1, ENO2, DDC, ITGB2, GZMB, MMP7, VEGFR-3, PTK7
|Receiver Operator Characteristic Area Under Curve: 0.96
For those unfamiliar with Receiver Operator Characteristic curves, they are a measure of both specificity and sensitivity captured simultaneously, measuring false positive and false negative rates, and an AUC of 0.96 indicates extremely high accuracy. They applied this custom panel to a second independent cohort (figure 3) and still achieved an AUC of 0.94.
Figure 3: The numbers of CSF samples and their sample type in the Validation Cohort, adapted from reference 3. MCI(Aβ+) = Mild cognitive impairment patients with positive amyloid-β; AD = Alzheimer’s Disease
It is not uncommon for an assay with potential diagnostic usage to perform very well with the set of samples the biomarkers were discovered with, but when the assay is developed for a focused set of biomarkers there is often a clear degradation in accuracy once an independent set of samples are tested. This is due to both a change in platform (from a wide biomarker screen to a narrowly-focused set of chosen biomarkers) as well as the independently collected and independently measured samples.
The authors show concordance data between the large Olink Target 96 panels and the performance of the same replicate samples assayed with the 15-marker Olink Focus panels. (See “Extended Data Fig.1: Protein replicates measured through different PEA panels showed comparable differences”4.)
A definitive conclusion and next steps
The authors address something they call the “cross-technology gap often encountered in biomarker studies”, where a technology used for discovery of a biomarker is not practical nor economical to consider for analysis of only a dozen or more biomarkers for routine analysis and potential diagnostic application. However, in this case they were able to take advantage of a broad panel of eleven Olink Target 96 panels, a total of 979 proteins, select the 15 proteins of interest and have the same Proximity Extension Assay technology applied on a much lower scale, with comparable sensitivity and specificity against an external cohort.
They state: “This study provides insights in to the multiple and specific protein changes underling the pathogenesis of AD and translates the multifactorial findings into practicable CSF biomarker panels.” They then conclude, “This study also highlights the effectiveness of our methodological workflow to discover and validate new biofluid-based biomarkers, leveraging the combination of large well-characterized cohorts with robust and translatable technologies.”
- Hebert LE, Beckett LA, Scherr PA, Evans DA. Annual incidence of Alzheimer disease in the United States projected to the years 2000 through 2050. Alzheimer Dis Assoc Disord (2001) 15(4):169-73. https://doi.org/10.1097/00002093-200110000-00002
- Fishman E. Risk of Developing Dementia at Older Ages in the United States. Demography. (2017) 54(5):1897-1919. https://doi.org/10.1007/s13524-017-0598-7
- del Campo M. and Teunissen C.E. et al. Nat Aging (2022) 1-14 CSF proteome profiling across the Alzheimer’s disease spectrum reflects the multifactorial nature of the disease and identifies specific biomarker panels. https://doi.org/10.1038/s43587-022-00300-1
- Extended Data Fig. 1: Protein replicates measured through different PEA panels showed comparable differences. From: CSF proteome profiling across the Alzheimer’s disease spectrum reflects the multifactorial nature of the disease and identifies specific biomarker panels. https://www.nature.com/articles/s43587-022-00300-1/figures/4