Blood tests in development for Alzheimer’s disease rely on Aβ and phosphorylated tau, markers of its hallmark pathologies. While powerful, these markers are only accurate in about 90 percent of cases. Now, a new study suggests that a broad protein panel can do better, and allow clinicians to determine disease stage. In the May 25 Alzheimer’s and Dementia, researchers led by Nancy Ip at the Hong Kong University of Science and Technology debuted a 19-protein panel that detected clinically diagnosed AD in two small cohorts with 97 percent accuracy. The panel includes several proteins that change with disease stage, allowing it to distinguish mild from severe dementia. Ip believes this set of diverse proteins might hold clues to underlying biological processes and mechanisms of disease as well.

  • A panel of 19 plasma proteins distinguishes AD from controls with 97 percent accuracy.
  • It also differentiates mild and severe dementia.
  • It is not yet clear if the panel is specific for AD.

Commenters agreed the panel has potential. “This is a very nice, first-step study that should be followed up in larger, independent samples,” Sid O’Bryant at the University of North Texas Health Science Center in Fort Worth wrote to Alzforum (full comment below). Pieter Jelle Visser and Betty Tijms at Amsterdam University Medical Center suggested comparing the panel to traditional CSF and PET biomarkers of AD, as well as testing its ability to detect preclinical disease. “A number of validation studies would be needed in order to further understand the clinical utility of this panel,” they wrote to Alzforum (full comment below).

Several groups have attempted to define plasma biomarkers of AD over the years, but these typically have not held up in replication studies (e.g., Mar 2014 news; Feb 2016 news). Among those efforts, a panel of 18 proteins identified by researchers in Tony Wyss-Coray’s lab at Stanford University and a set of four proteins found by William Hu and colleagues at Emory University, Atlanta, garnered some attention, but neither finding has been reproduced (Oct 2007 news; Aug 2012 news). 

To get more robust results, Ip turned to a relatively new methodology, proximity extension assay (PEA). In this approach, plasma proteins are detected by pairs of antibodies that recognize different epitopes on the same protein. Each antibody is linked to a complementary DNA strand, one being longer than the other. When the two antibodies bind their target protein, the strands hybridize, and DNA polymerase elongates the short strand, completing a unique “bar code” for each protein target. Lastly, a quantitative PCR step amplifies that signal (Lundberg et al., 2011). This assay has the advantage of being able to detect proteins across a broad concentration range, from 0.01 pg/mL to 1 mg/mL, Ip noted. This is important because many proteins are present in the blood in very tiny quantities, and older methods such as ELISAs cannot detect them. PEA is also suitable for high-throughput studies, assessing roughly 100 different proteins at a time.

With this method, first author Yuanbing Jiang profiled 1,160 proteins in the blood of 106 Alzheimer’s patients and 74 age-matched controls. All participants were seen at the Hong Kong clinic, and AD was diagnosed by clinical criteria. Excluding the known AD biomarkers—Aβ42/Aβ40, total tau, and NfL—the authors found 429 plasma proteins that were more, or less, abundant between AD patients and controls. Many of these differences were in line with previous work, with this set including three proteins from Wyss-Coray’s panel and one from Hu’s, and showing good agreement with a recent BioFINDER plasma protein study (Whelan et al., 2019). 

The authors clustered these 429 proteins into 19 distinct biological processes, for example cell adhesion, extracellular matrix disassembly, and apoptosis, and then selected the protein in each group that changed most dramatically in AD. These 19 “hub proteins” formed the new diagnostic panel.

The panel detected AD status in the cohort with an area under the curve (AUC) of 0.98. In an independent cohort of 36 AD patients and 61 controls, the panel identified AD with an AUC of 0.97. Notably, this accuracy is higher than that of plasma Aβ42/Aβ40, total tau, and NfL, which have a combined AUC of 87 percent, and about equal to the accuracy of plasma p-tau181 and p-tau217, which have AUCs of 0.98 and 0.96, respectively.

Commenters said the data are valuable, but cautioned that “overfitting,” i.e., assessing a diagnostic in the same cohort used to determine it, can result in misleadingly high accuracy. They also noted that the replication cohort was small, and only 16 of the 19 proteins repeated in this group. “The AUCs … are impressive but maybe slightly over-optimistic, and warrant further validation in biomarker-verified AD,” Sebastian Palmqvist at Lund University, Sweden, wrote to Alzforum, expressing the common view (full comment below).

Others pointed out that because AD status in these cohorts was determined clinically, it is unclear if the panel is specific for Alzheimer’s or detects dementia more generally. John Ringman at the University of Southern California, Los Angeles, noted that if the 19-protein panel picks up non-AD forms of dementia as well, that may explain why it appeared to have higher diagnostic accuracy than traditional AD biomarkers (comment below).

However, because the protein set reflects broader biology than do the classic AD biomarkers, it may convey additional information than just the presence or absence of disease. Ip and colleagues found that seven of the 19 proteins changed with disease stage, allowing the panel to separate mild from severe dementia. Three of these proteins, NELL1, centrin 2, and keratin 14, were altered even in people with relatively little cognitive impairment, and were even worse in severe dementia. NELL1 affects cell growth, centrin 2 is a structural protein, and keratin 14 is part of the extracellular matrix. Three others, the tyrosine kinase LYN, protein kinase C theta, and the receptor LIF-R, were altered in people with moderate cognitive impairment, and did not further change in severe disease. The seventh protein, kallikrein-related peptidase 4, was only altered in people with severe cognitive impairment. “Dysregulation of different tissues, cells, and biological processes may be involved during distinct stages of the disease,” Ip told Alzforum.

Others agreed that the use of multiple proteins could have added value. “Biomarker panels that reflect multiple aspects of AD pathophysiology have the potential to expand our understanding of AD heterogeneity and staging,” Nick Seyfried and Erik Johnson at Emory University wrote to Alzforum (full comment below).

In future work, Ip will investigate whether the panel can distinguish subtypes of AD, and whether it differentiates AD from other neurodegenerative diseases such as Parkinson’s. She will also conduct longitudinal studies to determine how these plasma proteins change with aging and disease progression.—Madolyn Bowman Rogers


  1. The authors have done a thorough job in identifying 19 candidate AD-associated proteins. A strength of the study is the comparison with the plasma biomarkers Aβ42/40, p-tau181, and NfL. However, the lack of biomarker-verified AD participants and only a comparison between clinical AD versus controls, without knowledge about their amyloid status or other dementias, make it difficult to interpret if this protein panel is specific to AD, or if it reflects more general alterations seen in cognitive impairment.

    The AUCs of 0.97-0.98 for differentiating clinical AD from controls are impressive, but maybe slightly overoptimistic, and warrant further validation in biomarker-verified AD versus other neurodegenerative diseases and in AD at different disease stages.

    Previous studies on complex biomarker panels for identifying AD have been notoriously difficult to replicate by other research groups, and there are, to my knowledge, no such protein panels used in clinical practice or in clinical trials. Future replication studies will tell if this panel of 19 proteins will be different.

  2. AD pathology can now be assessed in plasma with assays for Aβ42/40, total-tau, and p-tau. Jiang et al. studied a broader panel of plasma proteins, not including these key AD markers, as a predictor for a clinical diagnosis of AD relative to cognitively normal individuals. Of the 1,160 proteins, 426 differed between groups, from which 19 proteins were selected. This panel showed an extremely high predictive accuracy for clinical AD relative to controls, with an area under the curve (AUC) of 0.97. For comparison, the AUC of a clinical diagnosis of AD dementia relative to controls is for amyloid PET around 0.87, for CSF Aβ42 0.89, and for CSF p-tau 0.88 (Mattsson et al., 2014; Jung et al., 2020). One possible explanation for this high accuracy is overfitting or “double dipping” if feature selection and classification performance are determined within the same sample (Kriegeskorte et al., 2010; Kriegeskorte et al., 2009). 

    Overfitting can be tested by replication in an independent cohort, which the authors did in a cohort of 97 individuals, of whom 33 had clinical AD. However, it is not clear from the paper whether the original discovery fit was used to predict AD in new, unseen individuals, or if the same group of 19 proteins was refitted again in the validation cohort. If the discovery fit was used, it is remarkable that it performed with such a high accuracy (AUC=0.97) since in the replication cohort 13 of the 19 proteins did not differ between AD and controls, and proteins that did replicate showed smaller effect sizes for four proteins.

    The authors suggest that the profile can identify two disease stages, in addition to the control group. Still the accuracy of the 19-panel for prediction of clinical AD relative to controls, looks very similar across the range of MOCA scores with AUCs between 0.96-1 (supplementary figure 5).

    A number of validation studies would be needed in order to further understand the clinical utility of this panel. It would be great to see whether the panel can predict CSF or PET brain amyloid abnormalities in cognitively normal individuals as a proxy of preclinical AD, in individuals with MCI as a proxy of prodromal AD, and in individuals with dementia, in order to test discriminative value between AD and other dementias such as frontal temporal lobe dementia and Lewy body dementia.


    . Diagnostic accuracy of CSF Ab42 and florbetapir PET for Alzheimer's disease. Ann Clin Transl Neurol. 2014 Aug;1(8):534-43. Epub 2014 Jul 19 PubMed.

    . Comparison of Diagnostic Performances Between Cerebrospinal Fluid Biomarkers and Amyloid PET in a Clinical Setting. J Alzheimers Dis. 2020;74(2):473-490. PubMed.

    . Everything you never wanted to know about circular analysis, but were afraid to ask. J Cereb Blood Flow Metab. 2010 Sep;30(9):1551-7. Epub 2010 Jun 23 PubMed.

    . Circular analysis in systems neuroscience: the dangers of double dipping. Nat Neurosci. 2009 May;12(5):535-40. PubMed.

  3. This is a nice paper that looks into a very important topic. Blood-based biomarkers have significant advantages over CSF and imaging modalities in screening stages of AD, and can potentially be used to determine which patients require those confirmatory diagnostic procedures. The study is early discovery work conducted with small samples, so overfitting is a significant concern. This work will need to be cross-validated independently.

    The association of the hubs with disease status is not surprising and is consistent with a lot of prior work, though the specific markers and protein panel may or may not validate in the long run. The study used clinical diagnosis instead of confirmatory diagnostic methods (PET or CSF) and MoCA scores rather than full cognitive testing, so it is unknown how many of the patients actually had amyloid (A), tau (T), or neurodegenerative (N) pathology. Despite great progress in the field, plasma markers associated with these pathways are not proxies for cerebral levels. That means that the protein classification is of clinical dementia, not “Alzheimer’s disease” according to the ATN framework.

    Regardless of the limitations, this is a very nice first-step study that should be followed up in larger, independent samples. In future work, it would be helpful for the group to specifically define exactly what the context of use is for the intended “biomarker,” which is not defined here. That will enable others to clearly examine the findings within the intended use. 

  4. This is an interesting and important study that used the Olink PEA proteomic technology to investigate plasma AD biomarkers in discovery and replication cohorts totaling nearly 300 Chinese subjects, almost half of whom had AD. One of the many nice aspects of the study is that the authors used co-expression analysis to cluster the proteins found to be differentially abundant in AD into 19 modules, and then created a biomarker panel consisting of the strongest hub protein from each module.

    This panel represents all AD-related pathological changes that can be observed through measurement of the approximately 1,200 plasma proteins in this cohort. As the authors note, some of the modules and hub proteins likely represent central or peripheral pathophysiologies that may be stage-specific. A number of important questions arise from the results, including whether the observed changes are linked to brain-specific pathology, how they may change longitudinally in a given patient, whether certain AD subtypes can be differentiated by these 19 plasma markers, as has recently been suggested for protein markers in CSF (Tijms et al., 2020), and if they are specific tor AD. Clearly, biomarker panels that reflect multiple aspects of AD pathophysiology have the potential to expand our understanding of AD heterogeneity and staging beyond what may be possible through measurements of current plasma markers, such as amyloid, p-tau, and NFL (Higginbotham et al., 2020; Johnson et al., 2020). 

    A significant challenge when developing plasma biomarkers is the large variance in plasma protein levels across individuals and populations. The authors are commended for using genomic information to adjust for variation due to population structure in their analyses. Eventually, a better understanding of individual plasma marker variation, including how it relates to age, sex, and genetic background, will be important for implementation of precision medicine approaches using such panels in the clinic.


    . Pathophysiological subtypes of Alzheimer's disease based on cerebrospinal fluid proteomics. Brain. 2020 Dec 1;143(12):3776-3792. PubMed.

    . Integrated proteomics reveals brain-based cerebrospinal fluid biomarkers in asymptomatic and symptomatic Alzheimer's disease. Sci Adv. 2020 Oct;6(43) Print 2020 Oct PubMed.

    . Large-scale proteomic analysis of Alzheimer's disease brain and cerebrospinal fluid reveals early changes in energy metabolism associated with microglia and astrocyte activation. Nat Med. 2020 May;26(5):769-780. Epub 2020 Apr 13 PubMed.

  5. Jiang et al. assessed the plasma proteome in discovery and validation cohorts of Chinese persons in various stages of clinically diagnosed Alzheimer's disease or who were healthy controls. They initially identified 429 proteins that were differentially expressed in AD, which represented 19 clusters of proteins that were related in function. Within these clusters they identified one protein in each that was mostly correlated with AD diagnosis, and thereby created a panel of 19 proteins (four upregulated and 15 downregulated) that they further studied regarding their ability to differentiate AD from controls and their relation to disease stage markers. They found that this panel had a near-perfect Area Under the Curve (AUC) of 0.9816 in differentiating AD patients from controls. They then applied the panel to a smaller, independent sample (consisting of controls and 36 AD patients) and found an AUC of 0.969, which was better than the AUC of 0.8871 using other A/T/N markers in plasma. They also calculated an index based on these protein levels which classified participants into normal, mild, or severe AD, and this appeared to be correlated with disease stage measured by MoCA score, hippocampal volume, and gray-matter volume.

    Strengths of this study are that it appears to have begun as an unbiased hypothesis-free approach that nonetheless identified proteins that have been previously described in association with AD. They identified 19 clusters of proteins with no direct association with amyloid or tau pathways that were associated with clinical AD diagnosis. This is consistent with our widening perspective on the pathways involved in AD pathogenesis; that is, that there are multiple pathways involved that will likely need to be addressed to successfully treat the disease. Clusters identified include the inflammatory and immune response, platelet activation, apoptosis, cell adhesion, and other aspects of endothelial function. They employed both a discovery cohort and an independent validation cohort, strengthening the likelihood that the association of these protein levels with AD is a real one. The association of plasma markers with severity of disease is intriguing but should be considered preliminary.   

    A question I have about the study is that it indicates that “only individuals in whom the 19-protein biomarker panel and plasma ATN biomarkers were detectable (i.e., above the lower limit of detection; n = 172 and 97 for the discovery and validation cohort, respectively) were included in subsequent analyses.”

    As the paper reads, there were only 180 persons in the discovery cohort and 97 in the validation cohort to start with so, if true, only eight persons total were excluded. This is great if so, but it seems like a more straightforward way of saying it is, "Eight persons were excluded as the relevant proteins could not be measured in plasma." 

    Limitations of this study are that, like many studies of AD biomarkers, a clinical diagnosis of AD is used to classify patients. As we know this is subject to error. The ideal study would involve neuropathological diagnosis as the gold standard. They compared their measures to more well-established A/T/N markers and appear to have found good correspondence, which strengthens the likelihood of an association with AD pathology. However, to then claim the 19-marker panel is superior based on its (slightly) higher ability to predict clinical diagnosis is suspect.

    Might cases with discrepant 19-marker panel and A/T/N indicators of AD be those without AD neuropathology? Also, the validation cohort had a relatively small number of participants with AD, such that the study of these markers in further independent cohorts will be required to establish its utility. 

    The authors appropriately cite quite a few prior papers that have similarly found panels of plasma proteins purported to have diagnostic utility in AD that have not found their way into widespread use.

  6. We are very grateful for the helpful comments about the development of a blood-based biomarker panel for AD in the present stage. One comment is that clinically diagnosed AD is sometimes mixed with other types of dementia, including frontotemporal dementia and Lewy body dementia. To exclude non-AD dementia, AD biomarkers—including amyloid PET, CSF Aβ42, or CSF p-tau—are often used to verify clinical AD diagnosis; nonetheless, information on these biomarkers is not always available in some cohorts. Meanwhile, recent studies showed that changes in plasma p-tau181 are highly specific to AD, and correlate with amyloid and tau status in the brain (Karikari et al., 2020), can also be used to differentiate AD from healthy people as well as from other types of neurodegenerative diseases (AUC = 0.974) (Rodriguez et al., 2020). 

    Notably, in the present study, we showed that the 19 hub proteins exhibit more prominent changes in p-tau-positive AD cases compared with p-tau-negative healthy individuals, achieving accurate classifications between the two (AUC = 0.9863-0.9881).

    Moreover, 10 out of the 19 hub plasma proteins show significant correlation with plasma p-tau181 levels. These findings together support the notion that the 19-protein panel provides indication for the status of AD-specific tau pathology as a diagnostic paradigm. In the future, we will examine the performance of this 19-protein panel in identifying AD cohorts that are verified by other biomarkers such as amyloid PET, as well as its ability to differentiate AD from the other neurodegenerative diseases.

    Another comment is the “overfitting” during the discovery process, which often happens when the sample size of the discovery cohort is not big enough to represent the general AD population. In the present study, to reduce such overfitting effects, we evaluated the classification accuracy of the 19-protein panel in an independent AD cohort, which warranted high performance. To further ensure that the identified AD-associated plasma proteins can be replicated in other AD cohorts, we conducted a comprehensive comparison between our findings and previous AD plasma proteome studies, and we showed that 56 out of 77 previously reported proteins can be replicated in our study (Table A1), including well-known AD-associated plasma proteins such as EGF, sVCAM1, IGFBP2, and PPY (Ray et al., 2007; Sattlecker et al., 2014; Docket et al., 2012; Hu et al., 2012). 

    Moreover, we conducted a correlation analysis of 270 plasma proteins that had been measured in both our cohort and a Swedish MCI cohort (i.e., BioFINDER) using the same PEA technology (Whelan et al., 2019), and it showed strong correlation in the changes of the proteins between the two cohorts (r2 up to 0.49; Figure S3). These findings together suggest that those plasma proteins reported in our study may exhibit consistent alterations in AD across ethnic groups. To consolidate this, we will continue to examine the performance of this protein panel in other independent AD cohorts, which will help evaluate its applicability in the general population.


    . Blood phosphorylated tau 181 as a biomarker for Alzheimer's disease: a diagnostic performance and prediction modelling study using data from four prospective cohorts. Lancet Neurol. 2020 May;19(5):422-433. PubMed.

    . Plasma p-tau181 accurately predicts Alzheimer's disease pathology at least 8 years prior to post-mortem and improves the clinical characterisation of cognitive decline. Acta Neuropathol. 2020 Sep;140(3):267-278. Epub 2020 Jul 27 PubMed.

    . Classification and prediction of clinical Alzheimer's diagnosis based on plasma signaling proteins. Nat Med. 2007 Nov;13(11):1359-62. PubMed.

    . Alzheimer's disease biomarker discovery using SOMAscan multiplexed protein technology. Alzheimers Dement. 2014 Apr 24; PubMed.

    . Blood-Based Protein Biomarkers for Diagnosis of Alzheimer Disease. Arch Neurol. 2012 Jul 16;:1-8. PubMed.

    . Plasma multianalyte profiling in mild cognitive impairment and Alzheimer disease. Neurology. 2012 Aug 28;79(9):897-905. PubMed.

    . Multiplex proteomics identifies novel CSF and plasma biomarkers of early Alzheimer's disease. Acta Neuropathol Commun. 2019 Nov 6;7(1):169. PubMed.

Make a Comment

To make a comment you must login or register.


News Citations

  1. Do Lipids Hold the Key to Blood-Based Alzheimer’s Test?
  2. Replication a Challenge in Quest for Alzheimer’s Blood Test
  3. A Blood Test for AD?
  4. Plasma Markers for Alzheimer’s—Slowly But Surely?

Paper Citations

  1. . Homogeneous antibody-based proximity extension assays provide sensitive and specific detection of low-abundant proteins in human blood. Nucleic Acids Res. 2011 Aug;39(15):e102. Epub 2011 Jun 6 PubMed.
  2. . Multiplex proteomics identifies novel CSF and plasma biomarkers of early Alzheimer's disease. Acta Neuropathol Commun. 2019 Nov 6;7(1):169. PubMed.

Further Reading

Primary Papers

  1. . Large-scale plasma proteomic profiling identifies a high-performance biomarker panel for Alzheimer's disease screening and staging. Alzheimers Dement. 2021 May 25; PubMed.