In the March 11 Scientific Reports, researchers led by Noel Faux at IBM Research Australia in Southbank propose yet another blood test for Alzheimer’s disease, this one based on simple immunoassays. The scientists used machine learning to identify molecules in plasma that correlated with low Aβ42 in cerebrospinal fluid. Out of nearly 300 analytes, four proteins stood out. When combined with ApoE genotype, these predicted CSF Aβ status with 81 percent accuracy. Because the plasma proteins were measured by standard immunoassays, this test could be translated to the clinic quickly and cheaply, the authors claim.

  • Machine learning finds four plasma proteins that correlate with CSF Aβ42.
  • These proteins plus ApoE genotype predict amyloid positivity.
  • This method uses standard immunoassays, hence may translate quickly to the clinic.

Colin Masters at the University of Melbourne, Australia, agreed. “The performance of this approach may not be as high as more recent techniques … but may be relatively easier to introduce into clinical practice,” he wrote to Alzforum (full comment below). Masters was not involved in the research.

Viable blood tests for AD sprang onto the scene in 2017, with the first report that mass spectrometry was accurate enough to detect a small drop in plasma Aβ42 in people who were accumulating brain amyloid (Jul 2017 conference news). Additional mass spec methods quickly followed, as well as highly sensitive antibody-based single-molecule arrays, aka Simoas (Feb 2018 newsAug 2018 conference newsNov 2018 conference news). 

Adapting these sensitive techniques for routine clinical use may be expensive and technically challenging, however. Faux wondered if established immunoassay techniques could provide a simpler alternative. To test this, first author Benjamin Goudey used data from 358 participants in the Alzheimer’s Disease Neuroimaging Initiative. The cohort comprised 58 cognitively healthy controls, 198 people with amnestic mild cognitive impairment, and 102 people diagnosed with AD dementia. All had donated blood, which ADNI researchers analyzed using the Luminex xMap platform. This platform uses antibodies conjugated to beads to allow for the simultaneous capture and measurement of hundreds of different analytes.

Goudey and colleagues included Luminex plasma data from149 proteins and 138 metabolites in their analysis. First they assessed how well a person’s age and ApoE genotype predicted whether their CSF Aβ42 scores would fall above or below the threshold for AD. To this simple model they then added proteins, metabolites, or both. Proteins improved the predictive power, while metabolites slightly decreased it. Metabolites may add too much noise, the IBM scientists speculated.

Next, the authors used machine learning to find the smallest set of proteins that could equal the performance of the full complement. Four proteins, Aβ42, ApoE, chromogranin-A, and eotaxin 3, combined with ApoE genotype, predicted amyloid positivity with a sensitivity and specificity of 81 and 63 percent, respectively. For the first three proteins, low levels correlated with amyloid positivity; for eotaxin 3, high levels did. Chromogranin A is a synaptic protein previously associated with early AD (Mattsson et al., 2013; Duits et al., 2018). Eotaxin 3 is part of the innate immune system, and plasma levels rise in AD (Huber et al., 2016). 

To test their model, the authors applied it to a separate cohort of 208 ADNI participants, 198 of whom had amnestic MCI and the remainder AD. Among the former, those flagged as amyloid-positive progressed to AD faster than those predicted to be amyloid-negative. Because this test cohort had not undergone lumbar puncture, Goudey correlated plasma test results with amyloid positivity as judged by a PET SUVR of 1.5 or higher by PiB or 1.11 or higher by florbetapir. Plasma and PET predictions agreed 80 percent of the time. The model performed better than age and ApoE genotype alone, which correlated with PET only 71 percent of the time.

Charlotte Teunissen at VU University, Amsterdam, noted that the identification of plasma Aβ42 as a marker of AD fits with other recent findings. “One could even state that any final blood AD assay will include amyloid forms,” she wrote to Alzforum (full comment below). She recommended that replication experiments assess what added value the other three plasma markers bring, and also test different analysis methods besides the Luminex platform.

The authors could not be reached for comment before publication of this story. In their paper, they acknowledged the need for replication of these results in independent cohorts and with different immunoassay platforms. In several previous cases, putative AD blood tests based on small studies have failed to repeat (Oct 2007 news; Jun 2013 webinarFeb 2016 news). 

It remains to be seen which of the many blood tests under development will advance to clinical practice (Feb 2019 news). “Much progress has been made over the past two years,” Masters wrote. “It is very pleasing to note that all the major players are cooperating in head-to-head performance testing under the auspices of the Alzheimer’s Association. We look forward to results in the coming months.”—Madolyn Bowman Rogers

Comments

  1. These data add to the currently emerging evidence for an AD signature being detectable in blood. It is interesting that Aβ42 is identified by the random forest model, which is in line with other recent evidence of Aβ reductions in blood and CSF. It is not clear whether the ratio of Aβ42/40 has been included in the model, though this is the biomarker modality under current strong investigation. One could even state that any final blood AD assay will include amyloid forms, especially given the established pathological relation.

    The other markers in the panel are somewhat more difficult to understand. It is not clear what their added value is on the AUC compared to a model containing only Aβ42 (besides age, gender, and ApoE), but replication should show if they are useful. Replication should also include different analysis methods, because the Rules-Based Medicine platform running on a Luminex platform is not a routine laboratory technology. Moreover, comparison to other dementias and comorbidities, e.g. cardiac diseases (ApoE), and diseases involving inflammation (eotaxin-3) should show if the observed signature in blood is specific. But even if the panel shows specificity when the models are correct for these confounders in epidemiological large cohorts, implementation of such a test becomes a challenge: Should we not use the test when comorbidities are present? And what to do if these are not clinical visible?

    The lack of added value of the metabolome may not be surprising, again in view of the disease-specificity of these molecules. A signature that includes brain specific proteins would probably be a format to use in widespread clinical implementation.

  2. These data add to growing the evidence base that there are signals in plasma (and possibly in cellular blood elements) which are informative on the state of Aβ accumulation in the central nervous system. The technologies used in this report include the multiplexed immunoassays (Luminex xMAP) of Rules-Based Medicine (now owned by Myriad). Their machine-learning approach identified a combination of age, ApoE4 carriage, and the four plasma analytes chromogranin A (CGA), Aβ42, eotaxin 3 (CC chemokine ligand 26) and ApoE4, which delivered a reasonable predictor for CSF and PET Aβ status (AUCs high 70s to low 80s). The performance of this approach may not be as high as more recent techniques which utilize IP/MS or SIMOA (Ovod et al., 2017Nakamura et al., 2018; Verberk et al., 2018), but may be relatively easier to introduce into clinical practice.

    It appears that the CSF and plasma values of the brain-derived Aβ42 are in a more dynamic range than PET-Aβ signals at the lower cutpoint, when subjects are crossing the threshold from normality to abnormality. This is the window at which the field has to direct its attention for both primary and secondary prevention strategies. Hence the urgency for developing robust and high performance blood tests. Much progress has been made over the past two years. It is very pleasing to note that all the major players are cooperating in head-to-head performance testing under the auspices of the Alzheimer’s Association. We look forward to results in the coming months.

    References:

    . High performance plasma amyloid-β biomarkers for Alzheimer's disease. Nature. 2018 Feb 8;554(7691):249-254. Epub 2018 Jan 31 PubMed.

    . Plasma Amyloid as Prescreener for the Earliest Alzheimer Pathological Changes. Ann Neurol. 2018 Nov;84(5):648-658. Epub 2018 Oct 4 PubMed.

  3. Most previous blood tests have predicted amyloid PET measurements, whereas in our study we predict CSF Aβ1-42, as these have been shown to change potentially five to 10 years before the amyloid PET signal reaches the critical threshold (Palmqvist et al., 2016). The study by Nakamura et al., 2018, is very interesting in this regard. In this study, the principal prediction was for PET amyloid, but they also showed a strong association between plasma Aβ1-42, normalized for plasma APP669-771 peptide, and CSF Aβ1-42 in a small subcohort (n=46).

    While the predictive performance of our model is lower, the cohort we used was substantially larger (n=358). Further, the protein concentrations measured in ADNI were from an ELISA-based methodology, which is potentially easier to translate into a clinical setting compared to the IP-MALDI-TOF-MS method employed by Nakamura et al.

    We hope that our work will complement other approaches to developing blood-based tests, including the development of more sensitive assays for Aβ1-42 such as those by Nakamura et al., 2018, Verberk et al., 2018, or Ovod et al., 2017, in the effort to develop a screening tool for the identification of individuals at risk of AD. This could be followed up by lumbar puncture and/or amyloid PET scan if a positive result was predicted.

    Our work adds further evidence that plasma-based tests for Alzheimer’s risk factors are a very promising line of research. Approaches based on machine learning are likely to be complementary to other developments in plasma assays. We are showing that there are other proteins, in addition to plasma Aβ1-42, that add useful information to the signal from plasma.

    While we used an “internal validation cohort” (comprising individuals who were not used to train the model) to test the utility of the model, we need to validate the work in an independent external cohort. We are looking to use the Australian Imaging Biomarker and Lifestyle (AIBL) study of aging. We are also currently conducting research to apply these AI techniques to predict other important biomarkers that are associated with the progression of Alzheimer’s disease (e.g., CSF pTau and Tau) and are investigating the role that polygenic risk scores may play. The results from this work will be presented at AD/PD 2019 at the end of March.

    References:

    . Cerebrospinal fluid analysis detects cerebral amyloid-β accumulation earlier than positron emission tomography. Brain. 2016 Apr;139(Pt 4):1226-36. Epub 2016 Mar 2 PubMed.

    . High performance plasma amyloid-β biomarkers for Alzheimer's disease. Nature. 2018 Feb 8;554(7691):249-254. Epub 2018 Jan 31 PubMed.

    . Plasma Amyloid as Prescreener for the Earliest Alzheimer Pathological Changes. Ann Neurol. 2018 Nov;84(5):648-658. Epub 2018 Oct 4 PubMed.

    . Amyloid β concentrations and stable isotope labeling kinetics of human plasma specific to central nervous system amyloidosis. Alzheimers Dement. 2017 Aug;13(8):841-849. Epub 2017 Jul 19 PubMed.

Make a Comment

To make a comment you must login or register.

References

News Citations

  1. Finally, a Blood Test for Alzheimer’s?
  2. Closing in on a Blood Test for Alzheimer’s?
  3. With Sudden Progress, Blood Aβ Rivals PET at Detecting Amyloid
  4. Blood Tests for Amyloid Step Out at CTAD
  5. A Blood Test for AD?
  6. Replication a Challenge in Quest for Alzheimer’s Blood Test
  7. Blood Test Granted Breakthrough Status, To Be Tested in Trial

Webinar Citations

  1. Webinar: O Blood-Based Biomarker, Where Art Thou?

Paper Citations

  1. . CSF protein biomarkers predicting longitudinal reduction of CSF β-amyloid42 in cognitively healthy elders. Transl Psychiatry. 2013;3:e293. PubMed.
  2. . Synaptic proteins in CSF as potential novel biomarkers for prognosis in prodromal Alzheimer's disease. Alzheimers Res Ther. 2018 Jan 15;10(1):5. PubMed.
  3. . An emerging role for eotaxins in neurodegenerative disease. Clin Immunol. 2016 Sep 21; PubMed.

External Citations

  1. Luminex xMap

Further Reading

Primary Papers

  1. . A blood-based signature of cerebrospinal fluid Aβ1-42 status. Sci Rep. 2019 Mar 11;9(1):4163. PubMed.