CSF Biomarkers—Maybe You Can See the Forest for the Trees

The prospects of using cerebrospinal fluid biomarkers to diagnose Alzheimer disease may have just gotten a bit brighter. In this month’s Annals of Neurology online, Kelvin Lee and colleagues at Cornell University, New York, report that “random forest,” a type of multivariate statistical analysis, has identified a suite of peptide markers that can distinguish AD patients from controls with a high degree of accuracy. If confirmed in larger sample sets, the finding could form the basis of a much-needed diagnostic test.

Currently, diagnosis of AD requires a combination of brain imaging, expert neuropsychological testing, and good, old-fashioned review of patient history. But a definitive diagnosis of AD does not come until after death, when brain tissues can be examined for pathological hallmarks such as amyloid plaques and neurofibrillary tangles. While this situation may be tolerable at present, the availability of a reliable and early diagnostic test will take on new importance once treatments become available that can prevent or slow the progression of the disease.

Because the cerebrospinal fluid (CSF) is in direct contact with the brain, analysis of the fluid is an attractive means of charting the rise and fall of molecules that might serve as AD biomarkers. Many labs have already taken a protein/proteomic tack, identifying several likely candidates, including amyloid-β and phospho-tau (see ARF related news story), and proteomic profiles (see, for example, Carrette et al., 2003; Puchades et al., 2003), but most samples studied have been from unconfirmed AD cases. This study differs by combining multivariate analysis with a data set comprising antemortem samples from patients that were confirmed to have AD postmortem. It also uses control samples taken from patients with other brain disorders to help weed out AD-specific variables from those that reflect more general changes going on in the brain.

First author Erin Finehout and colleagues used two dimensional electrophoresis (2DE) followed by time-of-flight mass spectroscopy to quantify CSF peptides in samples from 34 AD patients and 34 controls. The latter included nine healthy volunteers, 10 Parkinson patients, and 15 subjects with various other neurological diseases. From 1,938 2DE spots, an initial random forest analysis correctly classified only 26 of the 34 AD samples—the random forest algorithm is borrowed from the machine learning branch of computer science and has been shown to be particularly useful for analysis of proteomic data sets (see review by Izmirlian, 2004) The researchers then started felling the less statistically significant spots from the analysis one by one, until a copse of just 23 spots remained. This correctly classified 32 each of the AD samples and controls—a predicted error rate of 5.9 percent. On a second validation set of samples (10 AD, 18 non-AD) the biomarker profile was not as accurate, correctly classifying nine of the AD and only 15 of the non-AD samples. Overall, based on the two data sets, random forest analysis of the 23 spots had a slightly higher classification error rate of 8.3 percent. “Nonetheless, this multivariate statistical study represents the largest cohort of pathologically characterized antemortem CSF samples used in an AD proteomic biomarker study to date and suggests the possibility of eventually developing clinically relevant diagnostic assays based on CSF proteomic analysis,” write the authors.

The mass spec analysis showed that some of the spots contained more than one protein. Eighteen spots yielded peptides that matched 21 known proteins, while in five of the spots no known protein could be identified. Notable by their absence were Aβ, tau, and phospho-tau. Finehout and colleagues classified the 21 detected proteins into four major categories: Aβ transport (vitamin D-binding protein; albumins 1, 2 and 3; ApoE; ApoJ 1, 2 and 3; transthyretin 1 and 2; retinol binding protein); inflammation (immunoglobulin light and heavy chains; complement component 3; plasminogen; fibrin β); proteolytic enzyme inhibition (α-1-antitrypsin 1 and 2; ProSAAS), and neuronal membrane proteins (contactin; neuronal pentraxin receptor). Reasonable arguments can be made for the involvement of many of these proteins in AD pathology. The Aβ transporters may help regulate Aβ flux in the brain—albumin polymorphisms have also been linked to AD in a Japanese cohort (see Alzgene database)—while ApoE and J have been strongly linked to AD (see ARF related news story and ARF news story). Plasminogen may accelerate Aβ degradation (see Melchor et al., 2003), while the complement component 3 and the immunoglobulins play key roles in inflammation, which may be a major pathological process in the AD brain (see Forum poster review by Keith Crutcher). Interestingly, Maja Puchades and colleagues at Goteburg University, Sweden, previously identified α-1-antitrypsin in the CSF or AD patients.

“The method presented here has shown promising results,” write the authors. The analysis of a bigger and more diverse set of samples may help get a more accurate estimate of error prediction and may also help to identify profiles that correspond with different stages of the disease.—Tom Fagan

Comments

  1. Among efforts to develop biomarkers for neurodegenerative diseases, research on Alzheimer disease (AD) biomarkers certainly is at the vanguard. While hypothesis-driven candidate biomarkers such as CSF tau, Aβ, and other analytes linked to mechanisms of AD have shown the most promise (1-3), it is timely to pursue the identification of biomarkers using unbiased strategies based on proteomics, metabalomics or related technologies. In fact, this is being done now as reported in studies from a growing number of laboratories (4-7) including the one here by Finehout et al. However, a curious aspect of this and other CSF proteomic studies is that they often do not pick up tau or Aβ. Yet, these are the most extensively studied AD biomarkers in CSF, and measures of total tau as well as species of phospho-tau detected by antibodies specific for tau phosphorylated at Thr181, Ser199, or Thr231, in addition to Aβ1-42 (rather than Aβ1-40 or total Aβ) in CSF correlate best with a diagnosis of AD and even MCI. Indeed, the combination of total tau levels and the Aβ1-42/phospho-tau(Thr181) ratio predicted progression to AD with a sensitivity and specificity of 95 percent and 87 percent, respectively, in one recent study of MCI subjects followed 4-6 years (8). Nonetheless, there is no perfect AD biomarker, and it is not certain there will be an "AD pregnancy test equivalent" diagnostic assay that identifies AD before it manifests clinical signs, as is the case with a pregnancy test. More likely, a panel of biomarkers will be needed for the diagnosis of incipient AD and for following responses of AD patients to disease-modifying therapy. Hopefully, further efforts like those of Finehout et al. will turn up many new potential AD biomarkers, but it also is critical to emphasize that after the initial "sighting" of potential candidate AD biomarkers, there is a need for considerable further study to know if these analytes have staying power, and if so, how best to use them for diagnosis and in clinical trials.

    References:

    Consensus report of the Working Group on: "Molecular and Biochemical Markers of Alzheimer's Disease". The Ronald and Nancy Reagan Research Institute of the Alzheimer's Association and the National Institute on Aging Working Group. Am J Hum Genet. 1985 Jul;37(4):827-9. PubMed.

    Frank RA, Galasko D, Hampel H, Hardy J, de Leon MJ, Mehta PD, Rogers J, Siemers E, Trojanowski JQ, . Biological markers for therapeutic trials in Alzheimer's disease. Proceedings of the biological markers working group; NIA initiative on neuroimaging in Alzheimer's disease. Neurobiol Aging. 2003 Jul-Aug;24(4):521-36. PubMed.

    Clark CM, Pratico D, Shaw LM, Leight S, Xie SX, Gu A, Lee VM, Trojanowski JQ. Commentary on "Optimal design of clinical trials for drugs designed to slow the course of Alzheimer's disease." Biochemical biomarkers of late-life dementia. Alzheimers Dement. 2006 Oct;2(4):287-93. PubMed.

    Castaño EM, Roher AE, Esh CL, Kokjohn TA, Beach T. Comparative proteomics of cerebrospinal fluid in neuropathologically-confirmed Alzheimer's disease and non-demented elderly subjects. Neurol Res. 2006 Mar;28(2):155-63. PubMed.

    Hu Y, Malone JP, Fagan AM, Townsend RR, Holtzman DM. Comparative proteomic analysis of intra- and interindividual variation in human cerebrospinal fluid. Mol Cell Proteomics. 2005 Dec;4(12):2000-9. PubMed.

    Puchades M, Hansson SF, Nilsson CL, Andreasen N, Blennow K, Davidsson P. Proteomic studies of potential cerebrospinal fluid protein markers for Alzheimer's disease. Brain Res Mol Brain Res. 2003 Oct 21;118(1-2):140-6. PubMed.

    Zhang J, Goodlett DR, Quinn JF, Peskind E, Kaye JA, Zhou Y, Pan C, Yi E, Eng J, Wang Q, Aebersold RH, Montine TJ. Quantitative proteomics of cerebrospinal fluid from patients with Alzheimer disease. J Alzheimers Dis. 2005 Apr;7(2):125-33; discussion 173-80. PubMed.

    Hansson O, Zetterberg H, Buchhave P, Londos E, Blennow K, Minthon L. Association between CSF biomarkers and incipient Alzheimer's disease in patients with mild cognitive impairment: a follow-up study. Lancet Neurol. 2006 Mar;5(3):228-34. PubMed.

    View all comments by John Trojanowski

  2. The novelty of the study by Finehout and colleagues lies in the statistical analyses, unfortunately an issue in which I have no expertise and thus cannot comment. However, I can comment on the experimental design. In this regard the study has a number of strengths including the use of autopsy-confirmed AD samples. Nevertheless, having autopsy-confirmed "controls" is equally important since a percentage of elderly controls would be expected to have preclinical AD pathology. This could be a confounding issue, especially if the subjects were of advanced age, but the authors provided no information on subject demographics other than clinical diagnosis. Their inclusion of non-AD neurological controls and assessment of a validation sample set are also strengths. Ultimately, as the authors acknowledge themselves, it will be necessary to perform validation experiments using an independent, quantitative method like ELISA. Comparing relative abundances on gels is not an accurate method of quantification.

    Regarding the interpretation of their results, it is curious that many of the spots identified as being important for differentiating clinical groups in their study are proteins of high abundance in CSF (albumin, immunoglobulin, transthyretin, etc.) and thus likely have their origins in plasma. This probably reflects the fact that samples of whole CSF were analyzed. Proteins found in CSF that are derived from the brain are typically found in very low abundance and as such would not be detected with their approach. Depleting CSF of high-abundance proteins prior to running them on a gel (as is done by some other groups including our own) would allow for the enrichment and subsequent analysis of proteins of low abundance. It would be very interesting to know which proteins would cluster using their random forest statistical method if depleted samples had been run instead of whole CSF. The physiological relevance of presumed plasma proteins as biomarkers of a CNS disease is unclear.

    Since AD, especially the late-onset form, is a multifactoral disease, it stands to reason that a panel of biomarkers (as opposed to a single marker) will best discriminate diseased from non-diseased groups. The results of the Finehout study clearly provide support for such a notion and offer encouragement for the use of multivariate statistical methods in analyzing such complex datasets.

    View all comments by Anne Fagan

  3. The stringent requirements for a diagnosis of definite Alzheimer disease (AD) make it unlikely that the measurement of a single biomarker in isolation can provide a sufficiently accurate antemortem test for AD. The gold standard of postmortem diagnosis requires documentation of multiple biomarkers (plaques, tangles, neuronal loss, etc.) in a characteristic density and distribution throughout the brain. Furthermore, these markers need to be interpreted in light of other factors, such as the person's age at death and whether he or she was demented during life. Neuropathologists would be hard-pressed to make a diagnosis of definite AD if they were restricted to observing just one element of AD neuropathology and blinded to the clinical history. Is it reasonable to expect that an antemortem biomarker test with similar restrictions will perform any better?

    Alzheimer diagnosis is a Bayesian process, with information from the clinical history, physical examinations, lab tests, and brain imaging contributing to the probability of an accurate diagnosis. These and other considerations led us to study the cerebrospinal fluid (CSF) proteome as a means of identifying multiplexed antemortem markers for AD. Following this line of reasoning, if our panel of 23 biomarkers is further validated in future studies and implemented on a platform suitable for clinical use, it may be best utilized in conjunction with other diagnostic tests or as a supplement to them.

    Our use of CSF from autopsy-proven AD cases is one of the distinguishing features of our proteomic study. Autopsy confirmation of diagnosis in all of our controls would have been desirable; however, CSF from autopsied normal controls is not readily available. One of the factors that has impeded progress in this field is the lack of availability of large numbers of CSF specimens from well-characterized controls and cases at different stages of dementia and dementia prodromes. The collection of CSF in initiatives such as ADNI should help overcome this bottleneck, as will the sharing of CSF specimens among the various groups carrying out proteomic and other types of biomarker studies.

    We are aware that some laboratories pursue a "complexity reduction" step prior to CSF analysis such as the depletion of albumin. We performed a large number of experiments to study the capacity of commercially available solutions to carry out such depletions. In each case, we observed that the kits were unable to reproducibly deplete the specified proteins without altering other proteins of potential interest. While depletion strategies may help access lower abundance proteins (perhaps including tau and Aβ42), they may compromise other components of the proteome. It’s far from clear that the only proteins of relevance to AD diagnosis are those that are brain-derived. As such, we chose not to remove or ignore any proteins in CSF regardless of whether they were peripheral or central in origin. We studied intact CSF and were gratified to find, for example, changes in specific albumin fragments that correlated with an AD diagnosis.

    While proteomic methods are becoming increasingly sophisticated, the current reality is that there is no single technology available that can visualize all of the proteins present in a given sample equally well. Several special classes of proteins (e.g., low abundance, small molecular weight, high isoelectric point, etc.) can be well-studied but often at the expense of less effectively characterizing other classes of proteins. In considering these trade-offs, we pursued a method that we believed would give us the best overall quantitative picture of the CSF proteome relative to AD diagnosis. The fact that Aβ42 and various tau isoforms did not emerge from this analysis doesn't mean that they are unimportant markers. To some extent this may reflect limits in current proteomics technology, but a more positive interpretation is that proteomics provides complementary information to that which is obtained through hypothesis-driven biomarker studies.

    View all comments by Kelvin H. Lee

Make a Comment

To make a comment you must login or register.

References

News Citations

  1. Biomarker Bonus: Phospho-Tau/Aβ Ratio Increase Sensitivity
  2. ApoE—Breaking Ties to Aβ Offers Potential Therapy
  3. Lipoproteins and Amyloid-β—A Fat Connection

Paper Citations

  1. Carrette O, Demalte I, Scherl A, Yalkinoglu O, Corthals G, Burkhard P, Hochstrasser DF, Sanchez JC. A panel of cerebrospinal fluid potential biomarkers for the diagnosis of Alzheimer's disease. Proteomics. 2003 Aug;3(8):1486-94. PubMed.
  2. Puchades M, Hansson SF, Nilsson CL, Andreasen N, Blennow K, Davidsson P. Proteomic studies of potential cerebrospinal fluid protein markers for Alzheimer's disease. Brain Res Mol Brain Res. 2003 Oct 21;118(1-2):140-6. PubMed.
  3. Izmirlian G. Application of the random forest classification algorithm to a SELDI-TOF proteomics study in the setting of a cancer prevention trial. Ann N Y Acad Sci. 2004 May;1020:154-74. PubMed.
  4. Melchor JP, Pawlak R, Strickland S. The tissue plasminogen activator-plasminogen proteolytic cascade accelerates amyloid-beta (Abeta) degradation and inhibits Abeta-induced neurodegeneration. J Neurosci. 2003 Oct 1;23(26):8867-71. PubMed.

Other Citations

  1. Forum poster review

External Citations

  1. Alzgene database

Further Reading

Primary Papers

  1. Finehout EJ, Franck Z, Choe LH, Relkin N, Lee KH. Cerebrospinal fluid proteomic biomarkers for Alzheimer's disease. Ann Neurol. 2007 Feb;61(2):120-9. PubMed.

Annotate

To make an annotation you must Login or Register.