subgroups of alzheimer's disease based on cerebrospinal fluid molecular markers

Subgroups of Alzheimer’s Disease Basedon Cerebrospinal Fluid

Molecular MarkersKhalid Iqbal, PhD,1 Michael Flory, PhD,1 Sabiha Khatoon, PhD,1 Hilkka Soininen, MD, PhD,2

Tuula Pirttila, MD, PhD,2 Maarit Lehtovirta, MD, PhD,2 Irina Alafuzoff, MD, PhD,3

Kaj Blennow, MD, PhD,4 Niels Andreasen, MD, PhD,5 Eugeen Vanmechelen, PhD,6

and Inge Grundke-Iqbal, PhD1

Alzheimer’s disease, the most common cause of dementia, is multifactorial and heterogeneous; its diagnosis remainsprobable. We postulated that more than one disease mechanism yielded Alzheimer’s histopathology, and that subgroupsof the disease might be identified by the cerebrospinal fluid (CSF) levels of proteins associated with senile (neuritic)plaques and neurofibrillary tangles. We immunoassayed levels of tau, ubiquitin, and A�1-42 in retrospectively collectedCSF samples of 468 clinically diagnosed Alzheimer’s disease patients (N � 353) or non-Alzheimer’s subjects (N � 115).Latent profile analysis assigned each subject to a cluster based on the levels of these molecular markers. Alzheimer’sdisease was subdivided into at least five subgroups based on CSF levels of A�1-42, tau, and ubiquitin; each subgrouppresented a different clinical profile. These subgroups, which can be identified by CSF analysis, might benefit differentlyfrom different therapeutic drugs.

Ann Neurol 2005;58:748–757

Alzheimer’s disease (AD) is multifactorial and clinicallyand histopathologically heterogeneous. In less than 1%of cases, the disease cosegregates with certain mutationsin �-amyloid precursor protein, presenilin-1, orpresenilin-2 genes.1 The cause(s) of the disease in theremaining more than 99% is not yet understood.2 In-dependent of cause, AD is characterized clinically byprogressive dementia and histopathologically by thepresence of numerous neurofibrillary tangles and neu-ritic (senile) plaques with neurofibrillary changes in thedystrophic neurites. Because of clinical heterogeneity3

the diagnosis of AD remains probable until postmor-tem histopathological examination and is made primar-ily by exclusion of other causes of dementia.4 The twomost common confounding diagnoses are cerebral vas-cular disease (multiinfarct dementia) and dementiawith Lewy bodies. In most AD cases, deposits of�-amyloid in the cerebral vessels, called congophilicamyloid angiopathy, are seen in varying degrees. Lewybodies are intraneuronal inclusion bodies found in

brainstem structures, limbic structures, and the cerebralcortex. Their major protein subunit is �-synuclein. Inthe substantia nigra, it is a hallmark of Parkinson’s dis-ease; it is seen in the cerebral cortex of AD cases invarying degrees.5

AD histopathology (ie, neurodegeneration associatedwith numerous neurofibrillary tangles and neuritic [se-nile] plaques) shows considerable qualitative and quan-titative heterogeneity. AD can be neocortical type, lim-bic type, and plaque-dominant type, and it maypresent with numerous neurofibrillary tangles exclu-sively confined to the hippocampus and entorhinal cor-tex.6,7 The histopathological heterogeneity of AD isalso reflected in the CSF levels of the proteins associ-ated with these lesions, that is, A� peptide (the majorcomponent of �-amyloid from plaques8) and tau/phos-photau9,10 and ubiquitin11–13 from tangles. Some ofthe A� in the CSF might also originate from�-amyloid deposits seen in cerebral vessels as congo-philic amyloid angiopathy, known to comprise mostly

From the 1New York State Institute for Basic Research in Develop-mental Disabilities, Staten Island, NY; Departments of 2Neurologyand 3Pathology, University of Kuopio and Kuopio University Hos-pital, Kuopio, Finland; 4Unit of Neurochemistry, Department ofClinical Neuroscience, Sahlgren’s University Hospital, Molndal;5Karolinska Institute, Department of NEUROTEC, Section of Ge-riatric Medicine, Karolinska University Hospital, Huddinge, Stock-holm, Sweden; and 6Innogenetics NV, Zwijnaarde, Belgium.

Received Jun 21, 2005, and in revised form Jul 22. Accepted forpublication Aug 2, 2005.

Published online Oct 24, 2005, in Wiley InterScience(www.interscience.wiley.com). DOI: 10.1002/ana.20639

Address correspondence to Dr Iqbal, Department of Neurochemis-try, New York State Institute for Basic Research in DevelopmentalDisabilities, 1050 Forest Hill Road, Staten Island, NY 10314-6399.E-mail: [email protected]

748 © 2005 American Neurological AssociationPublished by Wiley-Liss, Inc., through Wiley Subscription Services

A�1-40 and to occur in varying degrees in most cases.Studies consistently have shown an increase in CSFlevels of tau/phosphotau and ubiquitin and a decreasein A�1-42 level in AD, but there is a considerable over-lap between diseased and control cases (eg, see Kudoand colleagues,14 Hulstaert and colleagues,15 and Huand colleagues16). We postulate disease heterogeneityto be a major cause of the overlap in CSF markers. Inthis study, we report that AD divides into various sub-groups based on the levels of A�1-42, tau, and conju-gated ubiquitin in CSF.

Subjects and MethodsSubjects, Cerebrospinal Fluids, and Immunoassays forA�1-42, Tau, and UbiquitinLevels of total tau, conjugated ubiquitin, and A�1-42 wereassayed in retrospectively collected lumbar CSF samples of468 patients clinically diagnosed as AD of Lewy body type(AD/L) or AD (353 CSF samples) and as non-AD neurolog-ical or nonneurological subjects (115 CSF samples). AD sub-jects fulfilled the National Institute of Neurological andCommunication Disorders–Alzheimer’s Disease and RelatedDisorders Association criteria of probable AD4; AD/L diag-nosis was based on McKeith and colleagues’ criteria5(Table1). Neuropathological examination of autopsied brains wasperformed for 39 patients, confirming AD pathology17 in 37subjects (AD: n � 21; AD with vascular changes: n � 7;

and AD with synuclein pathology: n � 9). One patient hadPick’s disease, and one patient was diagnosed with fronto-temporal dementia.

Levels of total tau and A�1-42 were assayed by sandwichenzyme-linked immunosorbent assay using h Tau Ag and�-Amyloid(1-42) kits (Innogenetics, Ghent, Belgium). Conju-gated ubiquitin levels were assayed by competitive inhibitionenzyme-linked immunosorbent assay using as its primary an-tibody the monoclonal antibody 5-25 (Signet Labs, Dedham,MA), which recognizes the amino acid residues 64 to 76 ofubiquitin, preferably the conjugated site generated by glycine76 of ubiquitin with the substrate protein.13,14

Statistical AnalysisClustering of subjects by CSF levels of A�1-42, tau, andubiquitin was performed by Latent Profile Analysis (LPA)using Latent Gold 3.0 (Statistical Innovations, Belmont,MA).18,19 This approach has been used successfully for clas-sification of patients with chest pain and patients with longbone fracture.20,21

LPA, like cluster analysis and Latent Class Analysis (LCA),explains the clustering of cases on values of a set of variablesby assuming that patterns of values are determined by a la-tent (unobserved) characteristic. (LPA and LCA extend clus-ter analysis, yielding goodness-of-fit scores for evaluating andcomparing models. LPA, unlike LCA, allows continuous in-dicators.) Here, the latent characteristic represents a subtypeof AD. Cases are “clustered” by their indicator values. Model

Table 1. Sample Characteristics by Source

Characteristics Finland (n � 280) Sweden (n � 188) Total (N � 468)

Mean age, yr (SD) 69.6 (9.0) 73.7 (8.5) 71.2 (9.0)Mean age at dementia onset, yr (SD)a 70.3 (7.1) 72.9 (7.5) 71.4 (7.3)Mean duration of dementia, yr (SD)a 2.7 (2.6) 3.3 (2.5) 2.9 (2.6)Sex

Female 62.1% 63.8% 62.8%Male 37.9% 36.2% 37.2%

DiagnosisAD 68.9% 75.0% 71.4%AD/L 2.1% 6.9% 4.1%Control 28.9% 18.1% 24.6%

ApoE genotype3 � 2 2.9% 5.3% 3.8%3 � 3 25.4% 34.0% 28.8%4 � 2 1.1% 2.7% 1.7%4 � 3 35.7% 48.4% 40.8%4 � 4 15.7% 8.5% 12.8%Unknown 19.3% 1.1% 12.0%

Mean A�1–42, pg/ml (SD) 659.4 (246.7) 615.7 (262.9) 641.9 (254.0)Mean tau, pg/ml (SD) 689.6 (270.9) 608.2 (289.0) 656.9 (280.9)Mean ubiquitin, ng/ml (SD) 144.2 (83.8) 134.3 (85.0) 140.2 (84.4)

All samples were received in dry ice from Kuopio University, Finland, or University of Goteberg, Sweden, and were kept at �75°C until used. TheFinnish samples included 199 Alzheimer’s disease (AD) subjects (mean age, 73.0 � 6.8 years), of which 6 were diagnosed as AD of Lewy bodytype (AD/L) and 13 subjects were considered to have concomitant vascular dementia. The control group (n � 81; mean age, 61.1 � 8.3 years)included 14 mild cognitive impairment (MCI) patients, 12 patients with cerebrovascular disorders, 32 epilepsy patients, 11 patients with othercentral nervous disorders, 3 patients with peripheral nervous system disease, and 9 patients without neurological disease (headache, dizziness,depression). The Swedish samples included 141 AD, 13 AD/L, and 34 control subjects. Control subjects had no evidence of brain disease and wereseen for minor depression and stress but were otherwise healthy. No neuropathological evaluation on these patients was available.aNumber of subjects for age of onset and duration of dementia: Finland, 208; Sweden, 154; total, 362.

ApoE � apolipoprotein E.

Iqbal et al: Subgroups of AD 749

fit is assessed using the log likelihood of the model orinformation-theoretical measures derived from it. Derivedmeasures, unlike the log likelihood itself, reward parsimony,guarding against overfitting, or gaining a spurious level of fitwith excessive, ungeneralizable classes. We assessed fit usingthe most commonly used information measure, the BayesianInformation Criterion.

ResultsThe clinical diagnosis of AD was in agreement with 37of 39 subjects for whom the postmortem histopatho-logical diagnosis became available, suggesting a high(�95%) accuracy of the clinical diagnosis (Table 2).Consistent with previous reports, CSF levels of tau andubiquitin were higher and CSF levels of A�1-42 werelower in AD patients than control subjects (Fig 1). Pa-tients appeared to cluster into groups according to thecombination and extent of abnormalities in levels ofthe three marker proteins. LPA models were con-structed to test the hypothesis that differences corre-sponded to subtypes of AD, to determine whether theclustering of indicators among patients was not attrib-utable to chance, and to establish how many subtypesbest explained observed clustering. The values of themarkers for each subject were presumed to be deter-mined by AD subtype. Models were estimated inwhich the number of clusters (subtypes) was fixed atvalues from two to eight. Age was entered as a covari-ate in all models.

The three- and six-cluster models fit best to the

data(Table 3). The three-cluster model essentially di-vided subjects into cases and control subjects, with athird small cluster of apparent outliers. The six-clustermodel, however, fitted the data better with or withoutconsideration of parsimony and yielded clusters thatdiffered substantively within the cases.

Each indicator differed by subtype, as did mean age(Tables 4–6); age also affected indicator levels, witheffects reaching statistical significance for tau (p �0.05) and ubiquitin (p � 0.002). Analyses demon-strated that the observed clustering was extremely un-likely in the absence of underlying differences withinthe sample and indicated a strong likelihood that sub-jects differed in some way. We needed to know if thesedifferences were related to AD, not an unrelated factor.Our confidence that they represented different sub-types of AD was strengthened by strong associationsbetween these categories and other observed character-istics related to AD and its symptomatic manifesta-tions.

Standardized mean levels of each indicator for eachsubtype and values of potentially validating variables ineach class indicated that cluster characteristics corre-sponded in several respects to diagnosis and apoli-poprotein E (ApoE) genotype (Table 7).

Cluster 1 (AD with low A�1-42, high incidence ofApoE4, and late onset [AELO]), the largest cluster(48% of clinically diagnosed AD-AD/L subjects), wascharacterized by low levels of A�1-42 coupled with

Table 2. Sample Characteristics by Clinical and Neuropathological Diagnosis

Characteristics

Clinical Diagnosis Neuropathological Diagnosis

AD(n � 334)

AD/L(n � 19)

Control(n � 115)

Total(N � 468)

AD(n � 37)

Other(n � 2)

Total(N � 39)

Mean age, yr (SD) 74.3 (6.8) 75.2 (5.8) 61.7 (8.4) 71.2 (9.0) 75.4 (7.3) 74.5 (0.7) 75.3 (7.1)Mean age at dementia

onset, yr (SD)a71.3 (7.5) 73.3 (6.1) 70.8 (4.3) 71.4 (7.3) 71.1 (8.2) 71.0 (4.2) 71.1 (8.1)

Mean duration ofdementia, yr(SD)a

3.0 (2.6) 2.0 (1.2) 1.7 (1.3) 2.9 (2.6) 4.4 (4.5) 4.5 (4.9) 4.4 (4.5)

SexFemale 66.2% 42.1% 56.5% 62.8% 86.5% 50.0% 84.6%Male 33.8% 57.9% 43.5% 37.2% 13.5% 50.0% 15.4%

Source of caseFinland 57.8% 31.6% 70.4% 59.8% 100.0% 100.0% 100.0%Sweden 42.2% 68.4% 29.6% 40.2% 0.0% 0.0% 0.0%

ApoE genotype3 � 2 2.1% 5.3% 8.7% 3.8% 0.0% 0.0% 0.0%3 � 3 28.7% 31.6% 28.7% 28.8% 29.7% 50.0% 30.8%4 � 2 2.1% 5.3% 0.0% 1.7% 8.1% 0.0% 7.7%4 � 3 49.4% 52.6% 13.9% 40.8% 35.1% 50.0% 35.9%4 � 4 17.4% 5.3% 0.9% 12.8% 27.0% 0.0% 25.6%Unknown 0.3% 0.0% 47.8% 12.0% 0.0% 0.0% 0.0%

Mean A�1–42, pg/ml(SD)

564.5 (211.2) 626.6 (233.2) 869.1 (236.0) 641.9 (254.0) 512.9 (192.2) 464.5 (71.4) 510.5 (187.7)

Mean tau, pg/ml (SD) 753.7 (244.9) 423.5 (266.6) 414.4 (202.4) 656.9 (280.9) 847.7 (239.3) 641.0 (162.6) 837.1 (238.9)Mean ubiquitin,

ng/ml (SD)147.5 (86.1) 148.4 (136.5) 117.7 (62.1) 140.2 (84.4) 148.6 (80.5) 89.0 (21.2) 145.5 (79.6)

an � 9 for age of onset and duration of dementia among control subjects.

AD � Alzheimer’s disease; AD/L � Alzheimer’s disease of Lewy body type; ApoE � apolipoprotein E; SD � standard deviation.

750 Annals of Neurology Vol 58 No 5 November 2005

somewhat increased tau and unaffected ubiquitin levels(Fig 2). It comprised 177 subjects, 97% of whom wereAD-AD/L patients with relatively late-onset (age, 71.7years) dementia. Seventy-four percent of Cluster 1cases had one or two ApoE4 alleles (2 [1df] � 17.612;p � 0.001).

The 101 subjects in Cluster 2 (74% of the controlsubjects) had levels of A�1-42 greater than those of thesample as a whole and lower levels of tau (see Fig 2).This accorded well with that 84% of subjects in thiscluster were non-AD control subjects. The mean age(60.5 years) was �15 years younger than that for Clus-ter 1. Their ApoE allele distribution correspondedmore closely to that of the general population.

Cluster 3 (AD with low A�1-42, high tau, and earlyonset [ATEO]), which (like the first cluster) was com-posed overwhelmingly of AD subjects (96%), likewisehad low A�1-42 levels but also manifested (unlike thefirst cluster) considerably increased levels of tau, ap-proximately 1.5 standard deviations greater than themean (see Fig 2). Ubiquitin levels were not greatly dif-ferent from those of the sample as a whole. This cluster(22% of the clinically diagnosed AD subjects) was notsignificantly more likely to possess ApoE4 allele thanwas the rest of the sample (2 [1df] � 3.612; p �0.07). Among those for whom information was avail-able, age at onset of dementia was relatively early.

The fourth cluster (AD with high incidence of Lewybodies, low A�1-42, and late onset [LEBALO]), al-though still predominantly AD subjects, included pro-portionately about five times as many AD/L cases(63% of the total) than did the preceding clusters(15.6 vs �3% in all other clusters). Levels of all mark-ers were low, particularly that of tau (see Fig 2). Thiswas the oldest (age, 76.4 years) cluster, with the latestonset (age at onset, 73.6 years) of dementia.

The fifth and sixth clusters were considerably smaller(5 and 1% of clinically diagnosed AD-AD/L subjects,respectively) than the first four. Cluster 5 (AD with highA�1-42 and recent onset [HARO]) comprised cases withparticularly increased levels of A�1-42 (see Fig 2) and rel-atively recent onset. Although its size was insufficient tomake meaningful inferences about genotypic and othercharacteristics, these cases did not appear to have an un-usually high probability of possessing an ApoE4 allele.

Cluster 6 (AD with low A�1-42, high tau, high ubiq-

Š Fig 1.Box plots showing the degree of variations in the mark-ers across diagnosis and among autopsied cases. In these plots,the box represents the interquartile range, which contains50% of the cases. Lines that extend from the box indicate thehighest and lowest values, excluding outliers (0, 1.5–3.0 and*, 3.0 box lengths above the box). A line across the boxindicates the median. AD � Alzheimer’s disease; AD/L �Alzheimer’s disease of Lewy body type.


uitin, and recent onset [ATURO]) comprised only fourcases and was unusual in that it was the only oneshowing, together with low levels of A�1-42 and highlevels of tau, substantially heightened ubiquitin levelsthat were, on average, more than six standard devia-tions greater than the mean (see Fig 2).

If we take membership in any cluster except Cluster2 (control subjects) as an indicator of AD, its sensitivity(ie, ability to detect a true positive case) is 95% (Finns,

94%; Swedes, 97%). Its specificity (ie, ability to cor-rectly identify a true negative), however, was somewhatlower, but 74% (Finns, 75%; Swedes, 71%) of truenegatives would be identified as such. The remaining26% of true negatives fell into some other cluster.Some of these clinically healthy individuals might rep-resent preclinical cases. Interestingly, most of thesecases fell in Cluster 5/subgroup HARO (36.7%) andCluster 4/subgroup LEBALO (33.3%). These twoclusters, 4 (LEBALO) and 5 (HARO), which repre-sented less than 25% of all cases examined, had un-usual marker profiles. Cluster 5 (HARO) cases had thehighest levels of A�1-42 and high levels of tau. Cluster4 (LEBALO) cases had decreased levels of all threemarkers and represented most of the cases of AD withLewy bodies. The CSF marker profiles of Cluster 4suggest that the Lewy body pathology might play a sig-nificant role in the clinical development of the diseasein these patients.

The separation of the five AD clusters from one an-other and from the control cluster can easily be appre-ciated in a three-dimensional plot of the values of thethree molecular markers studied (Fig 3). To classify di-agnosed AD cases into the proposed subgroups, wesought a simple set of rules using the level of only oneindicator protein at any stage in the classification pro-cess. Ideally, it would classify cases with a sensitivityand a specificity of no less than 90% for each categoryand a comparable overall level of correct classification.

Table 3. Fit of Models by Number of Latent Clusters (subtypes)

No. ofClusters LL

No. ofParameters BIC (from LL)

Change inBIC p

2 �1804.6 17.0 3713.8 — —3 �1724.3 22.0 3583.9 �129.9 �0.00014 �1755.6 27.0 3677.3 93.4 �0.00015 �1736.0 32.0 3668.7 �8.6 �0.12616 �1654.9 37.0 3537.3 �131.4 �0.00017 �1650.4 42.0 3559.0 21.7 �0.00068 �1722.0 47.0 3733.0 174 �0.0001

A higher (less negative) log likelihood (LL) indicates a better fit but does not take the number of classes into account. A lower BayesianInformation Criterion (BIC) value indicates a better fit and, the degree of fit being equal, a more parsimonious model, that is, a model usingfewer parameters to explain the observed distribution of the markers. The statistical significance of the change in the BIC from the previousmodel is listed in the last column. With the exception of the nearly similar fit of the four- and five-cluster models, each model represents a quitesignificant change in fit from the preceding one.

Table 4. Intercept, Effects of Age, and Effects of Indicators on Cluster Membership Probabilities

Variables

Cluster

Wald p1 2 3 4 5 6

Intercept �3.65 13.09 1.76 �6.15 2.93 �7.98 56.9 �0.001A�1–42 �0.5957 0.9351 �0.7065 �0.6043 1.9256 �0.9543 288.6 �0.001Tau 0.0592 �1.08 1.32 �1.05 �0.32 1.07 772.9 �0.001Ubiquitin �0.9654 �1.23 �0.63 �1.56 �0.81 5.20 326.9 �0.001Age 0.07 �0.18 �0.01 0.09 �0.04 0.08 57.9 �0.001

Table 5. Intercept for Indicators in Cluster MembershipPrediction Model

Indicators

Intercept

Coefficient Wald p

A�1–42 �0.487 2.090 �0.150Tau �0.887 3.340 �0.068Ubiquitin �0.270 0.411 �0.520

Table 6. Direct Effect of Age on Indicators

Indicators

Effect of Age

Coefficient Wald p

A�1–42 0.008 3.205 �0.073Tau 0.015 4.937 �0.026Ubiquitin 0.018 9.507 �0.002


The algorithm must unambiguously categorize allcases. Figure 4 presents a decision tree based on thealgorithm we derived, based on examination of clustercharacteristics and experimental runs, that came closestto fulfilling these criteria. The respective sensitivitiesand specificities were: AELO: 90%, 92%; ATEO:90%, 95%; LEBALO: 88%, 99%; HARO: 100%,99%; and ATURO: 100%, 100%. Overall, 86% ofcases were classified correctly.

DiscussionCurrently, a major goal of AD research is the develop-ment of neuroprotective therapeutic drugs to inhibitthe underlying pathology, �-amyloidosis, and neurofi-brillary degeneration. Histopathological studies ofadults and on Down’s syndrome subjects have demon-strated that both �-amyloidosis and neurofibrillary de-generation develop several years before any detectableclinical symptoms.22,23 However, these two lesions areseen in AD in widely varying proportions; lumping to-

gether such AD patients in a clinical trial of a neuro-protective drug will make it difficult to evaluate thetherapeutic benefit if the drug is effective toward onlyone subgroup of the disease.

A major hurdle in developing anti-AD drugs hasbeen the lack of means to identify subgroups and thelack of reliable molecular markers of neurodegenerationin living patients. All existing anti-AD drugs were de-veloped based on improvement in clinical symptoms,that is, activities of daily living, cognition, or both.Whether these first-generation anti-AD drugs, com-monly referred to as symptomatic drugs, inhibit thedisease process is unknown. This study demonstratesthat there are various distinct patterns of neurodegen-eration in AD, that is, subgroups that can be identifiedby monitoring CSF levels of A�1-42, tau, and ubiq-uitin. Efficacy of therapeutic drugs might thus be mon-itored using CSF levels of these markers. The inclusionof additional CSF markers, such as abnormal phos-phorylation of tau at different specific sites and differ-

Table 7. Characteristics of Clusters

Characteristics

Cluster

1 2 3 4 5 6

Cluster size (% of sample) 177 (37.8%) 101 (21.6%) 79 (16.9%) 77 (16.5%) 30 (6.4%) 4 (0.9%)Number of AD-AD/L

cases (% of all AD-AD/L cases)

171 (48.4%) 16 (4.5%) 76 (21.5%) 67 (19.0%) 19 (5.4%) 4 (1.1%)

Indicator levelsA�1–42 (pg/ml) 532.5 895.0 490.3 513.3 1191.6 433.8Tau (pg/ml) 737.4 373.3 1089.1 391.6 632.5 1010.5Ubiquitin (ng/ml) 150.2 106.4 172.7 94.0 158.0 670.0

Age, yr 75.4 60.5 70.8 76.4 70.7 75.7Sex

Female 50% 45% 64% 47% 56% 75%Male 50% 55% 36% 53% 44% 25%

ApoE genotype3 � 2 0.6% 8.9% 0.0% 3.9% 13.3% 25.0%3 � 3 24.9% 28.7% 27.8% 31.2% 50.0% 25.0%4 � 2 0.6% 0.0% 5.1% 3.9% 0.0% 0.0%4 � 3 55.9% 15.8% 36.7% 53.2% 20.0% 0.0%4 � 4 17.5% 0.0% 30.4% 5.2% 0.0% 25.0%Unknown 0.6% 46.5% 0.0% 2.6% 16.7% 25.0%

DiagnosisAD 94.9% 12.9% 96.2% 71.4% 63.3% 75.0%AD/L 1.7% 3.0% 0.0% 15.6% 0.0% 25.0%Control 3.4% 84.2% 3.8% 13.0% 36.7% 0.0%

OriginFinland 58.2% 72.3% 67.1% 44.2% 50.0% 50.0%Sweden 41.8% 27.7% 32.9% 55.8% 50.0% 50.0%

Age at dementia onset, yr 71.7 —a 66.6 73.6 71.3b 71.5Duration of dementia, yr 2.6 —a 3.3 2.9 1.6b 1.5Cluster name AELO Control ATEO LEBALO HARO ATUROSensitivity/specificity of

assignment using deci-sion treec

90%/92% 90%/95% 88%/99% 100%/99% 100%/100%

aCluster 2, comprised mostly of control subjects, had no onset data for 80% of its members.bCluster 5 had no onset data for 33% of its subjects (all control subjects), allowing mean age of dementia onset to be greater than mean ageat assessment for the cluster.cThe sensitivity/specificity of the Swedish and Finnish cases separately was similar to those of the total cases.

AD � Alzheimer’s disease; AD/L � Alzheimer’s disease of Lewy body type; ApoE � apolipoprotein E; AELO � AD with low A�1–42, highincidence of ApoE4, and late onset; ATEO � AD with low A�1–42, high tau, and early onset; LEBALO � AD with high incidence of Lewybodies, low A�1–42, and late onset; HARO � AD with high A�1–42 and recent onset; ATURO � AD with low A�1–42, high tau, highubiquitin, and recent onset.


ent fragments of A�, in future studies may identify ad-ditional subgroups of AD. This study should encouragethe pharmaceutical industry to invest in developingdrugs to prevent or inhibit AD.

The most commonly used clinical criteria, the Na-tional Institute of Neurological and Communication

Disorders-Alzheimer’s Disease and Related DisordersAssociation protocol,4 which uses a combination of in-clusion criteria that characterize this particular demen-tia and exclusion criteria to rule out other dementias,has a highly variable accuracy rate of �60 to 90%.24,25

Although in specialized AD centers the accuracy rate is

Fig 2.Cluster profiles of five subgroups of Alzheimer’s disease (AD) named as AD with low A�1-42, high incidence of ApoE4, andlate onset (AELO), AD with low A�1-42, high tau, and early onset (ATEO), AD with high incidence of Lewy bodies, low A�1-42,and late onset (LEBALO), AD with high A�1-42 and recent onset (HARO), and AD with low A�1-42, high tau, high ubiquitin,and recent onset (ATURO) in comparison with the control group (Cluster 2). AD/L � Alzheimer’s disease of Lewy body type;ApoE, apolipoprotein E.


90% or better, it requires high cost, time, and person-nel, and early and preclinical diagnoses cannot bemade. An approach to overcome these problems is theuse of biomarkers in the biological fluids of living pa-tients. Of all the biomarkers reported to date, proteinsassociated with the histopathological hallmarks of AD,the plaques and the neurofibrillary tangles, have beenmost promising. This study shows that CSF levels ofA�1-42, tau, and ubiquitin could diagnose AD in fivedifferent subgroups at sensitivities and specificities ofgreater than 88%, and overall, 86% of cases were clas-sified correctly. This rate of diagnostic accuracy notonly is superior than using any one of these markersindividually or in combination of twos,15 but also ex-ceeds the biomarker criteria of the Consensus Report.26

The role of the physical state of A� in neurotoxicityand AD has been contradictory and the topic of nu-

merous previous studies (for review, see Selkoe27).Both soluble and aggregated (as oligomers or fibrils)forms of A� have been reported to be neurotoxic. InAD, the decreased levels of A� in the CSF are believedto reflect greater polymerization of this peptide intoamyloid in plaques in the brain. In this study, AD sub-groups AELO, ATEO, LEBALO, and ATURO, whichhad decreased levels of A�1-42, and the subgroupHARO, which had an increase in A�1-42 of more thantwo standard deviations, are consistent with the possi-bility of AD cases in whom the A�1-42 is mostly poly-merized and cases in whom it is mostly soluble in thebrain such as at early stages of the disease,28 respec-tively.

A large majority of the AD/L cases are classified inCluster 4/LEBALO. Interestingly, this is the only sub-group of AD in this study that did not have signifi-

Fig 3.Three-dimensional representation of the five Alzheimer’s disease (AD) subgroups (Clusters 1, 3, 4, 5, and 6) and the controlsubjects (Cluster 2). AELO � AD with low A�1-42, high incidence of apolipoprotein E4, and late onset; ATEO � AD with lowA�1-42, high tau, and early onset; ATURO � AD with low A�1-42, high tau, high ubiquitin, and recent onset; HARO � ADwith high A�1-42 and recent onset; LEBALO � AD with high incidence of Lewy bodies, low A�1-42, and late onset.


cantly increased CSF tau levels. A recent study hasshown that �-synuclein, the major protein subunit ofLewy bodies, is a microtubule-associated protein that,like tau, promotes assembly of tubulin into microtu-bules and stabilizes microtubules.29 It is intriguing topostulate that in this subgroup (LEBALO) of AD, thediseased brain responds to the abnormal hyperphos-phorylation of tau, which precedes the formation oftangles and inhibits assembly and depolymerizes micro-tubules30,31 by increasing the production of�-synuclein; the �-synuclein, a phosphoprotein, on hy-perphosphorylation or reaching a critical concentrationin the affected neurons, or both, probably leads to theformation of Lewy bodies. Such a relation betweenneurofibrillary degeneration of the abnormal hyper-phosphorylation of tau and Lewy bodies remains to beinvestigated.

Cluster 6 (ATURO), the smallest of the five sub-groups of AD, showed the most marked decrease inA�1-42 and increase in tau and especially elevated ubiq-uitin levels. This subgroup, otherwise similar to Cluster3 (ATEO) in A�1-42 and tau levels, differed fromATEO as a group with recent onset. It is possible thatthe expression of ubiquitin, which tags abnormal pro-teins for degradation by the nonlysosomal ubiquitinpathway, is highest at the early stages of the pathology.If this proves true, then the early cases of AD, espe-cially preclinical cases, might show relatively increasedlevels of ubiquitin in the CSF.

In agreement with previous studies,32 most caseswith ApoE4�3 in this study classified into one or an-other AD subgroup. The control cases had mostlyE4�2, E3�2, or E3�3 ApoE genotypes. Interestingly,none of the E4�4 patients showed up in the controlgroup by the cluster analysis, confirming E4�4 carriersas the group at highest risk for AD.

In conclusion, based on the CSF levels of proteinmarkers of the histopathological lesions, AD clearly di-vides into at least five subgroups. These subgroupsmight represent different causes and mechanisms ofneurodegeneration. Patients from different subgroupsmight respond differently to different therapeuticdrugs. Consideration of these different subgroups ofAD identifiable by CSF analysis should help in the de-velopment of therapeutic drugs specific to each sub-group of the disease.

Š Fig 4.Decision tree for identifying various subgroups of Alzhei-mer’s disease based on cerebrospinal fluid levels of ubiquitin,A�1-42, and tau (see Fig 2). AELO � AD with low A�1-42,high incidence of apolipoprotein E4, and late onset; ATEO �AD with low A�1-42, high tau, and early onset; ATURO �AD with low A�1-42, high tau, high ubiquitin, and recentonset; HARO � AD with high A�1-42 and recent onset;LEBALO � AD with high incidence of Lewy bodies, lowA�1-42, and late onset.


This study was supported by the New York State Office of MentalRetardation and Developmental Disabilities (K.I., M.F., I.G.-I.), theNIH (National Institute on Aging, AG19158, K.I.), the Instituutvoor de aanmoediging van Innovatie door Wetenschap en Tech-nologie (950309, E.V.), and the Kuopio University Hospital (EVO-5772720, H.S., T.P.).

We are grateful to J. Biegelson and S. Warren for secretarial assis-tance.

References1. Campion D, Dumanchin C, Hannequin D, et al. Early-onset

autosomal dominant Alzheimer disease: prevalence, genetic het-erogeneity, and mutation spectrum. Am J Hum Genet 1999;65:667–670.

2. Iqbal K, Grundke-Iqbal I. Metabolic/signal transduction hy-pothesis of Alzheimer’s disease and other tauopathies. ActaNeuropathol (Berl) 2005;109:25–31.

3. Blennow K, Wallin A, Gottfries CG. Heterogeneity of “proba-ble Alzheimer’s disease.” In: Iqbal K, McLachlan DRC, Win-blad B, Wisniewski HM, eds. Alzheimer’s disease: basic mech-anisms, diagnosis and therapeutic strategies. Chichester, UK:John Wiley & Sons, 1991:21–26.

4. McKhann G, Drachman D, Folstein M, et al, for theNINCDS-ADRDA Work Group under the auspices of Depart-ment of Health and Human Services Task Force on Alzheimer’sDisease. Clinical diagnosis of Alzheimer’s disease. Neurology1984;34:939–944.

5. McKeith IG, Galasko D, Kosaka K, et al, for the Consortiumon DLB International Workshop. Consensus guidelines for theclinical and pathologic diagnosis of dementia with Lewy bodies(DLB). Neurology 1996;47:1113–1124.

6. Mizutani T, Sakata M, Enomoto M, et al. Pathological heter-ogeneity of Alzheimer type dementia. In: Iqbal K, Winblad B,Nishimura T, et al, eds. Alzheimer’s disease: biology, diagnosisand therapeutics. Chichester, UK: John Wiley & Sons, 1997:247–255.

7. Bancher C, Jellinger KA. Neurofibrillary tangle predominantform of senile dementia of Alzheimer type: a rare subtype invery old subjects. Acta Neuropathol (Berl) 1994;88:565–570.

8. Glenner GG, Wong CW. Alzheimer’s disease: initial report ofthe purification and characterization of a novel cerebrovascularamyloid protein. Biochem Biophys Res Commun 1984;120:885–890.

9. Grundke-Iqbal I, Iqbal K, Quinlan M, et al. Microtubule-associated protein tau: a component of Alzheimer paired helicalfilaments. J Biol Chem 1986;261:6084–6089.

10. Grundke-Iqbal I, Iqbal K, Tung YC, et al. Abnormal phosphor-ylation of the microtubule-associated protein (tau) in Alzheimercytoskeletal pathology. Proc Natl Acad Sci U S A 1986;83:4913–4917.

11. Mori H, Kondo J, Ihara Y. Ubiquitin is a component of pairedhelical filaments in Alzheimer’s disease. Science 1987;235:1641–1644.

12. Perry G, Friedman R, Shaw GV. Ubiquitin is detected in neu-rofibrillary tangles and senile plaque neurites of Alzheimer dis-ease brains. Proc Natl Acad Sci U S A 1987;84:3033–3036.

13. Wang GP, Khatoon S, Iqbal K, et al. Brain ubiquitin is mark-edly elevated in Alzheimer disease. Brain Res 1991;566:146–151.

14. Kudo T, Iqbal K, Ravid R, et al. Alzheimer disease: correlationof cerebrospinal fluid and brain ubiquitin levels. Brain Res1994;639:1–7.

15. Hulstaert F, Blennow K, Ivanoiu A, et al. Improved discrimi-nation of AD patients using beta-amyloid(1-42) and tau levelsin CSF. Neurology 1999;52:1555–1562.

16. Hu YY, He SS, Wang XC, et al. Levels of nonphosphorylatedand phosphorylated tau in cerebrospinal fluid of Alzheimer’sdisease patients. Am J Pathol 2002;160:1269–1278.

17. Mirra SS, Heyman A, McKeel D, et al. The Consortium toEstablish a Registry for Alzheimer’s Disease (CERAD). Part II.Standardization of the neuropathologic assessment of Alzhei-mer’s disease. Neurology 1991;41:479–486.

18. Magidson J, Vermunt JK. Latent class models for clustering: acomparison with K-means. Can J Mark Res 2002;20:36–43.

19. Muthen BO. Beyond SEM: general latent variable modeling.Behaviormetrika 2002;29:81–117.

20. Bernstein L, Bradley K, Zarich S. GOLDmineR: improvingmodels for classifying patients with chest pain. Yale J Biol Med2002;75:1–16.

21. Audige L, Hunter J, Weinberg AM, Magidson J, Slongo T.Development and evaluation process of a pediatric long-bonefracture classification proposal. Eur J Trauma 2004;4:248–254.

22. Tagliavini F, Giaccone G, Verga L, Frangione B, Bugiani O.Down syndrome as a key to the time sequence of brain changesin Alzheimer disease. Prog Clin Biol Res 1992;379:143–158.

23. Braak H, Braak E. Frequency of stages of Alzheimer-related le-sions in different age categories. Neurobiol Aging 1997;18:351–357.

24. Galasko D, Chang L, Motter R, et al. High cerebrospinal fluidtau and low amyloid beta42 levels in the clinical diagnosis ofAlzheimer disease and relation to apolipoprotein E genotype.Arch Neurol 1998;55:937–945.

25. Growdon JH. Biomarkers of Alzheimer disease. Arch Neurol1999;56:281–283.

26. Consensus report of the Working Group on: “Molecular andBiochemical Markers of Alzheimer’s Disease.” The Ronald andNancy Reagan Research Institute of the Alzheimer’s Associationand the National Institute on Aging Working Group. Neuro-biol Aging 1998;19:109–116.

27. Selkoe DJ. Alzheimer’s disease: genes, proteins, and therapy.Physiol Rev 2001;81:741–766.

28. Assini A, Cammarata S, Vitali A, et al. Plasma levels of amyloid�-protein 42 are increased in women with mild cognitive im-pairment. Neurology 2004;63:828–831.

29. Alim MA, Ma QL, Takeda K, et al. Demonstration of a rolefor alpha-synuclein as a functional microtubule-associated pro-tein. J Alzheimers Dis 2004;6:435–442.

30. Alonso AC, Zaidi T, Grundke-Iqbal I, Iqbal K. Role of abnor-mally phosphorylated tau in the breakdown of microtubules inAlzheimer disease. Proc Natl Acad Sci U S A 1994;91:5562–5566.

31. Alonso AC, Grundke-Iqbal I, Iqbal K. Alzheimer’s disease hy-perphosphorylated tau sequesters normal tau into tangles of fil-aments and disassembles microtubules. Nat Med 1996;2:783–787.

32. Corder EH, Saunders AM, Strittmatter WJ, et al. Gene dose ofapolipoprotein E type 4 allele and the risk of Alzheimer’s dis-ease in late onset families. Science 1993;261:921–923.


subgroups of alzheimer's disease based on cerebrospinal fluid molecular markers

Documents