Confronting Complexity in Late-Onset Alzheimer Disease: Application of Two-Stage Analysis Approach Addressing Heterogeneity and Epistasis

Abstract

Common diseases with a genetic basis are likely to have a very complex etiology, in which the mapping between genotype and phenotype is far from straightforward. A new comprehensive statistical and computational strategy for identifying the missing link between genotype and phenotype has been proposed, which emphasizes the need to address heterogeneity in the first stage of any analysis and gene-gene interactions in the second stage. We applied this two-stage analysis strategy to late-onset Alzheimer disease (LOAD) data, which included functional and positional candidate genes and markers in a region of interest on chromosome 10. Bayesian Classification found statistically significant clusterings for independent family-based and case-control datasets, which used the same five markers in LRRTM3 as the most influential in determining cluster assignment. In subsequent analyses to detect main effects and gene-gene interactions, markers in three genes—PLAU, ACE and CDC2—were found to be associated with late-onset Alzheimer disease in particular subsets of the data based on their LRRTM3 multilocus genotype. All of these genes are viable candidates for LOAD based on their known biological function, even though PLAU, CDC2 and LRRTM3 were initially identified as positional candidates. Further studies are needed to replicate these statistical findings and to elucidate possible biological interaction mechanisms between LRRTM3 and these genes.

Introduction

Alzheimer’s disease (AD; MIM: 104300) is a neurodegenerative disorder characterized clinically by a decline in two or more areas of cognition, one of which is usually episodic memory, in the absence of acute causes (Pericak-Vance MA and Haines JL, 2002). Presenting symptoms range from memory impairment to visuospatial disorientation, language impairment, depression and psychotic episodes. AD is defined pathologically by the presence of two abnormalities in the cerebral cortex. The first is senile plaques that have an amyloid beta (Aβ) protein core, and the second is neurofibrillary tangles, which contain the microtubule-associated protein tau (Goedert M, 1999; Wisniewski T et al., 1993). It remains controversial whether the plaques and tangles are themselves pathogenic or whether they are merely “tombstones” of other pathogenic processes (Glabe C, 2000). Only a weak link between plaque load and severity of illness has been found, while the load of neurofibrillary tangles might be more strongly correlated with severity (Guillozet AL et al., 2003; Mufson EJ et al., 1999). Also, both plaques and tangles have been found in normal older adults, leading many to suggest that these abnormalities are secondary effects arising from the true pathological mechanisms underlying AD.

In addition, Lewy bodies, which contain fibrils of aggregated, insoluble alpha-synuclein (McKeith I et al., 2004), have been observed in up to 20% of AD cases in the substantia nigra (which is characteristic of PD) and elsewhere in the brain (Ditter SM and Mirra SS, 1987; Growden JH, 1995; McKeith IG et al., 1996). A growing body of literature suggests substantial overlap among AD, dementia with Lewy bodies, and Parkinson Disease (Metzler-Baddeley C, 2007; Meyer JS et al., 2007). It is possible that the developments of Aβ plaques, neurofibrillary tangles and Lewy bodies have common physiological pathways. However, it is also possible each one of these features is a distinct trait, suggesting that AD is a heterogeneous trait better defined as the coincident state of having both plaques and tangles. Likewise, AD with PD could then be better described as the concomitance of the three traits for plaques, tangles and Lewy bodies. To the extent that each of these traits is likely to have its own distinct genetic etiology, trait heterogeneity can be manifest statistically in a way similar to genetic heterogeneity.

While AD can occur as early as the third decade of life (Cruts M et al., 1995), it most commonly occurs after the sixth decade. The age of onset for late-onset Alzheimer disease (LOAD) is generally defined to be after age 60 or 65 but extends into the ninth decade. The only confirmed gene conferring risk for LOAD is apolipoprotein E (APOE). It is estimated that at least fifty percent of the genetic effect of LOAD remains unexplained (Daw EW et al., 2000; Roses AD et al., 1995; Slooter AJC et al., 1998). Over 115 LOAD candidate genes have been tested and have generated a positive main effect, but all except APOE have failed to be consistently replicated (Bertram L et al., 2007a). While the initial reports might have been false positive findings, alternatively, these inconsistencies could be indicative of heterogeneity and/or environmental interactions across the entire phenotype. Reported differences of incidence and prevalence between ethnic and gender groups are also indicative of interactions with environment and/or genetic background. The possibility of gene-gene interactions has been explored only superficially and most recently (Bagli M et al., 2000; Martin ER et al., 2005; Mateo I et al., 2006).

Late Onset Alzheimer Disease is just one example of a complex disease, in which traditional statistical methods of analysis such as linkage and association have failed to identify main effect genes. Among the possible reasons for this failure are false positives due to population stratification and true differences in genetic etiology between study populations (Hirschhorn JN et al., 2002). In addition, while a small number of supervised computational methods exist for discovering gene-gene interactions, the power of these methods drops dramatically when locus or trait heterogeneity is present (Ritchie MD et al., 2003). There is extensive evidence that heterogeneity exists in LOAD, within and between positional candidate regions (Liang X et al., 2007a; Lin PI et al., 2006b; Olson JM et al., 2002). Current statistical approaches for detecting genetic heterogeneity, such as the admixture test (Ott J and Hoh J, 2003; Smith CAB, 1963), are neither sensitive nor powerful and can merely account for, not resolve, any underlying heterogeneity.

It is possible that phenotypic data could be utilized to improve the performance of these methods in the face of genetic or trait heterogeneity by facilitating heuristic stratification of data. For instance, age of onset data was used to stratify AD patients, leading to the detection of association with the apolipoprotein ε4 allele in late-onset and sporadic cases (Corder EH et al., 1993; Saunders AM et al., 1993; Strittmatter WJ et al., 1993). However, for most diseases, particularly neurological diseases, little detailed phenotypic data has been consistently collected in combination with genotypic data. Postmortem histological data are rare for cases, even rarer for controls, and neuroimaging can be expensive and challenging with mentally ill patients. It is for these reasons that an unsupervised method, such as the Bayesian Classification cluster analysis method investigated elsewhere (Thornton-Wells TA et al., 2006) or the Grade of Membership latent class analysis approach (Corder EH and Beaumont H, 2007; Corder EH and Woodbury MA, 1993), which does not rely on phenotypic data, would be valuable to mine potentially heterogeneous genotypic data as a means of data stratification and hypothesis generation.

A comprehensive two-step approach to analysis has been proposed in which heterogeneity is first addressed and then main effects and interactions are subsequently investigated in the more homogeneous subsets discovered in the first stage (Thornton-Wells TA et al., 2004). Here we present an application of this two-stage approach to a LOAD dataset where cluster analysis is first used to uncover heterogeneity and to subdivide the data into more homogeneous groups. Then in the second stage, traditional linkage and association tests are used to detect main effects and a computational data reduction method is used to investigate gene–gene interactions within each of the subgroups.

Subjects and Methods

Specifics of Late-Onset Alzheimer Disease Dataset

The late-onset Alzheimer Disease dataset includes samples obtained by (1) the Collaborative Alzheimer Project (CAP) (Vanderbilt University, Duke University and UCLA), (2) the Indiana University (IU) Alzheimer Disease Center National Cell Repository, and (3) the National Institute of Mental Health (NIMH) Alzheimer Disease Genetics Initiative. Although the IU and NIMH samples represent a rich resource for generating hypotheses, they are used by multiple groups (including CAP). In contrast, the CAP-collected sample represents an independent set of families that can be used to confirm and extend initial findings. All subjects are Caucasian Americans. Written consent was obtained from all participants in agreement with protocols approved by the Institutional Review Board at each contributing institution. Alzheimer Disease was diagnosed according to the NINCDS-ADRDA criteria (McKhann G et al., 1984). Age of onset was recorded as the age at which the first symptoms were noted by the participant or family member. Only subjects with an age of onset of 65 or greater were included in this late-onset dataset.

From these three samples, two datasets were assembled—a family-based dataset and a case-control dataset. The family-based dataset, derived from all three ascertainment sources, consists of 654 families with 1422 subjects with possible, probable or definite LOAD and 744 cognitively normal elderly individuals. Of these families, 328 contain a total of 1279 discordant sibling pairs (DSPs), in which one sibling is affected with LOAD and the other is unaffected. The CAP dataset also includes a clinic-based unrelated case-control sample of 451 cases with possible, probable or definite LOAD and 699 cognitively normal elderly controls who were either spouses of AD patients or subjects recruited from outpatient clinics at the participating institutions.

Markers previously genotyped in over 25 positional or functional candidate genes and a region of interest (ROI) on chromosome 10 were initially included in the datasets, totally 185 markers. Markers with more than 10% missing data and subjects with more than 15% missing data were removed. The data were then processed to remove markers with more than 10 percent missing data and subjects with more than 15 percent missing data. This was an iterative process that resulted in a dataset with 148 markers in 22 genes on eight chromosomes—82 markers, covering 18 genes, that were common to both datasets. Thirteen of the 22 genes had the same markers genotyped in both the case-control and the family-based datasets. Of the remaining nine, GAPDH was only genotyped in the case-control dataset, and A2M, LRP1 and OLR1 were only genotyped in the family-based dataset. In addition, some genes had more markers genotyped in the case-control dataset (GAPDHS, UBQLN1) or the family-based dataset (ACE, LRRTM3, VR22). In results tables, when a marker is not present in one of the datasets, the appropriate cell is grayed out to indicate this. Many of these markers were chosen and genotyped prior to information about linkage disequilibrium was available from the HapMap Project (Gibbs RA et al., 2003), and in order to minimize loss of information, sometimes even when multiple markers were known to be in high LD, they were all genotyped.

Most of the functional candidate genes chosen are purported to have some role in LOAD through their involvement in the processing of amyloid precursor protein (APP; MIM: 104760), the secretion of its product, Aβ, and/or the phosphorylation of tau or regulation of microtubules within neurons. Table 1 lists alphabetically the 22 genes genotyped in one or both of the samples, along with their full names. Information about known function for these and other genes that have been reportedly associated with Alzheimer Disease can be found at ALZGENE.org.

Table 1.

Genes Covered by Markers Genotyped in One or Both Samples

Symbol	Location	Name	OMIM ID	Gene ID
A2M	12p13.3-p12.3	Alpha-2-macroglobulin	103950	2
A2MP	12p13.3-p12.3	Alpha-2-macroglobulin pseudogene	-	3
ACE	17q23.3	Angiotensin 1 converting enzyme (petidyl-dipeptidase A)	106180	1636
AGT	1q42-q43	angiotensinogen	106150	183
APOE	19q13.2	Apolipoprotein E	107741	348
CDC2	10q21.1	Cell division cycle 2	116940	983
COG2	1q42.2	component of oligomeric Golgi complex 2	606974	22796
GAPDH	12p13	Glyceraldehydes-3-phosphate dehydrogenase	138400	2597
GAPDHS	19q13.1	Glyceraldehydes-3-phosphate dehydrogenase, spermatogenic	609169	26330
IDE	10q23-q25	Insulin degrading enzyme	146680	3416
LIPC	15q21-q23	Lipase, hepatic	151670	3990
LRP1	12q13-q14	Low-density lipoprotein receptor-related protein 1	107770	4035
LRRTM3	10q21.3	Leucine-rich repeat transmembrane neuronal 3 protein	-	347731
LTA	6p21.3	Lymphotoxin alpha (TNF superfamily, member 1)	153440	4049
OLR1	12p13.2-p12.3	Oxidized density lipoprotein (lectin-like) receptor 1	602601	4973
PLAU	10q24	Urokinase-type plasminogen activator	191840	5328
PPM1H	12q14.1-q14.2	Protein phosphotase 1H (PP2C domain containing)	-	57460
PZP	12p13-p12.2	Pregnancy-zone protein	176420	5858
TNF	6p21.3	Tumor necrosis factor (TNF superfamily, member 2)	191160	7124
TNFRSF6/FAS	10q24.1	Necrosis factor receptor superfamily member 6	134637	355
UBQLN1	9q21.2-q21.3	Ubiquilin 1	605046	29979
VR22/CTNNA3	10q22.2	Catenin (cadherin-associated protein), alpha 3	607667	29119

Statistical Analysis

A comprehensive, two-stage approach to analysis was performed in which heterogeneity was first investigated in the dataset and then main effects and gene-gene interactions were investigated among the resulting subsets or clusters of data (Figure 1). Although all of the markers in the dataset had been previously tested for main effects and some even for interactions, this testing was performed at different time points over the past 10 years and, therefore, the samples on which they were tested vary to different degrees from the sample being analyzed in the current study. It is for this reason that a preliminary analysis of the complete datasets was performed prior to the two-stage analysis, using all the main effect and interaction-detection methods proposed for the subsets of data.

Figure 1.

Flowchart of Comprehensive Two-Stage Analysis Approach Applied to LOAD Data.

Analysis of deviations from Hardy-Weinberg equilibrium (HWE) and linkage equilibrium were tested using the Haploview program (Barrett JC et al., 2005) on the complete case-control and family-based datasets. Population stratification was investigated using the STRUCTURE program (Pritchard JK et al., 2000). The Bayesian Classification method (Cheeseman P and Stutz J, 1996; Hanson R et al., 1991), previously investigated in simulation studies described elsewhere (Thornton-Wells TA et al., 2006), was used to detect heterogeneity. For the family-based and case-control data, separately, the affected individuals in the dataset were subjected to cluster analysis (based on their marker genotypes) to create more homogeneous subsets of affected individuals. Each cluster subset was then recombined with the entire group of unaffected individuals from the respective dataset for subsequent analysis of main effects and interactions.

Bayesian Classification aims to find the most probable clustering of data given the data and the prior probabilities. In the case of genotypic data, prior probabilities are based on genotype frequencies. The most probable clustering of data is determined from two posterior probabilities. The first involves the probability that a particular individual belongs to its ‘assigned’ cluster, or otherwise stated as the probability of the individual’s multilocus genotype, conditional on it belonging to that cluster, with its characteristic genotypes. The second posterior probability involves the probability of a cluster given its assigned individuals, or otherwise stated as the probability of the cluster’s characteristic genotypes, conditional on the multilocus genotypes of the individuals assigned to that cluster.

In actuality, individuals are not ‘assigned’ to clusters in the hard classification sense but instead in the fuzzy sense they are temporarily assigned to the cluster to which they have the greatest probability of belonging. Thus, each individual has its own vector of probabilities of belonging to each of the clusters. A ranked listing is produced of all loci in the dataset with their corresponding normalized “attribute influence” values (ranging between 0 and 1), which provide a rough heuristic measure of relative influence of each locus in differentiating the classes from the overall dataset. This information that can then be used to more directly stratify affected (and/or unaffected) individuals, for instance, by using the top n most influential loci identified, and to enable meaningful interpretation of the clustering result.

For the family-based data, two-point heterogeneity lod score (HLOD) linkage analysis using FASTLINK and HOMOG (Ott J, 1999) and the pedigree disequilibrium test (PDT) for detecting main effect association (Martin ER et al., 2000; Martin ER et al., 2001)—were performed. While the cluster subsets are more homogeneous than the overall dataset, at least with regard to the markers with high influence values, it is not possible to know whether heterogeneity still exists in any one of those subsets with regard to the other loci. Therefore, HLOD scores were used both in the overall dataset and in the cluster subsets. The PDT is based on the transmission disequilibrium test and is robust to population stratification. For the linkage analysis, both recessive and dominant disease models were tested, and the maximum heterogeneity lod score, which is the highest lod score found for either model under the full range of possible theta values, is reported. For the case-control data, a chi-square test of independence was used to detect main effect associations. In each case, a genotype-based model was tested.

For both the family-based and the case-control datasets, the multifactor dimensionality reduction (MDR) method was used to detect gene-gene interactions (Hahn LW et al., 2003; Moore JH, 2007; Moore JH et al., 2006; Ritchie MD et al., 2001). MDR is a nonparametric and genetic model-free method that uses constructive induction to change the representation space of the data to make nonadditive interactions easier to detect using classification methods such as naïve Bayes or logistic regression. This is accomplished by pooling high-risk genotype combinations into one group and low-risk genotypes into another group thus defining a new single constructed variable. Here, we used a simple probabilistic classifier that is similar to naïve Bayes to model the relationship between variables constructed with MDR and our discrete endpoint of disease status. Using 5-fold cross validation, we measured the average balanced prediction accuracy (across the five cross-validation intervals) of every possible combination for the best one-, two- and three-way MDR models (Moore JH et al., 2006; Velez DR et al., 2007). For each of the one-locus, two-locus and three-locus combinations, the ‘best’ MDR model was chosen as the one with the best average balanced prediction accuracy. All ‘best’ MDR models were evaluated for statistical significance using an omnibus permutation test with 1000 permutations.

Since MDR works by comparing the ratio of affected to unaffected individuals but does not account for intrafamilial correlations only discordant sibpairs (DSPs) are used in the analysis of family-based data. Two datasets were created—the first with only one randomly chosen DSP per family (designated ‘1DSP’) and the second with all individuals who are part of one or more DSPs in a family (designated ‘AllDSPs’). An alternative method, the MDR-PDT, which uses the geno-PDT instead of balanced accuracy as the test statistic, does allow for inclusion of additional family members with the potential for increased power (Martin ER et al., 2006). However, the version of MDR-PDT that is currently available does not incorporate cross-validation, and as a result, there is some concern over whether its results are generalizable. In addition, since the majority of the families are made up of discordant sibpairs, not extended families, the MDR method is likely to produce very similar results to that of the MDR-PDT. A newer version of MDR-PDT that does include cross-validation is under development.

For each of the ‘best’ two- and three-marker MDR models achieving prediction accuracy of 55 percent or greater, the markers in those MDR models were used in logistic regression analyses to further characterize the underlying statistical models. For the case-control data, a logistic regression analysis was performed in SPSS, and for the family-based data, a multivariate logistic regression method, which controls for intrafamily correlation, was implemented in SAS (Martin ER et al., 2006; Siegmund KD et al., 2000) and applied to all discordant sibpairs.

Results

Analysis of Complete Datasets

Linkage and association results for the complete dataset are presented in Table 2. Linkage analysis detected the known effect of APOE (HetLOD = 7.96), but no other marker achieved a score above 1.5 (nominal significance). The PDT detected the known association with APOE (χ²=98.388, df=2, p<0.001), two other substantial effects—one in A2M (χ²=6.772, df=1, p=0.009) and one in ACE (χ²=7.104, df=1, p=0.008)—and 10 other marginal effects (χ²> 4.5, p<0.05). Analysis using the chi-square test of independence on the complete case-control dataset detected the known association with APOE (χ²=171.62, df=5, p<0.001) and seven other marginal effects in CDC2, VR22, LRRTM3 and GAPDH (χ²>6.2, p<0.05).

Table 2.

Main Effect Analysis Results for Complete Dataset.

					PDT		Pearson
Chrom	Gene	Marker	Minor Allele Freq	2-Pt Linkage Max HetLOD	Chi-Square	p-value	Chi-Square	p-value
10	CDC2	rs2448347	0.45	0.090	0.059	0.808	6.581	0.037
10	VR22	rs1786927	0.41	0.000	0.116	0.734	7.035	0.030
10	VR22	rs2441718	0.31	1.407	0.312	0.577	8.553	0.014
10	VR22	rs2456737	0.15	1.038	1.143	0.285	6.222	0.045
10	VR22	rs7909676	0.48	1.068	4.540	0.033	0.349	0.840
10	LRRTM3	rs942780	0.19	0.000	2.770	0.096	7.586	0.023
10	LRRTM3	rs1925617	0.44	0.343	1.604	0.205	6.465	0.039
10	PLAU	rs2227568	0.16	0.000	5.170	0.023	1.254	0.534
10	PLAU	rs4065	0.43	0.000	3.152	0.076	1.028	0.598
10		rs4933194	0.49	0.052	4.676	0.031
12	GAPDH	rs1060621	0.20				7.188	0.027
12	A2M	rs3832852	0.16	0.011	6.772	0.009
12	LRP1	rs1800154	0.34	0.000	4.017	0.045
12	LRP1	rs9669595	0.35	0.003	4.599	0.032
12	LRP1	rs7956957	0.37	0.000	4.059	0.044
17	ACE	rs4291	0.37	0.000	7.104	0.008	2.828	0.243
17	ACE	rs4646994	0.47	0.000	5.481	0.019	1.180	0.554
17	ACE	rs4343	0.48	0.000	4.516	0.034	3.175	0.204
17	ACE	rs4353	0.48	0.000	4.887	0.027
17	ACE	rs4978	0.48	0.000	6.503	0.011
19	APOE	rs440446	(E4) 0.35	7.963	98.388	< 0.001	171.62	< 0.001

MDR gene-gene interaction analysis was performed on both the complete family-based and complete case-control datasets. MDR detected the main effect of APOE in all three datasets (Case-Control, 1DSP and AllDSPs) by choosing APOE as the best one-locus model with perfect (5 of 5) cross-validation consistency and by including APOE in the best two- and three-locus models as well, all of which were statistically significant (p < 0.05). To give MDR the opportunity to detect other effects without interference of the APOE effect, we excluded APOE from the datasets and re-ran the analysis. In these subsequent analyses, none of the best one-, two- or three-locus models achieved average balanced prediction accuracies of greater than 53 percent or cross-validation consistency values of more than 2, and none were statistically significant (p > 0.20).

Detection of Heterogeneity

Bayesian Classification was applied to each of the complete case-control and family-based datasets. Only affected individuals were used in the cluster analysis. The family-based dataset produced twelve clusters, and the case-control dataset produced four clusters. To focus on heterogeneity that might be present in both datasets, we took the top 30 markers from each dataset with the highest influence values and selected those markers present in both datasets (31 markers). A marker’s influence value provides a rough heuristic measure of the relative influence that marker had in differentiating the clusters from the overall dataset. Then, we performed the cluster analysis again using only those 31 markers. This second analysis produced 15 clusters in the family-based dataset and 6 clusters in the case-control dataset.

After again ranking the markers by their influence values, it was discovered that the top 5 markers were the same in both datasets (Table 3). Therefore, in one final attempt to produce a clustering that was similar across both datasets and produced a more reasonable number of clusters, which could be subsequently investigated for main effects and interactions, we performed the cluster analysis using only these top 5 markers. This third and final round of clustering produced 5 clusters in the family-based dataset and 3 clusters in the case-control dataset (Table 4). Upon closer inspection, two of the five clusters in the family-based dataset contained only seven and five affected subjects, respectively, making subsequent analysis of those clusters inadvisable due to reduced sample size. Thus, only three resulting clusters were analyzed in each of the datasets.

Table 3.

Top Five Highest-Influence Markers from Second-Round of Cluster Analysis

Chrom	Gene	Marker	FamInfluValue	CCInfluValue
10	LRRTM3	rs1925632	0.938	0.792
10	LRRTM3	rs1952060	0.623	0.944
10	LRRTM3	rs2147886	1.000	1.000
10	LRRTM3	rs2251000	0.940	0.834
10	LRRTM3	rs2764807	0.673	0.890

Table 4.

Distribution of Affected Individuals in Final Clustering Results

	Number of Affecteds
Cluster	Family-Based Data	Case-Control Data
0	673	215
1	480	157
2	257	79
3	7	-
4	5	-

Permutation testing was performed to determine whether this final clustering was statistically significant. In the family-based data, the clustering results produced an average class strength value of −4.34 (p<0.002) and an average cross-class entropy value of 4.00 (p<0.002). In the case-control data, the clustering results produced an average class strength value of −2.71 (p<0.002) and an average cross-class entropy value of 4.43 (p<0.012). Thus, for each of the datasets, the clustering results were significant at our predetermined alpha of ten percent, as suggested by our simulation studies reported elsewhere (Thornton-Wells TA et al., 2006).

Since the top 5 markers were all in the same gene (LRRTM3), we investigated whether they were in linkage disequilibrium (LD) with each other and thus were encoding a single haplotype block. LD analysis using Haploview indeed showed that the five markers and four additional flanking markers were all in high LD (r²≥ 0.63) with each other, and it showed the first four markers to be in a haplotype block (Figure 2). Furthermore, inspection of the multi-locus genotypes at the top 5 markers across the three clusters in each dataset showed that a separate multi-locus genotype was predominant in each of the three clusters and that these three multi-locus genotypes were the same across the case-control and family-based datasets (Table 5). Investigation of population stratification using STRUCTURE provided no evidence of structure in the data; thus, it is unlikely that these clustering results are related to this phenomenon.

Figure 2.

Linkage Disequilibrium Plot of Top 5 High-Influence Markers in (a) Family-Based Dataset and (b) Case-Control Datasets. The top five markers are: rs1925632, rs1952060, rs2147886, 2251000, and rs2764807. Numbers in each square represent pair-wise R² values. The markers in bold are those in a haplotype block (all pair-wise R² > 0.80).

Table 5.

Predominant Genotypes for the Top Five High-Influence Markers by Cluster

	Cluster
Marker	0	1	2
rs1925632	A/C	C/C	A/A
rs1952060	C/T	C/C	T/T
rs2147886	C/T	C/C	T/T
rs2251000	A/G	A/G	A/G
rs2764807	C/T	C/C	T/T

Detection of Main Effects in Subsets of Data

For each of the three clusters (0, 1, 2) in the family-based dataset, linkage analysis and association analysis by the PDT were conducted. For each of the three clusters in the case-control dataset, the chi-square test of independence was performed. Since the three clusters in the family-based and case-control datasets correspond exactly, analysis results are presented in the following subsections by cluster number. In each of the clusters, the top five high-influence markers and some flanking markers from the LRRTM3 and VR22 genes were found significant by one or more statistical tests. These results are expected and uninteresting and, therefore, they are excluded from the following text and tables of results.

Cluster 0 Results

Table 6 presents results for cluster 0 for all markers with significant scores on at least one of the three statistical tests performed (two-point linkage, PDT and Pearson chi-square). APOE produced significant results for the linkage HetLOD and the PDT and Pearson chi-square statistics. The PDT also found very marginal association with 13 other markers in LRP1, ACE and the ROI on chromosome 10 (χ²>3.9, p<0.05). Five markers in the PLAU, IDE, A2MP and ACE genes showed at least marginal association (χ²>8, p<0.05) according to the Pearson chi-square test of independence. When comparing results across the family-based and case-control datasets for cluster 0, two markers were found significant (p < 0.05)—rs4291 in the ACE and rs440446 in APOE.

Table 6.

Main Effect Analysis Results for Cluster 0.

					PDT		Pearson
Chrom	Gene	Marker	Minor Allele Freq	2-Pt Linkage Max HetLOD	Chi-Square	p-value	Chi-Square	p-value
10		rs6482044	0.33	0.609	4.898	0.027
10		rs7089698	0.45	0.156	4.902	0.027
10	PLAU	rs2227568	0.16	1.506	0.6	0.439	9.118	0.028
10	IDE	rs7099761	0.47	1.197	0.509	0.475	10.815	0.013
10	IDE	rs1544210	0.47	0.034	0.006	0.939	19.355	< 0.001
10		rs2255901	0.49	0.000	4.447	0.035
12	A2MP	rs34362	0.27	0.012	0	1.000	8.182	0.042
12	LRP1	rs1800181	0.34	0.000	5.433	0.020
12	LRP1	rs1800154	0.34	0.015	4.306	0.038
12	LRP1	rs1800165	0.34	0.014	4.976	0.026
12	LRP1	rs9669595	0.35	0.091	5.371	0.021
12	LRP1	rs7956957	0.37	0.000	3.918	0.048
17	ACE	rs4291	0.37	0.000	6.339	0.012	8.414	0.038
17	ACE	rs4295	0.36	0.000	4.831	0.028
17	ACE	rs4646994	0.47	0.000	5.236	0.022	1.126	0.771
17	ACE	rs4343	0.48	0.000	4.67	0.031	2.242	0.524
17	ACE	rs4978	0.48	0.000	5.657	0.017
19	APOE	rs440446	(E4) 0.35	3.752	66.373	< 0.001	118.292	< 0.001

Cluster 1 Results

Table 7 presents significant results for cluster 1. Linkage analysis did not produce any HetLOD scores above 3 except for the expected LRRTM3 and VR22 markers. APOE produced significant results for the PDT and Pearson chi-square statistics. The PDT also showed strong association with rs2251101 in IDE (χ²=7.39, p<0.007) and rs3832852 in A2M (χ²=6.674, p<0.010) and marginal association with 8 additional markers in CDC2, PLAU, IDE, ACE, GAPDHS and the ROI on chromosome 10 (χ²>3.8, p<0.05). The chi-square test of independence found very strong association (χ²>9, p<0.01) with five markers in AGT, IDE and GAPDH. In addition, 18 other markers in AGT, UBQLN1, CDC2, PLAU, IDE, GAPDH, A2MP, LIPC and ACE showed marginal association (χ²>6, p<0.05).

Table 7.

Main Effect Analysis Results for Cluster 1.

					PDT		Pearson
Chrom	Gene	Marker	Minor Allele Freq	2-Pt Linkage Max HetLOD	Chi-Square	p-value	Chi-Square	p-value
1	AGT	rs2148582	0.43	0.091	0.135	0.713	11.144	0.004
1	AGT	rs5051	0.43	0.025	0.064	0.800	12.809	0.002
1	AGT	rs1326886	0.10	0.073	0.040	0.742	8.652	0.013
9	UBQLN1	rs2781003	0.37	0.036	1.607	5.000	6.305	0.043
9	UBQLN1	rs2780995	0.37				6.794	0.033
9	UBQLN1	rs12344615	0.20				7.624	0.022
9	UBQLN1	rs11140213	0.20	0.528	1.986	0.159	8.023	0.018
10		rs1319013	0.45	0.370	4.306	0.038
10	CDC2	1920	0.23	2.092	0.405	0.524	11.509	0.021
10	CDC2	rs2448341	0.34	0.446	4.469	0.035	6.269	0.044
10	CDC2	rs2448347	0.45	1.270	0.540	0.463	7.161	0.028
10	PLAU	rs1916341	0.44	1.099	1.047	0.306	6.591	0.037
10	PLAU	rs2227566	0.47	1.306	1.848	0.174	6.866	0.032
10	PLAU	rs2227568	0.16	0.291	6.470	0.011	4.622	0.099
10	PLAU	rs4065	0.43	1.202	1.594	0.207	7.260	0.027
10	IDE	rs2251101	0.30	0.000	7.388	0.007	2.068	0.356
10	IDE	rs1832196	0.14	0.270	5.028	0.025	7.803	0.020
10	IDE	rs7076966	0.45	0.257	0.091	0.763	9.976	0.007
10		rs4372378	0.46	0.407	5.704	0.017
12	GAPDH	rs7307229	0.01				8.618	0.013
12	GAPDH	rs1060620	0.20				15.79	< 0.001
12	GAPDH	rs1060619	0.24				14.489	0.001
12	A2M	rs3832852	0.16	0.084	6.674	0.010
12	A2MP	rs16918212	0.10	0.000	0.010	0.920	6.051	0.049
12	A2MP	rs17804080	0.09	0.000	0.095	0.758	8.722	0.013
15	LIPC	rs6083	0.36	0.000	0.087	0.768	8.748	0.013
17	ACE	rs4646994	0.47	0.000	0.778	0.378	6.004	0.050
17	ACE	rs4343	0.48	0.000	3.341	0.068	8.903	0.012
17	ACE	rs4353	0.48	0.000	4.265	0.039
17	ACE	rs4978	0.48	0.000	3.991	0.046
19	GAPDHS	rs4806173	0.38	0.250	4.464	0.035	2.907	0.234
19	APOE	rs440446	(E4) 0.35	2.158	40.994	< 0.001	91.857	< 0.001

When comparing across the family-based and case-control datasets for cluster 1, three markers—rs2448341 in CDC2, rs1832196 in IDE, and rs440446 in APOE—were found significant (p<0.05). Worth noting, one marker in PLAU (rs2227568) was significant according to the PDT in the family-based dataset, and in the case-control dataset that marker is in a haplotype block with rs1916341, which was significant by the Pearson chi-square test. Additionally, there are two markers in ACE (rs4353 and rs4978) that were significant by the PDT in the family-based dataset but are not present in the case-control dataset. In the family-based dataset, these markers are in high LD (r²≥0.92) with two other markers (rs4646994 and rs4343) that were found significant by the Pearson chi-square test in the case-control dataset.

Cluster 2 Results

Table 8 presents significant results for cluster 2. Linkage analysis produced no HetLOD scores greater than 3 except for markers in LRRTM3 and VR22. APOE produced significant results for the PDT and Pearson chi-square statistics. The PDT found strong association with rs2227568 in PLAU (χ²>7, p<0.009) and marginal association (χ²>4.9, p<0.05) with one marker in CDC2 and one marker in LRP1. The Pearson chi-square test found association with two markers in A2MP—rs17804080 (χ²=10.42, p=0.005) and rs16918212 (χ²=7.03, p=0.03). When comparing across the family-based and case-control datasets for cluster 2, only APOE was found significant.

Table 8.

Main Effect Analysis Results for Cluster 2 Family-Based Dataset.

					PDT		Pearson
Chrom	Gene	Marker	Minor Allele Freq	2-Pt Linkage Max HetLOD	Chi-Square	p-value	Chi-Square	p-value
10	CDC2	rs7919724	0.32	0.650	4.955	0.026	4.437	0.109
10	PLAU	rs2227568	0.16	1.000	7.042	0.008	0.642	0.725
12	A2MP	rs16918212	0.10	0.000	0.690	0.406	7.025	0.030
12	A2MP	rs17804080	0.09	0.000	0.352	0.553	10.425	0.005
12	LRP1	rs1800154	0.34	0.000	5.141	0.023
19	APOE	rs440446	0.35	0.567	36.984	< 0.001	67.132	< 0.001

Detection of Gene-Gene Interactions in Subsets of Data

For each of the three clusters in both the family-based and case-control datasets, an MDR gene-gene interaction analysis was conducted. APOE and the top 5 high-influence markers, plus the four flanking markers in linkage disequilibrium with those top markers, dominated the best MDR models. To allow other effects to be detected over these known effects, these ten markers were excluded and the MDR analyses were repeated. Cross-validation (CV) consistency was computed as the number of times (out of 5) that the reported best model was the best in the fold, or split, of the data. The average (across all five cross-validation intervals) of the balanced prediction accuracy and its corresponding significance level (p-value) was also determined.

Cluster 0 Results

For cluster 0, in the family-based 1DSP dataset, the best one-locus MDR model (CVC=4, p=0.10) was rs4291 in ACE and the best two-locus model (CVC=3, p=0.17) was rs4291 in ACE and rs7909676 in VR22, which is not in LD (r² ≤ 0.01) with any LRRTM3 marker in the dataset. A statistically significant full factorial model was fit to the cluster 0 family-based dataset using rs4291 and rs7909676 (χ² = 19.264, df=3, p = 0.0002), but the individual parameter estimates indicate that the significant effect in the model is primarily coming from marker rs4291 (Table 9). The heterozygote and the A/A homozygote for rs4291 increased risk for disease by 2.066 (p = 0.0106).

Table 9.

Logistic Regression Results for Cluster 0 Family-Based Data Using Markers from Significant Two-Locus MDR Model

						95% Odds Confidence	Ratio Limits
Factor	Minor Allele Freq	χ2	df	p Value	Odds Ratio	Lower	Upper
VR22(rs7909676)	0.48	3.68	1	0.0550	1.916	0.986	3.723
ACE(rs4291)	0.37	6.52	1	0.0106	2.066	1.184	3.606
rs7909676 * rs4291		0.35	1	0.5500	0.881	0.582	1.335

Cluster 1 Results

For cluster 1, in the case-control dataset, the best one-locus model (CVC=3, p=0.11) was rs3096244 in VR22, which is not in LD (r²≤ 0.04) with any LRRTM3 marker in the dataset, and the best two-locus MDR model (CVC=2, p=0.08) involved rs3096244 in VR22 and rs4343 in ACE. In the 1DSP family-based dataset, the best two locus model (CVC=3, p=0.13) was rs2255901 in the chromosome 10 ROI and rs922347 in VR22, which is not in LD with any LRRTM3 marker in the dataset.

A statistically significant full factorial model was fit to the cluster 1 case-control dataset using rs3096244 in VR22 and rs4343 in ACE from the best two-locus MDR model (χ² = 20.646, df=3, p<0.001) (Table 10). Both markers displayed significant main effects, and the interaction effect, which had the opposite effect on risk, was also significant. At marker rs3096244 in VR22, the heterozygote and T/T homozygote decreased risk by an odds ratio of 0.464, and at marker rs4343 in ACE, the heterozygote and G/G homozygote decreased risk by an odds ratio of 0.425. However, in reference to any genotype combination that included the A/A homozygote for rs3096244 or the A/A homozygote for rs4343, those same genotypes when considered together actually increased risk by 1.696. Using the two markers included in the best two-locus MDR model for the 1DSP family-based dataset, logistic regression was used to fit a full factorial model to the data. However, the full model was not statistically significant (χ² = 1.4917, df=3, p>0.68), nor were any of its factors.

Table 10.

Logistic Regression Results for Cluster 1 Case-Control Data Using Markers from Significant Two-Locus MDR Model

						95% Odds Confidence	Ratio Limits
Factor	Minor Allele Freq	χ2	df	p Value	Odds Ratio	Lower	Upper
VR22(rs3096244)	0.38	14.4	1	< 0.001	0.464	0.309	0.694
ACE(rs4343)	0.48	14.3	1	< 0.001	0.425	0.270	0.671
rs3096244 * rs4343		9.07	1	0.003	1.696	1.199	2.400

Cluster 2 Results

For cluster 2, the best one-locus MDR model (CVC = 5, p < 0.04) in the case-control dataset was rs10997591 in VR22, which is not in LD (r²≤ 0.12) with any LRRTM3 marker in the dataset, and the best one-locus MDR model (CVC = 5, p = 0.08) in the 1DSP family-based dataset was rs11816558 in the ROI on chromosome 10.

Discussion

The innovative two-stage approach we have applied is based on evidence that heterogeneity is a confounding factor in statistical genetic analysis and that the Bayesian Classification clustering method is capable of uncovering such heterogeneity in genotypic data. We primarily included markers in functional candidate genes (as opposed to markers from a genome wide scan, for instance), which increase the likelihood that any resulting cluster subsets will have a biological basis. An LD block in LRRTM3 was utilized to cluster cases into subsets, in which significant main effects were subsequently demonstrated. Just as with any other statistical result, one must ask whether it is also biologically meaningful. Does LRRTM3 serve as an effective proxy for genetic background? Does LRRTM3 directly or indirectly interact with the other genes that are associated with disease in cluster subsets? Perhaps the pertinent interactions involve VR22, which is the larger gene in whose intron LRRTM3 resides. Ultimately, whether the statistical results reported here are biologically relevant will be determined by whether they can be replicated, and, more importantly, whether functional molecular studies can demonstrate the biological plausibility of such interactions.

VR22 or CTNNA3 (catenin, alpha 3; MIM#607667) is a binding partner of beta-catenin (Janssens B et al., 2001), which interacts with presenilin 1. Presenilin 1 interacts with the gamma-secretase involved in processing the amyloid precursor protein (APP), and its mutations have been associated with increased levels of amyloid beta 42 (Citron M et al., 1997; Duff K et al., 1996; Qian S et al., 1998), the primary component of senile plaques found in Alzheimer disease. VR22 has been associated with the LOAD endophenotype of plasma Aβ42 levels (Ertekin-Taner N et al., 2003), and we have previously reported an interaction between VR22 and APOE with regard to risk of LOAD (Martin ER et al., 2005). Bertram et al. also reported an association between VR22 and LOAD, but it was not contingent on APOE status (Bertram L et al., 2007b).

Leucine-rich containing proteins, like LRRTM3, are involved in protein-protein interactions, and the family of leucine-rich repeat transmembrane proteins (LRRTMs) are involved in many cellular events during nervous system development and disease (Lauren J et al., 2003). Of particular relevance to Alzheimer disease pathology, LRRTM3 is highly expressed in the adult mouse hippocampus, in the granular layer of the dentate gyrus (Lauren J et al., 2003). Tau-mediated neurodegeneration in this area is thought to play a role in Alzheimer disease progression (Shahani N et al., 2006).

Recent evidence is mounting in support of an alternative hypothesis for Alzheimer disease pathology, which implicates cell cycle reactivation as a key early event that precedes and possibly is causally related to tau, APP phosphorylation and apoptotic cell death (Andorfer C et al., 2005; McPhie DL et al., 2003; Yang Y et al., 2006). Amyloid precursor protein has been purported to regulate activation of neuronal cell cycle proteins (McPhie DL et al., 2003); therefore, hypothetically, mutations in VR22 could indirectly affect cell cycle activation, through interactions with APP (by way of beta-catenin and presenilin 1). Additionally, since LRRTM3 is thought to be involved in neuronal development in some of the key areas that are later targets of neuronal cell death in Alzheimer disease, perhaps LRRTM3 is being re-activated in some way that facilitates the cell cycle re-entry of neurons. Thus, it would be interesting to learn whether VR22 and/or LRRTM3 are differentially expressed in the brains of AD patients versus controls.

For every cluster, the main effect and interaction subset analyses showed LRRTM3 markers exhibiting strong effects. This is an expected result. Essentially, the cluster analysis is introducing a bias toward association in the cluster subsets with regard to the five LRRTM3 markers it used to create the clusters. Likewise, flanking or nearby markers in LRRTM3 and the larger gene, VR22, within which LRRTM3 resides, might demonstrate effects that could be attributed to linkage disequilibrium with the LRRTM3 haplotype block. In contrast, however, any association found with a gene not linked to the five LRRTM3 markers would be evidence of heterogeneity or a gene-gene interaction. Looking across all the main effect and interaction analyses, there are a few genes for each cluster that deserve further investigation in relation to their LRRTM3 multilocus genotype (Table 5).

Complete Dataset

The analysis of the complete family-based and case-control datasets found only the APOE marker rs440446 was significant in both the case-control and family-based datasets. Many of the markers that were found significant by at least one main effect statistical test in either the complete case-control or complete family-based datasets were also significant in the analysis of specific subsets produced by the Bayesian Classification analysis. Table 11 indicates which of the markers initially found significant in the complete datasets were also found significant in one or more of the cluster subsets. Excluding markers from LRRTM3 and VR22, two markers were found significant across all three clusters as well as in their respective complete dataset—APOE marker rs440446 and PLAU marker rs2227568. Interestingly, the effect of APOE was less in each of the cluster subsets than it was in the complete datasets, perhaps simply due to smaller sample sizes and more unbalanced data.

Table 11.

Cluster Subset Results for Markers Found Significant in Complete Dataset. Marks in a cluster column indicate that the marker was found significant by at least one main effect test (linkage, PDT or Pearson chi-square) in that cluster subset.

			Cluster
Chrom	Gene	Marker	0	1	2
10	CDC2	rs2448347		x
10	PLAU	rs2227568	x	x	x
10	PLAU	rs4065		x
10		rs4933194
12	GAPDH	rs1060621
12	A2M	rs3832852		x
12	LRP1	rs1800154	x		x
12	LRP1	rs9669595	x
12	LRP1	rs7956957	x
17	ACE	rs4291	x
17	ACE	rs4646994	x
17	ACE	rs4343	x	x
17	ACE	rs4353		x
17	ACE	rs4978	x	x
19	APOE	rs440446	x	x	x

In contrast, the PLAU marker effects were enhanced in the cluster subsets. The PLAU marker rs2227568 is approximately 6.26 Mb away from the nearest genotyped VR22 marker and it exhibits no LD (r²=0) with any of the VR22 or LRRTM3 markers. Therefore, it is unlikely that the consistency of the PLAU marker’s results can be attributed to the LRRTM3 effect. PLAU (urokinase-type plasminogen activator; MIM#5328) converts plasminogen to plasmin, and plasmin is involved in the processing of the amyloid precursor protein and in the degradation of amyloid-beta (Finckh U et al., 2003). The PLAU marker rs2227564 is a C/T missense polymorphism that has been associated with plasma amyloid-beta-42 levels and with LOAD in a German sample (Finckh U et al., 2003) and in a United States Caucasian sample (Ertekin-Taner N et al., 2005). However, at least two subsequent studies have failed to replicate these results—in an Italian sample (Bagnoli S et al., 2005) and in a Scottish and Swedish sample (Blomqvist ME et al., 2004). Most recently, PLAU was found to be associated with LOAD only in APOE-ε4 non-carriers in an Italian sample (Pesaresi M et al., 2007).

Cluster Subsets

For cluster 0, one gene (ACE) showed interesting results in main effect and interaction analyses. ACE (angiotensin 1 converting enzyme; MIM#106180) has been shown to inhibit the aggregation of amyloid beta by degrading amyloid beta 40 into less toxic products (Hu J et al., 1999; Hu J et al., 2001). Studies have reported conflicting results regarding the association between the ACE insertion/deletion polymorphism and LOAD, with the risk allele reported sometimes as the insertion and sometimes as the deletion. Such ‘flip-flop’ associations might be due to multilocus effects, in which the risk allele is inversely correlated with another risk allele at another locus (Lin PI et al., 2006a). A recent analysis of our family dataset using the MDR-PDT suggested an interaction between LRRTM3, ACE and possibly A2M (Hahn LW et al., 2006). The marker rs4291 was significant by its PDT and Pearson chi-square statistics and appeared in the best one- and two-locus MDR models for the 1DSP family-based dataset. This two-locus MDR model was confirmed by logistic regression to be largely a main effect of rs4291. Five other ACE markers—rs4295, rs4311, rs4646994, rs4343 and rs4978—which were all in LD with the former marker and/or each other, were also significant by their PDT statistics.

For cluster 1, three genes (PLAU, ACE and CDC2) showed interesting results for main effect and/or interaction analyses. In PLAU, marker rs2227568 was significant by its HetLOD score and is in LD with PLAU marker rs1916341, which was significant by its PDT chi-square statistics. In ACE, markers rs4646994 and rs4343, which are in LD with each other, were significant by their Pearson chi-square statistics, and rs4343 appeared in the best two-locus MDR model in the case-control dataset, which was confirmed by logistic regression to have both a main and interactive effect involving rs4343. In addition, ACE markers rs4353 and rs4978, which are in LD with rs4646994, were found significant by their PDT chi-square statistics.

CDC2 (cell division cycle 2; MIM#116940) is a kinase involved in the abnormal phosphorylation of tau and the aggregation of tau into paired helical filaments (Pei JJ et al., 2006), which are present in the neurofibrillary tangles of Alzheimer disease. CDC2 has been associated with AD and with increased levels of tau in cerebrospinal fluid (Johansson A et al., 2003; Johansson A et al., 2005). However, we were not able to replicate these findings in a previous analysis of our complete dataset (Liang X et al., 2007b). In cluster 1, the CDC2 marker rs2448341 was significant by its PDT and Pearson chi-square statistics. Another CDC2 marker, rs2448347, which is in LD with rs2448341, was also significant by its Pearson chi-square statistic.

In cluster 2, there were no additional genes that showed evidence for association in both the case-control and family-based datasets. This subset was the smallest and most unbalanced from each of the case-control and family-based clusters, and it is possible that its overall size and/or the extent of the imbalance between affecteds and unaffecteds made these analyses too underpowered to detect an effect, if it were there. It is also possible that no interactions exist with the cluster 2 LRRTM3 genotypes and the other markers included in the datasets.

Methodology

Previously conducted simulation studies of the Bayesian Classification method reported elsewhere (Thornton-Wells TA et al., 2006) used markers that had no linkage disequilibrium (LD) with each other, while the current application involved markers with considerable LD, comprising multiple haplotype blocks. The clustering method chose to focus on a set of markers in LRRTM3 that were in high LD with each other to cluster affected subjects into more homogeneous subsets. This is very similar to simple stratification according to a single genetic locus or haplotype block in a candidate gene—the type of analysis that is often performed for known genetic risk factors like APOE. However, the clustering did not merely stratify by haplotype but instead clustered individuals based on their similarity across the five genotypes in the haplotype block. For instance, in cluster 1, the predominant multilocus genotype combination (see Table 5) was found in only 35 percent of the individuals in that cluster. Most remaining individuals in cluster 1 had genotype combinations, which differed from the predominant genotype at either rs1952060 and/or rs2251000, and yet the clustering method was able to determine that they were more similar to each other than they were to the groups of individuals in clusters 0 and 2. Such pattern recognition is not trivial, hence the plethora of clustering algorithms designed for the task.

No discussion about a large data analysis project such as this would be complete without mention of the multiple-testing problem. As one increases the number of tests performed, the likelihood of generating false positive results also increases, beyond the per-comparison significance level (alpha) established at the beginning of the study. Correction using either Bonferoni or False Discovery Rate (Benjamini Y and Hochberg Y, 1995; Storey JD, 2002) results in statistical significance only for APOE marker rs440446 in the overall dataset and also for some LRRTM3 markers in the cluster subsets. Since our predominant goal was not to miss any real effects, which we could subsequently investigate further, such caution at the expense of power would be imprudent. In addition, since we know there is considerable LD among our markers, the assumption that all the tests are independent is not valid. We would, in fact, expect that two markers in LD with each other would frequently produce similar results, in excess of how often two independent markers should do so. Furthermore, since all of the markers tested were chosen because they are functional and/or positional candidates for late-onset Alzheimer disease, the likelihood that a positive result is true is higher than it would be if the markers were chosen at random, for example in the case of a genomic screen. Finally and perhaps most importantly, it should be noted that we have tested two independent datasets, which serve as one test and one replication dataset, and are focusing only on those effects that were found significant (at the per-comparison level of 0.05) in both datasets. Thus, we have further reduced the chance that such a statistically significant result is a false positive.

Conclusions

Common diseases with a genetic basis are likely to have a very complex etiology, in which the mapping between genotype and phenotype is far from straightforward. A new comprehensive statistical and computational strategy for identifying the missing link between genotype and phenotype has been proposed. Numerous examples of heterogeneity and gene-gene or gene-environment interactions support the theoretical basis for such an approach, which emphasizes the need to address heterogeneity in the first stage of any analysis (Thornton-Wells TA et al., 2004). Uncovering any heterogeneity that might exist in a dataset removes a formidable source of noise, affording main effect and interaction analysis methods the best opportunity to detect effects that might be present only on particular genetic backgrounds or in individuals with particular environmental exposure(s). It is important to keep in mind that while whole genome studies are producing vastly greater amounts of genotypic data, such a brute force method is still subject to effect masking by heterogeneity and interactions. Thus, it will be important to determine how well methods for dissecting heterogeneity can scale up to such studies.

It is a reality that currently a majority of genetic studies, particularly those involving neurological diseases, do not have substantial phenotypic data available, even though the quality and volume of genotypic data might be excellent. Many factors, including cost, feasibility (invasiveness), and technical limitations (reliability and interpretation) of phenotyping technologies, make the collection of rich phenotypic data more challenging. Given the lack of methods for dissecting heterogeneity that do not rely on substantial phenotypic data, an ‘unsupervised’ clustering method—Bayesian Classification—was used to detect multilocus genotype patterns that might underlie or be a proxy for genetic heterogeneity, and possibly by extension, trait heterogeneity.

This method found statistically significant clusterings for both the family-based and case-control datasets, which used the same five markers as their most influential in determining cluster assignment. These markers were all in LRRTM3 and were in high linkage disequilibrium with each other. Each of the three resulting clusters could be characterized by their multilocus genotypes at these five markers, and the same genotypes defined the clusters in both the family-based and case-control data. In subsequent analyses to detect main effects and gene-gene interactions, markers in three genes—PLAU, ACE and CDC2—were found to be associated with late-onset Alzheimer disease in particular subsets of the data based on their LRRTM3 multilocus genotype. All of these genes are viable candidates for LOAD based on their known biological function, even though PLAU, CDC2 and LRRTM3 were initially identified as positional candidates. Further studies are needed to replicate these statistical findings and to elucidate possible biological interaction mechanisms between LRRTM3 and these genes.

Over the past twenty years, advances in genotyping technology have far outpaced those in statistical and computational methods for analyzing genetic data. Likewise, geneticists have given much less attention to phenotyping technologies. To most effectively leverage the massive amounts of genotypic data being produced, we must have comparably rich datasets of phenotypic information available for mapping genotypes to phenotypes. Thus, going forward, genetic studies will need to increasingly focus time and resources on collecting phenotypic data that can refine definitions or subcategories of traits or diseases and can serve as endophenotypes, which are more likely to have simple etiologies and to directly map to specific genetic markers.