Exploring the Genetic Heterogeneity of Alzheimer's Disease: Evidence for Genetic Subtypes

Abstract

Background: Alzheimer's disease (AD) exhibits considerable phenotypic heterogeneity, suggesting the potential existence of subtypes. AD is under substantial genetic influence, thus identifying systematic variation in genetic risk may provide insights into disease origins.

Objective: We investigated genetic heterogeneity in AD risk through a multi-step analysis.

Methods: We performed principal component analysis (PCA) on AD-associated variants in the UK Biobank (AD cases = 2,739, controls = 5,478) to assess structured genetic heterogeneity. Subsequently, a biclustering algorithm searched for distinct disease-specific genetic signatures among subsets of cases. Replication tests were conducted using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset (AD cases = 500, controls = 470). We categorized a separate set of ADNI individuals with mild cognitive impairment (MCI; n = 399) into genetic subtypes and examined cognitive, amyloid, and tau trajectories.

Results: PCA revealed three distinct clusters ("constellations") driven primarily by different correlation patterns in a region of strong LD surrounding the MAPT locus. Constellations contained a mixture of cases and controls, reflecting disease-relevant but not disease-specific structure. We found two disease-specific biclusters among AD cases. Pathway analysis linked bicluster-associated variants to neuron morphogenesis and outgrowth. Disease-relevant and disease-specific structure replicated in ADNI, and bicluster 2 exhibited increased cerebrospinal fluid p-tau and cognitive decline over time.

Conclusions: This study unveils a hierarchical structure of AD genetic risk. Disease-relevant constellations may represent haplotype structure that does not increase risk directly but may alter the relative importance of other genetic risk factors. Biclusters may represent distinct AD genetic subtypes. This structure is replicable and relates to differential pathological accumulation and cognitive decline over time.

Keywords: Alzheimer’s disease; biclustering; genetic risk; genetic subtypes; genotyping.

Figure 1.. Principal component analysis of UKB data restricted to Alzheimer’s disease-associated variants reveals three constellations.

A) Principal component analysis was applied to allele combinations of UK Biobank cases and controls restricted to variants with a p-value<0.05 in the Kunkle et al. Alzheimer’s GWAS [12]. The scatter plot displays participant loading on the first two principal components (PC1 and PC2) and colored by Alzheimer’s disease case-control status. Three distinct clusters, or constellations, are clearly present and each contains a mix of cases and controls. Heatmaps displaying the density of cases and controls in each constellation are also shown to demonstrate that, despite substantial overlap between the groups, there is some offset in the distributions. However, the directional bias is not consistent across constellations. B) Principal component analysis results when variants were restricted across a range of p-value thresholds from the Kunkle et al. Alzheimer’s GWAS [12]. The scatter plots are colored by constellation labels defined at the p<0.05 threshold. Participants from all three constellations are highly mixed when all variants (p<1.0) are included. The constellation structure begins to emerge along the first principal component at p<0.25, and further separate along the second principal component at p<0.05.

Figure 2.

A) Hex bin plot of loadings across allele combinations on first principal component from the PCA on UKB cases and controls restricted to variants with a p-value<0.05 in the Kunkle et al. Alzheimer’s GWAS [12]. Allele combinations are ordered by chromosome and base position across the x-axis. Color represents density of data points that fall within a given hex. B) The location of peak loadings on PC1 is shown in greater detail along with gene annotations from the UCSC database. The peak loadings occur in the region of 17q21.31, overlapping with a known region of extended LD surrounding the MAPT locus.

Figure 3.. Bicluster traces of observed data versus label-shuffled data.

The disease-related signal-strength associated with the remaining UKB Alzheimer’s cases relative to controls is plotted on the y-axis. At each iteration, allele combinations and cases that contribute least to this difference are removed. The proportion of remaining cases is shown on the x-axis. The red trace represents the original data and black traces represent label-shuffled data, corresponding to a null distribution. A red dot indicates the iteration with the maximum separation between cases and controls (ignoring signal in the first iterations), and is used to define the bicluster. The sharper peak of constellation 1 indicates that this bicluster has more distinct boundaries, whereas the bicluster in constellation 2 has “fuzzier” boundaries as indicated by the broad yet lower peak.

Figure 4.. Network plot of gene sets enriched among bicluster genes.

Network visualization of gene set enrichment results for biclusters 1 and 2. First, separate GWAS compared individuals in each bicluster to all controls from the same constellation in which the bicluster was found (e.g., bicluster 1 cases versus all controls from constellation 1). SNPs that were nominally associated with a given bicluster (i.e., positive regression coefficient and p<0.05) were mapped to genes based on position. Over-representation analysis of gene lists for each bicluster was used to identify gene sets associated with each bicluster. Nodes represent significantly enriched gene sets (pFDR<0.05) with color indicating the bicluster they are associated with. Edges represent the overlap in genes belonging to gene sets using a threshold of 0.5. Gene sets were clustered based on overlaps and automatically annotated based on the descriptions of each gene set cluster.

Figure 5.. ADNI data projected along principal components defined in UKB data replicate distribution of disease-relevant constellations.

Principal components were recalculated in UKB data using only AD-associated variants common to both datasets. Participants from both datasets were then plotted by participants loadings on the first two principal components. Colors represent constellations, UKB participants are plotted with squares, and ADNI participants are plotted with circles.

Figure 6.. Replication of disease-specific biclusters in ADNI data.

The similarity between case-control labels of individuals in the test set (ADNI) and the most frequent label among nearest neighbors in the training set (UKB) was used to assess replication of biclusters. Only individuals belonging to the constellation in which the given bicluster was found were considered. The fraction of individuals from the training set (i.e., UKB) considered as nearest neighbors is plotted along the x-axis. The y-axis in Panels (A) and (D) shows the average fraction of nearest neighbors with a matching label. The red line in Panels (A) and (D) shows values from the original data, while the black lines show values from label-shuffled trials drawn from the null-distribution. A trial-wise mean and variance can be defined from the null distribution to normalize values, with the associated z-scores shown in Panels (B) and (E). A rank-normalization of the scores compared to the null distribution is shown in Panels (C) and (F).

Figure 7.. Association of bicluster groups in ADNI MCI sample.

ADNI individuals diagnosed with MCI were assigned to bicluster groups (bicluster 1, bicluster 2, or non-bicluster). Differences in longitudinal cognitive and biomarker trajectories between groups were tested with age × group interactions in linear mixed effects models. Model predicted values are shown in the figure. A) Cognition was measured with the Preclinical Alzheimer’s Cognitive Composite (PACC). B) Amyloid was measures using florbetapir PET. C) Phosphorylated tau (p-tau) was measured using CSF p-tau.