Genome-wide Network-assisted Association and Enrichment Study of Amyloid Imaging Phenotype in Alzheimer’s Disease

Abstract

Background:

The etiology of Alzheimer’s disease remains poorly understood at the mechanistic level, and genome-wide network-based genetics have the potential to provide new clues of the disease mechanisms.

Objective:

To explore the collective effects of multiple genetic association signals on an AV-45 PET measure, which is a well-known Alzheimer’s disease biomarker, by employing a network-assisted strategy.

Method:

First, we took advantage of dense module search algorithm to identify modules enriched by genetic association signals in a protein-protein interaction network. Next, we performed statistical evaluations to the modules identified by dense module search, including a normalization process to adjust the topological bias in the network, a replication test to ensure the modules were not found by chance, and a permutation test to evaluate unbiased associations between the modules and amyloid imaging phenotype. Finally, topological analysis, module similarity tests and functional enrichment analysis were performed for the identified modules.

Results:

We identified 24 consensus modules enriched by robust genetic signals in a genome-wide association analysis. The results not only validated several previously reported AD genes (APOE, APP, TOMM40, DDAH1, PARK2, ATP5C1, PVRL2, ELAVL1, ACTN1 and NRF1), but also nominated a few novel genes (ABL1, ABLIM2) that have not been studied in Alzheimer’s disease but have shown associations with other neurodegenerative diseases.

Conclusion:

The identified genes, consensus modules and enriched pathways may provide important clues to future research on the neurobiology of Alzheimer’s disease and suggest potential therapeutic targets.

1. INTRODUCTION

Alzheimer’s disease (AD) is an irreversible, neurodegenerative dementia characterized by gradual memory loss and other cognitive impairments. In 2017, AD became the sixth leading cause of 121,404 death in the US according to the record of official death certificates [1]. Currently, the number of people affected by AD is about 50 million and is expected to be over 152 million by 2050 [2]. Genetic factors play an important part in AD with heritability estimates up to 79% [3]. To identify common variants associated with AD, efforts have been focused on conducting genome-wide association (GWA) studies [4–10] in the last few years. However, most variants identified so far explain only a small proportion of the disease inherited risk [11]. A possible explanation is that real but individually moderate associations may be lost due to the stringent significance level after multiple comparison correction for a large number of tests conducted. Therefore, new strategies are required to improve the power of GWA analysis.

It has been commonly accepted that genes are not randomly distributed in regard of their biological function, but instead clustered in common biological pathways and networks [12,13]. Thus, the recognition of genes directly or indirectly interacting in pathway or complex network is gradually emerging [14,15]. Network-based approaches, which are not limited to sort the genes into not fully annotated predefined gene sets, have been proved successful in the identification of subnetworks of markers with higher reproducibility and higher predictability than individual markers [16]. They have been widely applied to the analysis of high-throughput expression data for various diseases [17]. More recent studies adopted protein-protein interaction networks (PPIN) to the analysis of GWA data to identify genome-wide enriched pathways [18–24]. In combination with the biological knowledge, they are of great help to interpret the function of the genetic variants, and to further understand the molecular mechanism influencing the traits of interest [25].

In the current study, we integrated GWA studies and PPIN to identify modules that contribute to an AD biomarker. Briefly, we aimed to (1) perform a dense module search (DMS) within the PPIN to identify modules enriched with genes associated with an amyloid imaging phenotype that is an AD biomarker; (2) estimate the association of the modules with the phenotype across another independent sample; (3) examine the role of the same module in permuted datasets which were generated by swapping the disease labels; and (4) complete similarity tests and functional enrichment analysis to assess whether the identified modules have common biological function. Our analysis yielded 24 modules including 156 genes, which formed a highly-connected and statistically significant subnetwork. The genes, modules and enriched pathways identified in this analysis are promising targets for further studies of the mechanism underlying AD.

2. MATERIALS AND METHOD

Data used in the preparation of this article were obtained from the ADNI database (http://adni.loni.usc.edu/). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD).

As shown in Fig. 1, we first randomly split the subjects into two groups (G₁, G₂) while maintaining the same gender and diagnosis distributions in each group. Then, we performed a DMS algorithm [26] on G₁ to identify modules enriched with phenotype-associated genes. After that, significant modules generated by G₁ were assessed whether they were significant in the independent samples G₂. Next, we used a permutation-based association test on G₂ to evaluate empirically whether the significant modules identified in both G₁ and G₂ were associated with the amyloid imaging phenotype. Finally, we evaluated the selected promising modules for centrality measurements, similarity tests and functional enrichment analysis.

Figure 1.

Workflow of network-assisted strategy to identify modules associated with AV-45 measure, an Alzheimer’s disease biomarker.

2.1. Genotype Data and Quality Control

The genotype data of the ADNI GO/2 cohort were collected by either the Illumina 2.5M array or the Illumina OmniQuad array [27]. The single nucleotide polymorphism (SNP) data we used were from both arrays for the present study.

A consistent quality control process described in Li et al. [28] was performed on these genotype data using the PLINK [29] (http://pngu.mgh.harvard.edu/purcell/plink/). SNPs were excluded from the GWA analysis if they could not satisfy any of the following condition: (1) no SNPs from chromosome 0 and chromosomes 22–25 (14,247 SNPs were excluded); (2) call rate per SNP ≥95%; (3) minor allele frequency ≥5% (1,845,510 SNPs were excluded based on criteria 2–3); and (4) Hardy-Weinberg equilibrium test of p ≥10⁻⁶ using control subjects only (198 SNPs were excluded). Participants were excluded from the analysis if any of the following condition was not met: (1) call rate per participant ≥90% (0 participants were excluded); (2) gender check (1 participant was excluded); and (3) identity check for related siblings (8 sibling pairs and 1 sibling triplet were identified with PI_HAT > 0.45, one participant was selected randomly from each family). Population stratification analysis was performed using complete linkage agglomerative clustering based on pairwise identity-by-state distance. We excluded 89 study participants who did not cluster with the CEU HapMap samples who are primarily of European ancestry (non-Hispanic Caucasians). After quality control, 1,079 participants each with 563,980 SNPs remained available for the subsequent analysis.

2.2. Subjects and Quantitative Traits

We focused this study on an AV-45 amyloid PET measure called SUMMARYSUVR_WHOLECEREBNORM for these ADNI-GO/2 participants, which is a florbetapir composite standardized uptake value ratio (SUVR) created by averaging across the 4 cortical regions (frontal regions, anterior/posterior cingulate regions, lateral parietal regions, and lateral temporal regions) and dividing this average by the reference region whole cerebellum [30].

For the present work, we concentrated our analyses on subjects with both genotype data and AV-45 PET baseline data available. This sample G₀ (N₀ = 774) included five diagnostic groups: 155 cognitively normal, 72 significant memory concern, 278 early mild cognitive impairment, 144 late mild cognitive impairment, and 125 Alzheimer’s disease subjects. To facilitate validation of the findings, we randomly split the 774 participants into two groups: G₁ (N₁=387) and G₂ (N₂=387), while maintaining the same distributions of gender and diagnosis in each group.

2.3. Integration of Genome-wide Association Analysis with Human Protein-Protein Interaction Network

SNP-trait association was evaluated by PLINK [29], using an additive genetic model with age, gender and diagnosis as covariates. SNPs from GWA analysis were further assigned to specific genes according to their genomic coordinates. A SNP was assigned to a gene if it was located within the gene using the hg19 genome build, which generated a total of 18,614 genes. Then, the most significant SNP’s P-value on a gene was used as its gene-wise P-value. Finally, after removing isolated nodes and self-interaction edges, a node-weighted PPIN with 11,960 nodes and 163,307 edges was constructed by mapping the gene-wise P-values on the PPIN. The comprehensive human PPIN was downloaded from the Protein Interaction Network Analysis platform [31] (http://cbg.garvan.unsw.edu.au/pina/).

2.4. Dense Module Search and Assessment

A DMS algorithm [26] was introduced to our analysis on the discovery set G₁ to search for dense modules by the node-weighted PPIN. Briefly, a seed module only had a seed gene at the beginning and started to grew by adding the neighbor nodes with an increment of $z_m$. $z_m$ is a module score computed for the current module. Module growth is stopped until there is no increment greater than 10% of the current module score. Specifically, $z_m=\sum z_i/\sqrt{k}$, where k is the number of genes for the current module. And $z_i$ was computed by $z_i= {\varnothing }^{-1}\left(1-p_i\right)$, where ${\varnothing }^{-1}$ is the inverse standard normal cumulative distribution function.

To determine how likely the identified modules could be found by chance, a background distribution of module scores $z_m\left(\pi \right)$ were created by computing $z_m$ through selecting the same number of genes randomly from the entire network for 100,000 times. And then the module scores $z_m$ were standardized by $z_n=(z_m-mean\left(z_m\left(\pi \right)\right)/SD\left(z_m\left(\pi \right)\right)$, where SD is standard deviation. Normalized module scores $z_n$ were transformed to P-values by $P_{z_n}=1-\varnothing \left(z_n\right)$, where $\varnothing$ is the standard normal cumulative density function. This analysis calculates test statistics $P_{z_n}$ for the modules defined in the discovery dataset G₁ and then for the same modules in the replication dataset G₂.

To further evaluate whether the enriched modules were significantly associated with the amyloid imaging phenotype, we performed 1,000 permutation tests of the validation dataset G₂ by randomly shuffling the phenotypes and genotypes. For each permutation data set, we calculated $z_m$ of the same modules searched in the real data set G₂. Thus 1,000 instances of $z_m$ were acquired to get the empirical normal distribution with mean $\mu$ and standard deviation $\sigma$. Unlike the original DMS algorithm, the empirical P-values for the association of the modules with the amyloid imaging phenotype was estimated by $P_{eval}=1-F\left(z_m;\mu ,\sigma \right)$, where F is the normal cumulative distribution function. The final candidate modules were those modules passed the tests: (1) $P_{z_n}<0.05$ in both G₁ and G₂ dataset; and (2) $P_{eval}<0.05$ after Bonferroni correction. A final subnetwork was then constructed for further functional analysis by merging candidate modules.

2.5. Module Similarity Tests and Functional Enrichment Analysis

To understand the topological property of the final identified subnetwork, genes in the subnetwork were ranked by some centrality measures, including degree centrality, betweenness centrality, closeness centrality, and page-rank centrality.

A biological process is a cooperation of a set of genes. If two or more modules consist of similar sets of genes, those modules might have a similar biological function or be related in the biological process somehow. We adopted Jaccard similarity coefficient [32] (JSC) to quantitatively measure the similarity between two different modules. $\text{JSC}\left(\text{A},\text{B}\right)=\left|A\cap B\right|/\left|A\cup B\right|$, where $\left|A\cap B\right|$ represents the number of shared genes between module A and module B, $\left|A\cup B\right|$ represents the number of total genes in module A and module B. All pairs of significant modules were scanned.

We performed KEGG [33] pathway enrichment analysis for each of the identified significant modules by using the web-based tool Enrichr [34] (http://amp.pharm.mssm.edu/Enrichr). The enrichment significance for each KEGG term was computed by ‘fuzzy’ enrichment analysis and only terms with an adjusted P-value of less than 0.01 in at least one module were retained.

3. RESULTS

3.1. Significant Signals Identified by Genome-wide and Gene-based Association Analysis

GWA analysis detected some significant SNPs associated with the amyloid imaging phenotype. As we can see from Fig. 2, four SNPs (rs4420638, rs769449, rs2075650, rs157582) in G₁ and one SNP (rs4420638) in G₂ reached our genome-wide significance (P<8.87×10⁻⁸) after Bonferroni correction. In particular, rs4420638 was replicated in these two independent datasets. After GWA analysis, we conducted the gene-based association analysis for G₁ and G₂ separately using gene’s smallest P-value. Overall, 3 genes (APOE, TOMM40, PVRL2) and 2 genes (APOE, TOMM40) were identified to reach the significance level (P<2.69×10⁻⁶) after Bonferroni correction in G₁ and G₂ respectively.

Figure 2. Manhattan plot of the genome-wide association analysis (GWA) of AV-45 PET measure.

(a) Manhattan plot of GWA in G₁; (b) Manhattan plot of GWA in G₂. 563,980 SNPs were tested for association with Aβ burden of participants from group G₁ and G₂ separately under an additive model, with age, gender, and diagnosis as covariates. The blue line indicates the significance level P = 10⁻⁵ and the red line is the genome-wide significance level P = 8.87×10⁻⁸ (0.05/563,980) after Bonferroni correction in our analysis.

3.2. Modules Associated with Amyloid Imaging Phenotype

By performing DMS with each gene as a seed in the node-weighted PPIN, a total of 11,782 modules were initially obtained for the discovery dataset G₁ after removing the replicated modules. For each module in G₁, we evaluated whether the module contains a proportion of significant genes rather than randomly expected and whether the module can be found in another independent dataset G₂. Of the 11,782 modules identified in G₁, 8,630 ones were significant; and 94 of those were also significant in G₂ after Bonferroni correction ($P_{z_n}<0.05$ in both G₁ and G₂). Then, a permutation test (N =1000) was conducted to evaluate whether the enriched modules were highly related to the amyloid imaging phenotype. 24 modules with 156 non-redundant genes (shown in Table 1) were enriched for phenotype-associated signals after Bonferroni correction ($P_{eval}<0.05$. Therefore, these results provided an independent replication for our module identification results and showed significant enrichment with association signals to the amyloid imaging phenotype.

Table 1.

24 significant modules enriched for amyloid imaging phenotype-associated signals.

No.	#G	Pz1	Pz2	P	Start Node	Module Genes
1	13	1.19E-13	2.61E-09	3.46E-08	TAX1BP1	TAX1BP1;ARHGAP1;CALCOCO2;CNTN2;GABRR1;IKBKE;PARK2; STARD13;APP;ATP5O;ITPR1;TOMM40;APOE
2	13	2.89E-15	7.87E-09	3.78E-08	HSDL2	HSDL2;EEF2K;IKBKE;NRF1;PARK2;ABCA3;ABLIM2;ACTG1; ADCY9;DAGLB;DDAH1;TOMM40;APOE
3	11	8.91E-11	3.57E-07	4.78E-08	RNF141	RNF141;SMURF1;TRIM24;ACAP2;ACTR1A;APP;ARHGAP15; CBX1;CDK14;U2AF1;APOE
4	15	4.78E-12	3.38E-07	7.13E-08	PSKH1	PSKH1;CDC6;HSP90AA1;RDH13;SNTA1;AARS;ABCA1;ABCC4; ACTG1;ACTN1;APP;CDK14;FBXO18;TOMM40;APOE
5	11	3.57E-10	1.32E-06	1.29E-07	RET	RET;CASP3;CBL;CBLC;DOK5;DOK6;PTPRF;APP; CASP10;SREBF2;APOE
6	12	2.97E-10	7.85E-07	1.92E-07	DPPA4	DPPA4;BTBD2;MAPK6;PLEKHF2;PTPRO;SH3GL2;SH3GL3; ACTG1;AIDA;AMPH;PVRL2;APOE
7	16	8.36E-12	9.00E-07	3.55E-07	CISD1	CISD1;ATP6V1C1;NR4A2;NRF1;PARK2;PRKAB1;RNF2;ABCA3; ABCF1;ABLIM2;ACTG1;ADCY9;DAGLB;DDAH1;TOMM40;APOE
8	14	3.36E-12	2.08E-08	3.80E-07	PSME2	PSME2;APP;BAG3;CFTR;EPHA2;IKBKE;MGMT;PARD3; PARK2;PSMA1;ABHD10;ANKS1B;APOE;TOMM40
9	14	5.53E-13	3.20E-07	5.37E-07	KLHL6	KLHL6;HSP90AA1;AARS;ABL1;ACTG1;ACTN1;AHR;AHSA2; ALK;APP;CDK14;FBXO18;TOMM40;APOE
10	14	4.85E-13	2.31E-07	5.78E-07	MYLK4	MYLK4;HSP90AA1;AARS;ABL1;ACTG1;ACTN1;AHR;AHSA2; ALK;APP;CDK14;FBXO18;TOMM40;APOE
11	16	1.13E-09	2.38E-07	6.31E-07	SUCLG1	SUCLG1;CHMP1B;ECT2;OXCT1;SEPT11;SEPT9;SUCLA2;SUCLG2; ABR;AGGF1;APP;CUX1;ELAVL1;MYO1D;TOMM40;APOE
12	12	0	1.23E-07	6.75E-07	TPM4	TPM4;ARRB1;CDH2;DBN1;ILF3;PARK2;APP;ATP5O; CDH4;PVRL2;TOMM40;APOE
13	11	6.53E-14	2.92E-06	7.37E-07	AMY2B	AMY2B;FOS;PARK2;ACSL1;ACTA2;APP;ATF7;ATP5C1; ATP5O;TOMM40;APOE
14	11	3.33E-16	5.63E-08	9.93E-07	SLC11A2	SLC11A2;NDFIP1;NEDD4;NEDD4L;PARK2;ANXA13;APP;ATP5O; GRIA1;TOMM40;APOE
15	10	0	9.59E-08	1.10E-06	PARK2	PARK2;ACTA2;ANXA2;APP;ATP5C1;ATP5O;CUL1;ESRRG; TOMM40;APOE
16	14	5.13E-13	2.89E-06	1.20E-06	DMRTA1	DMRTA1;HSP90AA1;ORC5;STAT5B;AARS;ACTG1;ACTN1;AHSA2; ALK;APP;CDK14;FBXO18;TOMM40;APOE
17	13	2.29E-12	1.46E-06	1.27E-06	HTRA2	HTRA2;ALDH1A2;APC;APP;BIRC2;CDC40;MAPK3;PAG1; PARK2;ABHD10;ANKS1B;APOE;TOMM40
18	11	5.53E-13	3.12E-07	1.77E-06	SLC22A4	SLC22A4;PDZK1;SLC9A3R2;ABCA1;ABCC4;ADORA2B;APP; ATP2B2;SLCO3A1;TOMM40;APOE
19	11	3.57E-14	1.86E-06	1.81E-06	GPD2	GPD2;ALDOA;PARK2;ACSL1;ACTA2;APP;ATP5C1;ATP5O; CUL1;TOMM40;APOE
20	13	1.97E-13	1.48E-07	2.48E-06	SCYL1	SCYL1;CD93;CHD1L;APLF;ARHGAP15;CBX1;COL4A1;COL4A2; APP;PARK2;SMURF1;APOE;TOMM40
21	12	0	9.46E-07	2.58E-06	TGM3	TGM3;ATG12;PARK2;PRKAB1;USP53;WWOX;APP;ATP5C1; ATP5O;PVRL2;TOMM40;APOE
22	16	8.86E-14	2.32E-06	2.64E-06	ZNF787	ZNF787;FMNL1;IL7R;NRF1;PTEN;RNF2;ABCA3;ABCF1; ABLIM2;ACTG1;ADCY9;ANK3;DDAH1;INO80C;TOMM40;APOE
23	8	0	2.06E-06	2.69E-06	ESRRB	ESRRB;NCOA3;PARK2;APP;ATP5C1;ATP5O;TOMM40;APOE
24	16	6.51E-11	2.15E-06	2.84E-06	STYK1	STYK1;GSK3B;HSP90AA1;AARS;ABL1;ACACA;ACLY;ACTG1; ACTN1;AHSA2;ALK;APP;BZW2;FBXO18;TOMM40;APOE

No.: the identifier of the module; #G: number of genes in the module; Pz₁: P-value transferred from z-score of modules identified in G1; Pz₂: P-value transferred from z-score of modules identified in G₂; P: P_eval, the empirical P-value of the module; Start Node: source node from which the module search started; Module Genes: genes enriched in the module.

Table 2 shows 23 genes appearing in at least three modules. Detail about all 156 genes are shown in Table S1. Significant genes (APOE, TOMM40, PVRL2) identified by gene-base analysis approach were also included in several significant modules, and APOE, TOMM40, APP enriched in more than 20 modules. In addition, links between these genes were overrepresented in several modules, such as ATP50-PARK2 (8 modules), ATPG1-APOE (9 modules), APOE-APP (20 modules), APOE-APP-PARK2-TOMM40 (13 modules).

Table 2. Genes appeared in more than three modules.

P-values smaller than 0.001 were in bold.

Genes	Degree	PCT (%)	P-G1	P-G2	#M
APOE	55	7.35	6.86E-16	1.28E-06	24
TOMM40	41	10.54	5.39E-14	2.49E-06	21
APP	1649	0.01	5.27E-03	3.73E-02	20
PARK2	314	0.33	3.66E-03	1.21E-03	13
ACTG1	111	2.48	1.12E-02	1.90E-01	9
ATP5O	25	18.32	2.86E-03	4.25E-03	8
AARS	40	10.84	6.93E-02	3.31E-02	5
ACTN1	95	3.23	1.19E-02	1.20E-01	5
ATP5C1	44	9.77	5.34E-03	2.80E-01	5
CDK14	18	24.82	1.35E-03	4.63E-03	5
FBXO18	11	36.21	3.62E-04	9.14E-02	5
HSP90AA1	630	0.06	2.92E-01	9.35E-02	5
AHSA2	1	83.60	1.60E-02	2.28E-02	4
ALK	52	7.92	1.30E-02	9.71E-05	4
ABCA3	2	73.34	2.74E-02	1.07E-01	3
ABL1	287	0.42	1.24E-01	8.14E-02	3
ABLIM2	4	59.37	1.80E-02	5.08E-03	3
ACTA2	55	7.35	5.59E-02	2.43E-03	3
ADCY9	1	83.60	3.32E-03	2.67E-03	3
DDAH1	4	59.37	4.54E-04	4.40E-03	3
IKBKE	299	0.38	2.68E-02	4.44E-03	3
NRF1	2053	0	9.09E-02	2.51E-02	3
PVRL2	11	36.21	2.46E-06	1.92E-04	3

Degree: degree of the gene in PPIN; PCT: percentage, number of genes with degree larger than the gene divided by the number of all genes; P-G₁: P-value of the gene from GWA analysis of G₁ dataset; P-G₂: P-value of the gene from GWA analysis of G₂ dataset; #M: number of identified modules containing the gene.

3.3. Statistical and Functional Analysis of the Significantly Enriched Modules

Table 3 shows the top 10 % genes ranked according to their topological centralities in the subnetwork, four of which were also found in more than three identified modules. Please see Table S1 for the topological centralities of all 156 genes.

Table 3.

Top 10% genes in all four topological centralities.

Gene	Betweenness	Closeness	PageRank	Degree
APP	0.28	0.50	0.04	0.22
HSP90AA1	0.23	0.47	0.04	0.19
PARK2	0.19	0.49	0.03	0.18
NRF1	0.12	0.41	0.03	0.15
ELAVL1	0.10	0.42	0.02	0.14
RNF2	0.05	0.40	0.02	0.10

Betweenness: betweenness centrality, denotes the number of short communication paths that a node participates in; Closeness: closeness centrality, corresponds to the average distance (the length of the shortest paths) between a given node and the rest of the network; PageRank: PageRank centrality, takes into account the quantity and quality of connections to a given node; Degree: degree centrality, counts the number of interactions with a given node.

A crosstalk analysis was performed for these 24 enriched modules. $\text{JSC}\left(\text{A},\text{B}\right)=0$ when there are no shared genes between module A and module B. JSC(A,B)=1 when module A and module B share exactly the same genes. Thus the higher the JSC is, the more similar the two modules are. As shown in Fig. 3, there are three large clusters (S₁ = M_9,10,16,24,4; S₂ = M_{13,15,19,23,21}; S₃ = M_2,7,22), each of which consists of at least a pair of modules with JSC more than 0.5.

Figure 3. Heat map of Jaccard similarity coefficients (JSC) of 24 significant modules.

JSC ranges from 0 to 1. JSC(A,B)=0 when there are no shared genes between module A and module B. JSC(A,B)=1 when module A and module B share exactly the same genes. Thus the larger the JSC is, the stronger agreement the modules have. In order to make an intuitive understanding of JSC among 24 modules, the clusters were arranged according to hierarchical clustering algorithm performed in Matlab.

Fig. 4 shows the most significant module M₁ and three subnetworks composed of those clustered modules separately. All significant modules included about 58.3% nominally significant (gene-wise P-value, 0.05) genes in G₁ and 57.1% nominal significant genes in G₂, while 122 genes (78.2%) reached the nominal significance in either dataset. These modules were highly significant in terms of not only nominal P-value but also topological structure measurements. 23.1% genes were with large degree — the degree of at least 95% genes in the network was smaller than that of these genes. Some non-significant genes were included in these modules, mainly because they were surrounded by low-P-value genes, such as TAX1BP1 in Fig. 4(a), HSP90AA1 in Fig. 4(b), NRF1 in Fig. 4(d). This further demonstrated the capability of the network-assisted strategy to identify collective effect of multiple genes, although they may not show significant association with the amyloid imaging phenotype on their own. The subnetwork with genes appearing in more than three modules were shown in Fig. 4(e).

Figure 4. Subnetworks composed of clustered modules with the Jaccard similarity coefficients larger than 0.5.

(a) The most significant module; (b) Subnetwork composed of M₄,M₉,M₁₀,M₁₆,M₂₄; (c) Subnetwork composed of M₁₃,M₁₅,M₁₉,M₂₁,M₂₃; (d) Subnetwork composed of M₂,M₇,M₂₂; (e) Subnetwork with genes appearing in more than 3 modules. P-values from genome-wide association analysis of G₂ are color-mapped. Node shape represents the number of modules a gene appears in.

Pathway enrichment analyses were conducted for the 24 significant modules respectively to explore their biological functions. 16 modules with at least one significantly (adjusted P-value < 0.01) enriched KEGG term were shown in Fig. 5. We observed the following. (1) Alzheimer’s disease pathway was enriched by 10 modules. (2) The most significant module M₁ only enriched the Alzheimer’s disease pathway. (3) Six pathways with adjusted P-value < 0.001 were enriched by M_{1,7,11,13,15,19,21,23}, including three metabolism-related pathways (Citrate cycle, Propanoate metabolism and Carbon metabolism), two neurodegenerative diseases (Parkinson’s disease and Alzheimer’s disease) and Vibrio cholerae infection pathway. (4) The metabolism-related pathways in (3) were enriched by M₁₁. (5) The neurodegenerative diseases in (3) were commonly enriched by M_{13,15,19,21,23}. This result is consistent with the crosstalk-analysis results, suggesting M_{13,15,19,21,23} may contribute to Parkinson’s disease and Alzheimer’s disease. (6) Besides two neurodegenerative diseases mentioned in (3), there are five other pathways enriched by two or more modules. They are Ubiquitin mediated proteolysis, Apoptosis, Bacterial invasion of epithelial cells, Oxytocin signaling pathway and Oxidative phosphorylation.

Figure 5. Heat map of pathway enrichment analysis at the significance level of P<0.01.

–log₁₀(P-values) from each analysis are color-mapped in the heat map. Heat map blocks labeled with “x” reach the adjusted significance level of P<0.01. Each row (pathways) and column (modules) with at least one “x” block are included in the heat map. Modules on the left of the red dotted line are clustered modules. The remaining modules are on the right.

4. DISCUSSION

In the 24 identified modules, 23 genes appeared in more than three modules (Table 2). Six genes were among the top 10% genes ranked by their topological centralities (Table 3), four of which also appeared in Table 2 (PARK2, APP, NRF1, HSP90AA1). Among those 25 (23+6–4) genes, associations of several genes with Aβ were previously well-studied.

The amyloid plaques in the brain due to the imbalance between production and clearance of amyloid-β (Aβ) peptides is one major hallmark of amyloid pathology in AD brains [35]. It damages synapses and eventually leads to neurodegeneration and dementia [36]. As shown in Fig. 2, SNPs downstream of APOC1 and SNPs in APOE and TOMM40 genes displayed genome-wide significant associations with Aβ in GWA analysis. The recent cell type-specific and inducible mouse models support that APOE has the greatest influence on amyloid during the initial stage by affecting the clearance and aggregation of Aβ [37]. Genetic variants in the TOMM40 gene have been reported to greatly increased the precision in the estimation of late-onset AD age in APOE ε3 carriers [38]. Apolipoprotein C-I, the increased expression of which induced by the H2 allele of APOC1, co-localizes in plaques with Aβ and delays its aggregation in vitro [39]. Mutations in the APP gene could cause the excessive secretion of Aβ, which is a primary cause of early-onset familial AD [35,36]. The overexpression of PARK2 induces the clearance of Aβ by enhancing autophagy, increases tricarboxylic acid cycle activity, and improves mitochondrial function as well as a synaptic restoration function in a triple transgenic AD mice model [40,41]. ABL1 inhibition has an effect on Aβ deposits in the brain [42]. Overexpression of DDAH1 significantly decreases Aβ secretion in Aβ precursor protein cells, whereas knockdown of endogenous DDAH1 exacerbates cellular Aβ secretion [43].

Besides those gene related with Aβ, we have highlighted a panel of genes have already been implicated in the etiology of AD. ACTN1 has been identified as one of high ranking genes being associated with AD related gene expression changes in the hippocampus [44]. ATP5C1 has been identified to be susceptible to AD by the analysis of the differentially expressed genes in entorhinal cortex with AD [45]. Increased mRNA expressions and protein levels of NRF1 have been found in 6-month-old in double mutant APPXDrp1+/− mice relative to APP mice [46]. PVRL2 has exhibited significant association with AD in late onset AD family study data, through GWA studies using both logistic and Cox’s regression [47]. Amadio et al have found that neuronal ELAV proteins were decreased in postmortem patients along with AD progression, but the ELAVL1 levels did not change, thus indicating that only neuronal ELAV proteins are affected in AD pathology [48]. ABCA1 has been identified one of the hub genes associated with late-onset AD APOE4 non-carriers [49].

Among the genes appeared in more than three modules, there were also some novel genes that had not been studied in AD, yet have associations with other types of neurodegenerative diseases. A novel mutation in AARS has been reported causing a progressive, mild myeloneuropathy [50]. ABLIM2 has been found to take part in the neuron guidance processes in human, rat and mouse [51]. Other genes in those 25 genes require more evidences to demonstrate the role they played in biology process. In conclusion, those 25 genes could be promising novel candidate genes for AD risk.

In the 24 identified modules, five links between genes (ATP50-PARK2, ATPG1-APOE, APOE-APP-PARK2-TOMM40) were found for at least more than eight times, two of which (APOE-APP and PARK2-TOMM40) have been reported by in vivo and/or vitro experiments. It has been announced that Apolipoprotein E (ApoE) activates a non-canonical MAP-kinase cascade stimulating APP transcription and Aβ synthesis [52]. Mitochondrial clearance triggered by PARK2 are reversed by the overproduction of TOMM40 after mitochondrial transmembrane potential loss, and the molecular interaction between PARK2 and TOMM40 are weakened or disrupted by PARK2 mutations [53].

Pathway enrichment analyses were conducted for the 24 significant modules respectively to explore their biological function (Fig. 5). Please see Section 3.3 for detailed information of the observation. Briefly, several modules enriched in two significant neurodegenerative disease pathways (Alzheimer’s disease, Parkinson’s disease) and Ubiquitin mediated proteolysis pathway. Parkinson’s disease has been previously found to be associated with AD [40,41]. The dysfunction of ubiquitin proteasome system has been reported associated with the accumulation of hyper-phosphorylated tau oligomers at human AD synapses [54]. Another highly enriched pathway is carbohydrate metabolism-related Citrate cycle (TCA cycle) pathway. A lower conversion of glucose into TCA cycle-related metabolites is shown in the brain of 13-month-old triple transgenic mice (a mouse model of AD) [55]. Another two pathways are Apoptosis pathway and Oxidative phosphorylation pathway. It has been reported that genetic variation within oxidative phosphorylation genes have associations with AD clinical status and neuroimaging measures [56]. The increased levels of apoptosis proteins in amnestic MCI have been found between AD and normal controls [57]. In conclusion, the enrichment results may provide valuable information for the subsequent finding of functional similar signals associated with AD.

CONCLUSION

For the current study, we performed a genome-wide network-assisted association and enrichment study to explore the collective effects of multiple genetic association signals on an AV-45 PET measure. First, we took advantage of the DMS algorithm [26] to identify modules enriched by the genetic association signals in PPIN. Next, we employed statistical evaluations to the identified modules, including a normalization process to adjust the topological bias in network (e.g., nodes with high degree), a replication test to ensure the modules were not found by chance (i.e. high probability found in other AD-related datasets), and a permutation test to evaluate unbiased association between the modules and AV-45 PET measure. At last, 24 modules (156 genes in total) enriched by robust genetic signals in GWA analysis datasets were found. Several centrality measures were computed to rank those 156 genes, and module similarity tests and functional enrichment analysis were performed subsequently for those 24 modules. Those highly reproducible and biologically meaningful modules not only confirmed previously reported AD genes, but also highlight several new candidate genes, gene-gene functional relationship and molecular pathways underlying AD or other types of neurodegenerative diseases.

Our results derived from computational analysis in this study need to be validated by additional biological experiments more than literature survey. The identified genes and network modules may provide more insights into the neurobiology of AD and suggest potential therapeutic targets.