Identifying Alzheimer’s disease-associated genes using PhenoGeneRanker

Abstract

Alzheimer’s disease (AD) is a neurogenerative disease that affects millions worldwide with no effective treatment. Several studies have been conducted to decipher to genomic underpinnings of AD. Due to its complex nature, many genes have been found to be associated with AD. Despite these findings, the pathophysiology of the disease is still elusive. To discover new putative AD-associated genes, in this study, we integrated multimodal gene and phenotype datasets of AD using network biology methods to prioritize potential AD-related genes. We constructed a multiplex heterogeneous network composed of patient and gene similarity networks utilizing phenotypic and omics datasets of AD patients from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. We applied PhenoGeneRanker to traverse this network to discover potential AD-associated genes. To assess the impact of each network layer and seed gene, we also run PhenoGeneRanker on different variants of the network and seed genes. Our results showed that top-ranked genes captured several known AD-related genes and were enriched in Gene Ontology (GO) terms related to AD. We also observed that several top-ranked genes that are not in AD-associated gene list had literature supporting their potential relevance to AD.

1. Background

Alzheimer’s disease (AD) is a complex and progressive neurogenerative disease, and is one of the leading causes of death among older people. AD has become a growing public health crisis, affecting about 50 million people around the world, and six million people in the US, leading to huge burden to patients, caregivers and the healthcare sector. By the year 2040, the entire direct medical expenses for AD are anticipated to reach $259 billion^1–3. One of the hallmarks of AD is the accumulation of tau and amyloid-β (Aβ) proteins in the brain, which lead to neuron death^4,5. Growing studies suggest that some of the biological factors of AD could occur decades before the symptoms of the disease are apparent^6,7. Screening methods such as amyloid and tau position emission tomography (PET) could be effective for early AD detection⁸. Several computational methods such as VGG-TSwinformer⁹, PPAD¹⁰, TA-RNN¹¹, and PVTAD¹² have been developed for early prediction of AD, too.

To characterize AD biology, several consortiums such as Alzheimer’s Disease Neuroimaging Initiative (ADNI), Alzheimer’s Disease Sequencing Project (ADSP), and National Alzheimer’s Coordinating Center (NACC) have facilitated generation of vast amount on multimodal biomedical datasets such as transcriptomics, genomics, epigenomics (i.e., multi-omics), neuroimaging, and clinical. Genome Wide Association Studies (GWAS) have been conducted^13–15 to report several AD risk genes such as APOE, APP, ABCA7, MPO, PSEN2, and PSEN1, annotated in databases such as MedGen^16, GTR¹⁷, PheGenI¹⁸, and AlzGene¹⁹. Further research studied these genes in more detail. For instance, increased level of MPO expression is associated with increased the risk of AD²⁰. The ABCA7 levels have been observed to be increased in the AD brain causing amyloid plaque burden²¹. Mutations within amyloid precursor protein (APP) might cause 10% to 15% of early-onset familial AD²². APOE has been identified as a genetic risk factor for both familial late-onset and sporadic late-onset AD in several studies. Potential driver genes were reported from transcriptomic data by identifying differentially expressed genes (DEG) between AD and normal samples and utilizing machine learning- or deep learning-based methods^23,24.

Utilizing the growing amount of multimodal datasets for AD, new AD-associated genes could be identified. Gene prioritization techniques can be applied to rank these genes based on their potential relevance to AD. Such findings would help improve disease knowledge, diagnosis, and leads to effective treatment options to prevent, postpone, and eventually cure AD. Network-based methods are effective to rank genes with respect to their potential relevance to a phenotype (e.g., disease)²⁵ . Several network-based gene prioritization tools have been developed to prioritize disease related genes^26–28. Among them, the Random Walk with Restart (RWR) that considered both global network information and closeness to the central node was utilized to prioritize genes. One of the drawbacks of RWR is its biasness towards high degree nodes³⁹. Recently, PhenoGeneRanker²⁹ was introduced, a tool to perform RWR to rank genes associated with a phenotype (e.g., disease). PhenoGeneRanker analyzes multiplex (i.e., having multiple edge types) heterogeneous (i.e., having multiple node types) networks, which consist of gene and phenotype networks. PhenoGeneRanker also generates empirical p-values for the ranking to address the degree bias issue of RWR.

In this study, we ran PhenoGeneRanker on a multiplex heterogeneous network for AD to rank genes based on their potential association to AD. Utilizing datasets from ADNI and protein-protein interactions (PPI) from STRING, we built gene and patient similarity networks and applied PhenoGeneRanker (Figure 1). We found some well-known AD genes in our top rankings. By doing the GO enrichment analysis of our significant-top 200 genes, we observed that GO terms were highly associated with AD-related GO terms. We also conducted a literature search for the top genes that are not in AD databases, and found several genes reported as AD-related. We also checked the importance of utilizing different data modalities for the gene rankings.

Figure 1:

Overall pipeline to identify AD-associated genes. Three patient networks were generated from patients’ demographic, cognitive, and neuro-imaging data modalities. Two gene networks were built using gene expression and PPI data. PhenoGeneRanker was applied to the entire network to rank genes based on their potential association with AD seed genes. Finally, genes were ranked with associate p-value.

2. Methods

2.1. Building a multiplex heterogeneous network for AD

In this study, we built a multiplex heterogeneous network that consists of gene and phenotype node types and multiple relation (i.e., edge) types based on multiple data modalities. We obtained phenotypic and omics datasets of AD patients from the ADNI database³⁰ (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 MCI and early AD. Protein-protein interaction (PPI) dataset was collected from STRING database³¹ (https://string-db.org/). To make the network more AD specific, we relied on data from AD patients whenever available. Particularly, we obtained demographic, cognitive and neuroimaging data of AD patients using ADNIMERGE R package³⁰ and gene expression data of AD patients from ADNI. In the ADNIMERGE dataset, most patients had longitudinal data. For these cases, we utilized the most recent data of each patient. For missing values, we used data from earlier time points to impute it.

We created three layers of phenotype network and two layers of the gene network. In the phenotype network where each node is an AD patient, the layers were built by computing similarity between AD patients based on demographic, cognitive, and neuroimaging data. The two gene layers were built based on gene expression and PPI datasets. We also built a bipartite layer to connect gene nodes to patient nodes based on expression activity of genes in AD patients. The features to build each layer are listed in Table 1.

Table 1:

Details of features for each network from ADNI and STRING database

Network layer	Feature Name	Description
Demographic Layer	Age	Patient’s age
	Gender	Patient’s gender
	Education level	Patient’s education level
	Race	Race of a patient
	Ethnicity	Ethnicity of a patient
	APOE4	Number of APOE-ε4 alleles
Neuroimaging layer	Ventricles	Volume of ventricles
	Hippocampus	Volume of hippocampus
	WholeBrain	Volume of Whole Brain
	Fusiform	Volume of fusiform
	MidTemp	Volume of middle temporal gyrus
	Entorhinal	Volume of entorhinal
	ICV	Volume of intracranial
Cognitive Layer	CDR-SB	Clinical Dementia Rating Sum of Boxes
	ADAS11	Alzheimer’s Disease Assessment 11-items composite score
	ADAS13	Alzheimer’s Disease Assessment 13-items composite score
	ADASQ4	Score from Task 4 (Word Recognition) of the Alzheimer’s Disease Assessment Scale
	MMSE	Total Score of Mini-Mental State Examination
	RAVLT immediate	Rey’s Auditory Verbal Learning Test_Immediate Recall
	RAVLT learning	Rey’s Auditory Verbal Learning Test_Learning
	RAVLT_forgetting	Rey’s Auditory Verbal Learning Test_Forgetting
	RAVLT_perc_forgetting	Rey’s Auditory Verbal Learning Test_Percent Forgetting
	LDELTOTAL	Delayed total recall
	TRABSCOR	Trail Making Test Part B Time
	FAQ	Total Score of Functional Activities Questionnaire
Gene Expression	Expression values	Gene Expression Values of 120 AD patients
PPI	Combined Score	Combined Score from human protein physical links

In the demographic phenotype layer, we utilized demographic and genetic features, namely age, gender, education level, race, ethnicity, and APOE4 due to their relevance to AD risk. We applied one hot encoding on the categorical features, namely gender, race, and ethnicity. The demographic layer contains both binary and numerical values, thus the reciprocal of Gower distance ³² was applied to create the patient-patient similarity matrix. For neuroimaging layer and cognitive layer, we scaled each feature using min-max normalization then computed the reciprocal of Euclidean distance as patient-patient similarity value. When computing similarity scores, we ignored the features with missing values.

We created a gene network layer based on PPI. We converted STRING protein IDs to gene Entrez IDs and applied min-max normalization to the combined score. We considered the combined scores as edge weights for gene-gene interactions. The second gene network layer was constructed based on gene co-expression of 120 AD patients. The gene expression from ADNI was generated using the Affymetrix Human Genome U219 Array (Affymetrix (www.affymetrix.com), Santa Clara, CA) platform, which had 530,467 probes. We collected the normalized expression values from ADNI that were preprocessed using Robust Multi-chip Average (RMA)³³ normalization method. To further preprocess gene expression data, we converted it from probe level to gene level by taking the mean value of probes for each corresponding gene. Then, we converted gene symbols to Entrez gene ID as the primary key for the gene networks. Reciprocal of Euclidean distance between gene expression of each gene pair was computed to generate a gene co-expression network. Then, we eliminated weakly connected edges by applying a threshold on edge weights (i.e., Reciprocal of Euclidean distance). The threshold value was chosen to ensure a balance between the number of edges and number of genes exist in gene co-expression network.

To construct the gene-patient bipartite network, we utilized the gene expression dataset. We connected a patient node to a gene node if the gene was highly upregulated in that patient. To determine which genes were highly upregulated, we first computed the median expression of each gene across all samples and considered a gene highly upregulated if its expression value was three standard deviations above the median. We used the expression values as the edge weight. Since not all patients had gene expression available, bipartite layer was only between highly expressed genes and AD patients with available expression data.

As weak edges might affect the entire network making it densely connected and causing inferior performance, we removed the weakly connected edges from the network. We selected a cutoff value (Table 2) for each network to remove the weak edges. For all the layers in the phenotype network and the PPI network layer, the cutoff value was based on the median of the edge weights. For the gene co-expression network layer, the cutoff was. selected to build a balanced network by avoiding excessive interactions at lower thresholds while preserving enough genes and connections to prevent losses at higher thresholds. Table 3 shows the final number of nodes and edges in each layer of the multiplex heterogeneous network.

Table 2:

Correlation function and cutoff for each network layer

Network layer	Edge weight scoring function	Cutoff
Demographic	Reciprocal of Gower distance	0.2
Cognitive	Reciprocal of Euclidean distance	0.5
Neuroimaging	Reciprocal of Euclidean distance	0.6
Gene expression	Reciprocal of Euclidean distance	0.8
PPI	Combined Score	0.18

Table 3:

Multiplex heterogeneous network statistics.

Network layer type	Data type	Number of edges	Number of nodes
Patient	Demographic	146,631	781
Patient	Cognitive	260,858	805
Patient	Neuroimaging	241,891	795
Gene	Gene expression	3,961,187	12,411
Gene	PPI	405,199	16,872
Patient-Gene	Gene expression	9,426	6,816

2.2. Identifying Gold Standard AD-associated Genes

To evaluate our results, we collected a gold standard AD-associated gene set from seven different databases, namely Genetic Testing Registry (GTR)¹⁷ MedGen¹⁶ Phenotype-Genotype Integrator (PheGenI)¹⁸, GWAS catalog³⁴ , MalaCards³⁵, GeneCards³⁶, and NCBI Gene³⁷. We searched AD in GTR database and found eight genes related to AD. In PheGenI and MedGen, we found 28 and eight genes for AD phenotype, respectively (p-value < 1e-05). We collected 153 genes from GWAS catalog selecting trait as AD and p-value < 1e-08. In GeneCards, we searched by AD relevance score greater than 30 and collected 304 genes. In MalaCards, we collected 237 genes that have ≥3 PubMed papers supporting their relevance to AD. We searched for “Alzheimer disease[All Fields] AND “Homo sapiens”[porgn] AND (“has refseqgene”[Properties] AND alive[prop])” on NCBI Gene database and found 369 AD associated RefSeq genes in human. We converted all gene symbols to Entrez gene IDs and assigned a weight to each database. We assigned a weight of 2 to GTR, MedGen, PheGenI as they are more reliable sources and a weight of 1 to GWAS, MalaCards, GeneCards and NCBI Gene databases. We computed the total score for each gene, which is the total weight of the databases the gene was reported and chose 71 genes with a score ≥ 3 (see Supplementantary Table 1).

2.3. Running PhenoGeneRanker

PhenoGeneRanker ³⁸ is a Bioconductor R package for gene prioritization on multiplex heterogeneous networks. To run PhenoGeneRanker, we need to identify a seed gene, a gene that is known to be associated with the phenotype of interest. We also need to specify some hyperparameters to determine the influence of each component of the network.

2.3.1. Seed Genes and Hyperparameters for PhenoGeneRanker

Seed nodes are the initial nodes from which the RWR algorithm starts traversing the network and returns to when “restart” occurs during the random walk. Among the gold standard AD-associated genes (see Section 2.2), we selected the genes that had a score ≥ 7 as seed genes. These genes were APP, APOE, PSEN1, PSEN2, MPO, NOS3, PLAU, CLU, SORL1 and ABCA7. For each PhenoGeneRanker run, we used one of these genes as the seed. For our reference run, we used APP as the seed gene.

The hyperparameters for PhenoGeneRanker are r, ζ, δ, τ, η, and λ. The r parameter denotes the restart probability to jump back to the seed node(s). The δ and ζ parameter represents the inter-layer navigation probability for gene and phenotype networks, respectively. In other words, they control the probability of jumping to the other layers or staying on the same layer of the gene or phenotype network. For high values, the random walk will be more likely to traverse to the neighbor nodes in a different layer. The τ and φ parameter is set for assigning restart probability to each gene and phenotype network layer, respectively. These hyperparameters can be adjusted to give a higher or lower priority to restart from a particular network layer. The η parameter is used for restarting to a gene or phenotype seed in the multiplex network. A low value of η indicates that the RWR algorithm is more likely to restart on a gene seed. λ represents the probability of jumping between the gene and phenotype networks. The higher value of λ increases the dependence on the bipartite layer over gene and phenotype networks. All parameter values used in the reference run are listed in Table 4.

Table 4:

Hyperparameter values of PhenoGeneRanker used in the reference run.

Hyperparameter	Value
r	0.8
ζ	0.5
δ	0.5
τ	(1, 1)
φ	(1, 1, 1)
η	0.1
λ	0.5

3. Results

In this study, we built a multiplex heterogeneous network for AD using two gene layers, three phenotype layers, and one bipartite layer between phenotype and gene nodes. We ran PhenoGeneRanker on the entire multiplex heterogeneous network using APP as the seed gene (called “reference run” hereafter) to rank genes based on their association to AD. To assess the results, we performed Gene Ontology (GO) enrichment of the top 200 ranked significant genes (see Supplementary Table 2) and bottom-ranked genes. To assess the contribution of each network layer and seeds, we also ran PhenoGeneRanker with different network layer and seed gene combinations. We generated cumulative distribution function (CDF) plots to see what percentage of the gold standard AD-associated genes (see Section 2.2) were captured by PhenoGeneRanker. We also conducted literature search for the top-ranked significant genes that were not in the gold standard AD gene list. We give the details of these results in the subsequent subsections.

3.1. The Top-Ranked Significant Genes were Enriched in Neurodegenerative and AD related GO Terms

We performed GO enrichment for the top 200 significant genes (p-value < 0.05) using WebGestalt³⁹ . The results showed that the top 200 significant genes were enriched in AD-related GO biological process terms such as amyloid precursor protein metabolic process, lipid metabolic process, proteolysis, and neurogenesis ^40–43 (False Discovery Rate (FDR) < 0.05, Table 5). These results indicate that the top genes ranked by PhenoGeneRanker were associated with AD (see Supplementary Table 3). As a negative control we repeated the same GO enrichment procedure for the bottom 200 ranked genes and found no GO terms showing that PhenoGeneRanker was able to move the irrelevant genes to the bottom of the gene prioritization list.

Table 5:

GO Biological Process enrichment of top-ranked 200 significant genes.

GO: Biological Process Term*	FDR
GO:0006508 proteolysis	6.52E-16
GO:0010876 lipid localization	2.85E-16
GO:0042982 amyloid precursor protein metabolic process	6.97E-19
GO:0051604 protein maturation	1.33E-11
GO:0051241 negative regulation of multicellular organismal process	2.37E-09
GO:0051049 regulation of transport	1.13E-09
GO:0006915 apoptotic process	2.32E-09
GO:0022008 neurogenesis	1.02E-07
GO:0006629 lipid metabolic process	5.22E-07
GO:0045184 establishment of protein localization	5.14E-07

^*GO terms while maximizing gene coverage with Benjamini-Hochberg corrected False Discovery Rate (FDR) <0.05.

3.2. The impact of different gene layers and seed genes on the gene rankings

We compared different combinations of seed genes and network layers to check their contributions to the final gene rankings (Figure 2). Our network included 69 of 71 gold standard AD-associated genes, thus we evaluated the rankings based on these 69 genes. For each of the combinations, we kept everything same in the reference run expect for one modification. We observed that the seed gene had a high impact on the gene ranking results (Figure 2A). We checked the rankings with nine different seed genes and their impacts on the top rankings. Top 300 gene list computed using the reference run captured over 50% of genes of the 69 known AD-associated genes, whereas the coverage dropped for some other seed choices (e.g., NOS3, PLAU). When we ran PhenoGeneRanker using a random gene as seed gene, the coverage of the known AD genes was lowest among all seed choices. We also removed the gene co-expression layer and PPI from the networks to find their contribution in the top rankings and observed that PPI contributed higher than the gene co-expression layer (Figure 2B). We also generated a random gene network with the same number of genes and density as the original gene co-expression using Erdős–Rényi model⁴⁴. Similarly, we generated random PPI network from original PPI network and observed the random gene layers to assess the impact of entire gene layers. We observed that random layers and using only gene co-expression layer performed poorly capturing AD-associated genes.

Figure 2:

Percentage of known AD genes appearing in the top-ranked genes of PhenoGeneRanker with seed combinations and multiple networks. A. Percentage of known AD genes appearing in the top-ranked genes of PhenoGeneRanker with using different seed genes on the original network of the reference run. B. Percentage of known AD genes appearing in the top-ranked genes of PhenoGeneRanker with using different network components. The reference run uses two gene and three patient layers with APP as a seed gene. Without PPI layer, APOE is used as the seed gene because APP is not present in the gene co-expression network. Other runs are the same as the reference run except for the one change described in the figure legend. See Section 3.2 for the description of random layer generation.

3.3. Hyperparameter Analysis of PhenoGeneRanker

To check the effect of each PhenoGeneRanker hyperparameter on the rankings, we ran PhenoGeneRanker with multiple values for each hyperparameter for the reference run (Figure 3). We changed one hyperparameter value at a time and kept the default values for the other hyperparameters. We observed nearly identical CDF plots for different δ, ζ, and λ values (Figure 3A, 3B and 3C) indicating that the probabilities have minimal impact on the results. We observed slight differences between lower and higher r values, but no clear trend of improvement as r increases or decreases (Figure 3D). However, the CDF plots for different η and τ values were nearly identical to each other (Figure 3E and 3F).

Figure 3:

Effect of different hyperparameters on the percentage of AD gene coverage. CDF plots for different values of A) δ, B) ζ, C) λ, D) r, E) τ, and F) η. For each run, all the other hyperparameters were set to their default values (see Table 4) and APP was used as the seed gene.

3.4. Literature analysis of top-ranked novel AD genes

To assess if other high-ranked genes that do not appear in the list of known AD genes could be related to AD, we have selected top genes that occur in multiple runs and conducted a literature search. More specifically, we used ten different AD-related genes as a seed node (i.e., APP PSEN1, PSEN2, APOE, MPO, PLAU ABCA7, SORL1, CLU and NOS3) and ran PhenoGeneRanker to get ten different rankings. Then we selected the top common genes that appeared in top 200 significant gene lists from at least five out of ten rankings, which led to 10 genes. We removed the genes that were already in our gold standard list from the list and did literature analysis for the remaining 7 genes, which could be novel AD-related genes. We found supporting evidence for all of seven genes (Table 6). For instance, APOA4 was reported to be decreased expression in AD patients than normal⁴⁵, VLDLR slightly increased in AD conditions⁴⁶.

Table 6:

Novel AD genes, appeared in multiple rankings of PhenoGeneRanker using different seed genes, with number of supporting publications and their references.

Entrez ID	Gene Symbol	# of Supporting Publications	References
337	APOA4	2	45,47
8542	APOL1	1	48
7436	VLDLR	3	46,49,50
3931	LCAT	2	51,52
64759	TNS3	1	53
116519	APOA5	1	54
7941	PLA2G7	1	55

4. Discussion

In this study, we constructed a multiplex heterogeneous network for AD that consists of two gene layers, three phenotype layers, and a bipartite layer connecting patient phenotype nodes to gene nodes. In the phenotype layer, each node represents an AD patient, the links with the other AD patients were based on their similarity scores calculated using demographic, cognitive, and neuroimaging data obtained from ADNI. For the gene layer, each node represents a gene, and connections between the two genes were based on PPI and gene expression values. After building complete multiplex heterogeneous network, the PhenoGeneRanker tool was executed on the network to rank genes by their association with AD known genes. Finally, potential AD-related genes were prioritized through the analysis of multimodal AD associated datasets.

We performed GO enrichment analysis for the top 200 and bottom 200 ranked genes to observe whether overrepresented biological processes were related to AD. Our top 200 significant genes were enriched in GO biological process terms such as APP metabolic process, lipid metabolic process, proteolysis, apoptotic process and neurogenesis. Improper metabolism of APP leads to the accumulation of amyloid-beta plaques, and abnormal apoptosis leads to the loss of neurons in the brain areas like the cerebral cortex and hippocampus^56,57 .However, the bottom 200 ranked genes were found to have no significant GO terms which indicates that the PhenoGeneRanker tool effectively pushed irrelevant genes to the bottom.

We checked the impact of different gene layers and seed genes on the gene rankings. We observed that the PPI layer contributed more than the gene expression layer. We assumed that it was due to the small sample size of gene expression data. Although our phenotype layers included nearly 800 AD patients, the gene expression layer had available data from only 120 AD patients. We selected well-known AD-associated genes as seed genes to evaluate their effectiveness in capturing gold-standard AD-associated genes. Among them, our reference run captured over 50% of these genes within the top 300 significant ranked genes. This indicates that our proposed approach was able to prioritize relevant genes in the top list. However, there is a room for improvement to be able to capture additional genes that were not ranked in the top listed genes. We also noticed the PhenoGeneRanker’s robustness to changes in hyperparameters such as δ, ζ, and λ (Figure 3). The model’s performance remained consistent which indicated its stability and reliability across different parameter settings.

We generated 10 different gene rankings using ten different AD-known genes as a seed: APP, PSEN1, PSEN2, APOE, MPO, PLAU, ABCA7, SORL1, CLU, and NOS3, and selected the genes if they appeared in the top 200 gene list in at least five out of the ten rankings. We found seven genes that are not listed in our gold-standard AD-associated genes. The identified seven genes were consistently ranked highly across different seed-based analyses which increased the confidence that these genes could be relevant to AD. We performed literature search for these genes to examine their potential relevance to AD biology. We found out that reduced levels of APOA4 are linked to a higher risk of developing AD⁴⁵ . Additionally, increased levels of APOL1 have been observed in the frontal cortex of individuals with frontotemporal lobar degeneration (FTLD) that leads to dementia⁴⁸. Polymorphisms in the VLDLR gene may be associated with AD in Japanese individuals, particularly those with an earlier age of onset⁴⁶. Furthermore, increased levels of PLA2G7 are linked to a risk factor for developing AD⁵⁵, while CSF analysis has shown reduced LCAT activity in the plasma lipoprotein profiles of AD patients compared to controls⁵¹. Although TNS3 is not directly associated with AD, it interacts with APOE and CLU, which have both been linked to AD⁵³. Lastly, genetic vulnerability related to APOA5 may influence the association between triglyceride levels and the risk of AD in women where low triglyceride levels are linked to a decreased risk of AD⁵⁴.

In our current study, we used gene expression and PPI networks; however, in future work, we plan to expand our gene layers by integrating DNA methylation and SNP data. Additionally, we only considered the connections between genes that might cause the loss of some important information about the genes by not considering the values coming from AD patients. In the future, our network will be generated based on gene connection and their values coming from different data modalities. Another limitation of our study is that the data predominantly represents a white population (751 out of 805 participants). Utilizing datasets from more diverse populations would be useful to identify additional genes that may be associated with AD in different demographic groups. To address this issue, we aim to select racially balanced datasets in future analyses.

5. Conclusion

In this study, we generated a multiplex heterogeneous network for AD and applied PhenoGeneRanker on this network to identify putative AD-related genes. We used APP as seed gene because of having strong relationships with AD from several studies. Our results showed that the top-ranked genes were enriched in AD-related GO terms. Several known AD genes appear at the top of the ranking. Our top 200 gene list captured about 40% of genes that were reported as AD-associated. We also found some genes that were ranked highly in multiple PhenoGeneRanker runs, which had supporting literature for their potential association to AD. The integration of multimodal datasets for AD using multiplex heterogeneous network and PhenoGeneRanker tool may be utilized for finding disease-associated genes.