Interplay between stress-related genes may influence Alzheimer’s disease development: The results of genetic interaction analyses of human data

Abstract

Emerging evidence from experimental and clinical research suggests that stress-related genes may play key roles in AD development. The fact that genome-wide association studies were not able to detect a contribution of such genes to AD indicates the possibility that these genes may influence AD non-linearly, through interactions of their products. In this paper, we selected two stress-related genes (GCN2/EIF2AK4 and APP) based on recent findings from experimental studies which suggest that the interplay between these genes might influence AD in humans. To test this hypothesis, we evaluated the effects of interactions between SNPs in these two genes on AD occurrence, using the Health and Retirement Study data on white indidividuals. We found several interacting SNP-pairs whose associations with AD remained statistically significant after correction for multiple testing. These findings emphasize the importance of nonlinear mechanisms of polygenic AD regulation that cannot be detected in traditional association studies. To estimate collective effects of multiple interacting SNP-pairs on AD, we constructed a new composite index, called Interaction Polygenic Risk Score, and showed that its association with AD is highly statistically significant. These results open a new avenue in the analyses of mechanisms of complex multigenic AD regulation.

1. Introduction

Over the last twenty years Alzheimer’s Disease (AD) researchers demonstrated evident progress in clarifying many previously unknown aspects of pathological development of this health disorder. However, the failures of numerous clinical trials to find disease modifying treatment for AD indicate that the mechanisms of its multifactorial regulation are still not fully understood. This lack of knowledge results in erroneous strategies used in the search for, and implementation of candidate medications (Anderson et al., 2017; Bracken, 2009; Gauthier et al., 2016). Furthermore, inherent difficulties in translating evidence from experimental studies to humans, implementing AD treatments at the late disease stage, wrong drug targeting, and inappropriate drug dosage are often listed as specific causes of such failures(Bracken, 2009; Gauthier et al., 2016).

The objective of this paper is to show that information from experimental studies of AD can be used in the analysis of human data and produce useful insights about mechanisms of polygenic regulation of AD. Using this information, we identify two genes GCN2/EIF2AK4 and APP whose interplay is likely to influence AD. These genes are involved in sensing stress signals and in production of biomarkers that orchestrate cellular stress response. We hypothesized that the interplay between these genes may influence AD initiation and contribute to further AD development. To test this hypothesis, we evaluated associations of interacting SNP-pairs from these genes with AD using data collected in the Health and Retirement Study (HRS). To estimate collective associations of many interacting SNP-pairs with AD we extended the construction of the polygenic risk scores (Chasioti et al., 2019; Leonenko et al., 2019) for measuring collective associations of many interacting SNP-pairs with AD, and evaluated the associations of such measures with AD.

2. Selection of genes whose interplay may influence AD in human data

The strong belief that the proper treatment of AD at its early stage would beneficially effect patient health encouraged researchers to investigate disease pathophysiology at its preclinical stage and evaluate possibilities of medical interventions at this stage before any observed signs of cognitive decline (Jack et al., 2018). The preclinical AD stage is characterized by (i) the in vivo evidence of the amyloidosis in the brain detected by either positron imaging tomography (PET) or by measuring concentration of beta amyloid protein (Aβ) in cerebrospinal fluid (CSF); (ii) the in vivo evidence of both amyloidosis and a tau biomarker (CSF phosphorylated tau, or tau PET); and (iii) biomarkers of neurodegeneration (Dubois, 2018; Dubois et al., 2016; Jack et al., 2018). The changes observed in several preclinical AD biomarkers provide the researchers with the new definition of AD whose cognitive symptomatic (used in the traditional definitions of AD) will be manifested later in life (Jack et al., 2018).

2.1. Stress signals and associated neuronal response are likely to initiate preclinical AD stage

Note that studies of the preclinical AD stage and its biomarkers could benefit from addressing the question about factors and processes that initiate this stage. The results of recent studies on the connection between disease development and the stresses of life indicate that stressful events and the processes of body’s response to them could trigger the AD development process (Bisht et al., 2018; Caruso et al., 2018; Matos and Souza-Talarico, 2019; Canet et al., 2019; Escher et al., 2019; Lyons and Bartolomucci, 2020; Trammell et al., 2020). The stressors and responses to them may also continue influencing AD at later stages of its development. Thus, the use of stress-origin hypothesis of AD in the analyses of multigenic mechanisms of AD regulation could be beneficial for better understanding the nature and the dynamics of AD.

Stresses of life that are experienced by human body are transformed into challenges for the functions of organs, tissues, cells, and other systems of the human organism (McEwen, 2012). When such challenges (stressors) reach neuronal cells, they initiate a neuronal stress response (Anderson and Haynes, 2020; Bond et al., 2020; Clayton et al., 2017; Ishimura et al., 2016). Such response is a normal adaptive behavior of a cell to the presence of antagonizing factors which always accompany human life. The cell continues normal functioning when the stress is over, and cellular homeostasis is restored. However, persistent stress response is an indicator that something is wrong with the cell functioning. In this case the mechanism of stress regulation sends signals for self-destruction (apoptosis) to the neuronal cell (Ohno, 2014). The synaptic dysfunction and neuronal death are among the most important pathological manifestations of AD in the central nervous system (SNC) (Ankarcrona and Winblad, 2005; De-Paula et al., 2012; Donev et al., 2009; Niikura et al., 2002; Trujillo-Estrada et al., 2014). Persistent cellular stress experienced by many neurons is accompanied by changes in the preclinical and clinical AD biomarkers – the mediators of stress response – and Aβ is one of such mediators (Jack et al., 2018). Understanding whether the interplay of stress related genes with genes related to preclinical AD biomarkers is associated with this health disorder are useful for AD treatment and prevention advancements. Confirmation of such association would open new opportunities for studying the dynamic polygenic mechanisms of AD regulation at its different stages starting from disease initiation stage.

2.2. Two genes whose interplay may contribute to AD in human data

To select one of the genes that drives the stress regulation machinery we considered factors involved in the stress related processes in the cell. These factors include cellular stressors, genes that sense these stressors, and the cellular stress response genes. The latter are activated by the signals from the stress sensing genes. Although each gene from the stress regulation pathways might be an important contributor to AD, in this study we selected the GCN2 gene (EIF2AK4 in humans) whose product is involved in the process of sensing stress signals. This gene plays one of the key roles in the integrated stress response (ISR) (Anderson and Haynes, 2020; Bond et al., 2020; Clayton et al., 2017; Harding et al., 2003) (Box 1).

Box 1. The Integrated Stress Response (ISR).

The integrated stress response (ISR) is an evolutionarily conserved adaptive pathway that helps the cell to survive and restore normal functioning after exposure to variable pathological conditions (stressors) that lead to alterations in cellular homeostasis, among them defects in protein folding, nutrient deprivation, viral infection, and oxidative stress (Bond et al., 2020; Girardin et al., 2020; Oliveira and Lourenco, 2016). In ISR, these stressors are sensed by at least four different kinases HRI, PKR, PERK, and GCN2.

HRI is a heme-regulated inhibitor kinase. It is an enzyme that in humans is encoded by the EIF2AK1 gene. Heme is an iron containing compound which forms non-protein part of hemoglobin, the substance inside red blood cells, that binds to oxygen in the lungs and carries it to the tissues. HRI is an intracellular heme sensor that coordinates heme and globin synthesis in erythropoiesis by inhibiting protein synthesis of globin and heme biosynthetic enzymes during heme deficiency. HRI is also activated by denatured proteins and oxidative stress (Donnelly et al., 2013).

PKR is an interferon-induced, double-stranded RNA-activated protein kinase. It is an enzyme that in humans is encoded by the EIF2AK2 gene. PKR is activated by double-stranded RNA (dsRNA) that can be introduced to the cells by a viral infection. It can also be activated by the protein PACT (that in humans is encoded by the PRKRA gene), or by heparin, cytokines, growth factor and other cellular stress signals (Chesnokova et al., 2017; Taylor et al., 2005; Gal-Ben-Ari et al., 2018).

PERK is a protein kinase R (PKR)-like endoplasmic reticulum kinase. It is an enzyme that in humans is encoded by the EIF2AK3 gene. This protein kinase is induced by endoplasmic reticulum (ER) stress signals caused by misfolded (unfolded) proteins. It initiates the unfolded protein response (UPR) (Balsa et al., 2019; Endres and Reinhardt, 2013).

GCN2 is a general control nonderepressible 2 kinase. It is an enzyme that in humans is encoded by the EIF2AK4 gene. GCN2 plays the key role in modulating amino acid metabolism as response to nutrient deprivation(Chikashige et al., 2020; Falcón et al., 2019; Hamanaka et al., 2005; Masson, 2019; Yuan et al., 2020). GCN2 can also be activated by other stressors, including ultraviolet irradiation, viral infection, serum starvation, glucose deprivation, and oxidative stress. Recent work shows that GCN2 strongly activates by binding to ribosomal protein suggesting that GCN2 actively monitors mRNA translation(Chikashige et al., 2020; Falcón et al., 2019; Hamanaka et al., 2005; Masson, 2019; Yuan et al., 2020).

Activation of any of the four kinases lead to phosphorylation of the alpha subunit of eukaryotic translation initiation factor-2, EIF2A. This results in decrease in global protein synthesis and in increase production of the transcription factor ATF4. Phosphorylated EIF2A causes translational activation of the β-secretase (called β-site APP-cleaving enzyme 1, BACE1) a key enzyme responsible for triggering the production of Aβ peptides. Together these processes promote cellular recovery. However under persistent stress conditions the attempts to restore normal cellular functioning fail, and ATF4 promotes the induction of apoptosis.

ISR serves as a molecular control center for long-term memory formation and is involved in a wide range of disorders of the brain. Its activation may also contribute to different diseases (Andhavarapu et al., 2019; Bond et al., 2020; Rashidi et al., 2020; Rodrigues et al., 2018).

The polymorphisms in the GCN2/EIF2AK4 gene might contribute to the variability in the strength and duration of the stress response by modulating how affected cells perceive stress. It is the stressor perception, the signals sent from the stress sensing genetic mechanism, not the initial stress signal, that determines strength and duration of cellular stress response, cell’s functioning quality and its survival outcome. Incase of neurons the malfunctioning and neuronal death may indicate the beginning of AD or other neurodegenerative health disorder.

Most of the properties of this gene were obtained in experimental studies of yeasts and other laboratory animals where they use the name GCN2 (Anda et al., 2017; B’Chir et al., 2013; Bond et al., 2020; Falcón et al., 2019; Jiang and Wek, 2005; Masson, 2019). So, we use this name when discussing these properties. In the description of the results of genetic analysis of human data we use the EIF2AK4 name because this is how this gene appears in human genes databases. In the discussion we will refer to it as GCN2/EIF2AK4.

Since Aβ is the key preclinical AD biomarker (Jack et al., 2018) which results from the proteolytic processing of the transmembrane amyloid precursor protein (APP), by β- and γ-secretases (Beyreuther et al., 1992; Dar and Glazner, 2020; Hefter and Draguhn, 2017) we selected the APP gene as an interacting partner for EIF2AK4. Aβ is also the main component of extracellular amyloid plagues (one of the key hallmarks of AD) which accumulates in the brains of AD patients. Soluble Aβ oligomers are toxic. They may spread throughout the brain and contribute to AD progression. Other fragments of the APP proteolytic processing also play important roles in AD (Grimm et al., 2012; Wu et al., 2020). APP expressed in many tissues and concentrated in the synapses and neurons. Although the APP implications in many activities including antimicrobial protection (Carter, 2010), synapse formation (Fanutza et al., 2015; Octave et al., 2013), and iron homeostasis (Rogers and Cahill, 2020; Tsatsanis et al., 2019) were intensively studied – its primary function remains unknown. The testing of possible connection between the interacting genes described above and AD is performed by stimating associations of the pairs of interacting single nucleotide polymorphisms (SNPs) taken from each gene (SNP-pairs) with AD.

Note that although the connection of selected genes with AD pathological development has been established in experimental studies of AD it is unclear whether the interplay between them may be associated with AD in human data. Currently there are no studies of human population data that show how ISR pathway genes and other genes interact to influence AD. Evaluating patterns of these interactions may contribute to better understanding multifactorial regulation of AD. Because of notable biological roles of the EIF2AK4 gene in cellular stress response and the APP gene in manifestation of its product (Aβ) at the preclinical and clinical AD stages, we hypothesize that the association of interactions of these genes with AD will be detected in genetic analysis of human data.

3. Data and methods

To test this hypothesis, we used data on white males and females, both combined and as independent strata, drawn from the Health and Retirement Study (HRS). Education, smoking status, and sex were included as observed covariates in the association analyses of genetic interactions. The summary statistics for white HRS study subjects is shown in Table 1.

Table 1.

Summary Statistics for White HRS Respondents with Genetic Data.

	Males		Females		Total
	AD	No AD	AD	No AD	AD	No AD
N	294	563	431	759	725	1322
Continuous Variables: Mean (Standard Deviation)
Entry Age	66.78(8.22)	72.21(4.45)	67.88(8.21)	72.38(3.69)	67.44(8.23)	72.31(4.03)
End of Follow-Up Age	84.00(7.03)	89.28(3.50)	85.78(7.55)	89.94(3.75)	85.08(7.39)	89.66(3.66)
Dichotomous Variables: N (Percentage)
Education, Highschool+	260(86.95)	529(91.68)	412(90.95)	732(93.97)	672(89.37)	1261(92.99)
Ever Smoked	215(71.91)	392(68.77)	196(43.27)	298(38.50)	411(54.65)	690(51.34)

The genetic data – 2.5 million single nucleotide polymorphisms (SNPs) – were produced on the Illumina platform using Illumina’s Human Omni2.5-Quad (Omni2.5) BeadChip methodology on 15,620 individuals (6472 males and 9148 females). Quality control (QC) has been performed before running analyses using two procedures. The first, based on the protocol proposed in reference (Anderson et al., 2010), resulted in: 129 individuals dropped because of failure in the check of duplicates, missingness rate (5 %), heterozygosity outlier (±3SD), sex mismatch, and divergent ancestry outlier ( 8SD); 538,451 variants were removed due to minor allele frequency (MAF)<1 %, 69,947 variants were removed due to genotyping missing rate higher than 5 %; and 411,945 variants were removed based on Hardy-Weinberg test (HWE, p-value < 1.0E-7). The second procedure used the protocol proposed in reference (Marees et al., 2018). Total of 15,492 individuals and 1,266, 057 variants with MAF> 1 %, missing rate <5 % and p-value of HWE > 1E-7 passed this procedure. Individuals and SNP which passed either the first QC step or the second step remained in the final dataset. This led to 1,267,439 variants cleaned and mapped to the human reference genome GRCh38 for further analysis.

The AD trait was defined as a binary variable (bn_AD_2) using HRS survey data. In the waves from 4 through wave 9, the survey asked the question “whether or not a doctor has told the respondent s/he had a memory related disease.” In wave 10, the health questionnaire changed and asked the respondents “if they have ever been told they have Alzheimer’s disease or dementia” rather than inquiring about memory-related disease as in previous waves. If AD onset was observed at any age, bn_AD_2 was set to 0. For people who survived to 85 and did not get AD during their life, the bn_AD_2 was set to 1. Otherwise, bn_AD_2 was set to missing. To estimate association of interacting SNP-pairs with AD the logistic regression model with the interaction term of INTERSNP (Herold et al., 2009), has been used.

3.1. Collective association of interactive SNP-pairs with AD

Because of polygenic nature of AD, the interactions of SNPs from EIF2AK4 with genetic variants from many other genes may also contribute to AD. The interacting SNP-pairs associated with AD can be detected in the genome-wide association studies in which one SNP from the EIF2AK4 gene is fixed and others include all SNPs that passed the quality control procedure. The results of this analysis can be used for constructing a measure of collective association of many interacting SNP-pairs with AD. This measure extends the notion of polygenic risk score (PRS), widely used in the genetic epidemiological studies of AD (Chasioti et al., 2019; Cruchaga et al., 2018; Leonenko et al., 2019) and other complex traits (Yashin et al., 2010; Yashin et al., 2012c; Yashin et al., 2012b; Ikeda et al., 2020; Ala-Korpela and Holmes, 2020). A simple measure of such capacity for a given SNP* might be the number of detected associations of interacting SNP-pairs (with a fixed SNP*) whose p-values did not exceed a given p-value threshold (similar to the genetic dose index described in (Yashin et al., 2010, 2012c; Yashin et al., 2012b)). More general measure of interactive capacity of a given SNP* involved in the AD regulation is a composite index called “SNP-specific Interaction Polygenic Risk Score” (SI_PRS). The construction of such index follows the procedure outlined below.

3.2. The construction of the SNP-specific interaction polygenic risk score (SI_PRS)

The construction of the SI_PRS involves several steps. The first step deals with selection of a suitable SNP1 whose capacity to interact with other SNPs in the process of AD regulation is a matter of interest. In step two the interaction effects of SNP1 with all other SNPs (SNP2_i i = 1,2,..N_SNP −1) available in the data are estimated by performing GWAS-like analysis of SNP1x SNP2_i –pairs (i 1,2,..N_SNP −1). This analysis uses a regression model that contains an interaction term with a fixed SNP1. Here N_SNP is the total number of SNPs that passed the quality control procedure. As a result, we obtain estimates of the interactions effects of SNP1 with all other available SNPs on AD and corresponding p-values. Then all SNP1x SNP2_i -pairs (i = 1,2,.. N_SNP −1) are ranked with respect to p-values of their associations with AD (from smallest to the largest).

The construction of SNP-specific Interaction PRS (SI_PRS) includes only conditionally independent interaction effects. This means that if there is a group of n dependent SNP2_j j = 1,2,..n, in linkage disequilibrium (LD) with the LD measure R² > = 0.1 and a corresponding group of SNP1x SNP2_j -pairs, j = 1,2,..n, then only one SNP-pair (SNP1xSNP2_j*) which has most significant interaction effect on AD among n SNP1x SNP2_i –pairs will represent this group in the construction of SI_PRS.

Thus, the constructed SNP1-specific interaction PRS summarizes interaction effects of a given SNP1 with all other independent SNPs (SNP2_i i = 1,2,.. N_SNP −1) on the trait of interest. Using such SI_PRS one gets an opportunity to evaluate collective associations of different numbers of interacting SNP-pairs of a given SNP1 with other SNPs (ordered with respect to p-values of their interaction effects (from smallest to largest) on AD). Such ordering allows for construction of many SI_PRS indices for a given SNP1 that consist of different numbers of interacting SNP-pairs. These indices can be ranked with respect to p-values of the last SNP-pairs included in them (the largest p-value among all SNP-pairs included in the index). Each such p-value threshold determines the number of SNP-pairs in the index. Such ranking is also convenient for representing properties of all such indices (e.g., percent of explained phenotypic variance) at one diagram where the horizontal axis shows the values of the p-value threshold (see diagrams in Figs. 1 and 2).

Fig. 1.

On the left: The diagram of the proportion of phenotypic variance (R²) explained by the SI_PRS for rs8035396. The horizontal axis shows p-value threshold that determines the number of SNP-pairs included in this composite index. The numbers on the top of each pillar describe: (a) the number of SNP-pairs corresponding to the p-value threshold (shown below this pillar at the horizontal axis), (b) the value of R², and (c) the significance (p-value) of the association of this SNP-specific SI_PRS with AD. The analysis included 9488 SNP-pairs ranked with respect to p-values of their interaction effects on AD (from smallest to the largest). Both positive and negative effects on AD were used in the construction of SI_PRSs. The HRS data on 2095 genotyped ((Illumina-2.5mln) males and females combined were used in the analysis. The logistic regression model with the interaction term has been used for evaluating interaction effects of SNP-pairs. Education, smoking, sex and rs8035396 were used as observed covariates in regression model. On the right: The ROC curve characterizing the fit of the logistic regression model describing the association of the rs8035396-specific SI_PRS that included 217 SNP-pairs for whom the significance of interaction effects on AD did not exceed the p-value threshold 1.0E-2. The area under the curve (AUC) characteristics is 0.64 with 5% confidence interval (62–67). The logistic regression model of generalized linear models (GLM) in R package with education, smoking, sex, and rs8035396 as observed covariates have been used. The R package pROC (Robin et al., 2011) has been used to calculate and plot the receiving operating characteristic (ROC) curve.

Fig. 2.

On the left: The diagram of the proportion of phenotypic variance (R²) explained by SI_PRS for rs7175913. All notations and explanations are the same as used in description of Fig. 1. On the right: The receiver operating characteristics (ROC) curve characterizing the fit of the logistic regression model describing the effect of the SI_PRS for rs7175913 that included 183 SNP-pairs for whom the significance of interaction effects on AD did not exceed the p-value threshold 1.0E-2. The area under the curve (AUC) characteristics is 0.70 with 5% confidence interval (68–73). Other notations and explanation are the same as used in description of Fig. 1.

4. Results

4.1. Associations of interactions between the EIF2AK4 and APP genes with AD

After quality control procedure 59 and 119 SNPs remained from the EIF2AK4 and the APP genes, respectively, which corresponds to 7021 SNP-pairs. The interaction analysis among other things produced the values of the type I errors for each SNP-pair. The analysis of dependence between SNPs from each of the two genes showed that many SNPs are in high linkage disequilibrium (LD) with each other. We used this property together with information about the values of the type I errors available for each SNP-pair to reduce the number of interacting SNP-pairs by selecting appropriate independent (tag) SNPs. This clumping procedure available in PLINK allowed us to select such SNPs from each gene and corresponding SNP-pairs keeping the SNP-pairs with the smallest values of the type 1 error. The use of LD threshold R2 0.1 resulted in 6 SNPs from the EIF2AK4 gene and 15 SNPs from the APP gene. This corresponds to 90 SNP-pairs prepared for evaluating associations of genetic interactions with AD. The Bonferroni correction for multiple testing corresponds to p-value threshold 5.55E-04 which is substantially larger than the smallest p-value, 9.94E-06, obtained in our analysis of interacting SNP-pairs. Interaction of one SNP-pair rs8035396 and rs2829973 was significantly associated with AD with p-value 9.94E-06 vs Bonferroni correction threshold 5.55E-04.

Thus, the results of our analysis of human data confirm the presence of association between interacting SNPs from the two genes and AD. Note that such connection may be mediated by actions of many other genes which aere not included in this statistical analysis.

The most significant results of analysis performed with all 7021 SNP-pairs are shown in Table S1 (Supplementary data).

4.2. Properties of the SNP-specific interaction polygenic risk scores

We selected two independent (not in LD) SNPs, rs8035396 and rs7175913, from the EIF2AK4 gene for construction SI_PRS from Table S1. Then we constructed SI_PRS for these SNPs, evaluated their associations with AD and investigated their properties. The properties of the rs8035396 SI_PRS indices are illustrated in Fig. 1.

The properties of the rs7175913 SI_PRS indices are illustrated in Fig. 2.

The vertical axis at the diagram on the left of Fig. 1 (and Fig. 2) shows the proportion of phenotypic variance, R², explained by the SI_PRS describing interactive capacity of the rs8035396 SNP. The horizontal axis shows the p-value threshold (see Section 3.2) which determines the number of SNP-pairs included in construction of each rs8035396 SI_PRS.

One can see from this diagram that 217 SNP-pairs are included in the SI_PRS determined by determined by the threshold value 0.01. This SI_PRS explains 7.8 % of phenotypic variance and the association of this SI_PRS with AD is statistically significant with the p-value 2.6E-21. This index can be called the rs8035396 SI_PRS-217, or the rs8035396 SI_PRS_0.01 depending on whether one wants to emphasize the number of interactive SNP-pars included in this index, or the value of the threshold used to restrict the number of ordered interacting SNP-pairs included in this index. The SI_PRS with the threshold value 0.05 includes 1023 SNP-pairs and explains 15.2 % of phenotypic variance. The association of this SI_PRS with AD is highly statistically significant with p-value 3.1E-39. This index can be called rs8035396 SI_PRS-1023, or rs8035396 SI_PRS_0.05.

Similarly, the diagram on the left in Fig. 2 shows that 183 interacting SNP-pairs are included in the rs7175913 SI_PRS with the threshold value 0.01. This SI_PRS explains 15.5 % of phenotypic variance and the association of this SI_PRS with AD is highly statistically significant with the p-value 2.0E-34. This index can be called rs7175913 SI_PRS-183, or rs7175913 SI_PRS_0.01. The SI_PRS with the threshold value 0.005 includes 78 interacting SNP-pairs and explains 13 % of phenotypic variance. This index can be called rs7175913 SI_PRS-78, or rs7175913 SI_PRS_0.005. The association of this SI_PRS with AD is highly statistically significant with p-value 1.5E-32.

Comparison of the SI_PRSs corresponding to the threshold 0.01 (horizontal axis) on Figs. 1 and 2 shows that the rs8035396 SI_PRS has higher number of interacting SNPs than corresponding the rs7175913 SI_PRS. However, the rs7175913 SI_PRS explains larger percent of phenotypic variance than the rs8035396 SI_PRS (15.5 % vs 7.8 %). This observation indicates that the level of interactive ability represented by SI_PRS may characterize role of this SNP in multigenic regulation of AD.

The comparison of results of genetic association studies of AD and human longevity showed that many highly significant genetic variants associated with AD are also associated with longevity traits (Lu et al., 2014; Yashin et al., 2018b; Shadyab et al., 2017). It is interesting to know whether SI_PRS indices constructed for AD are also associated with longevity traits. To address this problem, we estimated associations of these indices with longevity trait defined as “1” if individual’s lifespan exceeds or is equal to age 85 years, ans as “0” if his/her lifespan does not reach age 75 years. The significant associations of the rs8035396 SI_PRS indices with longevity trait are observed starting with the SI_PRS-104 index constructed from the first 104 interacting SNP-pairs, with p < 2.9E-02. The associations of indices rs8035396 SI_PRS-217 and rs8035396 SI_PRS-1023 constructed from 217 and 1023 interacting SNP-pairs with longevity trait are statistically significant, with p < 1.2E-03 and p < 5.2E-11, respectively.

Similarly, significant associations of the rs7175913 SI_PRS indices with longevity trait are observed starting with the SI_PRS-78 index constructed from the first 78 interacting SNP-pairs, with p < 4.2E-04. The associations of indices rs7175913 SI_PRS-183 and rs7175913 SI_PRS-933 constructed from 181 and 933 interacting SNP-pairs with longevity trait are significant, with p < 2.3E-07 and p < 2.4E-14, respectively. The three indices constructed for each of the two SI_PRSs correspond to the p-value thresholds 0.001, 0.01 and 0.05 shown at the horizontal axis of the diagrams on the left in the Fig. 1 (Fig. 2), respectively.

The diagrams on the right of the Figs. 1 and 2 also show that the model with the rs7175913 SI_PRS is a better fit for data than model with the rs8035396 SI_PRS_0.01. Although the estimated values of AUCs (0.644 for the rs8035396 SI_PRS, and 0.703 for the rs7175913 SI_PRS_0.01) do not look very impressive – in each case this is a contribution from an interaction component related to one SNP. The construction of the gene-specific SI_PRS or pathway-specific SI_PRS will better present the data and explain the larger percent of phenotypic variance.

5. Discussion

5.1. New insights for further research

The results of analyses of human data performed in this paper indicate that the interactions of the GCN2/EIF2AK4 gene with the APP gene may play crucial role in the initiation, development, and progression of AD. The detected associations may result from the interplay among multiple genes from many signaling and metabolic pathways that, together with these two genes, are involved in the AD regulation. The statistical analysis performed in this paper allowed us to estimate the role of interplay between two selected genes. Although these results do not explain the entire polygenic mechanisms of AD regulation, the detected associations provide us with useful insights concerning directions of further research that could clarify the roles of other players that contributed to detected association of the interactions between GCN2/EIF2AK4 and APP with AD.

For example, several studies emphasized an important role of the GCN2/EIF2AK4 gene in the cell cycle regulation pathways (Grallert and Boye, 2007; Menacho-Marquez et al., 2007; Silva et al., 2015). Other studies provide evidence that the aberrant cell cycle reentry in neurons is observed in AD (Arendt, 2012; Currais et al., 2009; Koseoglu et al., 2019). At the same time, it was also known that APP promotes cell proliferation (Lim et al., 2014). This information in combination with our findings suggests a way to learn more about the details of detected connections by identifying genes from pathways mentioned above and estimating associations of corresponding genetic interactions with AD using human data. These analyses may reveal new features of polygenic regulation of this health disorder.

5.2. Interaction analysis may identify new genetic players in AD

The association analyses of genetic interactions with AD performed in this study showed that most SNPs from the interacting SNP-pairs that have significant associations with AD (e.g., all SNPs from Table S1 do not have nominally significant individual associations with AD (results not shown)). As a result, the involvement of these SNPs and respective genes in AD regulation would not be detected in the traditional genome- wide or candidate genes association studies of this health disorder. Thus, the genetic interaction analyses allow for detecting new genetic factors that contribute to multifactorial AD regulatio

5.3. Aβ mediates the influence of stress related genes on AD

A substantial part of useful information about connection of AD with cellular stressors, and biomarkers mediating the influence of stress response genes on AD has been obtained in experiments with transgenic laboratory animals and with cellular cultures performed in vitro. In particular, it was found that the integrated stress response (ISR) in which GCN2/EIF2AK4 is involved, initiates expression of the BACE1 gene that codes for β-secretase (Chami and Checler, 2012). This enzyme together with γ-secretase performs the proteolytic processing of the APP to pro- duce Aβ (De-Paula et al., 2012). The extracellular accumulation of the beta amyloid plague is an important AD hallmark. Soluble Aβ oligomers, the major toxic substances associated with the AD pathology, can be found in cerebrospinal fluid, interstitial fluid, and plasma (Menendez--Gonzalez et al., 2018; Wang et al., 2017). This means that the preclinical AD biomarker, Aβ, is a mediator of the effects of stress related genes on AD. This biomarker may continue mediate effects of these genes on AD at any stage of disease development.

5.4. Translation of information from experimental studies to humans

Disappointing results followed numerous attempts to find efficient medication against AD manifest difficulties in the translation of experimental research findings to applications in humans. This translation difficulty motivated the researchers to study mechanisms of AD development using human data. The idea to treat disease at its early stage before its clinical manifestation resulted in the new, biological definition of AD at its preclinical stage. The new definition is based on changes in the preclinical AD biomarkers, and Aβ is one of the key AD biomarkers (Jack et al., 2018). The desire to better understand the early stage of AD development resulted in the collection of human data on preclinical AD biomarkers (Petersen et al., 2010) and in their use in the studies of dynamics of AD development. Although the number of data in such studies is rapidly increasing (Weiner et al., 2015b, 2013; Weiner et al., 2017) their sample sizes are relatively small compared with other available human data on AD.

Although most of these data do not have information about preclinical AD biomarkers they have high research potential for generating useful insights and ideas concerning the nature of AD (Arbeev et al., 2020; Bodily et al., 2016; Li, 2019; Romero-Rosales et al., 2020). In particular, the genome wide association studies of HRS data showed highly significant associations of genes located at chromosome 19 with AD, and these findings were replicated using data from several other datasets (Yashin et al., 2018b). Interestingly, the same genes were associated with human longevity (Yashin et al., 2018a). Most of such analyses, however, test the connections of individual SNPs with AD. The eQTL connections of detected genetic variants with other genes (Xu et al., 2017; Yashin et al., 2018b; Zhbannikov et al., 2017) indicate the importance of studying epistatic effects on AD. The analyses of such data might be useful for translating research findings from animal studies to information appropriate for making conclusions about humans.

5.5. The interplay between cellular stress related genes may influence AD at its any stage

Our earlier studies of complex health related traits showed that the unexpected results of genetic association studies of health and longevity related traits could often be explained by the presence of epistasis (Yashin et al., 1999), a phenomenon in which the effect of genetic mutation depend on the presence or absence of mutations in one or more other genes. The examples, and comprehensive review of paradoxical situations that can be explained by the presence of epistasis are dis- cussed in (Ukraintseva et al., 2016, 2017). The epistasis in AD is a part of multifactorial AD regulation, and its studying using human data requires efficient methods of data analyses. Several studies of AD were focused on evaluating effects of genetic interaction (epistasis) on AD (Ebbert et al., 2015, 2016; Chuang et al., 2012; Cole et al., 2017; Fanutza et al., 2015; Floudas et al., 2014; Gusareva et al., 2018; Kitzlerová et al., 2018; Ritchie, 2015; Wang et al., 2020). To reduce the number of statistical tests many researchers prefer hypothesis-based analysis using candidate genes. In this paper we found that the interplay between GCN2/EIF2AK4 a sensor for cellular stress signals—and the APP – whose proteolytic processing results in beta-amyloid production – is associated with AD. These results indicate that the interaction between other stress related genes and the genes involved in regulation of stress mediating bio- markers may influence AD. Since the AD definition used in our study was based on the results of longitudinal survey (i.e., when clinical disease stage was manifested), the detected association of the interplay between these genes with AD may take place at any stage of disease development including the moment of disease initiation, its preclinical, and clinical stages.

5.6. Other cellular stress response genes

The GCN2/EIF2AK4 gene – whose interaction with the APP gene is associated with AD – is not the only gene from signaling and metabolic pathways involved in cellular stress response. Many other stress related genes are likely to participate in the polygenic AD regulation. The tyrosine kinases EIF2AK1, EIF2AK2, and EIF2AK3 sense other cellular stressors. The response of cell to signals from these sensors phosphorylates EIF2A which initiates changes in translational activity of some genes and in transcriptional activities of others (Ohno, 2014). The polymorphisms in each of these genes could be responsible for variability of responses to the same stress signals in different people. The strength, the duration of cellular stress response and, ultimately, the future fate of the cell depend to a large extent on signals obtained from such sensors. The long-lasting (persistent) stress response may exhaust the cell’s ability to adapt and switch on the program of its self-destruction (apoptosis), contributing to AD initiation, development, and progression. The variability in stressors and polymorphisms in stress related genes may contribute to difference in AD onset among individuals.

5.7. Medical treatment as a stressor

A better understanding of stress related processes of polygenic AD regulation will facilitate detection of its efficient medical treatment. Studying the details of functioning of stress response adaptive machinery in connection with changes in preclinical and clinical AD biomarkers, will generate useful insights and ideas for appropriate preventive and treatment strategies against this health disorder. This is because any medical intervention (e.g., radio- or chemo-therapy) is a stressor, and any treatment procedure such as regularly used medications (prescribed by medical doctors) initiates an organism’s adaptive response (allostatic adaptation (Korte et al., 2005; Yashin et al., 2012a)). The polygenic nature of AD regulation indicates the need for developing multi-targeting approach for AD prevention and treatment.

5.8. Limitations

In this paper we investigated how the interplay between genes implicated in regulation of cellular stress response may influence AD. In the interest of reducing the number of statistical tests and gaining significant results in conditions of multiple testing, our analysis only considered two genes. The results of experimental studies suggest that many other stress related and unrelated genes may contribute to such interplay. Detecting effects of many interacting genes on AD remains a challenging problem that was partly addressed by construction of SNP-specific interaction polygenic risk scores (SI_PRS). The problems with SI_PRS are the same as with the polygenic risk scores (PRS) that summarize the results of genome-wide (or candidate genes) associations studies (Harrison et al., 2020; Choi et al., 2020). Additional issues for SI_PRS are related with more possibilities of its alternative calculations than in the traditional PRS, and with the evaluation of percent of explained phenotypic variance in logistic regression model (Mittlböck and Schemper, 1999).

The development and progression of AD is a dynamic process in which the sets of interacting genetic and non-genetic factors affecting it, as well as the strength of the influence may change with age, time, and the disease stage. Most of genetic interaction analyses performed so far, including this study, did not include these properties of AD into consideration because of the lack of appropriate data and efficient tools for their analysis. The methods of more comprehensive statistical analyses of dynamic polygenic regulation of AD are needed to deepen the understanding of this health disorder. The extension of biodemographic methods of statistical modeling might be a promising way to addressing this problem. These methods allow for linking resistance to stresses, allostatic adaptation and allostatic load (Yashin et al., 2012a, 2016) as well as robustness (vulnerability), resilience, and physiological dysregulation (Arbeev et al., 2014, 2019) with health related traits and longevity.

The analyses performed in this paper evaluated associations of interacting SNP-pairs with AD which does not necessarily describes causal connections. This is, however, the common drawback of statistical methods used in genetic association analyses of observational data. Detecting causal connections could be more useful for better understanding the nature of disease and for developing candidate medications. Testing causality of detected connections requires additional efforts using methods of mediation analyses (Agler and De Boeck, 2017) and Mendelian randomization (Grover et al., 2018). These methods, howeer, have their own limitations (Allman et al., 2018).

The use of the stress origin hypothesis as a conceptual guideline in AD studies would benefit from information about potential cellular (e.g., neuronal, microglial) stressors, as well as about factors and conditions that can affect the variability (e.g., strength, duration) of cellular stress response. However, available human datasets about AD have limited information about such stressors, factors, and conditions. In would be a good idea to include such information in the studies collecting data on preclinical AD biomarkers (Dubois, 2018; Weiner et al., 2015a). In the absence of such data some useful insights and ideas about possible cellular stressors can be obtained indirectly using information about properties of genetic sensors of cellular stress whose involvement in AD regulation can be detected in data analyses. Each sensor recognizes and responds to specific groups of stress signals. The information about many such groups of cellular stress signals is available from experimental data.

The estimates of the effects of individual genetic variants on health and longevity related traits may produce paradoxical results of association analyses. Many these results may be explained by the effects of genetic epistasis (Ukraintseva et al., 2016) which motivates studies of the effects of genetic interactions on AD (Combarros et al., 2009; Ebbert et al., 2015; Gusareva et al., 2018; Raghavan and Tosto, 2017). Studying epistatic effects on AD using human population data is often performed by including the genetic interaction terms in the regression models and using these models in genetic association studies. These methods, however, are completely different from those used for detecting epistasis in biological experimental studies. For this reason, the genetic interactions identified in statistical analyses of human data may have different meaning than those detected in experimental studies (Ebbert et al., 2015).

Nevertheless, the genetic interactions detected in the analysis of population data may be extremely useful for understanding synergistic connections between genetic variants which aren’t always detected in experimental studies. In contrast to biological epistasis that may be detected as a result of proteins’ interactions in biochemical reactions, the statistical interaction can capture connection between genes which is likely to be mediated by the chain of intermediate biomarkers. The new experimental studies can also be initiated to clarify possibilities and details of biological connections consistent with the results of statistical analyses of available human data.

It is important to note that traditional PRS does not necessarily includes SNPs from SNP-specific interaction PRS (SI_PRS) constructed in this paper, and therefore the traditional PRS will not provide an upper bound to the value added from GxG analysis to PRS prediction as one could expect. This is because most SNPs from the SNP-pairs whose interaction showed highly significant associations with AD in our analysis did not show statistically significant individual associations with this trait. This means that many SNPs that contribute to the SI_PRS will not be included in the traditional PRSs constructed from SNPs having significant individual associations with AD. Many currently used traditional PRSs are constructed from SNPs whose individual associations with the trait have p-values smaller than 0.001. Moreover, by construction, the SI_PRSs characterize interactive properties of a given SNP. It makes sense to compare these properties for different SNPs. This is what we did for SNPs rs8035396 and rs7175913. Since SI_PRS is being trained and tested in the same data the association results represent an upper-bound of the true associations. More analyses are needed to replicate these results using other dataset that has similar design and genetic obtained from the same genotyping platform.

6. Conclusions

The results of analyses of human data obtained in this study support the idea that genes whose products deal with sensing cellular stress signals participate in multifactorial mechanisms of AD regulation. They also show how results of experimental studies could be used for testing hypotheses about specific aspects of multifactorial AD regulation in analysis of human data.

The associations of the SI_PRS with AD detected in the analysis of the role of collective genetic interactions in AD support the idea that polygenic AD regulation includes interplay among stress related genes. This connection indicates that stress-related conceptual framework can serve as a convenient research guideline for studying mechanisms of AD initiation and development.

Experimental data provided us with compelling evidence about fundamental role of the interplay of genes from the cellular stress response pathways in AD (Ohno, 2014). However, to our knowledge, there are no studies of human population data that evaluate association of such interplay with AD or with its hallmarks in the analysis of human data. The detection of such associations in our study is an important “proof of principle” and a step forward in translation of evidence about possible AD regulation in transgenic laboratory animals into useful information about some aspects of multifactorial regulation of AD in humans.

The properties of newly constructed composite indices may be used in detection of SNPs and genes which demonstrate high interactive capacity in AD regulation. These properties may play important role in clarifying the details of genetic network involved in AD regulation in humans. Note that such networks may differ in animals and humans.

The fact that the effects of medical interventions on AD are likely to be mediated by cellular stress response indicates the importance of understanding of how this response works (e.g., how medication’s presence is sensed, the expression of which genes are modified in response to stress, and how this process could be controlled to promote beneficial and avoid deleterious effects of interventions on AD.

Differences in the exposure to stressful conditions, in the access to health care, the polymorphisms in genes that sense cellular stress signals, and in other genes involved in cellular stress response, together with aging related changes in vulnerability and resilience are likely to be important sources of disparity and heterogeneity in AD. Being evaluated, these characteristics can be used in predicting future AD burden under different scenarios/strategies of reducing disparities, improving access to health case facilities, using medical advances for the groups of individuals exposed to different environmental and living conditions, and having different genetic backgrounds.