Neurobiological substrates underlying the effect of genomic risk for depression on the conversion of amnestic mild cognitive impairment

Depression increases the risk of conversion from amnestic mild cognitive impairment (aMCI) to Alzheimer’s disease. Xu et al. show that the polygenic risk score for major depressive disorder-specific genetic variants predicts conversion from non-depressed aMCI to Alzheimer’s disease. The underlying mechanisms may involve modulation of hippocampal development and amyloid-beta binding.

Abstract

Depression increases the conversion risk from amnestic mild cognitive impairment to Alzheimer’s disease with unknown mechanisms. We hypothesize that the cumulative genomic risk for major depressive disorder may be a candidate cause for the increased conversion risk. Here, we aimed to investigate the predictive effect of the polygenic risk scores of major depressive disorder-specific genetic variants (PRS_sMDD) on the conversion from non-depressed amnestic mild cognitive impairment to Alzheimer’s disease, and its underlying neurobiological mechanisms. The PRS_sMDD could predict the conversion from amnestic mild cognitive impairment to Alzheimer’s disease, and amnestic mild cognitive impairment patients with high risk scores showed 16.25% higher conversion rate than those with low risk. The PRS_sMDD was correlated with the left hippocampal volume, which was found to mediate the predictive effect of the PRS_sMDD on the conversion of amnestic mild cognitive impairment. The major depressive disorder-specific genetic variants were mapped into genes using different strategies, and then enrichment analyses and protein–protein interaction network analysis revealed that these genes were involved in developmental process and amyloid-beta binding. They showed temporal-specific expression in the hippocampus in middle and late foetal developmental periods. Cell type-specific expression analysis of these genes demonstrated significant over-representation in the pyramidal neurons and interneurons in the hippocampus. These cross-scale neurobiological analyses and functional annotations indicate that major depressive disorder-specific genetic variants may increase the conversion from amnestic mild cognitive impairment to Alzheimer’s disease by modulating the early hippocampal development and amyloid-beta binding. The PRS_sMDD could be used as a complementary measure to select patients with amnestic mild cognitive impairment with high conversion risk to Alzheimer’s disease.

Introduction

The amnestic mild cognitive impairment (aMCI) is a state of cognitive deficit that is not severe enough to fulfil the criteria of dementia (Winblad et al., 2004) and shows a much higher probability of developing into Alzheimer’s disease (Palmer et al., 2008). Identifying biological measures with the potential to predict the conversion from aMCI to Alzheimer’s disease is clinically important for early interventions of Alzheimer’s disease. In the past decades, a variety of demographic (Tokuchi et al., 2014), clinical (Mazzeo et al., 2016), cognitive (Julayanont et al., 2014), neuroimaging (Yuan et al., 2009), and genetic measures (Rodriguez-Rodriguez et al., 2013; Adams et al., 2015) have been proposed as candidate measures for the prediction. Among these measures, a lifetime history of major depressive disorder (MDD) (Jorm, 2001), the presence of depressive symptom (Kida et al., 2016; Mourao et al., 2016; Sacuiu et al., 2016; Barca et al., 2017), or the coexistence of a diagnosis of MDD (Modrego and Ferrandez, 2004) has been found to increase the conversion risk from aMCI to Alzheimer’s disease, despite of non-significant findings (Palmer et al., 2010; De Roeck et al., 2016). However, few studies have investigated why depression could increase the conversion risk from aMCI to Alzheimer’s disease.

Besides social-psychological factors, such as stressful life events (Park et al., 2015) and interpersonal dysfunction (Hames et al., 2013), genome-wide association studies (GWAS) reveal a role of genetic factors in MDD (Ripke et al., 2013). In terms of genetics, one possibility for the association between MDD and the conversion of aMCI is that MDD and Alzheimer’s disease may share common genetic variants. This idea is supported by candidate gene studies (Skoog et al., 2015; Ye et al., 2016), but is not supported by the polygenic score analyses (Bulik-Sullivan et al., 2015a; Power et al., 2017). An alternative possibility is that the MDD- and Alzheimer’s disease-specific genetic variants may affect the same biological pathways, processes, cells, or structures via different mechanisms at different developmental periods. Thus, this study aimed to investigate if the cumulative MDD-specific genetic risk could predict the conversion from aMCI to Alzheimer’s disease. If so, we explored the possible neurobiological mechanisms underlying the predictive effect further.

The cumulative genetic risk for MDD could be assessed by the polygenic risk scores (PRS) for this disorder (PRS_MDD) (Milaneschi et al., 2016) and the cumulative genetic risk for Alzheimer’s disease could be evaluated by the PRS_AD (Sabuncu et al., 2012). Although many studies have used the PRS_MDD to predict MDD (Holmes et al., 2012; Milaneschi et al., 2016; Gibson et al., 2017; Qiu et al., 2017) and applied the PRS_AD to predict Alzheimer’s disease (Escott-Price et al., 2015; Mormino et al., 2016) and the conversion from aMCI to Alzheimer’s disease (Rodriguez-Rodriguez et al., 2013; Adams et al., 2015), no studies have used the PRS_MDD to predict the conversion from aMCI to Alzheimer’s disease. To exclude the possibility that the predictive effect of the PRS_MDD on the conversion is driven by genetic variants common to MDD and Alzheimer’s disease, we only used genetic variants specific to MDD to calculate PRS_sMDD by excluding common genetic variants of the two disorders. If a significant predictive effect existed, we investigated neuroanatomical substrates underlying such a prediction. The identified genetic variants were fine-mapped into genes using different strategies, and then enrichment and protein–protein interaction network analyses were performed to identify potential functions of these genes. Temporal- and cell type-specific expression analyses were finally used to explore in which developmental periods and cell types these genes are significantly expressed. A schematic summary of the study design is shown in Fig. 1.

Figure 1.

A schematic summary of the study design. ADNI = Alzheimer’s Disease Neuroimaging Initiative; CHR = chromosome; eQTL = expression quantitative trait loci; GMV = grey matter volume; GWAS = genome-wide association studies; IGAP = International Genomics of Alzheimer’s Project; OR = odd ratios; PGC-MDD = major depressive disorder working group of Psychiatric Genomics Consortium; P_T = P-values threshold of genome-wide association studies.

Materials and methods

Discovery and target samples

The PRS calculation requires a discovery sample and a target sample. The discovery sample was used to identify effect size of a set of genetic variants that were nominally associated with the disease status at a predefined P-value. Then, the PRS was calculated for each subject in the target sample to estimate cumulative genetic risk of this subject for the disease. GWAS data of the Psychiatric Genomics Consortium (PGC) (Sullivan, 2010) and International Genomics of Alzheimer’s Project (IGAP) (Lambert et al., 2013) were used as the discovery samples to calculate PRS_MDD and PRS_AD, respectively. The target sample included 398 non-depressed (with a geriatric depression scale < 6) aMCI patients provided by the first stage of Alzheimer’s disease Neuroimaging Initiative (ADNI-1). The detailed information about the discovery and target samples is provided in the Supplementary material.

Genotyping

For the 757 subjects from ADNI-1, the genome-wide single-nucleotide polymorphisms (SNPs) were genotyped using the Illumina Human610-Quad Bead chip.

Quality control

In individual-level quality control, we excluded subjects with a missing genotyping rate >0.05, sex inconsistency, possible relative relationships, and being European population outliers identified by multidimensional scaling (MDS). In SNP-level quality control, we excluded SNPs with a missing call rate >0.05, minor allele frequency <0.01, significant deviation from Hardy-Weinberg equilibrium, and ambiguous strand. Finally, imputation was performed using MaCH (Li et al., 2010) and MiniMac (Howie et al., 2012). Detailed information of quality control and imputation is shown in the Supplementary material.

Polygenic risk score calculation

The PRS is used to assess the cumulative genetic risk for a certain disorder (Purcell et al., 2009). In the discovery sample, we calculated the association between SNPs and disease status at a predefined P threshold (P_T). Under a certain P_T, we then removed the effects of SNPs in linkage disequilibrium (LD) in each clumped region (excluding SNPs with r² > 0.25 within a 250-kb window) and selected the index SNPs with the most significant P-value from each clumped association region based on the discovery sample. Thus, we obtained the information of the risk alleles and effect sizes of the index SNPs for each P_T value. In the target sample, the PRS was calculated for each individual as the sum of the count of risk alleles multiplied by the corresponding effect sizes (natural log of the odds ratio) across these index SNPs.

To identify the P_T value that could generate PRS with the best prediction for the conversion from aMCI to Alzheimer’s disease, the PRSice software (http://prsice.info) (Euesden et al., 2015) was used to generate 1000 PRS values for P_T ranging from 0.001 to 1 with an increment of 0.001, while controlling for sex, age and educational years at baseline, the number of APOE ɛ4, and the first four MDS components for population stratification. P_T = 1 indicates that all SNPs of the discovery sample are included to calculate the PRS in the target sample. By evaluating the predictive abilities, we could obtain the best P_T values for calculating PRS_MDD and PRS_AD in the target sample. Using the best P_T values, we could obtain the risk alleles and effect sizes of the index SNPs for calculating the PRS_MDD and PRS_AD in the target sample. The PRS_MDD and PRS_AD were computed for each aMCI patient and then z-transformed for visualization.

To exclude the possibility that the predictive effect of PRS_MDD on the aMCI conversion is driven by genetic variants common to MDD and Alzheimer’s disease, we calculated PRS specific to MDD (PRS_sMDD) only using genetic variants specific to MDD by excluding common variants of the two disorders. In the same way, we also calculated PRS specific to Alzheimer’s disease (PRS_sAD). LD score regression (Bulik-Sullivan et al., 2015b) and co-localization analyses (Giambartolomei et al., 2014; Pickrell et al., 2016) were additionally performed to validate the specificity of index SNPs to MDD rather than Alzheimer’s disease (Supplementary material).

Image acquisition

Subjects were scanned with a standardized MRI protocol developed for ADNI (Jack et al., 2008). Details about the rationale and development of the standardized MRI datasets have been previously described (Wyman et al., 2013). The high resolution structural MRI data were acquired at 59 sites using 1.5 T MRI scanners with a sagittal 3D magnetization prepared rapid acquisition gradient echo sequence (http://adni.loni.ucla.edu).

Grey matter volume calculation

All structural images were visually checked by two radiologists. The preprocessing processes of structural images included segmentation, normalization, modulation, and smoothing (Supplementary material). The specific reasons for exclusion for patients with aMCI are shown in Supplementary Table 1. Finally, 322 aMCI patients with qualified genotyping and neuroimaging data were included and divided into conversion (aMCI-C, n = 187) and stable (aMCI-S, n = 135) groups. The demographic and genetic data of these patients are shown in Table 1.

Table 1.

Demographic and genetic characteristics of the target sample

Demographic variables	aMCI-S (n = 135)	aMCI-C (n = 187)	Statistics	P
Age at baseline, y	75.64 (7.16)	74.50 (7.21)	1.99	0.16
APOE ɛ4 carriersa, n	62	120	10.62	1.10 × 10−3
Educational years	15.50 (3.12)	15.73 (2.89)	0.46	0.50
Males, n	90	120	0.22	0.37
GDS	1.63 (1.38)	1.59 (1.36)	0.07	0.79
PRSADb	−0.41 (0.88)	0.30 (0.98)	44.04	1.37 × 10−10
PRSMDDb	−0.20 (0.97)	0.28 (0.97)	19.35	1.48 × 10−5
PRSsADb	−0.40 (0.88)	0.29 (0.98)	42.59	2.65 × 10−10
PRSsMDDb	−0.16 (0.97)	0.22 (1.00)	11.96	6.18 × 10−4
PRSsMDD+ADb	−0.40 (0.89)	0.31 (0.98)	43.63	1.29 × 10−10

Data are shown as mean (SD). AD = Alzheimer’s disease; GDS = Geriatric Depression Scale.

^aAPOE ɛ4 carriers include subjects with one or two copies of ɛ4 allele at the APOE locus.

^bThe PRS are z-transformed for visualization.

P-values in bold indicate there are significant differences between groups.

Statistical analyses

Demographic data

The demographic data were analysed using the Statistical Package for the Social Sciences version 18.0 (SPSS Inc., Chicago, Illinois, USA). The chi-square or t-test was used to compare differences in sex, age, educational years, geriatric depression scale, and PRS measures between the aMCI-S and aMCI-C groups.

Polygenic risk analysis

The logistic regression was used to predict odds of the aMCI-C using the PRS calculated under each of the 1000 P_T thresholds. The permutation test (P < 0.05) was used to correct multiple comparisons. Nagelkerke’s pseudo R² was calculated to measure the proportion of variance explained by the PRS. A Cox proportional hazard model was used to explore the relations between the PRS and the conversion at the different time points, with the age and educational years at baseline, the sex, the number of APOE ɛ4, and the PRS as independent variables. For the best-fitting PRS, hazard ratios (HRs) for these variables were calculated using the prediction models.

The 322 aMCI patients were bisected according to their PRS_sMDD or PRS_sAD; we defined 161 patients with relatively low PRS as low PRS group and another 161 patients as high PRS group. The chi-square test was used to compare differences in conversion rate of aMCI among the four hierarchical risk groups (double low risk, low PRS_sAD but high PRS_sMDD, high PRS_sAD but low PRS_sMDD, and double high risk groups). The 322 aMCI patients were also trisected according to their PRS_sMDD or PRS_sAD, the bottom third (107 patients) was defined as the low-risk group, the middle third (107 patients) as the middle-risk group, and the upper third (108 patients) as the high-risk group. We also compared differences in conversion rate among the nine hierarchical risk groups.

Imaging data

The voxel-wise multiple regression analysis was performed to identify brain regions whose grey matter volumes were significantly correlated with PRS_sMDD using Statistical Parametric Mapping software package (SPM8, http://www.fil.ion.ucl.ac.uk/spm), while controlling for the effect of neuroimaging sites. Multiple comparisons were corrected using a voxel-level family-wise error (FWE) method (P_c < 0.05, cluster size >200 voxels). The grey matter volumes of brain regions with significant correlations with the PRS_sMDD were extracted for further analysis.

Mediation analysis and Mendelian randomization

The mediation analysis was performed to test whether the grey matter volume of each significant region mediates the association between PRS_sMDD and the conversion of aMCI (Preacher and Hayes, 2008; Hayes, 2013). The PRS_sMDD was defined as an independent variable, the grey matter volume of each significant region as a mediator variable, and the aMCI group assignment (aMCI-S versus aMCI-C) as a binary dependent variable. In addition, both a conventional two-stage Mendelian randomization method and a Mendelian randomization-Egger sensitivity analysis (http://www.mendelianrandomization.com/index.php/software-code) were applied using PRS_sMDD as instrumental variable to make causal inference between the grey matter volume of each significant region and aMCI conversion (Burgess and Thompson, 2013, 2015; Burgess, 2014) (Supplementary material).

Fine-mapping MDD-specific genetic variants into genes

The context-dependent epigenomic weighting for prioritization of variants (CEPIP) (Li et al., 2017) was used to estimate the brain region-specific regulatory probability of each PRS_sMDD genetic variant by integrating the expression quantitative trait loci (eQTLs) (Brown et al., 2013) and epigenomic features in the specific human brain tissues (Kundaje et al., 2015). The PRS_sMDD genetic variants were weighted by the brain region-specific regulatory probability, and were then fine-mapped into genes using gene-based association test (GATES) (within a 5-kb window) in the human reference assembly (GRch37/hg19) (Li et al., 2011). As a result, we could identify genes that increase MDD susceptibility through brain region-specific functionally regulatory mechanisms. To exclude the bias derived from introducing brain region-specific biological information, we also fine-mapped PRS_sMDD genetic variants into genes only based on physical position of each variant (within a 5-kb window) and re-performed the analyses.

Gene enrichment analyses

The functions of these genes were annotated by identifying significant enrichments using the WebGestalt (http://www.webgestalt.org/option.php) (Wang et al., 2017) and Gorilla (http://cbl-gorilla.cs.technion.ac.il) (Eden et al., 2007, 2009). The Benjamini and Hochberg method for false discovery rate (FDR-BH correction) (q_c < 0.05) was applied to correct for multiple comparisons. The reference gene list included 18 774 genes derived from fine-mapping of all imputed SNPs of PGC-MDD and ADNI datasets.

Network topology-based analysis

The WebGestalt software was also used to perform the network topology-based analysis based on the human protein–protein interaction (PPI) of the Biological General Repository for Interaction Datasets (BIOGRID). All fine-mapping genes were firstly mapped to the PPI network of the BIOGRID; and the tightly connected genes (a portion of the candidate genes) formed a new PPI network based on the assumption that the mechanistically important genes are likely to form tightly connected clusters whereas other genes tend to be randomly distributed in the network. The network topology-based analysis could create a score for each gene in the PPI network of the BIOGRID based on its overall proximity to the seed genes, where the proximity was measured by the random walk analysis (Kohler et al., 2008). In the PPI network of the BIOGRID, the top 10 genes with the most functional similarity with seed genes were identified as neighbouring genes and included in the new PPI network. The constituent genes of the resulted network, including the tightly connected candidate genes and the top 10 neighbouring genes, were enriched in gene ontology (GO) items using the hypergeometric test (q_c < 0.05).

Temporal expression analysis

The precise regulation of the spatial and temporal gene expression plays a key role in brain development, maturation and ageing. The CSEA was used to explore in which developmental periods these PRS_sMDD-related genes are specifically expressed (Xu et al., 2014) (http://genetics.wustl.edu/jdlab/csea-tool-2/). The analysis was performed using the Fisher’s exact test (q_c < 0.05) across developmental stages at a specificity index threshold (pSI) threshold of 0.05. If any genes showed specific expression in any developmental stage under the most stringent threshold of pSI = 0.001, the expression pattern of these genes would be depicted using Human Brain Transcriptome (Kang et al., 2011). In our analysis, we only focused on temporal feature of these genes expressed in the significant brain regions of mediation analysis and Mendelian randomization.

Cell type-specific expression analysis

With the single-cell RNA-sequencing technique, a prior study has provided the detailed cell type-specific expression data of nine major cell types in the mouse brain regions (Zeisel et al., 2015). The Fisher’s exact test was used to identify in which cell types these PRS_sMDD fine-mapping genes were specifically expressed in a certain brain region. These tests resulted in the pSI for each cell type, and the P threshold represents how likely the gene set is specifically expressed in a given cell type relative to the other cell types.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Results

PRS_MDD could predict the conversion from aMCI to Alzheimer’s disease

When the PRS_MDD were calculated using seven broad P_T values for associations between SNPs and MDD, we found the PRS_MDD (P_T = 0.05) could predict the status of the aMCI-C (P = 0.002) (Fig. 2A) and explained 4.08% of variance for the aMCI-C based on 10 618 index SNPs. When the PRS_MDD were computed using a high resolution method, the PRS_MDD calculated at P_T = 0.009 showed the highest prediction (P = 7.49 × 10⁻⁵), explained 6.86% of variance based on 2559 index SNPs (Fig. 2B and Table 2). Compared with aMCI-S patients, aMCI-C patients had a significantly higher adjusted PRS_MDD (P = 1.48 × 10⁻⁵) (Fig. 2C and Table 1). Compared to the low risk group, the middle [HRs = 2.45, 95% confidence interval (CI) = 1.48–4.03, P = 4.56 × 10⁻⁴] and high (HRs = 2.68, 95% CI = 1.24–5.82, P = 3.23 × 10⁻⁴) risk groups had more than two times of risk for the aMCI-C. The Cox survival analysis showed that the middle and high risk groups had a higher proportion of the aMCI-C than the low risk group (Fig. 2D). Moreover, the middle and high risk groups (mean time to conversion 35.91 months) progressed more rapidly into Alzheimer’s disease than the low risk group (mean time to conversion 42.75 months).

Figure 2.

The predictive effects of PRS_MDD (A–D) and PRS_AD (E–H) on the conversion from aMCI to Alzheimer’s disease, conversion rate among the bisected PRS groups (I) and mediation analysis (J–K). (A and E) The bar plots show the predictive effects of the PRS constructed by the best-fit P_T and other seven broad P_T values (0.001, 0.05, 0.1, 0.2, 0.3, 0.4 and 0.5) on the conversion of aMCI. The numbers above the bars are P-values for the logistic regression between the PRS and aMCI-C. The y-axis shows proportion of variance explained by the PRS. (B and F) The high resolution PRS analyses identify the best-fit P_T for constructing the PRS_MDD (P_T = 0.009) and PRS_AD (P_T = 0.352). The x-axis shows the P_T from 0.001 to 1 with an increment of 0.001. The black line shows the P-value of the logistic regression (left y-axis) at different P_T; the dashed line shows numbers of SNPs (right y-axis) at different P_T; and the red line connects points at the P values of the broad PRS analysis. (C and G) Distribution of the PRS measures for aMCI-S and aMCI-C. The y-axis shows the counts of each PRS bin (x-axis). (D and H) Cox proportional hazard model shows the associations between the PRS and the conversion from aMCI to Alzheimer’s disease at different time points (x-axis). The y-axis shows the cumulative proportion of aMCI-C for any given follow-up period on the x-axis. The red, green and purple lines show the low, middle and high PRS groups, respectively. (I) Differences in conversion rate among the bisected PRS groups. (J) Left: The brain region with significant negative correlations between the PRS_sMDD and GMV. Right: The scatter plots show correlations between the PRS_sMDD and GMV in the aMCI-S and aMCI-C groups. (K) The mediation analysis shows that the left hippocampal volume mediates the predictive effect of the PRS_sMDD on the conversion of aMCI. AD = Alzheimer’s disease; P_T = P-values threshold of genome-wide association studies; *P < 0.01; **P < 0.001.

Table 2.

The predictive effects of PRS on the conversion of aMCI (n = 322)

	PT	SNPs	iSNPs	P	R2, %	Specificity, %	Senstivitity, %	Accuracy, %	AUC
PRSAD	0.352	2 622 845	49 831	7.10 × 10−10	18.55	71.85	87.17	80.75	0.72
PRSMDD	0.009	13 472	2559	7.49 × 10−5	6.86	68.89	83.43	77.33	0.65
PRSsAD	NA	2 618 134	49 504	1.05 × 10−9	18.09	70.24	85.63	80.12	0.70
PRSsMDD	NA	8761	1806	1.74 × 10−3	4.19	65.11	81.74	76.01	0.62
PRSsMDD+AD	NA	2 631 606	50 527	2.26 × 10−10	19.17	70.37	88.24	81.06	0.75

AUC = area under curve of receiver operating characteristic curve; iSNPs = numbers of index single-nucleotide polymorphisms that constitute PRS; NA = not applicable; P_T = P-values threshold of genome-wide association studies; R² = Nagelkerke’s pseudo R² of logistic regression; SNPs = numbers of single-nucleotide polymorphisms that constitute PRS.

We also investigated the predictive effect of the PRS_AD on the aMCI-C to validate the PRS method and to provide a reference for the predictive effect of the PRS_MDD. The broad analysis showed that the PRS_AD calculated at P_T = 0.4 could predict the aMCI-C (P = 1.20 × 10⁻⁹) (Fig. 2E), explained 18.02% of variance in the aMCI-C based on 54 106 index SNPs. The high resolution analysis revealed that the PRS_AD computed at P_T = 0.352 showed the highest prediction (P = 7.10 × 10⁻¹⁰), explained 18.55% of variance based on 49 831 index SNPs (Fig. 2F and Table 2). The aMCI-C group showed higher PRS_AD than the aMCI-S group (P = 1.37 × 10⁻¹⁰) (Fig. 2G and Table 1). Compared to the low risk group, the middle (HRs = 1.55, 95% CI = 1.04–2.31, P = 0.032) and high (HRs = 3.35, 95% CI = 2.30–4.88, P = 3.25 × 10⁻¹⁰) risk groups had increased risk for the aMCI-C. Cox survival analysis showed that the high risk group had a much higher proportion of the aMCI-C than the low risk group (Fig. 2H). Sensitivity, specificity, accuracy, area under curve were also provided (Table 2) and the accuracy of the models was confirmed by 10-fold cross validation (Supplementary material and Supplementary Table 2). We also found that aMCI patients carrying APOE ɛ4 showed an increased risk for the conversion than APOE ɛ4 non-carriers (HRs = 1.78, 95% CI = 1.31–2.42, P = 2.21 × 10⁻⁴); however, the age, educational years at baseline, and sex were not associated with the conversion.

PRS_sMDD could independently predict the conversion from aMCI to Alzheimer’s disease

After excluding overlapping SNPs (n = 4711) between PRS_MDD and PRS_AD, the PRS_sMDD could significantly predict the aMCI-C (P = 0.002) and explained 4.19% of variance for the aMCI-C based on 1806 index SNPs, suggesting that MDD-specific genetic variants have an independent contribution to the aMCI-C (Table 2). Similarly, the PRS_sAD could significantly predict the aMCI-C (P = 1.05 × 10⁻⁹) and explained 18.09% of variance for the aMCI-C based on 49 504 index SNPs (Table 2). Additionally, PRS for MDD-specific and Alzheimer’s disease-related genetic variants (PRS_sMDD+AD) could predict the aMCI-C (P = 2.26 × 10⁻¹⁰) and explained 19.17% of variance for the aMCI-C based on 50 527 index SNPs (Table 2). Compared to PRS_AD that explained 18.55% of variance for the aMCI-C, only 0.33% unique MDD variants were added to the PRS_sMDD+AD; however, the explained variance increased by 3.34%. To exclude the potential effects of known Alzheimer’s disease genome regions, we removed SNPs located in 500 kb regions centred on the top 10 Alzheimer’s disease loci (http://www.alzgene.org/) (Supplementary Table 3) and the predictive model of each PRS was still significant (Supplementary Table 4).

To validate the specificity of these index SNPs to MDD rather than Alzheimer’s disease, LD score regression showed that there was not any significant genetic correlation [r_g = 0.69, standard error (SE) = 1.32, P = 0.60] between MDD-specific genetic variants and Alzheimer’s disease-related genetic variants. Using a Bayesian posterior probability of 0.90 as a cut-off threshold, there are 1582/1806 (87.5%) index SNPs that are highly associated with MDD (Supplementary Fig. 1A), but none of them are associated with Alzheimer’s disease (Supplementary Fig. 1B), and none of these SNPs show significant colocalization between MDD and Alzheimer’s disease (Supplementary Fig. 1C).

To balance the number of index SNPs used to construct the PRS, we created (i) PRS_tsAD, PRS_AD for the top 1806 Alzheimer’s disease-specific index SNPs with the same numbers of index SNPs as the PRS_sMDD; and (ii) PRS_sMDD and PRS_sAD under the P_T = 0.009 (the same P_T as the PRS_MDD). We found that the predictive effect of PRS_tsAD and PRS_sMDD on conversion risk of aMCI were also significant (Supplementary material and Supplementary Table 5). These results indicate that the conversion from aMCI to Alzheimer’s disease is related to only a small number of MDD-specific genetic variants but to a large number of Alzheimer’s disease-specific genetic variants.

When the aMCI patients were divided into the low and high PRS_sMDD groups, the high PRS_sMDD group (65.38%) showed 16.25% higher conversion rate than the low PRS_sMDD group (49.13%) (P = 0.002). When patients were divided into the double low risk group, low PRS_sAD but high PRS_sMDD group, high PRS_sAD but low PRS_sMDD group, and double high risk group. There were significant differences in the conversion rate of the aMCI among these hierarchical PRS groups (P = 4.26 × 10⁻⁷). In the low PRS_sAD group, the aMCI patients with high PRS_sMDD showed marginally higher conversion rate than those with low PRS_sMDD (36.84% versus 51.81%, P = 0.05). In the high PRS_sAD group, the aMCI patients with high PRS_sMDD showed significantly higher conversion rate than those with low PRS_sMDD (63.21% versus 80.26%, P = 0.002) (Fig. 2I and Table 3). When PRS_sMDD and PRS_sAD were trisected into the low, middle and high risk, there were significant differences in conversion rate among the nine hierarchical PRS groups (P = 3.23 × 10⁻⁶) (Supplementary materials andSupplementary Table 6).

Table 3.

Conversion rates of aMCI in the bisected PRS groups

PRS groups	aMCI-S, n	aMCI-C, n	Conversion rate, %
Low PRSsAD and low PRSsMDD	48	28	36.84
Low PRSsAD and high PRSsMDD	40	43	51.81
High PRSsAD and low PRSsMDD	32	55	63.21
High PRSsAD and high PRSsMDD	15	61	80.26
All	135	187	58.07

Table 4.

Top 10 neighbouring genes of the PPI network from PRS_sMDD fine-mapping of 1860 genes and PRS_sMDD fine-mapping of 1608 genes

Ranking	PRSsMDD fine-mapping of 1860 genesa		PRSsMDD fine-mapping of 1608 genesb
	Gene symbol	Random walk probability	Gene symbol	Random walk probability
1	APP	6.84 × 10−3	APP	6.78 × 10−3
2	ELAVL1	4.49 × 10−3	ELAVL1	4.49 × 10−3
3	NTRK1	2.21 × 10−3	NXF1	2.25 × 10−3
4	NXF1	2.03 × 10−3	NTRK1	2.17 × 10−3
5	CUL3	1.70 × 10−3	CUL3	1.71 × 10−3
6	MOV10	1.50 × 10−3	MOV10	1.50 × 10−3
7	TP53	1.45 × 10−3	UBC	1.46 × 10−3
8	EWSR1	1.43 × 10−3	TR53	1.32 × 10−3
9	UBC	1.42 × 10−3	EWSR1	1.19 × 10−3
10	TMEM17	1.33 × 10−3	COPS5	1.17 × 10−3

^aPRS_sMDD genetic variants were fine-mapped into 1860 genes based on the hippocampal-specific regulatory probability between eQTLs and epigenomic features (within a 5-kb window).

^bPRS_sMDD genetic variants were fine-mapped into 1608 genes based on physical position of each variant (within a 5-kb window).

The abbreviation of genes is referred to at https://www.ncbi.nlm.nih.gov/gene/.

Hippocampal volume mediates association between PRS_sMDD and aMCI-C

To identify brain regions whose grey matter volumes were associated with PRS_sMDD, voxel-wise correlations were performed between PRS_sMDD and grey matter volume in the whole brain (P_c < 0.05, voxel-level FWE correction, cluster size > 200 voxels). The PRS_sMDD was negatively correlated with the grey matter volume of the left hippocampus (Brodmann area 36; peak MNI coordinate: x = −27, y = −13.5, z = −21; peak intensity = −5.329; 227 voxels; Fig. 2J).

A total of 1806 index SNPs specific to MDD (r² < 0.25 within 250-kb window) were included in the calculation of PRS_sMDD after excluding genetic variants common to PRS_MDD (P_T = 0.009) and PRS_AD (P_T = 0.352). We set the PRS_sMDD as independent variable in the mediation analysis and instrumental variable in the Mendelian randomization analysis. In the mediation analysis, we found a significant direct effect from the PRS_sMDD to the aMCI group assignment (P < 0.001); from the PRS_sMDD to the left hippocampal volume (P < 0.01); and from the left hippocampal volume to the aMCI group assignment (P < 0.001). A significant indirect effect was also found in the left hippocampal volume (P < 0.001), which accounted for 19.20% of variance for the aMCI-C in this mediation model (Fig. 2K). In the conventional two-stage Mendelian randomization analysis, we found a negative association between PRS_sMDD and hippocampal volume (β = −0.15, SE = 0.06, P = 0.01); the predicted value of the left hippocampal volume was significantly associated with the aMCI-C (β = −2.41, OR = 0.09, SE = 0.78, P = 0.002). Mendelian randomization-Egger regression indicated no unbalanced horizontal pleiotropy (intercept = 0.011, P = 0.07) in the association between left hippocampal volume and aMCI conversion using PRS_sMDD as the instrumental variable. We also found significant causal effect of the left hippocampal volume on the aMCI conversion in the Mendelian randomization-Egger regression analysis (β_Egger = −2.12, OR = 0.12, SE = 0.77, P = 0.002). These results confirmed a causal chain from PRS_sMDD to hippocampal volume to conversion risk of aMCI.

Enrichment analyses using genes prioritized with hippocampal biological information

The 8762 SNPs used for calculating the PRS_sMDD were prioritized by integrating the hippocampus-specific eQTL and epigenomic features. The weight of the prioritization and the PGC-MDD GWAS P-value of each SNP were imported into the GATES software to fine-map these SNPs into 1860 significant genes (within a 5-kb window, 5705 SNPs located inside genes, P < 0.001).

To functionally annotate the 1860 genes, WebGestalt and GOrilla were applied to identify significant enrichment in GO and pathways (Fig. 3A–C and Supplementary Tables 7–9). In the annotation of GO, 505/1860 genes were enriched in the development process, 855/1860 in the protein binding, and 738/1860 in membrane part (Fig. 3A and Supplementary Table 7). Specifically, these genes mainly over-represented in biological processes of the anatomical structure morphogenesis (q_c = 4.27 × 10⁻⁵), cellular developmental process (q_c = 7.08 × 10⁻⁴) and development process (q_c = 1.14 × 10⁻³) (Fig. 3B and Supplementary Table 8), in the molecular function of the amyloid-β binding (q_c = 8.64 × 10⁻⁵), and in the cellular components of the neuron part (q_c = 3.10 × 10⁻¹¹) and neuron projection (q_c = 1.69 × 10⁻⁵) (Fig. 3C and Supplementary Table 8). In the KEGG pathway analysis, these genes were enriched in neuronal development-related axon guidance (q_c = 3.95 × 10⁻³) (Fig. 3C and Supplementary Table 8). These functional annotations suggest that the PRS_sMDD genes were mainly related to the developmental process, and involved in the molecular binding between amyloid-β protein and its precursor, which is a well-known neuropathology of Alzheimer’s disease.

Figure 3.

Gene enrichment of PRS_sMDD fine-mapping 1860 genes. (A) Enrichment of the PRS_sMDD genes in GO items. The x-axis shows the numbers of genes enriched in each item (y-axis). The red, purple, green bars denote the biological process, molecular function and cellular component, respectively. (B) Top 40 significant enriched GO biological process items of the PRS_sMDD genes. The x-axis shows enrichment factor of each GO item (y-axis). The size of the spheres reflects the number of genes (labelled beside the balls) enriched in each item. The colour of the spheres demonstrates the significance of the enrichment analyses. (C) Top 20 significant enriched GO molecular function and cellular component items, and all significant KEGG pathway items of the PRS_sMDD genes. The purple, green and grey background colours show the molecular function, cellular component and KEGG pathway, respectively. (D) Top 40 significant enriched GO biological process items of the PPI network. (E) Left: PRS_sMDD genes were highly expressed in the middle-late foetal developmental period in the hippocampus. Right: CDC20B was highly expressed in the early and late mid-foetal development periods in the hippocampus. Periods 1–15 are described in Supplementary Table 11. AMY = amygdala; BP = biological process; CBC = cerebellar cortex; CC = cellular component; CDC20B = cell division cycle 20B; cp = cell projection; GO = gene ontology; HIP = hippocampus; MF = molecular function; MD = mediodorsal nucleus of the thalamus; NCX = neocortex; pr = positive regulation; rg = regulation; STR = striatum; tt = transmembrane transporter.

To combine protein-level information, we mapped 1860 genes to the PPI network (642 unmatched genes were excluded). To illustrate potential functional connectivity of the remaindering 1218 genes, we used a network-based approach to identify a tightly connected PPI network (only 846 genes were included in the network). Using the network topology-based analysis, the top 10 neighbouring genes with the most functional similarity with seed genes were also included in the construction of the final PPI network consisting of 856 genes. As the most functionally neighbouring genes (Fig. 3D), the amyloid-β precursor protein (APP) formed the protein basis of the amyloid plaques in the brain of Alzheimer’s disease patients. The 856 genes in the final PPI network were enriched mainly in development processes of the nervous system, such as nervous system development (q_c = 2.83 × 10⁻⁵), neuron projection development (q_c = 9.08 × 10⁻⁵), neuron projection guidance (q_c = 4.20 × 10⁻⁴), neurogenesis (q_c = 6.51 × 10⁻⁴) and generation of neurons (q_c = 9.37 × 10⁻⁴) (Fig. 3D and Supplementary Table 10).

The prior analyses indicate that PRS_sMDD fine-mapping genes are involved in brain developmental processes. We further explored in which developmental periods these genes were over-represented in the hippocampus. The CSEA online tool was used to explore the temporal-specific expression of these genes in the hippocampus. Under a pSI threshold of 0.05, 60 genes showed temporal-specific expression in the hippocampus in the middle-late foetal developmental period (q_c = 6.02 × 10⁻⁴) (Fig. 3E). Under the most stringent threshold (pSI = 0.001), only cell division cycle 20B (CDC20B) exhibited a temporal-specific high expression in the hippocampus in the middle-late foetal stage. The Human Brain Transcriptome dataset showed that the CDC20B was highly expressed in the early and late mid-foetal development period (Fig. 3E).

Among the nine major cell types in the mouse cortex and hippocampus, eight cell types exist in the hippocampal CA1 (Fig. 4A). The PRS_sMDD genes showed substantial over-representation mainly in the hippocampal pyramidal neurons (q_c = 1.60 × 10⁻⁷) and interneurons (q_c = 3.61 × 10⁻⁷) (Fig. 4B). However, these genes also demonstrated specific expression in the microglia cells (q_c = 0.01) and astrocytes (q_c = 0.01) (Fig. 4B).

Figure 4.

Cell type-specific expression analyses. (A) The frequency of the nine major cell types in the CA1 and S1 (top) and only in the CA1 (bottom) of the mouse brains (Zeisel et al., 2015). (B) The first bar plot shows the numbers of genes in eight labelled cell types of hippocampus from the Zeisel et al. study. The second bar plot shows the numbers of overlapping genes between PRS_sMDD fine-mapping 1860 genes and significantly expressed genes in each cell type of hippocampus. The last three bar plots show the cell type-specific expression results of the PRS_sMDD-related genes in the eight labelled cell types of the hippocampus. The dashed red line shows the significant level under q_c < 0.05 in FDR-BH correction. CA = cornu ammonis; GMV = grey matter volume; OR = odd ratios; S1 = somatosensory cortex.

Enrichment analyses using genes fine-mapped based on physical position

We also re-mapped 8762 MDD-specific SNPs into 1608 genes only based on the physical position of each variant (within 5-kb window). We found that 1468 genes were common between the two fine-mapping methods (with and without hippocampal biological information), suggesting that most genes related to these MDD-specific SNPs are associated with hippocampal gene expression. We performed enrichment analyses for the 1608 genes and found significant enrichment in developmental process and amyloid-β binding (Supplementary material, Supplementary Fig. 2 and Supplementary Tables 9 and 10).

Discussion

To our knowledge, this is the first study that integrates cross-scale neurobiological analyses and functional annotations to investigate the independent predictive effect of MDD-specific genomic variants on the conversion from aMCI to Alzheimer’s disease and the underlying neurobiological mechanisms. The PRS_sMDD could predict the conversion by affecting the hippocampal volume. These genes were functionally related to the hippocampal development and amyloid-β binding. In addition, these genes showed temporal-specific expression in the hippocampus in the foetal developmental period and cell type-specific expression in the hippocampal pyramidal neurons and interneurons. These findings support a model of double attacks of hippocampus to explain for the neurobiological mechanisms of the independent contributions of MDD- and Alzheimer’s disease-specific genomic variants to the conversion from aMCI to and Alzheimer’s disease.

PRS_sMDD could independently predict the conversion from aMCI to Alzheimer’s disease

That Alzheimer’s disease is a progressive polygenic disorder (Dubois et al., 2007) and no effective therapies at the late stage makes early diagnosis and treatment to be the only way to improve prognosis (Chu, 2012). As the putative premorbid state of Alzheimer’s disease (Mariani et al., 2007), patients with aMCI have a high risk of developing into Alzheimer’s disease (Palmer et al., 2008). Because not all aMCI patients would convert to Alzheimer’s disease, it is clinically important to predict the conversion from aMCI to Alzheimer’s disease for early intervention. Consistent with previous studies (Rodriguez-Rodriguez et al., 2013; Adams et al., 2015), the PRS_AD was found to predict conversion. More importantly, we found that the PRS_MDD could also predict the conversion of non-depressed aMCI patients. Thus, the predictive ability of the PRS_MDD cannot be explained by parallel depressive symptoms (Kida et al., 2016; Mourao et al., 2016; Sacuiu et al., 2016; Barca et al., 2017) and the coexistence of MDD (Modrego and Ferrandez, 2004). To exclude the possibility that the predictive effect of the PRS_MDD is driven by genetic variants common to MDD and Alzheimer’s disease, we found that the PRS_sMDD could also predict the status of aMCI-C. LD score regression and co-localization analysis were performed to validate that these SNPs calculated for PRS_sMDD are specific to MDD rather than Alzheimer’s disease. Moreover, aMCI patients with high PRS_sMDD showed 14.97% and 17.05% higher conversion rates than those with low PRS_sMDD in the low and high PRS_sAD groups, respectively. These findings suggest that the PRS_sMDD has an independent contribution to the conversion from aMCI to Alzheimer’s disease.

In terms of the clinical prediction for the conversion from aMCI to Alzheimer’s disease, aMCI patients with the double high risk (high PRS_sMDD and high PRS_sAD) showed a much higher conversion rate than those with the double low risk (low PRS_sMDD and low PRS_sAD) (89.65% versus 33.33%). With the increased availability and decreasing cost of the sequencing technique, the genomic data of aMCI patients could be easily obtained, which could be used to construct the PRS_sMDD and PRS_sAD using the method of this study. In clinical practice, one could jointly use the PRS_sMDD and PRS_sAD to select aMCI patients with a much greater risk for Alzheimer’s disease. Appropriate intervention procedures for these aMCI patients may prevent or delay them from progressing into Alzheimer’s disease.

Hippocampal volume mediates the prediction of PRS_sMDD in aMCI-C

Correlation, mediation and Mendelian randomization analyses revealed that hippocampal volume mediated the predictive effect of the PRS_sMDD on the conversion from aMCI to Alzheimer’s disease. It has been recognized that reduced hippocampal volume is the most consistent subcortical abnormality in MDD patients in a meta-analysis of 1728 patients and 7199 controls (Schmaal et al., 2016). The genetic variants have been proposed as an important cause for the reduced hippocampal volume in MDD. For example, MDD patients with risk alleles show a smaller hippocampal volume than those without risk alleles (Egan et al., 2003; Frodl et al., 2004, 2007, 2012; Kohli et al., 2011). The reduction in hippocampal volume has been attributed to reduced gene expression in the hippocampus (Egan et al., 2003; Kohli et al., 2011). In children with few life stress events, the high PRS constructed by depression-related stress system genes could predict reduced hippocampal volume (Pagliaccio et al., 2014). Moreover, in healthy individuals without depressive episodes, the depression-related risk-allele carriers also displayed a reduced hippocampal volume compared with non-carriers (Frodl et al., 2008). These findings suggest that genetic variants for MDD could result in the reduction in the hippocampal volume in different populations.

It is well known that the volumetric reduction in the hippocampus is the most prominent early pathological feature of Alzheimer’s disease (Yang et al., 2012) and one of the most reliable predictive measures for the conversion from aMCI to Alzheimer’s disease (Brueggen et al., 2015). The additional burden to the hippocampus by MDD-specific genetic variants may exacerbate hippocampal atrophy and facilitate the conversion from aMCI to Alzheimer’s disease. Thus, the model of double attacks on the hippocampus may explain the increased predictive value of the PRS_sMDD on the conversion from aMCI to Alzheimer’s disease.

Neurobiological mechanisms underlying the predictive effect of PRS_sMDD on aMCI-C

To understand why hippocampal volume could mediate the predictive effect of the PRS_sMDD on conversion, the MDD-specific genetic variants were fine-mapped into the most common genes between the two different strategies (with and without hippocampal biological information), which suggested that most genes related to these MDD-specific SNPs are associated with hippocampal biological information.

The GO annotation revealed that MDD-related genes were related to amyloid-β binding, which is also supported by the PPI network analysis, showing that APP was the most important neighbouring gene in the network constructed by the MDD-related genes. The increased hippocampal amyloid-B, which is an important neuropathological change in Alzheimer's disease, is also related to depression (Rapp et al., 2006) or depressive symptoms (Donovan et al., 2018). These findings suggest that MDD-related genes may be involved in amyloid-β binding, an important neuropathological process of Alzheimer’s disease.

The enrichment analysis of the biological process revealed that MDD-related genes were mainly involved in the nervous developmental process. The temporal-specific expression analysis further indicated that these genes were mainly expressed in the middle to late foetal periods in the hippocampus, which is in line with the critical developmental period of hippocampal neurons (Tran and Kelly, 2003). For example, CDC20B showed the highest expression in the hippocampus in the middle-late foetal periods (Fig. 3E). This gene is a developmental regulator of hippocampal neurons (Kim et al., 2009; Yang et al., 2010) and plays an essential role in dendritic morphogenesis (Kim et al., 2009), axon growth (Yamada et al., 2013) and presynaptic differentiation (Yang et al., 2009), all of which are associated with hippocampal volume (Stockmeier et al., 2004; Duman et al., 2016). In addition, cell type-specific expression confirmed that these genes were mainly expressed in the hippocampal pyramidal neurons and interneurons. These findings suggest that MDD-related genes affect the hippocampal volume via modulating the gene expression of hippocampal neurons, which are the main cell types with prominent pathological impairment in Alzheimer’s disease (Blazquez-Llorca et al., 2011).

Limitations

Two limitations should be mentioned regarding this study. First, only volumetric measure was used as the intermediate phenotype. Consequently, we cannot exclude the possibility that other intermediate phenotypes may also mediate the association between PRS_sMDD and aMCI conversion. Second, the cell type-specific expression analysis was based on the hippocampus of mouse rather than human, because there are no publicly available cell type-specific expression data in the human hippocampus.

Conclusion

In this study, we found that the PRS_sMDD could independently predict conversion from aMCI to Alzheimer’s disease, and the combined use of the PRS_sMDD and PRS_sAD could select aMCI patients with a much higher risk for conversion. The predictive effect of the PRS_sMDD on conversion may be mediated by the hippocampus via affecting its early developmental process and amyloid-β binding.

Glossary

Abbreviations

aMCI(‐C/S)

amnestic mild cognitive impairment (‐conversion to Alzheimer’s disease/stable diagnosis)

GO

gene ontology

MDD

major depressive disorder

PPI

protein–protein interaction

PRS_{AD/MDD/sAD/sMDD/stAD}

polygenic risk scores for Alzheimer’s disease/major depressive disorder/Alzheimer’s disease-specific/major depressive disorder-specific/top 1806 Alzheimer’s disease-specific genetic variants

SNP

single-nucleotide polymorphism

Funding

This work has been supported by the Natural Science Foundation of China (81425013, 81501451, 81501454, 81601476, 81701676, 81701668 and 81801687), Tianjin Key Technology R&D Program (17ZXMFSY00090), National Key Research and Development Program of China (2018YFC1314301), Postgraduate Innovation Fund of 13th Five-Year comprehensive investment of Tianjin Medical University (YJSCX201719), and Science&Technology Development Fund of Tianjin Education Commission for Higher Education (2016YD11).

Competing interests

The authors report no competing interests.