During the PRS calculation, a base dataset was needed to calculate the genetic effect size associated with the disease status at a predefined threshold. In this study, GWAS data of the Diabetes Genetics Replication and Meta-analysis (DIAGRM) (Mahajan et al., 2018) and International Genomics of Alzheimer’s Project (IGAP) (Lambert et al., 2013) were used as the base dataset to calculate PRS_T2DM and PRS_AD in a target dataset, respectively. The target dataset provided by all stages of Alzheimer’s Disease Neuroimaging Initiative (ADNI1/GO/2/3).¹ There are 975 patients in total with the baseline diagnosis of aMCI. Diagnosis of aMCI was made according to the criteria by Petersen (Petersen et al., 1999). Among them, 522 patients had no whole-genome sequencing information, five patients had no follow-up information, one patient’s follow-up was too short (≤3 months), 52 patients converted into normal controls during the follow-up, eight patients converted into AD then back to aMCI during the follow-up, and one patient’s structural MRI image could not be downloaded. After excluding the above subjects, 386 aMCI patients’ data were retained in the subsequent statistical analyses. According to the follow-up outcome in March 2019 (the follow-up time range is 6–156 months, an average of 50.69 months), 386 aMCI patients were divided into conversion (aMCI-C, n = 164) and stable (aMCI-S, n = 222) groups. The final clinical diagnosis will be based on the last follow-up results if the subject is missing during the follow-up.

The whole-genome sequencing was performed on Illumina HiSeq2000 platform. The genome-wide SNPs were genotyped using the Illumina Omni 2.5M Bead Chip for the ADNI subjects.

Quality control was performed at both the individual level and SNP level. Subjects with a missing genotyping rate of >0.05, sex inconsistency, possible relative relationships, and European population outliers identified by multidimensional scaling (MDS) were excluded. The first four components of MDS analysis were used as covariates in subsequent analysis. SNPs with a missing call rate of >0.05, minor allele frequency <0.01, a significant deviation from Hardy-Weinberg equilibrium (P < 5 × 10^–6), and ambiguous strands were excluded. Finally, 9,845,494 SNPs from 386 subjects were included in the subsequent analyses.

In the base dataset, the associations between the SNPs and disease status were calculated at predefined P threshold (P_T) values ranging from 5 × 10^–5 to 0.5 with an increment of 5 × 10^–5. Under each P_T value, we removed the effects of SNPs in linkage disequilibrium (LD) in each clumped region (excluding SNPs with r² > 0.1, within a 250-kb window) and selected the index SNPs (iSNPs) with the most significant P-value from each clumped association region. Thus, the information of the risk alleles and effect sizes of the iSNPs were obtained for each P_T value. In the target dataset, the PRSice-2 (v2.2.6) software (Choi and O’Reilly, 2019) was used to calculate the PRS according to Equation 1.

S_i is the number of risk alleles of iSNPs, G_i is the effect sizes (natural logarithm of the odds ratio) of the iSNPs, and M is the number of iSNPs.

In this study, the logistic regression method was used to assess the predictive effects of the PRSs on the conversion from aMCI to AD after controlling for sex, age, educational years at baseline, the number of APOEε4, and the first four MDS components for population stratification. Nagelkerke’s R² was used to estimate the percentage of the variance predicted by the PRS in the regression model. Thus, we could obtain the best P_T values for calculating PRS with the best predictive abilities in the target dataset. We calculated both PRS_T2DM and PRS_AD to explore the predictive effects of the two groups of risk genes on the conversion from aMCI to AD.

In this study, we aimed to evaluate the predictive effect of the PRS specific to T2DM (PRS_sT2DM) on the conversion from aMCI to AD. Therefore, we removed the common genetic variants of T2DM and AD and reconstructed PRS_sT2DM and PRS_sAD. Finally, the four groups of PRS values were z-transformed before being used in the statistical analyses.

All imaging data are publicly available and were downloaded from the ADNI website (see text footnote 1). Following the ADNI acquisition protocol, three-dimension T1 weighted imaging (3D-T1WI) was scanned by 1.5T (ADNI1) or 3.0T (ADNIGO/2) MR scanner with a magnetization prepared rapid gradient echo (MPRAGE) sequence at 54 sites. Image corrections involved calibration, geometry distortion, and reduction of the intensity of non-uniformity applied to each image by the ADNI. We downloaded the corrected 3D-T1WI images of 386 aMCI patients for subsequent analysis.

Statistical Parametric Mapping software package (SPM12²) and the voxel-based morphometry (VBM) toolbox CAT12.6-rc1³ were used to preprocess the 3D-T1WI images. The preprocessing processes included bias-field-corrected, tissue-classified, DARTEL-based spatial normalization, segmentation, modulation, and smooth with a Gaussian kernel with a full width at half maximum of 8 mm.

FreeSurfer (version 5.3.0⁴) was used to reconstruct the cerebral cortex using the SBM method. In brief, 3D-T1WI images of all subjects were registered to the MNI305 template with an affine way, and then the skulls were stripped. White matter and pial surfaces were constructed with a triangle area called a vertex unit. We visually inspected all images and segmentation quality and manually edited them as necessary.⁵ Vertex-wise cortical thickness was obtained by calculating the shortest distance between the pial and white surface. Vertex-wise surface area was calculated by assigning one-third of each triangle’s area to each of its vertices. Vertex-wise cortical volume was calculated by multiplying the surface area by cortical thickness (Fischl and Dale, 2000; Han et al., 2006). The cortical thickness, surface area, and cortical volume were resampled at 1 mm resolution and smoothed with a Gaussian kernel with a full width at half maximum of 10 mm.

The demographic and PRS results are shown in Table 1. There was no significant difference in gender, education, blood glucose level, and diabetes condition between the aMCI-C and aMCI-S groups (P > 0.05). The age of the aMCI-C group was slightly older than that of the aMCI-S group (P < 0.05). The proportion of APOEε4 carrier and the four groups of PRS were higher in the aMCI-C group than in the aMCI-S group (P < 0.001). There were significant differences in the Aβ and tau levels in the CSF between aMCI-S and aMCI-C groups (Supplementary Table 1).

When the T2DM-GWAS data was taken as the base dataset, the PRS_T2DM calculated at P_T = 0.0088 exhibited the best predictive effect on the aMCI conversion to AD (P = 6.86 × 10^–4) (Figure 1A, the detailed information at different P_T values are listed in the Supplementary Table 2) and explained 3.74% of variance based on 22,763 index SNPs (Table 2). In addition, the PRS_T2DM in the aMCI-C group was significantly higher than in the aMCI-S group (Table 1 and Figure 1B).

The PRS_AD calculated at P_T = 0.03645 showed the best predictive effect on the aMCI conversion to AD (P = 1.45 × 10^–5) (Figure 1D) and explained 6.15% of variance based on 29,321 index SNPs (Table 2). The PRS_AD in the aMCI-C group was significantly higher than in the aMCI-S group (Table 1 and Figure 1E).

After the common genetic variants (n = 237) of the two diseases were removed, the PRS_sT2DM still exhibited a significant predictive effect on the aMCI conversion (P = 5.06 × 10^–4) and explained 3.93% of the variance, which indicated T2DM-specific genetic variants could independently predict the conversion from aMCI to AD. After extra controlling the Aβ and tau level in the CSF, logistic regression analysis showed the PRS_sT2DM still exhibited a significant predictive effect on the aMCI conversion to AD (β = 0.461, P = 0.001). The PRS_sAD could also independently predict the aMCI conversion (P = 9.26 × 10^–6) and explain 6.44% of the variance.

The ROC curves of the four PRSs predictive models are shown in Figure 2, and the AUC, sensitivities, and specificities are listed in Table 2.

According to the PRS_sT2DM and PRS_sAD, 386 aMCI patients were divided into low-, middle-, and high-risk subgroups, respectively. The three PRS_sT2DM subgroups exhibited a significant difference in the aMCI conversion rate (χ² = 14.30, P = 7.87 × 10^–4). A significant difference in aMCI conversion rate in PRS_sAD subgroups was also found (χ² = 19.18, P = 6.83 × 10^–5) (Figure 3). Cox survival analysis showed that the risks of aMCI conversion to AD in the middle PRS_sT2DM group (HRs, 0.67; 95% CI, 0.46–0.97; P = 0.034) and the low PRS_sT2DM group (HRs, 0.60; 95%CI, 0.41–0.88; P = 0.009) were lower than in the high PRS_sT2DM group (Figure 1C). Additionally, aMCI patients in the high PRS_sT2DM group converted to AD about 5 months on average earlier than the middle PRS_sT2DM group and a mean of 7 months earlier than the low PRS_sT2DM group (average 33.67, 38.25, and 40.10 months, respectively). Accordingly, the risk of aMCI conversion to AD in the middle PRS_sAD group (HRs = 0.66, 95%CI = 0.47–0.95, P = 0.024) and the low PRS_sAD group (HRs = 0.50, 95%CI = 0.34–0.75, P = 0.001) was lower than in the high PRS_sAD group (Figure 1F). aMCI Patients in the high PRS_sAD group converted to AD about 13 months on average earlier than the middle PRS_sAD group and a mean of 15 months earlier than the low PRS_sAD group (average 29.10, 42.00, and 44.05 months, respectively).

In the aMCI-C group, the PRS_sT2DM was negatively correlated with the GMV of the right superior frontal gyrus (MNI coordinates: x = 18, y = 36, z = 55.5; t = −3.97; 458 voxels, cluster-level P < 0.05) (Figure 4A). The ROI-based correlation analysis is shown in Figure 4B. A significant negative correlation was found between the PRS_sT2DM and the cortical volume of the right superior occipital gyrus in all aMCI patients (cluster size: 108.05 mm³; cluster-level P < 0.05) (Figure 5A). The ROI-based correlation analysis is presented in Figure 5B.

The mediation analysis showed that the cortical volume of the right superior occipital gyrus significantly mediated the association between the PRS_sT2DM and aMCI conversion (P < 0.05), and the mediation effect could explain 5.8% of the variance (Figure 6).

Using MAGMA, the 22,526 T2DM-specific genetic variants were fine-mapped into 6,238 genes based on genomic location (within a 5 kb window both upstream and downstream, 19,427 among the 22,526 SNPs were located inside genes). The genetic variants of the GWAS data of the Diabetes Genetics Replication and Meta-analysis (DIAGRM) were fine-mapped into 6,330 genes with a threshold of P < 0.05 using MAGMA. The overlapped 4,667 genes were used in the following analysis. Detailed fine-mapped genes are listed in Supplementary Table 3.

By performing GO and KEGG enrichment analyses, the 4,667 fine-mapped genes were significantly enriched in multiple GO items and pathways (P < 0.05, Bonferroni corrected) including molecular function GO items: ion binding (P = 6.49 × 10^–14), calcium ion binding (P = 3.52 × 10^–13), and protein binding (P = 1.72 × 10^–12) (Figure 7A); biological processes GO items: nervous system development (P = 6.84 × 10^–38), anatomical structure morphogenesis (P = 1.47 × 10^–31), and generation of neurons (P = 2.45 × 10^–31) (Figure 7B); cellular component GO items: cell junction (P = 2.11 × 10^–35), cell projection (P = 1.08 × 10^–30), and plasma membrane bounded cell projection (P = 6.30 × 10^–30) (Figure 7C). In addition, the 6,267 genes were significantly enriched in oxytocin signaling pathway (P = 6.66 × 10^–8), PI3K-Akt signaling pathway (P = 3.17 × 10^–6), focal adhesion (P = 1.84 × 10^–7), circadian entrainment (P = 1.12 × 10^–7), and MAPK signaling pathway (P = 2.76 × 10^–5) in KEGG (Figure 7D). Detailed enrichment analysis results are listed in Supplementary Table 4.

To assess cell-specific expression of the 4,667 fine-mapped genes, cell type-specific enrichment analysis was performed. Under different pSI thresholds (pSI = 0.05, 0.01, 0.001, and 0.0001, permutation corrected), which represents how likely a gene was specifically expressed in a given cell type relative to other cell types, the fine-mapped genes showed substantial overrepresentation, mainly in the cortical neuron (P = 0.011).

Alzheimer’s disease is a multifactorial neurodegenerative disorder that lacks a curative treatment. Therefore, early discovery, diagnosis, and treatment could significantly improve the patients’ prognosis. The aMCI is a type of syndrome between normal aging and AD (Winblad et al., 2004), patients with aMCI are at a high risk of converting to AD. AD risk genes such as APOEε4 can induce the deposition of tau protein and Aβ in and out of cells of brain regions, lead to reduced neuronal activity and loss of synapses, and accelerate brain atrophy in related areas (Kim et al., 2014), further promote the conversion of aMCI to AD. Our results confirmed Adams’ finding that the PRS_AD could predict the conversion of aMCI to AD (Adams et al., 2015).

It has been found that AD and T2DM patients share some pathological characteristics of the central nervous system. Hoyer (2000) found that brain metabolism changes, including impaired glucose utilization and energy metabolism, occurred in AD patients after the first clinical symptoms appeared. The metabolic impairment gradually worsens with AD’s progression (Hoyer et al., 1991). Therefore, some researchers have proposed that abnormal energy metabolism in AD patients may be caused by insulin resistance or weakened insulin action in the brain (Blass et al., 2002; Hoyer, 2004a,b). In addition, some studies have shown that T2DM risk genes could predict the conversion of aMCI to AD. For example, a longitudinal follow-up ADNI study showed that the SNP (rs391300) located on the serine racemic enzyme (SRR, a risk gene of T2DM) could predict the conversion of aMCI to AD (Girard et al., 2018). In this study, for the first time, we proved that PRS_sT2DM could predict the aMCI’s conversion after the shared iSNPs between PRS_T2DM and PRS_AD were removed, suggesting that the polygenic genetic risk of T2DM could be one of the risk factors for the conversion of aMCI to AD.

Type 2 diabetes mellitus risk genes can lead to impaired insulin signaling (Meur et al., 2010). The relationship between impaired brain insulin signaling and AD’s pathological changes may involve the following aspects: impaired insulin signaling transduction may lead to elevated glycogen synthase kinase-3β (GSK-3β) activity and oxidative stress, promoting tau protein’s hyperphosphorylation; impaired insulin signaling will inhibit the PI3-K-Akt signaling pathway, resulting in decreased neuronal activity and loss of synapses; insulin or IGF-1 can activate the Erk-MAPK pathway and promote the physiological processing and intracellular transport of β-APP to the plasma membrane (de la Monte and Wands, 2005). In addition, insulin receptors are usually highly expressed in brain areas related to cognition and memory, such as the cortex and hippocampus (Craft, 2009; Nisticò et al., 2012). Therefore, impaired insulin signaling-related tau protein increase and Aβ deposition in the brain may be the plausible neurobiological mechanism that the PRS_T2DM predicts the conversion of aMCI to AD.

Our study found that the PRS_sT2DM was significantly negatively correlated with the GMV of the right superior frontal gyrus in the aMCI-C group. A meta-analysis showed that the GMV of the medial superior frontal gyrus of T2DM patients was significantly smaller than that of the controls (Liu et al., 2017). Furthermore, the GMV of the superior frontal gyrus in patients with T2DM combined with aMCI is significantly smaller than that of controls (Zhang et al., 2014). These results suggested that the superior frontal gyrus is prone to atrophy in T2DM patients. Our finding also indicated that the T2DM-related genetic risks might modulate the GMV of the superior frontal gyrus in T2DM patients.

In all aMCI patients, the PRS_sT2DM was significantly negatively correlated with the cortical volume of the right superior occipital gyrus. Many studies showed structural changes in the occipital lobe of T2DM patients, including gray matter atrophy (Zhang et al., 2014; Moulton et al., 2015) and decreased topological attributes (Qin et al., 2019), which demonstrated that the occipital lobe is a brain area prone to structural damages in T2DM patients. The occipital lobe can transmit visual information and cooperate with other brain regions to process and integrate visual and verbal information. It had been found that GMV of the occipital lobe in AD patients was significantly associated with attention function (Cromarty et al., 2018). Kunst et al. (2019) found that the atrophy of the occipital cortex can be used to distinguish between AD, aMCI, and healthy controls. Our study showed that the T2DM-related genetic risks might modulate the cortical volume of the right superior occipital gyrus in aMCI patients. Mediation analysis demonstrated that the cortical volume of the right superior occipital gyrus might mediate the predictive effect of the PRS_sT2DM on the conversion of aMCI to AD.

To explore the neurobiological mechanism behind the predictive effect of T2DM-specific genetic variants on the aMCI conversion, we fine-mapped the T2DM-specific genetic variants to the genes and then performed the enrichment analyses.

Cell type enrichment analysis showed that these genes are mainly expressed in the cortical neuron. Gene GO enrichment analyses showed that these genes were primarily associated with ion and protein binding, neural development and generation, cell junction, and projection. These biological processes are known to be associated with AD. For example, metal ions binding to the Aβ peptide can promote aggregation of Aβ (Wallin et al., 2016, 2017). Insulin-like growth factor binding protein-2 (IGFBP-2) is associated with AD and brain atrophy (Lane et al., 2017). Neurogenesis represents an integral part of AD pathology (Mu and Gage, 2011). The projection of Hippocampal neurons may affect learning and memory (Milner and Veznedaroglu, 1993). The KEGG annotation revealed that fine-mapped genes were mainly related to the PI3K-Akt and MAPK signaling pathways. These enrichment analysis results suggested that these T2DM-related genetic variants are highly expressed in the cortical neurons, modulate ion and protein binding, neural development and generation, cell junction and projection, the PI3K-Akt and MAPK signaling pathway, and further affect Tau phosphorylation (Salkovic-Petrisic et al., 2006; Ke et al., 2009) and Aβ accumulation (Yang et al., 2014), accelerating the conversion of aMCI to AD.

Several limitations should be mentioned in this study. First, the PRS was calculated with an additive effect by default; we did not consider the influence of other genetic effects on the results. Second, we did not control some clinical situations, such as vasculopathy, during the predicting analysis because of insufficient data. Further study should be performed to clarify the impact of clinical status on the results. Third, the cell type-specific expression analysis was based on the cortex of the mouse because no public cell type-specific expression data in the human cortex can be obtained.