A total of 155 self-declared Han Chinese college students aged 18–24 were recruited from Southwest University using advertisements. We have focused on Han population because previous studies showed that the APOE ε4 allele is significantly more prevalent in Han than non-Han ethnic Chinese, making Han Chinese at higher risk of developing dementia.^20‐22 All participants provided written informed consent. This study was approved by the Ethic Committee of Psychological Research at Southwest University, Chongqing City, China. Of the 155 participants, 11 were excluded from the MRI analysis because of excessive head motion.

Whole-brain T₁-weighted scan [MPRAGE, 160 slices, voxel size 1.0 mm,³ repetition time (TR) = 2300 ms, echo time (TE) = 2.98 ms, flip angle (FA) = 9°] was acquired using a Siemens Verio 3T MRI scanner. Structural MRI images were processed using FreeSurfer v5.3 (http://surfer.nmr.mgh.harvard.edu/) to calculate cortical area, thickness, curvature, folding index and subcortical volume. Specifically, imaging processing included motion correction and averaging²³ of multiple volumetric T₁-weighted images (when more than one is available), removal of non-brain tissue using a hybrid watershed/surface deformation procedure,²⁴ automated Talairach transformation, segmentation of the subcortical white matter and deep grey matter volumetric structures (including hippocampus, amygdala, caudate, putamen and ventricles)^25,26 intensity normalization,²⁷ tessellation of the grey/white matter boundary, automated topology correction^28,29 and surface deformation following intensity gradients to optimally place the grey/white and grey/cerebrospinal fluid borders at the location where the greatest shift in intensity defines the transition to the other tissue class.^30‐32

Then, the whole-brain MRI images were parcellated into 34 regions per hemisphere according to the Desikan–Killiany Atlas.³³ Surface area was measured at the grey/white matter boundary, and thickness was measured as the average distances in a region between the white matter and pial surfaces. At each vertex, FreeSurfer measured the mean curvature as follows:

and the folding index as follows:

where K1 and K2 denote the maximum and minimum normal curvature.³⁴ Then, the regional mean curvature and folding index were calculated.

To construct models that can identify individuals carrying increased genetic risks for Ad, we trained support vector machine (SVM) models and evaluated their performance with 5-fold cross-validation which is a standard method used in machine learning to avoid over-fitting. The pipeline consisted of the following three steps: (i) feature extraction and selection, (ii) model training and evaluation and (iii) identifying the most discriminative features, which we will explain in turn. The below pipeline was performed for classifying APOE ε4 positive (APOE+) individuals from negative (APOE−) individuals and rs242557 A positive (MAPT+) individuals from negative (MAPT−) individuals, respectively.

To investigate the gene–gene interaction between APOE and MAPT on the imaging features, we used clustering model rather than classification that is unsuitable to evaluate interaction effect. To cluster the regions, we trained a growing hierarchical self-organizing maps (GHSOM) model, which is an artificial neural network and provides a means of representing multidimensional data set as a 2D map.³⁷ The features we used here was the same as those in discriminative analysis, and their Z-scores were input into model for training. Once trained, this map is a model of the original input data, with individual measures represented as individual weight planes (or layers). Each node corresponds to a weight vector with the same dimensionality as the input data. The nodes in the top layer can be regarded as the centroids of the clusters, i.e. the typical pattern of the class. Then, we assigned every region to the cluster that has the closest distance from the centroids/region. For each participant and each cluster, the intra-cluster distance was calculated by averaging distances from regions to the cluster centroid.

Group comparisons of demographics were performed with analysis of variance and chi-square test. GLM analyses were used to statistically explore the effect of APOE and MAPT on cortical morphology characteristics including thickness, area, volume, mean curvature and folding index and subcortical volume controlling for age, sex and education. When comparing volume between groups with different genotypes, total intracranial volume was also controlled for. False discovery rate was used to correct for multiple comparisons in the demographics analysis and univariate analysis.

We used GLM analyses to statistically test for the effects of APOE, MAPT and their interactions on the four intra-cluster distances controlling for age, sex and years of education. Bonferroni correction was used to correct for multiple comparisons in the clustering analysis. For the cluster showing a significant interaction effect on the intra-cluster distance among groups, we also compared the mean cortical metrics from all regions belonging to that cluster. Threshold for statistical significance was set at P < 0.05, two-sided in all analyses.

The samples consisted of 83 individuals without neither genetic risk factor studied here, 11 individuals only carrying APOE ε4, 34 individuals with genetic risk due to the MAPT locus alone and 16 individuals with both genetic risks. Their age (P = 0.784), sex (P = 0.393) and education (P = 0.536) were not statistically different between groups. Detailed descriptions and the demographics of the samples are provided in Table 1.

We compared cortical thickness, area, volume, mean curvature and folding index and subcortical volume between APOE ε4 carriers and non-carriers using univariate tests. Similarly, we also compared these individual structural characteristics between MAPT A carriers and non-carriers. However, no result survived false discovery rate correction.

We used 5-fold cross-validation to estimate the generalizability of the classifier between APOE ε4 carriers and non-carriers. As shown in Table 2, the classifier achieved a classification accuracy of 0.81, a sensitivity of 0.52, a specificity of 0.88 and an area under curve of 0.66. As we mentioned above, the model’s accuracy was also assessed with positive predictive value, negative predictive value and F-score that are more robust for our unbalanced data set. These metrices (positive predictive value: 0.50, P = 0.011; negative predictive value: 0.89, P = 0.001; F-score: 0.51, P = 0.001) were statistically significantly higher than those of random models. We compared the performance of model using all cortical features and the models using only a single feature. As shown in Fig. 1A, the model using all the features outperformed all other models based on single cortical feature demonstrating the advantage of multivariate approach including multiple cortical features in a single analysis.

The cortical distribution of positive and negative weights is presented in Fig. 1B and C, respectively, and no subcortical volumes contributed to the classification. Overall, the left superior frontal cortex, left precuneus, left isthmus of cingulate, right precentral cortex, right insula and right caudal middle frontal cortex were the most discriminative regions (Fig. 1D). In addition, the thickness of most brain regions was negatively relative to the presence of APOE ε4 and the folding index of most affected brain regions was positively relative to the presence of APOE ε4 (Fig. 1E).

Similarly, we used 5-fold cross-validation to estimate the generalizability of the classifier between MAPT A allele carriers and non-carriers. As shown in Table 2, the classifier achieved a classification accuracy of 0.63, a sensitivity of 0.46, a specificity of 0.71 and an area under curve of 0.61. The positive predictive value, a negative predictive value and an F-score of the classifier were 0.47, 0.71 and 0.46, respectively. The last three metrices were more robust and were all statistically significantly higher than that of random models. We compared the performance of model using all features and the models using just a single feature. As shown in Fig. 2A, the model using all features also outperformed other models similar to our previous analysis.

The distribution of positive and negative weights is presented in Fig. 2B and C, respectively, and volume of bilateral hippocampus, bilateral amygdala and right caudate also contributed to the classification. Overall, the left middle temporal cortex, left pericalcarine, left lateral orbitofrontal cortex, rostral anterior cingulate cortex, right pars orbitalis and banks of the superior temporal sulcus were the most affected regions (Fig. 2D). In addition, the thickness of most brain regions was negatively relative to the presence of MAPT A allele and the folding index of most affected brain regions was positively relative to the presence of MAPT A allele (Fig. 2E).

As shown in Fig. 3A, the four nodes in the top layer represent the centroids of four clusters: (i) the brain regions with medium area, large thickness and small curvature; (ii) the brain regions with large area, small thickness and small curvature; (iii) the brain regions with small area, large thickness and large curvature; and (iv) and the brain regions with medium area, small thickness and large curvature. For ease of comparison, the weight profiles of the top layer of nodes are also plotted in Fig. 3B. The spatial distribution of the four clusters across all brain regions is shown in Fig. 3C.

GLM analyses showed a significant interaction (T = −2.600, P = 0.009) between APOE and MAPT such that individuals with both genetic risks had the significantly increased distance from the centroid of the first cluster comparing with the other groups (Fig. 4A). This deviation was mainly caused by the increase in curvature (T = −2.379, P = 0.019) and folding index (T = −2.567, P = 0.011) (Fig. 4B). While there was no difference in distance from the centroid in other three clusters, we also compared their mean cortical measurements. Similar gene–gene interaction on the folding index in regions from the second cluster (T = −2.034, P = 0.044) (Fig. 5B) and the curvature of the regions in the fourth cluster (T = −2.258, P = 0.026) (Fig. 5D) was found.

The SVM model separating APOE status based on cerebral structural features showed a significant discriminability between APOE ε4 carriers and non-carriers, which indicated that young healthy adults carrying APOE ε4 had a different pattern of cerebral structure from non-carriers. By analysing the model’s weights, the individuals carrying APOE ε4 have higher folding index and curvature and lower thickness across widespread regions of the cerebral cortex. Regarding the most discriminative regions, the right precentral cortex, right insula, left superior frontal cortex, right superior temporal cortex and right caudal middle frontal cortex were the top five regions that contributed most to the classification. This is consistent with previous studies showing that the superior temporal cortex¹¹ and insula³⁸ have reduced grey matter volume in amnestic mild cognitive impairment with APOE ε4 compared with non-carriers. Furthermore, a longitudinal study showed that APOE ε4 accelerated brain atrophy in superior temporal gyrus in healthy older participants.³⁹ The right precentral cortex, left superior frontal cortex and right caudal middle frontal cortex showed thinner cortex or lower grey matter volume already in middle-aged healthy individuals with APOE ε4.^40,41 Our results extend the existing knowledge about the relationship between APOE ε4 and altered cortical morphology from middle-aged and older people to young healthy adults, when Ad pathology was unlikely.

Although the model discriminating the MAPT status had a slightly lower performance than the model classifying the APOE status, it significantly outperformed random models in terms of the F-score test. Across the cerebral cortex, the model for MAPT showed a similar pattern as the APOE analysis, that is, individuals carrying MAPT A allele have higher folding index and lower thickness. The difference between two models is that the number of regions contributing to classifying MAPT status is less than that for APOE. However, most of these regions identified in the SVM analysis were those implicated in the early phase of Braak pathological staging for tau and amyloid beta, such as the entorhinal cortex, fusiform, lingual, middle temporal, cingulate and temporal pole.⁴² Although we did not have biomarker evidence for tau and amyloid pathology in this study, it is reasonable to assume that our young healthy cohort was free from Ad pathology as well as other comorbidities found in patients with established Ad. So, the current findings might represent a pure genetically determined neuroanatomical trait which makes these individuals more vulnerable to Ad pathology in the future.

The GHSOM clustering analysis showed that in our cohort, the cerebral cortex could be separated into four major morphologically distinct networks; each is composed of several morphologically similar regions: (i) larger but thinner and less convoluted regions, (ii) smaller and thicker but more convoluted regions, (iii) medium sized and thicker but less convoluted regions and (iv) medium sized and thinner but more convoluted regions. The various morphological pattern might be contributed to large-scale and regionally heterogeneous development in thickness, area and convolution during adolescence and early adulthood.^43‐48 As expected, the spatial patterns identified in this study are consistent with previous studies.^49‐51

Further analysis based on these clusters/networks revealed that the individuals carrying both APOE ε4 and MAPT A had a significant deviation from the typical morphology in the medium sized and thicker but less convoluted brain regions by increasing curvature and folding index. These regions overlapped with the regions discriminating APOE and MAPT status in the SVM analysis, such as the superior, middle and medial temporal cortex and insular and caudal middle frontal cortex. The supervised classification and unsupervised clustering analysis gave convergent evidence supporting individuals with APOE ε4 and MAPT A tend to have higher curvature and folding index in these key brain regions. It is worth noting that these regions showed similar development and aging-related cortical thickness change and had a shared genetic influence.¹² Not only individuals with both genetic risk but also patients with schizophrenia,⁵² 22q11.2 deletions,⁵³ autistic and Asperger disorders⁵⁴ and Williams syndrome⁵⁵ showed the similar morphological characteristics, which may suggest that this kind of cortical morphology may be adverse for brain and mental health and increases vulnerability due to Ad pathology in older age.

The process of forming cortical folding during brain development is complex and has not been comprehensively studied yet. However, there are several widely investigated hypotheses arguing that cortical folding is caused by the growth processes during cortical development and/or the mechanical tension within axons.⁵⁶ Based on the axonal hypotheses, it has been speculated that synaptic pruning and reduction of connectivity during normal neurodevelopment may lead to the folding reduction from about 75th week of gestation to almost 23 years of old.⁵⁷ So, genetic risks of Ad may interfere with such pruning process leading to higher folding. Recent studies have also suggested prion-like mechanisms of propagation of Aβ and tau proteins in AD,⁵⁸ so the individual carrying higher risk of Ad may have remaining unpruned connections thus more convoluted cortex. Hyperconnectivity may promote the spread of neurodegenerative pathology when it initiates later in life. Combining with existing knowledge about the effect of APOE ε4 and MAPT A on Ad, these two genes may already play a significant role in shaping the brain across the lifespan by changing the cerebral structural to be more vulnerable to molecular pathology in the future.

When considering the present study, there are several limitations. First, despite using multiple metrics to evaluate the performance of classification models, the performance of the model might be affected by the unbalance of risk groups.⁵⁹ Although statistically significant, the discriminability of models, especially the model classifying MAPT, was modest, so the results await future validation and replication based on larger cohorts. In addition, although our ratios of APOE and MAPT carriers are consistent with the natural frequency of these genes in non-selected population, the unbalanced groups defined by genetic risks may also impede the generalization of our conclusions to other data sets where genetic risks are enriched by including patients with dementia. Second, the cross-sectional nature of our current study makes it difficult to interpret the results and make causal inference. Although this represents a limitation of our study, it also encourages the development of more comprehensive longitudinal studies to be conducted in the healthy population for brain health related to Ad in the future. Third, all the analyses in this study were based on regions of interest. Considering that the cerebrally structural variation in young adults is subtle, vertex-wised analyses may be more flexible to discover these subtle differences and unbiased by the selection of cortical parcellation. However, the high-dimensionality of imaging data and multiple comparisons problem in vertex-wised analyses introduce computational and other methodological challenges. Finally, future research is needed to consider diversity and inclusivity to cover other non-Han Chinese ethnic minorities because they have both unique and shared genetic and environmental risks for Ad.