Multivariate investigation of aging in mouse models expressing the Alzheimer’s protective APOE2 allele: integrating cognitive metrics, brain imaging, and blood transcriptomics

Abstract

APOE allelic variation is critical in brain aging and Alzheimer’s disease (AD). The APOE2 allele associated with cognitive resilience and neuroprotection against AD remains understudied. We employed a multipronged approach to characterize the transition from middle to old age in mice with APOE2 allele, using behavioral assessments, image-derived morphometry and diffusion metrics, structural connectomics, and blood transcriptomics. We used sparse multiple canonical correlation analyses (SMCCA) for integrative modeling, and graph neural network predictions. Our results revealed brain sub-networks associated with biological traits, cognitive markers, and gene expression. The cingulate cortex emerged as a critical region, demonstrating age-associated atrophy and diffusion changes, with higher fractional anisotropy in males and middle-aged subjects. Somatosensory and olfactory regions were consistently highlighted, indicating age-related atrophy and sex differences. The hippocampus exhibited significant volumetric changes with age, with differences between males and females in CA3 and CA1 regions. SMCCA underscored changes in the cingulate cortex, somatosensory cortex, olfactory regions, and hippocampus in relation to cognition and blood-based gene expression. Our integrative modeling in aging APOE2 carriers revealed a central role for changes in gene pathways involved in localization and the negative regulation of cellular processes. Our results support an important role of the immune system and response to stress. This integrative approach offers novel insights into the complex interplay among brain connectivity, aging, and sex. Our study provides a foundation for understanding the impact of APOE2 allele on brain aging, the potential for detecting associated changes in blood markers, and revealing novel therapeutic intervention targets.

Introduction

It is crucial to identify strategies that mitigate the risk of age-associated decline and prevent irreversible brain damage caused by pathology and neurodegeneration. Among neurodegenerative diseases, Alzheimer’s disease (AD) is particularly prevalent, affecting approximately 10% of the population aged 65 and above, with this percentage increasing as the population ages (“2023 Alzheimer’s disease facts and figures,” 2021). The APOE gene, a lipid transporter gene, plays a significant genetic role in AD risk, with allelic variations being key determinants. The APOE4 allele has been extensively studied and is known to confer the highest risk. In contrast, the APOE2 variant is associated with longevity and neuroprotection against AD, reducing the risk of AD by nearly 50% (Conejero-Goldberg et al. 2014). However, the APOE2 allele has not received as much attention in research compared to other alleles. While the protective role of APOE2 in longevity and AD has been investigated, studies exploring its impact on cognition and imaging markers have shown inconsistent findings (Kim et al. 2022), and the underlying mechanisms remain largely unknown.

To understand the mechanisms that modulate biological pathways during aging, it is essential to develop methods that integrate multivariate data, including behavioral, imaging, and omics data. Such approaches have the potential to reveal imaging markers that predict cognitive decline and to uncover the underlying biological pathways. Genome-wide association studies (GWAS) have identified multiple genes associated with an increased risk of pathological brain aging (Ballard et al. 2011; Harold et al. 2009; Lambert et al. 2013; Saykin et al. 2015). Traditionally, researchers have relied on brain tissue samples to understand how aging affects brain network integrity and to uncover associated gene expression changes. However, obtaining brain tissue samples from younger individuals or during specific life transitions is challenging. In contrast, MRI and peripheral markers, such as blood, provide more accessible data. Recent studies have shown that gene expression levels in blood have been linked to memory function (Niculescu et al. 2020), and GSEA pathways across APOE genotypes involved overlapping pathways with different set of genes (Panitch et al. 2022), which motivates the need for studying the relationship between blood-based, brain-based transcriptomic markers, and brain circuits in aging.

Mouse models that replicate human APOE alleles through targeted replacement of the murine APOE gene serve as valuable tools for investigating the role of APOE in aging. While mouse models expressing the high-risk APOE4 allele and the APOE3 allele have been extensively studied, our focus is on the less explored model expressing the human APOE2 Alzheimer’s protective allele (Conejero-Goldberg et al. 2014). We investigate cognitive, imaging, and transcriptional changes during the transition from middle to old age to gain insights into the mechanisms underlying neuroprotective aging and cognitive resilience against AD, with particular emphasis on exploring the unique contributions of the APOE2 allele. The age range is selected due to its relevance to the onset of neurodegenerative diseases, the understanding of underlying mechanisms, and the potential for implementing interventions at prodromal stages before the typical cognitive decline of AD. Additionally, the transition from middle to old age is a pivotal period in aging, characterized by significant changes in brain structure and function observed through imaging metrics (Fjell et al. 2014; Raz and Rodrigue 2006).

To assess cognition in this critical transition period, we utilized the Morris water maze (MWM), a widely used behavioral test for rodents, and introduced a metric, the absolute winding number (AWN), to evaluate spatial learning and navigation strategies (Badea et al. 2022). We have used the traditional MWM metrics in addition to AWN that accounts for spatial navigation, which can be an early and translational biomarker for AD (Coughlan et al. 2018). We employed diffusion MRI for detecting atrophy and structural changes, and supplemented fractional anisotropy with estimates of cellularity from mean apparent propagator (MAP) MRI. MAP MRI enhances the diffusion MR signal using orthogonal basis functions related to the eigenfunctions of Fourier transform, improving robustness to noise and capturing intricate tissue structure details (Avram et al. 2016). We estimated cellularity and restricted diffusion using the return-to-origin probability (RTOP) (Avram et al. 2016; Özarslan et al. 2013). We hypothesized that MRI-detected structural changes are accompanied by alterations in brain gene expression. Identifying these changes is crucial for understanding the mechanisms underlying neurodegenerative diseases, and may unveil factors contributing to longevity.

Integration of brain imaging and omics data has gained significant attention due to its ability to provide a multiview perspective on complex neurological conditions, such as Alzheimer’s and Parkinson’s diseases (Donini et al. 2019; Markello et al. 2021). Combining spatial transcriptomic data from atlases with connectome data has enabled studies to investigate the correlation between gene expression patterns and variations in brain structure and function (Arnatkeviciute et al. 2021; Fornito et al. 2019). This approach differs from research focused mostly on genetic, mRNA transcriptomic, or proteomic associations with other phenotypes, including disease risk, onset age, cognitive aging measures, and imaging and biofluid markers. Spatial studies have mainly concentrated on integrating connectomes and transcriptomes, but in contrast, research using mouse models has explored the correlations between gene expression patterns and endophenotypes, such as cortical thickness, in neurodegenerative diseases like Alzheimer’s disease (Grothe et al. 2018; Timonidis et al. 2021).

In this study, we propose a multifaceted modeling approach, integrating omics data acquired by blood transcriptomics, MRI-derived connectomes representing brain networks, and cognitive metrics from behavioral tests. Our study design prioritizes translation to humans by investigating the interplay between age, sex, and phenotypes of aging using easily accessible blood biomarkers based on gene expression and diffusion MRI. This approach is especially relevant considering the challenges associated with obtaining brain and cerebrospinal fluid markers. Our canonical correlation model identifies blood gene expression markers correlating with changes in brain connectivity, biological traits, and behavior, while pinpointing relevant brain sub-networks. Additionally, we use graph neural network to predict biological traits using connectome data and identify key brain regions for each trait. This integrative approach facilitates a more comprehensive understanding of aging, including vulnerabilities and resilience, and provide insights on the interplay of genotype, sex, and environmental factors, with the aim of contributing to the development of targeted strategies for combating age-related neurodegenerative disorders.

Methods

Animals

We utilized mice homozygous for human APOE2 alleles, known to confer protection during aging and against Alzheimer’s disease (P. M. Sullivan et al. 1998a, b). The study was approved by Duke IACUC committee. Our initial cohort comprised 13 mice (8 males, 5 females) aged ~ 12 months (mean: 12.64 ± 0.70 months) and 28 mice (11 males, 17 females) aged ~ 16 months (mean: 16.42 ± 0.72). We conducted behavioral experiment on these mice. We imaged 33 (13 middle age, 20 old; 16 males, 17 females) of these mice and extracted RNA data for 27 mice (7 middle age, 20 old; 13 males, 14 females).

The investigation focused on age- and sex-associated variances spanning from the mouse equivalent of human middle age, where 10–14 months in mice equate to 38–47 human years, and to old age, where 18–24 months in mice corresponds to 56–69 human years (Ackert-Bicknell et al. 2015; Flurkey et al. 2007). After behavioral studies, brain specimens were prepared for imaging. Mice were anesthetized with ketamine/xylazine (100 mg/kg ketamine, 5–10 mg/kg xylazine) and perfused via the left cardiac ventricle. Blood was cleared with 0.9% Saline at 8 ml/min for approximately 5 min. Fixation was achieved using 10% neutral buffered formalin phosphate containing 10% (50 mM) Gadoteridol (ProHance, Bracco Diagnostics Inc., United States) at 8 ml/min for about 5 min. The specimens were then fixed in formalin for 12 h and subsequently stored in phosphate-buffered saline with 0.05% (2.5 mM) Gadoteridol until imaging. Animal age and sex were used as biological traits for our predictive models.

Behavior

We evaluated spatial memory and learning of mice using the MWM test, as described in previous studies (Badea et al. 2022, 2019). Mice underwent a 5-day acclimation period involving handling and water exposure. Training took place in a 150 cm-diameter pool with opaque water, with a ceiling-mounted camera tracking swim paths using ANY-maze software (Stoelting, United States). Daily trials included four 1-min trials in two blocks, with mice introduced from varying positions to locate a submerged platform using visual cues for 5 consecutive days. Learning was gauged by swim distance to the platform, percentage of swim distance in the platform’s quadrant (normalized distance), and the AWN, which quantifies the circularity of the path (Badea et al. 2022). The AWN summed the absolute values of winding numbers for each loop, capturing both loop number and size. If a mouse failed to find the platform, it was guided there and allowed a 10-s stay. On day 5, an hour post-last training, and on day 8, platform-removed probe trials assessed navigation strategies through total swim distance, distance within the target quadrant (normalized distance), and AWN. We used R and the packages lme4, 1.1–27.1, and lmertest 3.1–3 to build mixed effect models for the learning trials with fixed effects for age, sex, and time (Stage), and random effects for animal identity, e.g., Distance−Age*Sex*Stage + (1| AnimalID). We used linear models for the probes, e.g., Distance−Age*Sex. We used ANOVA to determine the effects of age and sex on behavioral markers including the total swim distance, normalized swim distance in the target quadrant, and the AWN. ANOVA was followed by post hoc tests (using Sidak adjustments), and significance was set at p < 0.05. Cohen’s F effect sizes were estimated using the package effsize 0.8.1. AWN and normalized SW distance were used as behavioral traits for our integrative models.

Imaging

Mouse brain specimens were imaged at 9.4 T, as described in a previous study (Badea et al. 2019). We used a 3D spin echo diffusion weighted imaging (SE DWI) sequence with TR/TE: 100 ms/14.2 ms; matrix: 420 × 256 × 256; FOV: 18.9 mm × 11.5 mm × 11.5 mm, BW 62.5 kHz; reconstructed at 45 μm isotropic resolution. Diffusion weighting was applied along 46 directions, using 2 diffusion shells (23 at 2,000 and 23 at 4,000 s/mm²); and we acquired 5 non-diffusion weighted images (b0). The max diffusion pulse amplitude was 130.57 Gauss/cm; duration 4 ms; separation 6 ms. We used eightfold compressed-sensing acceleration (Uecker et al. 2015; Wang et al. 2018). Diffusion parameters were reconstructed using MRtrix3 (Tournier et al. 2019) producing −2 million tracts and MAP MRI parameters were reconstructed using DIPY (Garyfallidis et al. 2014; Tournier et al. 2012). We have used pipelines implemented in a high-performance computing environment for voxel-based analysis and to calculate connectomes (Anderson et al. 2019), with reference to a symmetrized mouse brain atlas (Calabrese et al. 2015). Connectomes were normalized, so that the diagonal was removed and then the connectome and elements were divided by the sum of the remaining entries. We conducted regional and voxel-wise analyses as in (Badea et al. 2019). The Statistical Parametric Mapping SPM toolbox (Friston et al. 1994), version 12 was used to produce cluster-based statistical maps, at 5% cluster false discovery rate correction. Connectomes were used as the imaging traits in our models.

Transcriptomics

We conducted RNA-Seq experiments at the Duke Sequencing and Genomic Technologies Shared Resource Core to identify genes differentially expressed concerning age and sex. The RNA was extracted from whole blood using QIAGEN kits, and quality was assessed with a NanoDrop 2000 spectrophotometer. We processed the samples on an Illumina NovaSeq600 S2 platform, with 50 bp PE full flow cell, utilizing NuGEN mRNA-Seq with Any Deplete Globin. Sequencing data underwent quality control through FastQC, adhering to a Phred score cutoff above 20. As no adapter sequences were found adjacent to transcript reads, data trimming was not required. Consequently, raw data were directly employed in alignment and quantification stages. We aligned raw RNA-seq reads to a Mus musculus reference transcriptome, and the Salmon program (Patro et al. 2017) was utilized to quantify relative transcript abundance. Principal component analysis was executed to identify any confounding metadata factors that could affect differential expression analysis, such as batch and qubit. DESeq2 package (Love et al. 2014) was employed for normalization and evaluating gene expression differences. Genes were annotated with equivalent human gene names (mouse orthologs) using BioMart (Durinck et al. 2005). Additionally, gene set enrichment analysis (Mootha et al. 2003; Subramanian et al. 2005) was conducted to identify gene ontology terms and pathways associated with altered gene expression, utilizing GOnet (Pomaznoy et al. 2018). We used the Sumer package that implements affinity propagation, which significantly condenses gene sets using a weighted set cover algorithm after enrichment analysis, and clusters similar gene sets into groups and identifies one representative gene set that best represents each group as exemplars (Savage et al. 2019).

Integrative modeling

To reveal the complex relationships among disparate datasets, we employed sparse multiple canonical correlation analysis (SMCCA) that was designed to integrate inputs from multiple sets of data (Mahzarnia et al. 2022; Witten and Tibshirani 2009). In our model, we included connectomes from diffusion MRI, biological traits including age and sex, behavioral metrics from MWM experiments, and blood-based RNA-seq data.

The SMCCA method identifies sparse canonical variables by optimizing linear combinations of three or more sets of multivariate variables. The objective function of this optimization problem is to maximize the pairwise correlations among datasets by determining the optimal weighting vectors for each variable. Incorporating sparsity constraints enhances interpretability. Additionally, we filtered the connectomes, omitting the bottom 2% of unique non-zero values, which corresponds to around 70 streamlines. Consequently, only cells with over 70 streamlines were included in the analysis. This acts as a noise reduction, decreasing the occurrence of false positives. With multiple data matrices ${\text{X}}_1,{\text{X}}_2,\dots ,{\text{X}}_{\text{m}}$, the SMCCA method finds vectors ${\text{w}}_1,{\text{w}}_2,\dots ,{\text{w}}_{\text{m}}$ in the respective variable spaces to maximize the following objective function:

$$argma{x}_{w_1,w_2,\dots ,w_m}\sum _{i=1}^m\sum _{j=1}^m\left(w_i^T{X}_i^T{X}_j{w}_j\right),$$

subject to:

$$w_k^T{X}_k^T{X}_k{w}_k=1forallkfrom1tom.$$

The algorithm produces sparse and interpretable sets of coefficients for biological traits and behavior metrics such as the winding number at day 8, connectomes, and/or RNA-seq data. To estimate confidence, we employed a bootstrap resampling method with 1000 iterations and replacement to construct the 95th percentile confidence interval for the sum of the correlations.

Following SMCCA analysis, we adopted Graph Neural Networks (GNN), utilizing the BrainGNN framework (Li et al. 2021), to exploit the graph-structured nature of our connectome datasets and to identify which subgraphs are relevant for age and sex prediction. Connectomes were fed into a Graph Convolutional Network (GCN) layer succeeded by a topK pooling layer (Gao and Ji 2022). This configuration was repeated and subsequently appended by two batch normalization layers and dropout layers followed by a SoftMax layer for the classification task. Within the GCN layers, the ReLU activation functions were employed, and sigmoid activation functions were used in the topK pooling layers. The utilization of node scores extracted from the topK pooling layers facilitated the identification of salient brain regions pertinent to the classification task (Li et al. 2021). To estimate the predictive accuracy, we adopted a fivefold cross-validation approach with a data partition ratio of 80:20.

Our integrative modeling approach, which coalesces biological traits, cognition, brain networks, and peripheral gene expression changes, is illustrated in Fig. 1.

Fig. 1.

Schematic of our multivariate study. Data on RNA-seq, behavior and imaging were acquired and analyzed individually. These data were then subjected to integrative modeling approaches, including SMCCA and GNN prediction

Results

We have tested the dynamics of mouse brain networks in relation to cognitive and gene expression changes during the transition between middle and old age in those carrying the APOE2 alleles. Our results have revealed sex-specific alterations in brain networks associated with memory function and gene expression changes during aging.

Cognitive performance changes in aging APOE2 mice

In the MWM test involving mice aged 12 and 16 months, learning and memory were evaluated through three metrics: the distance swam to find a submerged platform, the percentage distance in the target quadrant, and the winding number of swim trajectories, as shown in Fig. 2. There were no significant differences due to sex and age in learning trials, but a time effect was significant in all metrics: total distance (F(4, 763) = 6.74, p = 2.49e-05), normalized distance (F(4, 758.32) = 7.59, p = 5.32e-06), and AWN (F(4, 763) = 5.86, p = 0.0001). This indicated that mice were able to learn and modify their swim trajectory or shape. A significant age–stage interaction was found only for the winding number (F(4, 763) = 2.85, p = 0.02). In the first probe trials, age significantly impacted total distance (F(1, 37) = 8.36, p = 0.0064), but no significant effect for normalized distance was found. Age was also significant for the AWN (F(1, 37) = 8.93, p = 0.005). In the second probe trials, age had significant effects on total distance (F(1, 37) = 5.03, p = 0.03), normalized distance swam in the target quadrant (F(1, 37) = 5.95, p = 0.02), and the AWN (F(1, 37) = 8.39, p = 0.006). The results highlight age as a significant factor influencing learning and memory, with no observable differences related to sex.

Fig. 2.

Learning and memory in APOE2 mice transitioning from middle to old age. We detected significant age but not significant sex effects for learning and memory based on the MWM test. A There was a significant interaction of age × stage (days 1 to 5) for the AWN during the learning trials. The winding number increased with age, indicating more complex trajectories in old relative to middle-aged mice. B The total distance did not change with age during the learning trials but decreased in day 5 and day 8 probes. C The normalized distance in the target quadrant during learning trials and first probe was similar for the two groups but increased in aged mice for the second probe. In the learning trials, for each day, p < 0.05 is as asterisk (*). Global p values are reported for each comparison, and p < 0.05 was considered significant

Image‑based metrics in aging APOE2 mice

Morphometry

The total brain volumes did not differ by sex, but declined with aging from 527.91 ± 9.86 mm³ to 512.35 ± 17.41 mm³; this small 3% change was significant F(1, 28) = 7.47, p = 0.01. A regional analysis revealed that males had larger volumes relative to females for the bed nucleus of stria terminalis, subbrachial nucleus, postsubiculum (−10%), and claustrum (> 5%); while females had larger volumes for the brachium of the superior colliculus, orbitofrontal cortex, frontal association cortex, and the longitudinal fasciculus of pons > 9%). The largest age-associated atrophy was observed for the ventral hippocampal commissure (15%), fornix, and cingulate cortex areas 24 (9%). Opposite effects were noted for the parietal cortex, parietal association cortex, and primary somatosensory cortex−18% (Supplementary Table 1).

Voxel-based analysis (VBA) revealed age- and sex-specific spatial distribution of brain volume (Fig. 3). Males had larger olfactory areas, CA3, and dentate gyrus of hippocampus, hypothalamus, and cerebellar vermis compared to females, who had larger CA1 of hippocampus, subiculum, and entorhinal cortex. Younger mice exhibited larger volumes in olfactory areas, cingulate cortex, hypothalamus, amygdala, hippocampus, substantia nigra, and ventral thalamic nuclei compared to older mice. Thus, a shared age and female sex vulnerability was noted for the olfactory areas, hypothalamus, and hippocampus. Additionally, there was a significant age-by-sex interaction. Males showed faster age-related changes in insula, dorsal hippocampus, substantia nigra, and perifascicular and ventral thalamic nuclei. In contrast, females showed faster atrophy in olfactory areas, amygdala, ventral hippocampus, entorhinal cortex, and subiculum.

Fig. 3.

Morphometric differences in sex and age of APOE2-targeted replacement mice and their interaction. Clusters larger in males than in females are shown in blue and larger regions in females are shown in red. Clusters larger in middle age are shown in red. The positive interaction of age by sex, indicating faster changes in males are shown in blue, and the negative interaction of age by sex indicating faster changes in females are shown in red. FDR-corrected (5%) parametric maps show t and F values

Diffusion metrics

Among white matter tracts, the region-based analysis showed the most significant age-related differences in fractional anisotropy (FA) values for the inferior cerebellar peduncle, lateral olfactory tract, and fornix; where younger mice showed > 5% larger values than older mice. Among the gray matter brain regions, the cingulate cortex and primary somatosensory cortex were significant, with middle-aged mice having > 10% higher in FA values compared to older mice. Regarding sex-based differences, females showed > 7% and > 5% larger FA values in the cingulate cortex and ventral hippocampal commissure, respectively. Detailed results can be found in Supplementary Table 1.

Voxel-based FA analyses revealed white matter fiber density differences related to sex and age (Fig. 4). Male mice showed a higher FA for the corpus callosum, motor and olfactory areas, anterior cingulate, septal nucleus, hypothalamus, hippocampus, somatosensory areas, thalamus, caudate-putamen, amygdala, and subiculum. Female mice instead showed higher FA in the entorhinal cortex, colliculi, and cerebellum. Old mice have larger FA in the anterior commissure, motor cortices, partial cingulate area, and somatosensory areas. Middle-aged mice exhibited much more widespread higher FA various areas of the brain, including the corpus callosum, olfactory areas, nucleus accumbens, globus pallidus, hypothalamus, thalamus, hippocampus, colliculi, and cerebellum. Thus, a shared age and female sex vulnerability was noted for the hippocampus, S1 and M1. A positive interaction of age by sex which indicates faster decline in males was observed in the colliculi and midbrain reticular nucleus. Conversely, a negative interaction supportive of a faster decline in females was seen in the olfactory areas, taenia tecta, orbital areas, hippocampus, thalamus, M1 areas, amygdala, and reticular nucleus.

Fig. 4.

Fractional anisotropy differences in sex and age of APOE2-targeted replacement mice and their interaction. Clusters with higher FA estimates in males than in females are shown in blue and larger regions in females are shown in red. Clusters with higher FA in middle age are shown in red, and higher FA in old animals are shown in blue. The positive interaction of age by sex, indicating faster change in males, and is shown in blue and the negative interaction of age by sex indicating faster change in females is shown in red. FDR-corrected (5%) parametric maps show t and F values

MAP MRI metrics

No regional differences for sex or age survived FDR correction in the region-based analysis.

Voxel-based analysis results for the return-to-origin probability (RTOP) indices are shown in Fig. 5. Sex-specific differences with males showing larger values RTOP than females, suggesting larger axonal diameters, in the corpus callosum, fimbria, and optic tract. No regions displayed greater RTOP in females. Middle-aged animals had significantly higher RTOP values in the cingulum, corpus callosum, stria terminalis, internal capsule, fasciculus retroflexus, as well as the substantia innominata, amygdala, and hypothalamus. Thus, a shared age and female sex vulnerability was noted for the corpus callosum. No regions with greater RTOP were observed in older individuals compared to middle-aged ones. Furthermore, a positive interaction between age and sex, indicative of a faster decline in males, was seen in the primary and secondary motor areas, nucleus accumbens, caudate-putamen, primary somatosensory areas, globus pallidus, and amygdala, suggesting combined age and sex effects in these regions. No negative interaction was found in the RTOP analysis.

Fig. 5.

Differences in MAP metric, RTOP, based on sex and age in APOE2-targeted replacement mice and their interaction. Clusters with higher RTOP estimates in males than in females are shown in blue. Clusters with higher RTOP estimates in middle age are shown in red. The positive interaction of age by sex, indicating faster change in males, is shown in blue. FDR-corrected (5%) parametric maps show t and F values

Connectivity

SMCCA detection of networks involved in learning and memory

SMCCA detected brain networks associated with biological traits (age and sex), and behavior (spatial learning and memory) (Fig. 6), with a sum of the three correlations pairs of 1.69, and 95% bootstrap interval [1.34, 2.33]. The weights associated with these traits were 0.51 for sex and 0.85 for age; 0.75 for the normalized distance in the target quadrant and 0.65 for the absolute winding number (during probe trials at day 8), denoting a large importance of age in shaping brain networks, and for the winding number in describing memory and learning. Furthermore, the distribution of subgraph weights, as depicted in Supplementary Figs. 1 and 2, suggests a more pronounced difference between middle-aged and old-aged groups compared to the difference between males and females. The average median difference between males and females is 138, whereas the average median difference for age-related weights between middle-aged and old-aged groups is significantly higher at 258.72. The largest subgraph weights were identified for a large sub-network including the parvicellular reticular nucleus and principal sensory trigeminal nucleus, gigantocellular reticular nucleus, and pontine reticular nucleus (Fig. 6C). The sum of network weights is negative (−1.506), but the absolute sum of weights is 2.032, indicating that this complex network includes both negative and positive connectivity weights within the network.

Fig. 6.

SMCCA detected brain networks associated with age, sex, and spatial memory as a composite risk factor. A Positive weights are shown in red, and negative weights are shown in blue. B Connectivity for selected sub-networks is shown using violin plots for both age and sex comparison. The full set of violin plots can be viewed in Supplementary Figs. 1 and 2. C List of SMCCA selected sub-networks and associated weights

The sub-network 4 in Fig. 6C consists of the amygdala, medial entorhinal cortex, and globus pallidus, and has more weight for females and old animals compared with males and middle age, indicating that it may have a compensatory role with aging. Sub-networks 6, 10, and 11, which include regions, such as the orbital cortex, frontal cortex, cingulate cortex, olfactory tract, amygdala, and piriform cortex, had more weight for middle age than old, which indicate connectivity loss during aging.

SMCCA detection of networks involved in peripheral/blood markers

We investigated the impact of age and sex differences on serum gene expression, and the results are presented in Fig. 7A for sex, and Fig. 7B for age, and in Supplementary Table 2. Next, we used SMCCA on biological traits, connectome, and RNA-Seq data showing differential expression for either age or sex comparisons. This results in effectively filtering gene subsets linked to age and sex (Fig. 7C, D), which were associated with connectome subgraphs (Figs. 8 9).

Fig. 7.

Differentially expressed genes (DEG) for sex and age, before and after connectome-based filtering via SMCCA. A Volcano plot illustrating the genes with significant differential expression (adjusted p < 0.05) related to sex. B Volcano plot illustrating the genes with significant differential expression (adjusted p < 0.05) related to age. C SMCCA analysis applied to RNA-seq and connectome data, resulting in age-related gene subsets associated with connectome subgraphs. D SMCCA analysis applied to RNA-Seq and connectome data, resulting in sex-related gene subsets associated with connectome subgraphs

Fig. 8.

Brain networks associated with serum genes and aging in APOE2 mice indicate networks that decline and also compensatory networks. A Resulting sub-networks. Positive weights are shown in red, and negative weights are shown in blue. B Connectivity differences between middle age and old age for selected sub-networks are shown using violin plots. The full set of violin plots can be viewed in Supplementary Fig. 3. C List of surviving sub-networks and associated weights

Fig. 9.

Brain networks associated with serum genes and sex in APOE2 mice. A Resulting sub-networks. Positive weights are shown in red, and negative weights are shown in blue. B Connectivity differences in males and females for selected sub-networks are shown using violin plots. The full set of violin plots can be viewed in Supplementary Fig. 4. C List of surviving sub-networks and associated weights

The sum of correlations between age-related genes, connectomes, and biological traits was 2.45, with weights of − 0.04 for sex and − 0.99 for age, and a 95% confidence interval of [2.09, 2.70]. Of the initial 42 genes, 39 remained after hyperparameter tuning. In the case of sex-related genes, the sum of correlations was 2.33 with weights of 0.99 for sex and 0.04 for age. The 95% confidence interval was [2.05, 2.67]. After tuning, 249 out of the initial 436 genes were retained in the SMCCA output.

Age‑related blood transcriptomics

Out of 15,934 genes analyzed for age-related blood transcriptomics, 44 exhibited significant fold changes with an FDR (False Discovery Rate) below 0.05. Among these, 42 genes had positive fold changes, with log2FC ranging from 0.8 to 7.1 (Fig. 7A). Cpt1c showed the highest fold change (log2FC = 7.1, adjusted p = 0.04), and is associated with the transport of long-chain fatty acids into mitochondria and neuronal oxidative metabolism (Lee and Wolfgang 2012). Other genes with notable fold changes included Nat14, Rangrf, Arg1, Nbeal1, Olfr1195, Pex1, and Myo1e. Most of these genes are implicated in processes, such as the urea cycle, nervous system processes, cholesterol transport, and immune responses. Two genes, Ptprt and Zfp865, showed negative fold changes and are involved in signal transduction and synapse organization. After filtering based on SMCCA connectome associations (see Fig. 7D), genes like Ankzfp1 (related to cellular response to hydrogen peroxide and mitochondrial integrity), as well as Pex1, Cep250, Nat14, Arg1, and Rangrf, exhibited high absolute weights and were linked with connectomes for aging.

Sex‑related blood transcriptomics

Among the genes related to sex, Maoa exhibited the highest male:female fold change (log2FC = 3.6, adjusted p = 0.0008). Maoa plays a role in catabolic and cellular nitrogen compound metabolic processes (Gaweska and Fitzpatrick 2011). Additionally, several olfactory receptors like Olfr365 and Olfr646, and the gene TMPT, associated with thiopurine S-methyltransferase, showed significant fold changes based on sex. TMPT had a log2FC of 4.1 and an adjusted p value of 0.02 (log2FC = 4.1, adjusted p = 0.02). Ahrr gene had a significant negative fold change (log2FC = −1.5, adjusted p = 0.003). Ahrr is known to be involved in the regulation of cell growth and differentiation. Other genes showing negative fold changes include Krt8, which is involved in maintaining cellular structural integrity and has roles in signal transduction and cellular differentiation; Lurap1, which is associated with the positive regulation of cytokine production; and Oosp1, which may be involved in cell differentiation. Connectome-based filtering through SMCCA for sex-associated changes revealed (Fig. 7C) overexpressed genes, including Uhmk1 involved in phosphorylation of myelin basic protein and Synapsin I, Oosp1, Fcrl5, Pex1, Serping6b, Fcho2, and Peg10.

Brain subgraphs associated with age‑related blood transcriptomics

The brain subgraphs associated with age-related differential gene expression by SMCCA, which integrated biological traits, connectome, and age-related blood RNA-seq data, are presented in Fig. 8, together with violin plots for top sub-networks. Sub-network 1 had the highest absolute total weight of 1.398, and includes the cerebellar peduncle, cerebellar white matter, vestibular nuclei, and the fastigial medial nucleus of the cerebellum. This network exhibited stronger connectivity in older individuals, suggesting a compensatory mechanism in aging. Conversely, sub-networks 2, 10, and 12, involving regions such as the amygdala, globus pallidus, and hippocampus, showed deteriorating connectivity with aging as shown in Fig. 8C.

The Gene Set Enrichment Analysis (GSEA) for the 44 age-significant genes identified annotations related to stress, signal transduction, transport, cellular nitrogen compound metabolism, immune system processes, and anatomical structure development (details in Supplementary Table 3).

Brain subgraphs associated with sex‑related blood transcriptomics

The brain subgraphs associated with sex-related differential gene expression by SMCCA, which integrated biological traits, connectome, and sex-related blood RNA-seq data, are presented in Fig. 9. Sub-network 1 had the highest total weight of 2.206, and included regions like the ventral thalamus, reticular thalamic nucleus and rest of thalamus, superior and inferior colliculus, hippocampus, and entorhinal cortex. This sub-network had stronger connectivity in females. On the other hand, sub-networks 2 and 4, including the dorsal tegmentum, hypothalamus, hippocampus, presubiculum, and insular cortex had stronger connectivity in males.

The genes that were significant for both age and sex after FDR correction were Myo1e (log2FC = −1.5, SE = 0.3, adjusted p = 0.003) involved in intracellular trafficking and endocytosis, and part of disease-associated microglia genes (Sobue et al. 2021), and Pex1 involved in peroxisome biogenesis and lipid metabolism (Reuber et al. 1997) (log2FC = 3.6, SE = 0.7, adjusted p = 0.0008).

Gene Set Enrichment Analysis (GSEA) of the 546 sex-related genes with significant differential expression identified associations with biological pathways and Gene Ontology (GO) terms, such as anatomical structure development (144 genes), transport (107 genes), cell differentiation (103 genes), stress response, and cellular nitrogen compound metabolic processes (Supplementary Table 4).

The SMCCA filtered genes differentially expressed with age or sex pointed to several enriched processes, with a role for biological and metabolic processes, cellular component organization and biogenesis, transport, immune response, and cell communication (Fig. 10). Our findings indicate that by combining blood-based transcriptomics with imaging data and considering traits, such as age, sex, and cognition, we can gain insights into the mechanisms that underlie the transition from middle to old age in mouse models that express the APOE2 allele.

Fig. 10.

Enriched processes and pathways (n = 154) included regulation of cellular component organization or biogenesis (p = 2.1 × 10⁻⁷, N = 111 genes), localization (p = 1.2 × 10⁻⁵, N = 96 genes), metabolic process (p = 0.04, N = 93 genes), response to stimulus (p = 5.4 × 10⁻⁵, N = 77 genes), signaling (p = 5.12 × 10⁻⁵, N = 72 genes), transport (p = 7.6 × 10⁻⁴, N = 70 genes), and response to stress (p = 0.001, N = 64 genes). Regulation of cell communication (p = 4.7 × 10⁻⁵, N = 72 genes) was highly significant

Affinity propagation was employed to delve deeper into the core processes of resultant SMCCA gene pathways. The outcomes of this analysis are illustrated in Fig. 11. We discerned two distinct pathway clusters, each comprising seven pathways. The central process of the first cluster pertains to the regulation of localization. In contrast, the core process of the second cluster revolves around the negative regulation of cellular processes. Within the first cluster, there was a pronounced positive regulation of cellular processes and localization. The second cluster predominantly involves negative regulatory processes, and notable pathways are response to stress and immune system processes.

Fig. 11.

Selected biological processes and pathways. A The enriched biological processes were analyzed using the Sumer package, and trimmed down to the highest 14 scores. B Gene ontology ID and the corresponding definition. C Affinity propagation clustering resulted into two clusters, each with seven pathways. The color indicates the scores, and the size of the circles indicates the number of genes involved in each pathway

GNN prediction

GNN was used to predict the age and sex based on APOE2 mouse brain structural connectomes, as well as the associated brain subregions. We validated the results through fivefold cross-validation, equivalent to an 80:20 data split for each fold. The sex prediction accuracy was estimated by BrainGNN (Li et al. 2021) with prediction accuracy of 0.794. In contrast, binary age prediction (middle age or old) accuracy was 0.765. We examined the scores assigned by the topK pooling layer for each node of the connectome to find important structures during training.

Of the top 20 regions with high weights in the model, most regions were shared for age and sex (16 exact matches, and 4 neighboring subdivisions for the visual, somatosensory, and cingulate cortex). These shared regions involved subdivisions of the primary somatosensory cortex, the primary motor cortex, frontal association cortex, the secondary visual and auditory cortex, as well as the cingulate cortex. For sex prediction, graph neural networks highlighted the corpus callosum and striatum as unique. In contrast, age prediction identified unique regions, e.g., the insular cortex and frontal cortex area 3.

We identified specific brain regions that are either unique to or shared in the prediction of both age and sex. This underscores the intertwined relationships of these factors in influencing brain structure and functionality.

Discussion

Understanding the aging process helps identify factors that contribute to successful aging and longevity, or to age-related conditions such as Alzheimer’s disease (AD), a neurodegenerative disorder that involves the accumulation of amyloid-β (Aβ) and tau proteins in the brain, and which leads to cognitive impairment and memory loss. Among the genetic risk factors for AD, the APOE gene, which encodes for apolipoprotein E involved in cholesterol transport, plays a significant role. Our study focuses on the less studied APOE2 allele, which has been associated with protective effects against AD and preserved cognitive function, but also with cerebral amyloid angiopathy, risk for PTSD, AMD, supranuclear palsy, and argyrophilic grain disease (Li et al. 2020). The aim of our study was to examine the relationship between cognitive changes, brain connectivity, and peripheral markers of gene expression during aging, focusing on critical middle age to old age transitions. Additionally, considering that females are more susceptible to neurodegeneration, we aimed to investigate how the aging trajectories of APOE2 carrying females compare to males.

The MWM results revealed a significant interaction of age by stage for the AWN and not a significant difference in sex, indicating that age was the major determinant of group differences. Imaged-based analysis revealed a small but significant age-related atrophy in total brain volume (3%), but no sex differences. Males had larger regional volumes in the orbitofrontal cortex, involved in sensory integration, decision-making, and emotional processing (Rolls 2019), and brachium of the superior colliculus, involved in visuospatial orientation (Comoli et al. 2012). Females displayed a larger bed nucleus of stria terminalis, involved in anxiety and stress responses (Bangasser and Wicks 2017), and claustrum, which has a role in sensory integration (Smith and Alloway 2010). The largest age-associated atrophy was observed for the ventral hippocampal commissure (15%), fornix, and cingulate cortex areas 24 (9%), supporting that these regions have sensitivity not just to pathological but also normal aging (Dutar et al. 1995; Good et al. 2001; Leech and Sharp 2014).

Voxel-based analysis revealed sex-specific variations in morphometry. Males had a larger volume in the CA3 of the hippocampus, associated with memory encoding (Rolls and Kesner 2006), and the cerebellar vermis, linked to motor control (Stoodley and Schmahmann 2010). In contrast, females showed a larger volume in the CA1 of the hippocampus, crucial for long-term memory formation and retrieval (Maren 1999), and the entorhinal cortex, integral for memory and navigation (Knierim et al. 2013). Younger mice exhibited larger olfactory areas, cingulate cortex, hypothalamus, amygdala, hippocampus, parafascicular and ventral thalamic nuclei, which supports age-related decline in the sensory and memory processing (Doty and Kamath 2014). Males showed faster age-related changes in the insula, dorsal hippocampus, substantia nigra, and thalamic nuclei. The dorsal hippocampus is notable because of its role in spatial memory and navigation (Eichenbaum 2017). The substantia nigra is involved with reward and movement, suggesting that males might be more vulnerable to the impairment of certain reward and movement system (Obeso et al. 2017). In white matter fiber analysis, middle-aged mice displayed higher FA throughout the brain relative to old mice, especially in areas of the corpus callosum, olfactory areas, nucleus accumbens, globus pallidus, hypothalamus, thalamus, hippocampus, colliculi, and cerebellum. Finally, MAP MRI metrics of RTOP, indicative of axonal diameter, showed larger values predominantly in males in the lateral septal nucleus, involved in stress and anxiety responses (Sheehan et al. 2004), and the hippocampal commissure, crucial for interhemispheric communication (Bloom and Hynd 2005). Middle-aged animals had significantly higher RTOP values in the cingulum, corpus callosum, stria terminalis, internal capsule, and fasciculus retroflexus, denoting specificity for age changes in white matter bundles. An age-by-sex interaction was seen in the primary motor and somatosensory areas and the amygdala. Interestingly, the primary motor and sensory areas were also detected by our GNN method to predict age and sex based on connectomes.

The SMCCA identified sub-networks related to biological and behavioral traits, and these included regions such as the hippocampus, which had higher weight for middle-aged subjects, signifying network deterioration with aging. This supports the VBA results, and hints at potential changes in memory processing, aligning with earlier research (Fanselow & Dong 2010). Our SMCCA focusing on cognitive traits showed the cingulate cortex having more weight for middle age compared to old animals, implying a reduction in connectivity with aging, which is linked to cognitive functions (Fjell et al. 2014). Furthermore, GNN analysis identified the cingulate cortex as a critical region for predicting both age and sex.

Our analyses of blood-based transcriptomics aimed to detect the effects of aging and sex on peripheral markers in APOE2 mice. Our study uncovered a more extensive array of genes affected by sex (546) compared to aging (44). A noteworthy finding was the significant positive fold change of Cpt1c with aging. Cpt1c plays a crucial role in the regulation of beta-oxidation and the transport of long-chain fatty acids into mitochondria, essential for neuronal oxidative metabolism (Lee and Wolfgang 2012). Moreover, several genes, such as Pex1, Myo1e, and Arg1, exhibited differential expression with aging, with roles in protein import into peroxisomes, intracellular movement, and immune responses, respectively (Karch and Goate 2015; Neuner et al. 2020; Romero-Molina et al. 2022). Maoa displayed the most significant positive fold change associated with sex. Maoa is involved in the catabolic processes and the breakdown of neurotransmitters, such as serotonin, epinephrine, norepinephrine, and dopamine (Cho et al. 2021). In contrast, Ahrr, Krt8, and SAAL1, which are involved in the regulation of cell growth and differentiation, maintenance of cellular structural integrity, and responses to proinflammatory stimuli, respectively, displayed large negative changes (Kawajiri and Fujii-Kuriyama 2017; Lim & Ku 2021; Yang et al. 2022). Further, our results highlighted genes, such as Uhmk1, Serping6b, Fcho2, Peg10, and Rida, which are implicated in anatomical structure development, sensory perception of sound, cell proliferation, differentiation, apoptosis, and metabolism of amino acids (Arfelli et al. 2023; Irons et al. 2020; Mulkearns and Cooper 2012; Turato and Pontisso 2015; Xie et al. 2018).

Importantly, through SMCCA, we identified Myo1e and Pex1 as genes that display differential expression in relation to both sex and aging. These genes are linked to immune processes, apoptosis, efferocytosis, and stress responses, which are crucial factors in the development of Alzheimer’s disease (Karch and Goate 2015; Neuner et al. 2020; Romero-Molina et al. 2022). Enrichment analysis further revealed involvement in biological and metabolic processes, cellular component organization and biogenesis, transport, and immune responses. Previous studies examining human blood and rat brain transcriptomic changes in aging also highlighted gene pathways involved in immune and defense response (Harris et al. 2017; Shavlakadze et al. 2019). Our gene affinity propagation analysis further supported to role of such processes, including regularization of localization, negative regulation of cellular process, response to stress, and immune system process. Our results built on previous studies that demonstrated preserved pathways, particularly immune-related, between humans and mouse models with human APOE TR genes through brain transcriptomics and blood metabolomics (Zhao et al. 2020).

The brain subgraphs corresponding to age-related differential gene expression included amygdala, globus pallidus, and hippocampus, which showed deteriorating connectivity in aging, while interestingly cerebellar connectivity appeared to play a compensatory role. The brain subgraphs associated with sex-related differential gene expression revealed higher connectivity in males in the dorsal tegmentum, hypothalamus, hippocampus, presubiculum, and insular cortex. In contrast, a subgraph including the ventral and reticular thalamic nuclei, colliculus, hippocampus, and entorhinal cortex exhibited stronger connectivity in females. Our analysis suggests blood markers for tracking various traits and brain-related changes.

In our study, we incorporated a multivariate approach, integrating behavioral, imaging, and connectomic and blood-based transcriptomic analyses, which highlighted the significance of particular brain regions across various dimensions/domains of vulnerability. Collectively, the cingulate cortex, somatosensory areas, olfactory areas, and hippocampus demonstrated significance in morphometry, diffusion, and MAP metrics, and were also identified as relevant regions in behavior-associated and gene-associated SMCCA.

We acknowledge several limitations of our study. First, the utilization of a narrow age range, spanning middle to old age, though significant, imposes limitations on the generalizability of the results. Another limitation is the utilization of mouse models with targeted replaced human APOE genes. While these models have been instrumental in providing critical insights, they exhibit limitations in accurately replicating human aging. The APOE2 allele, which has been less studied compared to APOE4 and APOE3, poses a specific challenge, in that mice with the APOE2 allele exhibit a higher incidence of hyperlipoproteinemia compared to humans (Patrick M. Sullivan et al. 1998a, b), and there is a lack of consensus regarding their performance in learning and memory tests (Badea et al. 2022; Watson et al. 2021). In the future, we aim to include studies of APOE3 and APOE4 alleles with integrative deep learning-based models using multivariate data sets, to help understand networks’ vulnerabilities and resilience conferred by APOE genotypes and the underlying molecular substrates.

In conclusion, this study presents a comprehensive analysis that connects behavior, brain structure, and peripheral transcriptomics in a mouse model carrying the APOE2 allele. By linking peripheral molecular markers to brain and behavior, our research provides avenues on how to monitor multivariate markers to assess risk during aging. Notably, we pinpoint pathways governing immune regulation, as well as cellular stress response and repair processes. Future studies accounting for all three major APOE alleles could provide further understanding on APOE-allele-specific processes that play a role in human aging and Alzheimer’s disease (Karch et al. 2014; Karch and Goate 2015; Neuner et al. 2020; Romero-Molina et al. 2022). Such insights are crucial for advancing toward more targeted and effective interventions for age-related cognitive decline and dementia, and to establish factors that support longevity and successful aging.

Data availability

Code and data necessary to reproduce the original analyses are available from https://github.com/AD-Decode/APOE2_Mouse. The raw RNA-seq data are available from ENA database (Accession PRJEB59982).