All experiments involving mice were conducted with approval and accordance described in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All experiments were approved by the Animal Care and Use Committee at The Jackson Laboratory. Mice were kept in a 12/12-h light/dark cycle and fed ad libitum 6% kcal fat standard mouse chow.

Humanized APOE mice were created in collaboration with the Genetic Engineering Technologies core at The Jackson Laboratory. The mouse Apoe gene is located on chromosome 7 at 19,696,109–19,699,166. An APOE^ε4 gene-targeting construct was made that included 4980 bp of mouse sequence, which defined the mouse 5’ homology arm including exon 1 of mouse Apoe, 4292 bp of human APOE^ε4 sequence including human protein coding exons 2–4 of the human gene as well as an additional 1.5 kb of flanking human sequence after the 3′UTR to include any potential regulatory sequences. Exon 4 contained sequence that encoded the APOE^ε4 isoform (nucleotide sequence for arginine at R130 and R176). The predicted protein sequence in our humanized APOE^ε4 mouse model is 317 amino acids and includes the18 amino acid signal peptide at the N-terminus. As in all APOE isoforms, the 18 amino acid leader sequence is cleaved resulting in the mature 299 amino acid protein, with the amino acids that define APOE^ε4 located at positions 112 and 158 (Mahley and Rall, 2000; Rohn and Moore, 2017). A Frt-neo-Frt (FNF) selection cassette was inserted after the human sequence followed by a Nde1 restriction site (for ease of Southern screening). The FNF cassette was followed by 5166 bp of mouse sequence, the 3’ homology arm. The resulting 14,438 bp synthesized construct was cloned into pBlight vector using recombineering techniques, producing a construct called mApoE_hAPOE4_PGKneo_mAPOE for gene targeting in embryonic stem cells. The APOE^ε4 gene-targeting construct was introduced into cultured embryonic stem (ES) cells of a C57BL/6J (B6) mouse strain by electroporation. Homologous recombination produced loci that retained all normal mouse regulatory sequences (plus non-coding exon one) together with the human APOE^ε4 protein-encoding exons 2–4 (Figure 1A). Transfected ES cells were screened by Southern blot to ensure correct targeting. Three clones were identified that were correctly targeted. ES cells containing the correctly targeted locus were introduced into albino B6 embryos, and the resultant chimeric mice were bred with B6 mice. Offspring carrying the modified locus in the germline were interbred to generate the homozygous genetically modified genome. All F1 matings produced normal litter sizes with a Mendelian distribution of the locus. Heterozygous animals were crossed to FLP recombinase expression mice (JAX Stock No. 005703) to remove the FRT site flanked PGK-neo cassette. Mice that no longer contained the FRT flanked PGK-neo cassette were then backcrossed to B6 at least once to remove the FLP recombinase transgene. SNP analysis was performed to validate the B6 background. Offspring that were negative for the FLP recombinase transgene were then interbred and maintained as APOE^ε4/ε4 homozygous mice (JAX Stock No. 000664).

The APOE^ε3 model was generated using CRISPR/Cas9-mediated gene targeting in APOE^ε4 zygotes (Figure 1A). An APOE^ε4 specific sgRNA (GCGGACATGGAGGACGTGCG CGG; PAM site is underlined) was used to target the human knock-in allele just downstream of the valine codon that defines the APOE^ε4 allele, and a 129 nt single stranded oligonucleotide (GAGACGCGGGCACGGCTGTCCAAGGAGCTGCAGGCGG CGCAGGCCCGGCTGGGCGCGGACATGGAGGACGTcTGC GGCCGCCTGGTGCAGTACCGCGGCGAGGTGCAGGCCA TGCTCGGCCA) including the CGC->TGC substitution (bold) to change arginine to cysteine and silent mutation GTG-GTC (underlined) to prevent re-cutting. Putative founders were bred to B6 mice for at least four generations and then interbred to be maintained as APOE^ε3/ε3 homozygous mice (JAX Stock No. 029018).

To create experimental cohorts, APOE^ε3/ε3 mice were crossed to APOE^ε4/ε4 mice to generate APOE^ε3/ε4 mice, which were then intercrossed to give cohorts of litter-matched mice of all required genotypes (APOE^ε3/ε3, APOE^ε3/ε4, and APOE^ε4/ε4). Three separate cohorts of mice were used in this study. The first was aged to 2 mos for strain validation including western blotting. The second was aged to 2 mos for peripheral and cortical assessments. The third was aged to 1 mo, when half were provided a running wheel, and subsequently assessed at 4 mos along with sedentary controls.

Genotyping to differentiate between APOE^ε3 and APOE^ε4 alleles for our experimental cohorts was performed via ear punch at 1 mo of age. A gel-based PCR assay was used to determine insertion of the humanized construct (Figure 1B). Primers included a common forward spanning over intron 1 and exon 2 (AATTTTTCCCTCCGCAGACT), a wild type (WT) mouse reverse in intron 2 (ACAGCTGCTCAGGGCTATTG), and a humanized reverse (AGGAGGTTGAGGTGAGGATG). A band for WT (wild type – mouse control, no humanized insertion) shows at 244 bp, while the humanized insertion results in a 148 bp band. To differentiate between the APOE^ε3 and APOE^ε4 alleles, Sanger sequencing was used to identify CT for APOE^ε3, and GC for APOE^ε4, and a double peak signifying the presence of both APOE^ε3 and APOE^ε4 (GT) (Figure 1C).

Snap-frozen hemispheres were homogenized by hard tissue homogenizer (USA Scientific, Ocala, FL, United States) and lysed in 700 mL RIPA buffer (R0278, Sigma, St. Louis, MO, United States) supplemented with 100× protease and phosphatase inhibitor reagents (1861281, Thermo Fisher Scientific, Waltham, MA, United States). Lysates were incubated for 1 h at 4°C before pelleting insoluble proteins by spinning at 4°C, 11,000 × g for 15 min. Protein concentration was determined by Bradford protein assay (Biorad, Hercules, CA, United States), according to manufacturer’s instructions. Samples were mixed with 10× Laemlli buffer (42556.01, Amsbio, Cambridge, MA, United States), boiled for 10 min, and run on 12% SDS PAGE gels (456-1044, BioRad) with colorimetric ladder (RPN800E, GE, Boston, MA, United States). Gels were transferred to PVDF membranes for immunoblotting and imaging using an iBlot2 dry blotting system (Thermo Fisher). Membranes were blocked in 5% non-fat dry milk in 1XPBS + 0.1% Tween20 for 1 h prior to incubating with primary antibodies diluted in 5% non-fat dry milk in 1XPBS + 0.1% Tween20 for 1 h at room temperature. Membranes were washed in 1XPBS + 0.1% Tween20 before incubating with secondary antibodies diluted in 5% non-fat dry milk in 1XPBS + 0.1% Tween20. HRP-conjugated secondary antibodies targeting primary antibody host IgG were incubated at 1 h at room temperature. Membranes were washed in 1XPBS + 0.1% Tween20 before digital imaging with SuperSignal West Pico PLUS chemiluminescent substrate (34579, Thermo Fisher). Antibodies used include: Pan human APOE (AB947, Millipore, Burlington, MA, United States); Mouse Apoe (NB100-240, Novus Biologicals, Centennial, CO, United States); human APOE^ε4 (NBP1-49529, Novus Biologicals); Actin (ab179467, Abcam, Cambridge, United Kingdom).

For APOE^ε4 quantification, western blots were run as stated above with modifications. Protein extraction was performed on snap frozen cerebral cortex tissue in 700–1000 mL RIPA buffer supplemented with cOmplete Mini Protease Inhibitor Cocktail (Millipore Sigma, 11836153001), 100 uM PMSF, and 100 μM Na3V4O. Protein was quantified by Bradford Protein Assay as stated above. Samples were mixed with 2× Laemmli buffer and RIPA buffer to dilute all proteins to 100 μg/μl/sample. Samples were heated for 5 min at 95°C, and run on 4–20% gradient gels (BioRad, 4561096) with Full Range Rainbow Recombinant Protein Molecular Marker ladder (RPN800E). Gels were transferred to nitrocellulose membranes (Thermo Fisher, IB301001) using the iBlot system (Thermo Fisher) for immunoblotting. Membranes were blocked in 5% non-fat dry milk in 1XPBS + 0.1% Tween20 for 1 h. Primary antibodies were diluted in 5% non-fat dry milk in 1XPBS + 0.1% Tween20 and incubated overnight at 4°C moving. Membranes were washed in 5% non-fat dry milk in 1XPBS + 0.1%Tween20 and then incubated. Secondary antibodies were also diluted in 5% non-fat dry milk in 1XPBS + 0.1% Tween20 for 2 h at room temperature. Membranes were washed again in 5% non-fat dry milk in 1XPBS + 0.1%Tween20 and exposed to Amersham ECL Western Blotting Detection Reagent (RPN2109, Cytiva, Marlborough, MA, United States) just prior to digital imaging on the Azure Biosystems c600 (Azure Biosystems, Dublin, CA, United States). Actin and APOE^ε4 were quantified in ImageJ (FIJI).

At first, mice were group housed (two or three per pen) and were provided access to low profile saucer wheels (Innovive Inc.) 24 h a day from 1 mo to 4 mos. Sedentary mice were not provided access to running wheels. At 4 mos, just prior to harvest, mice were separated and individually housed with a trackable low-profile running wheel (Med Associates Inc.) or no wheel access. Rotations per minute during lights out (12 h, 6:00 pm – 6:00 am) were quantified. This time period was used as it is when mice are naturally awake and active. Running wheel rotations were measured in 1-min bins to allow for distance traveled (sum of rotations) calculated per mouse each night. Average rotations were calculated per mouse. Average speed while active was calculated by isolating the minute intervals where activity was measured (>0) and averaging the number of rotations for the minutes active. Percent of time at each speed was calculated by totaling the number of minute bins that mice ran between 0, 1–30 rotations, 31–70 rotations, 71–100 rotations, and 100+ rotations and dividing by the total amount of minutes tracked. Any nights that had fewer than 700 min tracked were excluded from analysis.

All mice were euthanized by intraperitoneal injection of a lethal dose of Ketamine (100 mg/ml)/Xylazine (20 mg/ml) and blood was collected in K2 EDTA (1.0 mg) microtainer tubes (BD, Franklin Lakes, NJ, United States) through approved cardiac puncture protocols. Mice were perfused intracardially with 1XPBS. Brains were carefully dissected then hemisected sagittally, and the cortex (Ctx) was then carefully isolated and snap frozen in solid CO₂ for RNA-sequencing and cholesterol profiling. Blood was kept at room temperature for at least 30 min to prevent clotting, and then centrifuged at 21°C for 10 min at 5000 rpm. Plasma was carefully collected. Plasma and brain lipid concentrations were characterized on the Beckman Coulter AU680 chemistry analyzer.

RNA sequencing (RNA-seq) was performed by The Jackson Laboratory Genome Technologies Core. RNA extraction involved homogenization with TRIzol (Invitrogen, Waltham, MA, United States) as previously described (Soto et al., 2015). RNA was isolated and purified using the QIAGEN miRNeasy mini extraction kit (Qiagen, Hilden, Germany) in accordance with manufacturer’s instructions. RNA quality was measured via the Bioanalyzer 2100 (Agilent, Santa Clara, CA, United States) and poly(A) RNA-seq sequencing libraries were compiled by TruSeq RNA Sample preparation kit v2 (Illumina, San Diego, CA, United States). Quantification was performed using qPCR (Kapa Biosystems). RNA-seq was performed on the HiSeq 4000 platform (Illumina) for 2 × 100 bp reads for a total of 45 million reads according to the manufacturer’s instructions. Quality control for each sample was completed using NGSQCToolkit v2.3 which removed adaptors and trimmed low quality bases (Phred < 30) (Patel and Jain, 2012). To quantify gene expression of the trimmed reads, we used RSEM v1.2.12 which uses Bowtie2 v2.2.0 for alignment of these reads (Langmead et al., 2009). We used a custom mouse genome including mm-10 based upon the B6 reference genome, with the addition of the human APOE gene for continuity.

Genes were filtered by (1) removing all genes that did not vary in expression (gene count change across all samples was 0) and (2) removing all genes that did not have at least five reads in 50% of the samples. Remaining genes were normalized using DEseq2 (Love et al., 2014). Variance stabilizing transformation (vst) was then performed. Principal component analysis (PCA) identified three outliers from the 2 mo dataset, and five outliers from the 4 mo dataset which were excluded from further analysis. APOE genotype will be referred to as ‘genotype.’ A linear model was used on the normalized counts to identify genes significantly fit by the predictors of our model: sex, genotype, and sex-genotype in the 2 mo dataset, and sex, genotype, activity, and genotype-activity in the 4 mo dataset. Reference data in the model were ‘female APOE^ε3/ε3’ for 2 mo and ‘female sedentary APOE^ε3/ε3, for 4 mo data. The linear model was performed on 18,402 genes for 2 mos and on 18,987 genes for 4 mos. Nominal significant genes were identified per predictor term (sex, genotype, etc.) based upon Pr(>|t|< 0.05). Functional enrichment for each linear model predictor term was performed using the clusterProfiler package with p < 0.10 to determine significant GO terms (Yu et al., 2012). Bonferroni Hochberg (BH) conditions were chosen for pAdjustMethod. Background lists consisted of all genes prior to significant geneset filtering.

Previously published cortical transcriptional data was accessed through the AD Knowledge Portal. Gene expression from Zhao et al. (2020) utilized conditional quantile normalization (CQN)-normalized log2RPKM values. Samples from 3 mo mice were filtered for, and then used for PCA clustering. To match previous publications, no scaling or centering was used in this PCA. Linear modeling was run on 19,120 genes across 32 samples with factors for sex, genotype, and sex-genotype. Significant genes were identified per predictor term (sex, genotype, etc.) based upon Pr(> | t| < 0.05). Functional enrichment for each linear model predictor term was performed using the clusterProfiler package with p < 0.05 to determine significant GO terms (Yu et al., 2012).

Publicly available cell type gene expression was downloaded from the brainrna-seq.org dataset generated by the Barres laboratory (Zhang et al., 2014, 2016). Significant genes from linear model results from 2- and 4- mo data were cross referenced to this dataset. Some genes in our datasets were not found in the brainRNA-seq dataset and were excluded. Number of genes per cell type was expressed as a percentage of a specific cell type divided by all genes that were able to be cross referenced for the linear model predictor terms.

Weighted gene co-expression network analysis required the WGCNA package by Horvath and Langfelder (Langfelder and Horvath, 2008, 2012). Normalized and variance transformed data (vst) without outliers was used to build weight gene co-expression networks. First, all samples passed the function goodSamplesGenes to check for incomplete sample data. Samples were clustered to identify outliers using hierarchical clustering; two samples were excluded from further analysis. Next, the soft-thresholding power (b) was chosen by calculating scale-free topology through the relationship between power and scale independence which resulted in a softPower of 5. Then 18 modules were clustered based upon minModuleSize = 30, and a mergeCutHeight = 0.4. These values were chosen based off previous studies with similar methods to achieve non-redundant smaller modules (Zhao et al., 2020). Module eigengenes for each module were identified and the correlation to sex, genotype, and activity was computed. The p values were then corrected for multiple testing both within and across terms by calculating false discovery rates. Genes in each module were annotated by using the AnnotationDbi package (Hervé Pagès et al., 2020). Module membership (module eigengene correlated with gene expression) was regressed against gene significance (the correlation between the linear model predictor term and each individual gene) to identify highly correlated modules (Pearson correlation). Genes in a module were processed through clusterProfiler to identify functional enrichment.

To quantify neuronal cell number, 4 mo female sedentary mice were evaluated as previous studies would suggest they would be the most susceptible to neuronal loss. Two of each genotype (six samples total) were stained in replicates of three (three brain slices) and then evaluated for NEUN+ DAPI+ cells in the cortex above the CA1 in the hippocampus. Brains were cryosectioned at 20 mm onto slides. Sections were blocked for 1 h in 2%PBT + 10% Normal Goat Serum. Primary antibody NEUN (abcam, ab104225) was diluted in 2% PBT + 10% Normal Goat Serum and kept at 4°C overnight. Slides were washed 3 × 10 min in 2% PBT. Secondary antibody (goat anti rabbit 488) was diluted in 2%PBT and kept for 3 h at room temperature. A representative cortical section was imaged on a Leica SP8 confocal microscope at 1 um step size for 20 um. Images were quantified identically with the spots feature through IMARIS software for NEUN+ and DAPI+ cells.

Statistical analysis for 2 mo biometric data including cholesterol composition, triglyceride, unfasted glucose, and non-esterified fatty acids (NEFA) were calculated by two-way ANOVA (interactions tested: sex-genotype, main effects: sex, genotype), and a post hoc Tukey test in GraphPad Prism v7.0a. Outliers were identified by ROUT (Q = 1.0%) using GraphPad Prism v7.0a. Statistical analysis for 4 mo biometric data was performed by three-way ANOVA (interactions tested: sex-genotype-activity, genotype-activity, sex-activity, sex-genotype, main effects tested: sex, activity, genotype) followed by a TukeyHSD post hoc test in R v1.2.1335. Terms were considered significant if p < 0.05. Post hoc tests were used to determine differences between genotypes per sex, significant effects between genotypes and sexes were not reported (i.e., female-sed-APOE^ε3/ε4 to male-sed-APOE^ε3/ε3). All weight and biometric data included groups of 5–16 mice per sex/genotype/activity. Transcriptional profiling was performed on groups of six mice per sex/genotype/activity. Details of statistical analyses of linear modeling, functional enrichment, and WGCNA data are provided above.

Humanized APOE^ε3/ε3 and APOE^ε4/ε4 mice were created on C57BL/6J (B6) mouse strain using a similar design to that described previously (Supplementary Figure 1A, see the section “Materials and Methods”) (Sullivan et al., 1997; Knouff et al., 1999). Briefly, exons two through four of the mouse Apoe gene were replaced by sequence encoding the equivalent region in human the APOE^ε4 gene through homologous directed repair (HDR) (Figure 1A and Supplementary Figures 1B,C). The APOE^ε3 allele was produced by CRISPR-Cas9 of APOE^ε4/ε4 mice (Figures 1B,C). A PCR-based assay combined with Sanger sequencing was used to confirm humanized APOE alleles (Figures 1B,C). Protein was confirmed by western blotting with human APOE protein present in all humanized APOE mice. APOE^ε4 protein only detectable in APOE^ε3/ε4 and APOE^ε4/ε4 mice in a relative abundance of 0.38 and 0.49, respectively (Figure 1D and Supplementary Figures 2,3).

To evaluate whether APOE^ε4 dosage affected body weight and circulating lipids, two translatable biometric measurements, litter-matched male and female APOE^ε3/ε3, APOE^ε3/ε4 and APOE^ε4/ε4 mice were assayed at 2 mos. As expected, body weight was significantly different between sexes, with males weighing more than females. However, body weight was not affected by APOE genotype (Figure 2A). Total cholesterol, HDL, and unfasted glucose levels in plasma showed a significant main effect of sex, with males showing higher concentrations than females (Figures 2B,D,F). There was a significant interaction between sex and APOE genotype for LDL with the sexes showing differing trends in APOE^ε3/ε4 mice (Figure 2C). Plasma triglyceride levels, and brain total cholesterol, HDL, and LDL concentrations did not show an effect of sex or APOE genotype (Figure 2E and Supplementary Figure 4).

We hypothesized that compared to APOE^ε4/ε4, the APOE^ε3/ε4 genotype may differentially modulate cerebral processes. To assess this, RNA-seq was performed on cortical tissue collected from litter-matched male and female mice at 2 mos. Principal component analysis (PCA) showed clear separation between the sexes. There was no separation based on APOE genotype (Figures 3A,B) which agreed with previous studies (Zhao et al., 2020). Using APOE^ε3/ε3 as the ‘control’ genotype, linear modeling identified genes that varied as a function of sex, APOE genotype, and the interaction between sex and APOE genotype (Figure 3C, see the section “Materials and Methods”). Only 127 genes were in common between the 487 APOE^ε3/ε4 and the 518 APOE^ε4/ε4 genes (Figure 3D). GO terms enriched for the 360 unique APOE^ε3/ε4 genes included ‘regulation of hormone levels,’ ‘regulation of hormone secretion,’ and ‘insulin secretion’ (Figure 3E and Supplementary Figure 7). Hormone signaling, particularly in the glucose/insulin pathway, is disrupted in APOE^ε4 individuals and can increase the severity of dementia (Peila et al., 2002; Irie et al., 2008). No GO terms were enriched for either the 127 intersection genes or the 391 genes unique to APOE^ε4/ε4, suggesting multiple unrelated processes may be more subtly affected. Interestingly, APOE^ε4/ε4 unique genes included angiogenesis-related Acvrl1, Ang, Eng, Vegfc, and cell adhesion-related Fermt3, Itga2b, Itga5. Ang can function by endothelial nuclear translocation to stimulate endothelial angiogenesis, while Vegfc is involved in vascular permeability (Moroianu and Riordan, 1994; Gao and Xu, 2008; Tammela et al., 2008; Miyake et al., 2015; Nagai and Minami, 2015; Apte et al., 2019). Both Itga2b and Itga5 are receptors for fibrinogen and fibronectin, involved in the blood clotting cascade, while Fermt3 is involved in leukocyte adhesion to endothelial cells (Varner et al., 1995; Oksala et al., 2015; Botero et al., 2020). RNA in situ hybridization (Allen Brain Atlas) and single cell sequencing of vascular-related cells confirm both Acvrl1 (activin A receptor like Type 1) and Eng (endoglin) are expressed by vascular cells, including endothelial cells (Supplementary Figures 5, 6; Lein et al., 2007; Vanlandewijck et al., 2018).

Cell type analysis revealed that while the APOE^ε3/ε4 unique genes are expressed in multiple cell types, the majority of these genes are expressed by endothelial cells and astrocytes. In contrast, the majority of APOE^ε4/ε4 unique genes are expressed by neurons (Figure 3F). This further supports subtle but unique effects of APOE genotype to the cerebral cortex at a young age that may predispose for AD and dementia later in life.

Physical activity, especially running, has been widely considered a prevention, reducing risk for age-dependent cognitive decline and dementia. However, exercise may not prevent dementia in all individuals. We hypothesized that this may be due in part to genetics. To begin to test this we determined whether the effects of exercise are influenced by APOE genotype in young mice. Litter-matched male and female APOE^ε3/ε3, APOE^ε3/ε4 and APOE^ε4/ε4 mice were given access to a voluntary running wheel for 12 weeks from 1 mo to 4 mos (Figure 4A, see the section “Materials and Methods”). There was expected variation between mice for amount of time spent running and there were no differences in average rotations or running speed between APOE genotypes for both sexes, indicating activity may be compared across genotypes as it is not a confound (Supplementary Figures 8A–F). As has been shown before, female mice tended to run farther than male mice (Foley et al., 2019). As expected, there was a significant difference in weight between sexes at 4 mos, but there were no differences between the APOE genotypes or running and sedentary groups for either sex (Figures 4B–D).

Apolipoprotein E genotype did not significantly influence total cholesterol, LDL or HDL which agrees with previous findings in other APOE mouse models (Knouff et al., 1999; Mann et al., 2004). There was also no effect of APOE genotype on unfasted glucose, or NEFA concentrations in the plasma (Figures 4E–G,I,J). However, there was an effect of sex for total cholesterol, LDL, HDL and unfasted glucose levels, with males having higher total cholesterol, HDL, and unfasted glucose and lower LDL than females (Figures 4E–G,I). There was also an interaction between sex and activity for total cholesterol and HDL, suggesting the effect of activity on plasma cholesterol may be sex-specific at a young age (Figures 4E,G and Supplementary Table 2). There was a significant sex-activity-genotype interaction on triglyceride concentration (Figure 4H). Across total cholesterol and HDL concentration, there was an interesting sex-activity interaction, with activity appearing to have differential effects per sex. In total cholesterol and HDL there was an overall decrease in lipid concentration with running in males, however in females, there was an overall increase in these lipids. These results suggest that activity may have different effects on circulating lipid concentrations due to sex-based differences. Together, the effects of exercise on cholesterol composition were subtle and not influenced by APOE genotype, whereas triglyceride concentrations were influenced by sex, activity, and APOE genotype.

To determine whether the effects of APOE^ε4 on the brain are modified by running in an APOE genotype-specific manner, cortical transcriptional profiling and linear modeling were performed (see the section “Materials and Methods”). Genes significantly associated with sex, APOE^ε3/ε4, APOE^ε4/ε4, activity and the interactions were identified (Figures 5A–C and Supplementary Figures 8C,D; Foley et al., 2019). The largest number of genes were significant for sex which enriched for ‘response to interferon beta,’ ‘response to virus,’ and ‘innate immune response,’ including known AD and dementia risk genes Clu, and Plgc2, suggesting a difference in immune function between the sexes (Supplementary Figure 10). There was no enrichment for the 540 genes significant for activity, suggesting multiple processes affected. Of the 540 genes, Cx3cr1 and Csf1r were significant and have been previously implicated in maintaining microglia homeostasis, suggesting microglia response may be affected by activity across APOE genotypes (Keren-Shaul et al., 2017).

Of the 327 genes significant for APOE^ε3/ε4 and the 448 genes significant for APOE^ε4/ε4 only 74 intersected (22% from APOE^ε3/ε4 genes, 17% from APOE^ε4/ε4 genes) (Figure 5D) which was similar to our findings at 2 mos. Of the 253 genes unique for APOE^ε3/ε4 63 genes were downregulated (negative β value) and 190 were upregulated (β value). The 253 APOE^ε3/ε4 unique genes enriched for GO terms such as ‘extracellular matrix,’ ‘collagen-containing extracellular matrix,’ and ‘wound healing’ (Figure 5E and Supplementary Figure 12A). The 374 APOE^ε4/ε4 unique genes enriched for ‘branching morphogenesis of an epithelial tube,’ ‘chemeotaxis’ and ‘taxis,’ including genes such as Gli3, Cx3cr1, and Ccr1 (Figure 5C and Supplementary Figure 11A). Of the 374 genes there were 183 downregulated and 191 genes upregulated, indicated by a negative and a positive β-value, respectively. The 74 intersecting genes enriched for ‘MHC protein complex’ and ‘BMP receptor binding’ including genes associated with these GO terms included Bmp5, Bmp6, and Spp1 (Supplementary Figure 11B). BMP (bone morphogenetic protein) signaling has been previously implicated in vascular development and more importantly, abnormal BMP protein expression can lead to vascular dysfunction, suggesting another pathway by which the APOE^ε4 allele influences AD and dementia pathologies (Garcia de Vinuesa et al., 2016). Of these 74 intersecting genes, 19 genes were downregulated and 55 genes upregulated, although the degree of regulation (β value) was different.

While many cell types can contribute to the extracellular matrix (ECM)- and coagulation-related pathways, endothelial cells contributed to 33% of the unique genes that were in the APOE^ε3/ε4 dataset (Supplementary Figure 12B). The cell types enriched for intersection genes included both oligodendrocyte precursor cells, as well as microglia (Supplementary Figure 12B). No specific cell types appeared to be responsible for APOE^ε4/ε4 unique genes. While there is no single cell type contributing to the APOE^ε4/ε4-dependent transcriptional differences, endothelial cells may be critical drivers of transcriptional differences in APOE^ε3/ε4 mice (Supplementary Figure 12B). Interestingly, our data for APOE^ε4/ε4-specific changes only minimally overlapped with previous studies (Supplementary Figure 13), differences likely due to dissimilarities in study design and environment (Zhao et al., 2020). The APOE^ε3/ε4 genotype was not studied previously (Zhao et al., 2020).

In examining genes unique to APOE^ε3/ε4 genotype and expressed by endothelial cells we observed several annexin genes. Both Anxa1 and Anxa2 contributed to the enrichment terms ‘extracellular matrix,’ ‘response to wound healing,’ and ‘coagulation.’ Anxa2 has direct effects on the plasminogen/fibrin clotting response (Supplementary Figure 12C; Ling et al., 2004). While we expected a linear increase in gene expression with increasing APOE^ε4 gene dose, for endothelial-expressed Anxa1, Anxa2, and Anxa3, there was a significant increase in the APOE^ε3/ε4 genotype when compared to APOE^ε3/ε3, but not a significant increase in the APOE^ε4/ε4 genotype when compared to APOE^ε3/ε3. This pattern suggests that expression of some genes was differentially affected by the combination of the APOE^ε3 and APOE^ε4 allele.

To determine whether exercise effects the cortex in an APOE genotype-specific manner, we compared the genes for activity-APOE^ε3/ε4 with activity-APOE^ε4/ε4. Only 17% of activity-APOE^ε3/ε4, and 12% of activity-APOE^ε4/ε4 genes overlapped (Supplementary Figure 14A). Three intersection genes enriched for ‘anoikis’ (consisting of Src, Tfdp1 and Bmf), suggesting a common cell death response between genotypes in response to activity (Supplementary Figures 14B,C). Together, these data show that exercise and APOE genotype do not appear to overtly interact on a transcriptional level in 4 mo mice.

While linear modeling indicated subtle changes at the gene level, to ensure these changes were robust we employed a second approach, weighted gene co-expression network analysis (WGCNA) on the same 4 mo dataset. We used WGCNA to create modules of groups of genes across the entire dataset, thus reducing complexity of the large number of input genes to produce biologically relevant modules. This method differs from linear modeling by grouping genes into modules, including genes that may not be significant on the gene level, but still contribute to the coordination of a significant biological process. Nineteen modules were identified, and two modules were significantly correlated with sex (grey60, gray), two modules significantly correlated for genotype (yellow and greenyellow), and six modules significantly correlated for activity (cyan, lightgreen, tan, brown, pink, and red) (Figure 6A, see the section “Materials and Methods”). No modules were significant across multiple terms. Module eigengenes were calculated and dissimilarity plotted to visualize the relationship between modules and terms (Figure 6B). Both the yellow and greenyellow modules were significantly correlated to APOE genotype (Figures 6C,D). The yellow module showed no enrichment, however the genes in the greenyellow module enriched for ‘extracellular matrix,’ ‘embryonic morphogenesis,’ and ‘collagen-containing extracellular matrix’ and showed distinct differences between APOE^ε3/ε3 and both APOE^ε3/ε4 and APOE^ε4/ε4 (Figure 6C and Supplementary Figure 14D). Confirming our previous findings, these data suggest that APOE genotype affects the extracellular matrix, collagen, and morphogenesis (Figure 4).

The cyan and tan modules were significantly correlated for activity, and cyan enriched for ‘ensheathment of neurons,’ ‘myelination,’ and ‘gliogenesis,’ while the tan module enriched for ‘skeletal muscle cell differentiation,’ and ‘skeletal muscle organ development’ (Supplementary Figures 15A–D). However, no modules were significant for both APOE genotype and activity. These data support our previous findings by linear modeling that suggest the effects of exercise on cortical gene expression are not significantly impacted by APOE genotype, at least at this young age.