Novel differentially expressed genes and multiple biological pathways for Alzheimer's disease identified in brain tissue from African American donors

Abstract

INTRODUCTION

Few African American (AA) donors have been included in post mortem Alzheimer's disease (AD) studies compared to European‐ancestry (EA) individuals.

METHODS

We generated transcriptome‐wide bulk pre‐frontal cortex (PFC) gene expression data from 125 AA donors with neuropathologically determined AD and 82 AA controls.

RESULTS

Transcriptome‐wide significant differential expression was observed with 482 genes. The most significant, ADAMTS2, showed 1.52 times higher expression in AD cases (p = 2.96x10⁻⁸). Comparison of findings with those from a recent gene expression study of EA brain donors revealed substantial concordance, including ADAMTS2. Other associations not observed in EA results may be especially relevant to AD risk in the AA population. Examination of AA AD GWAS‐implicated variants identified several expression quantitative trait loci.

CONCLUSION

This first large‐scale AA brain AD gene expression study identified many differentially expressed genes, including ADAMTS2, and supports gene expression as a molecular pathway underlying the impact of several AA AD risk variants.

Highlights

• We performed the largest African American brain tissue Alzheimer's disease (AD) gene expression study.

• Expression differences for 482 genes, notably ADAMTS2, were study‐wide significant.

• Many significant differentially expressed genes are involved in energy metabolism.

• Several previously known AD‐associated variants in African Americans are eQTLs.

• These results advance knowledge of the genetic basis of AD in the AA population.

1. BACKGROUND

The prevalence of Alzheimer's disease (AD) is about two times higher in the Black/African American (AA) population compared to White/European‐ancestry (EA) individuals living in the United States. ¹ Some of this difference is due to social determinants of health such as disparities in health care access and quality of education, ² , ³ biases in testing, ⁴ , ⁵ and higher rates of AD risk factors such as cardiovascular disease ⁶ and diabetes ⁷ in those who identify as AA, although the mechanisms underlying these risk factors and contribution to differential rates of AD are not completely known. ⁸ Group differences have also been noted in the pattern of association with AD risk variants. ⁹ For example, apolipoprotein E (APOE) ε4 is more frequent in the AA population, but the reported AD risk associated with APOE ε4 heterozygosity and especially homozygosity is smaller in AA compared to EA cohorts. ⁹

Much of our understanding of the genetic basis of AD in AA individuals is derived from candidate gene and genome‐wide association studies (GWAS), although with much smaller sample sizes than those included in genetic studies of EA cohorts. ⁹ For example, a recent large AD GWAS in EA cohorts including approximately 85,000 AD cases and 296,000 controls as well as information about proxy AD (reported parental dementia) cases identified genome‐wide significant associations with 75+ loci. ¹⁰ By comparison, the largest GWAS of AD and AD‐related dementia (ADRD) in AA cohorts ¹¹ included 4012 ADRD cases, 18,435 controls, and 52,611 AD proxy cases and controls from the Million Veteran Program ¹² and an additional 2784 AD cases and 5222 controls assembled by the Alzheimer's Disease Genetics Consortium (ADGC). ¹³ That study confirmed associations previously established in EA studies (i.e., APOE, ABCA7, CD2AP, and TREM2) and identified highly significant novel associations with variants in ROBO1 and RP11‐340A13.2. In addition, associations have been identified with rare pathogenic variants in AA cohorts that are rare or absent in EA populations, suggesting that some AD‐related mechanisms may be more easily discernable in particular ancestry groups. ⁹

Although many studies have examined differential gene expression (DGE) in brain tissue from AD cases and controls in EA or mixed ancestry cohorts using bulk tissue, single‐cell sequencing, and individual cell types (e.g., ¹⁴ , ¹⁵ , ¹⁶ ), the number of AA individuals in these studies was unspecified or too small to identify significant findings within this group alone. In this study, we evaluated RNAseq data derived from post mortem prefrontal cortex (PFC) tissue from 207 AA brain donors (125 pathologically confirmed AD cases and 82 controls). The most significant differentially expressed gene, ADAMTS2, was also one of the top‐ranked genes in a concurrent study comparing gene expression in brain tissue from EA AD cases and controls, ¹⁷ and the most significant differentially expressed gene between the pathologically confirmed AD cases with and without clinical AD symptoms prior to death. ¹⁸ We also found evidence suggesting modulation of gene expression as a potential mechanism for some AD risk variants implicated in a recent AA AD/dementia GWAS. ¹¹

2. METHODS

2.1. Participants, specimens, and diagnostic procedures

We obtained frozen brain tissue specimens from 229 donors identified as AA individuals from 13 Alzheimer's Disease Research Center (ADRC) brain banks: (1) Boston University ADRC (BU), (2) Goizueta ADRC at Emory University (GADRC), (3) Johns Hopkins ADRC (JH), (4) University of Kentucky Alzheimer's Disease Center Tissue Bank (UKY), (5) Florida Brain Bank and ADRC at the Mayo Clinic (Jacksonville, FL), (6) Rush Alzheimer's Disease Core Center (RUSH), (7) University of Pittsburgh ADRC (PITT), (8) Carroll A. Campbell, Jr. Neuropathology Laboratory at the Medical University of Southern California (USC), (9) The Charles F. and Joanne Knight ADRC at Washington University in Saint Louis (WASHU), (10) Massachusetts ADRC (MGH), (11) University of California‐Davis ADRC (UCD), (12) Bryan Brain Bank and Biorepository at the Duke‐UNC ADRC (DUKE), and (13) the Northwestern ADC Neuropathology Core (NW). In addition, we received previously generated RNAseq data for 28 donors (18 cases and 10 controls) from the Columbia University ADRC (CU). Scores measuring the severity of neurofibrillary tangle involvement (Braak) and neuritic amyloid deposition (Consortium to Establish a Registry for Alzheimer's Disease [CERAD]) scores obtained from the brain banks were harmonized using the ADRC data dictionary to the standard 0–6 coding for Braak and 0–3 coding for CERAD. We used National Institute on Aging—Alzheimer Association guidelines ¹⁹ and standard NIA Reagan scoring criteria to classify the donors as AD cases (intermediate likelihood or high likelihood of AD) and controls (not AD or low likelihood of AD). This study was approved by the Boston University Institutional Review Board. Informed consent obtained from the brain donors covers the research performed in this study using their post mortem tissue.

2.2. RNA sequencing data generation and quality control

RNA was extracted from the PFC tissue using the Maxwell RSC simplyRNA Tissue Kit from Promega according to the manufacturer's instructions. RNA integrity number (RIN) values were assessed using Agilent 2100 Bioanalyzer RNA Chip (Santa Clara). RNAseq was performed on samples with RIN scores > 2 by the Yale Center for Genome Analysis in two batches of 90 and 99 samples, respectively. RNAseq data were successfully generated for all but one sample. Initial QC was performed using FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Adapter trimming and removal of low‐quality reads was performed using Trimmomatic ²⁰ based on the Trueseq 2 Universal Adapter reference sequence. BAM files containing paired‐end RNAseq read data were obtained for the CU participants. Generation of RNAseq data for the CU donors is described in detail elsewhere. ²¹ Reads were extracted using BEDTools ²² and aligned to the genome jointly with the other cohorts as follows. Paired‐end reads were aligned to the hg38 human reference genome using STAR. ²³ Quantification at the gene and transcript level was performed using RSEM. ²⁴ Post‐alignment QC was performed using RSeQC, ²⁵ and aggregation of the QC across samples was performed using MultiQC. ²⁶ Two samples with mapping rates < 50% were excluded from subsequent analysis. All of the remaining samples had mapping rates > 80%. A principal component (PC) analysis of the rLog samples using R did not reveal any outliers (> 6 SD), but examination of the PCs indicated clustering by batch (Batch 1, Batch 2, CU); hence, batch was included as a covariate in the differential expression analyses. Sex concordance was checked by examining the total number of reads mapped to four Y‐chromosome genes: KDM5D, DDX3Y, USP9Y, and UTY. We identified one mismatch, which was excluded. Cell‐type estimates for use as covariates were generated using the DeTREM method ²⁷ with single‐nuclei reference data from a PFC AD study. ²⁸ The estimated cell proportions in AD cases and controls are shown in Table S1.

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using traditional (e.g., PubMed) as well as preprinted (e.g., medRxiv) sources on Alzheimer's disease (AD), focusing on genetic associations in African American (AA) ancestry and mechanisms of cognitive resilience in individuals with AD pathology.

2. Interpretation: We showed that 482 genes are differentially expressed in the pre‐frontal cortex of AA brain donors with pathologically confirmed AD compared to those without neurodegenerative disease. These include ADAMTS2, a finding that is supported by a recent differential gene expression study of cognitive resilience in EA brain donors comparing pathologically confirmed AD cases with and without clinical symptoms prior to death.

3. Future directions: These results suggest genes and pathways as potential intervention targets. ADAMTS2 was previously linked to AD pathology and many of the identified differentially expressed genes are involved in mitochondrial energy production. Additionally, this study highlights the need to include large ancestrally diverse cohorts in AD research.

2.3. Genotype data

DNA was extracted from brain tissue using Maxwell RSC Blood DNA capture kits l. Samples were sent to The Children's Hospital of Philadelphia Center for Applied Genomics for genotyping using the Illumina Global Diversity Array (∼1.8 million variants). Post‐genotyping QC was performed using plink v1.90b3.36 and v2.00a2.3. ²⁹ Two samples were excluded for high missing rates (> 5%). X‐Chromosome heterozygosity was examined to identify potential sex mismatches. Two samples were removed due to discordant sex. Imputation of variants not on the array was performed using the Michigan Imputation Server (https://imputationserver.sph.umich.edu/) based on 1000 Genomes phase 3v5 reference data. ³⁰ Genotype calls were made with an 80% certainty threshold. After cleaning, genotype data, valid RNAseq data, and non‐missing covariates were available for 177 donors. PC analysis was performed using Plink to identify potential ancestry outliers in the autopsy cohort based on comparison with multi‐ancestry GWAS data from the ADGC. This analysis included 100K randomly selected non‐ambiguously genotyped single nucleotide polymorphisms (SNPs) with minor allele frequency (MAF) > 1% and missing rate < 1%.

2.4. Differential gene expression analysis

DGE analysis was performed using DEseq2 ³¹ for genes with ≥ 1 read in half of the sequenced cohort as previously suggested. ³² A Benjamini Hochberg false discovery rate (FDR) corrected p‐value (p _adj, often called a q‐value) ³³ was computed to adjust for 33,611 genes examined. Regression models included covariates for age at death, sex, batch, RIN, and the estimated proportions of inhibitory neurons, excitatory neurons, microglia, oligodendrocytes, astrocytes, and endothelial cells. Post mortem interval (PMI) was not included as a covariate because this information was missing for 20 samples, however, we did perform sensitivity analyses to confirm that excluding PMI did not substantially bias the results. Volcano plots were generated using the EnhancedVolcano R package (https://github.com/kevinblighe/EnhancedVolcano). Expression differences are reported as log₂ fold change (L2FC). We also examined differential expression of APOE and other genes implicated in EA AD GWAS, ¹⁰ AA dementia GWAS, ¹¹ and in other AA AD genetic studies. ⁹ DGE results from the AA cohort were then compared to those obtained from frontal cortex tissue of 526 autopsy‐confirmed AD cases and 456 controls from four EA cohorts ¹⁷ using a χ² test.

2.5. Over‐representation and gene network analyses

Over‐representation analysis was conducted to identify gene ontology (GO) terms (biological processes and pathways) that are enriched for the top‐ranked DEGs using Goseq, ³⁴ a method that adjusts for any differential likelihood of significance due to gene size. To increase interpretability and limit multiple testing, we only tested for overrepresentation within GO Biological Process and GO Molecular Function categories. Gene network analysis was performed using Weighted Gene Co‐Expression Network Analysis (WGCNA). ³⁵ Because WGCNA does not allow for covariates, prior to performing gene network analyses, RNA‐seq data were corrected using COMBAT‐SEQ ³⁶ to adjust for batch effects but preserve age, sex, and AD effects. The batch‐corrected counts were normalized and log transformed using the DEseq2 ³¹ package, and networks of correlated genes were identified by analysis of the resulting rLog values using WGCNA. This analysis was performed using the blockwiseModules function in the WGCNA R package to generate signed networks with power = 15 and options cut height = 0.75, min module of 30, max block of 4000, reassign threshold = 0, mergeCutHieght = 0.25, and the rest of the options left at the default. Only highly expressed genes (> 10 reads in half of the cohort) were included in this analysis to reduce computational complexity.

2.6. eQTL analysis of AD‐associated variants

We assessed by linear regression potential regulatory effects of variants previously associated with AD in AA cohorts ¹¹ with MAF > 5% and associated on expression of genes within 200 kb of the variant according to the ensemble v 75 (GRCH37) database accessed via the R biomaRt library. The model included covariates for age at death, sex, batch, RIN, and the estimated cell‐type proportions. FDR‐corrected p values were calculated for each variant adjusting for the number of genes examined. We computed the association between gene expression and all of the MAF > 5% SNPs within 200 kb of the top‐ranked findings in order to compare the associations findings from AA ADRD GWAS ¹¹ with the strength of the eQTL effect on genes under the association peaks. These comparisons were visualized using R, LocusZoom (https://my.locuszoom.org/), and the UCSC Genome Browser (http://genome.ucsc.edu/).

3. RESULTS

3.1. Demographics

After excluding 34 samples that were obtained from non‐PFC regions or had RNA quality or sequencing issues, and 1 sex mismatched sample, 212 participants (including 82 controls, 125 AD cases, and 5 subjects lacking sufficient neuropathological data to determine AD status) remained for subsequent analyses (Table 1). Participants with undetermined neuropathological AD status were excluded from the DGE analysis but included in eQTL analysis. The sex ratio is approximately 1 among the controls, but a higher proportion of cases are women. The mean age at death was 5.3 years greater in AD cases than controls (p = 0.0040). PC analysis of genotyped subjects in the AA autopsy cohort in conjunction with ADGC GWAS data indicated that the population structure of the AA cohort is concordant with the ADGC AA cluster (Figure S1).

TABLE 1.

Subject characteristics.

	AD cases				Controls				AD status unknown				Total
Available data	N	% Women	% ε4 carriers	Mean age (SD)	N	% Women	% ε4 carriers	Mean age (SD)	N	% Women	% ε4 carriers	Mean age (SD)	N	% Women	% ε4 carriers	Mean age (SD)
RNA‐seq only	125	61.6	NA	83.0 (9.9)	82	52.4	NA	77.7 (14.4)	0	–	–	–	207	58.0	NA	80.9 (12.2)
RNA‐seq + SNP array	104	61.5	71.2	82.4 (10.2)	68	51.5	33.8	76.6 (15.1)	5	20.0	60.0	61.6 (26.3)	177	56.5	56.5	79.5 (13.5)

Abbreviations: AD, Alzheimer's disease; APOE, apolipoprotein E; NA, APOE genotype information not available for subjects missing SNP array data; SNP, single nucleotide polymorphism.

3.2. Genes differentially expressed between AD cases and controls

Transcriptome‐wide significant (TWS) differences were observed for 482 of the 33,611 sufficiently expressed genes, among which 174 had higher and 308 had lower expression in AD cases than controls (Figure 1, Table S2). The most significant DEG, ADAMTS2, was expressed 1.52 times higher in AD cases than controls (L2FC = 0.60, p = 2.96x10⁻⁸, Table 2, Figure S2). Sensitivity analyses showed that excluding PMI as a covariate and excluding younger (age < 60 at time of death) donors did not substantively impact the results (Figure S3, Table S3). By comparison, none of the AD loci established by the largest GWAS in EA and AA cohorts ¹⁰ , ¹¹ were differentially expressed at a level that would survive a multiple test correction (Table S4). The most significant differentially expressed gene among the previously known AD loci, IG1FR, was implicated in an AA AD GWAS ¹³ due to a rare variant association (rs570487962), had higher expression in AD cases (L2FC = 0.11, p = 0.0013). 

FIGURE 1.

Volcano plot of a transcriptome‐wide differential gene expression analysis of 125 neuropathologically determined African American AD cases and 82 neuropathologically determined African American controls. Color‐coding denotes genes that passed multiple testing corrected significance thresholds (blue dots), L2FC > |0.5| (green dots), or both (red dots). AD, Alzheimer's disease.

TABLE 2.

Twenty genes most significantly differentially expressed between African American AD cases and controls.

Gene	L2FC	p	p adj
ADAMTS2	0.60	2.96E‐08	0.0010
AP002360.1	−0.35	8.71E‐08	0.0012
EFR3B	−0.25	1.11E‐07	0.0012
IRS4	0.47	2.18E‐07	0.0014
CA12	−0.78	2.33E‐07	0.0014
ITPKB	0.31	2.70E‐07	0.0014
PDE10A	−0.47	2.89E‐07	0.0014
TDRKH	−0.17	9.62E‐07	0.0040
LINC01182	−0.64	1.13E‐06	0.0040
TFCP2	−0.17	1.31E‐06	0.0040
AL158071.3	0.52	1.34E‐06	0.0040
EHHADH	−0.30	1.42E‐06	0.0040
PIEZO2	0.41	1.84E‐06	0.0047
P2RY1	−0.43	1.97E‐06	0.0047
AC018647.2	−0.25	2.64E‐06	0.0059
NOX4	−0.42	3.02E‐06	0.0062
ECHS1	−0.18	3.21E‐06	0.0062
ZCCHC14	0.14	3.38E‐06	0.0062
LINC01094	0.43	3.48E‐06	0.0062
AP001180.1	1.35	3.89E‐06	0.0063

Abbreviations: AD, Alzheimer's disease; L2FC, log₂ fold change.

Four of the most significant DEGs in our AA autopsy cohort (ADAMTS2, ITPKB, TDRKH, and LINC0194) were also differentially expressed between AD cases and controls at a TWS level in a large EA sample. ¹⁷ Additionally, the most significant DEG in the EA study, EMP3, was also significant in the AA cohort with a similar effect direction (L2FC = 0.33, p = 0.00037). The overlap in TWS associations was more than would be expected by chance (p < 2.2x10⁻¹⁶), with 6% of TWS DEGs in the EA sample also TWS in the AA cohort (65/1092) versus 1.4% TWS significant in the overall AA cohort. Importantly, the direction of effect is the same for the 65 DEGs that are TWS in both studies.

3.3. Enrichment of DEGs involved in mitochondrial function, phosphorylation, DNA binding, and transcription

Enrichment analysis of the differentially expressed genes yielded 71 significant Enriched GO terms, especially those involving oxidative reduction processes and mitochondrial energy generation. All of the top 10 GO terms include 11 nuclear‐encoded NDUF (NADH:ubiquinone oxidoreductase core subunit) genes (NDUFA1, NDUFA2, NDUFA3, NDUFA6, NDUFA7, NDUFB1, NDUFB2, NDUFB3, NDUFB8, NDUFS3, and NDUFS5) and cytochrome c oxidase genes (COX15, COX4I1, COX6B1) which were all expressed less in AD cases compared controls (Table S5). Because NADH‐ and COX‐gene encoded molecules form parts of the mitochondrial respiratory chain, we also examined but did not find differential expression of 16 mitochondrial DNA genes (p > 0.07). Gene network analysis identified 27 networks ranging in size from 44 to 2944 genes. Eigengenes (PCs representing gene activity) for three networks (light yellow, magenta, salmon) remained significant after correction for multiple testing (Table S5). The light yellow network included 111 genes and was enriched for those involved in regulation of developmental processes or phosphorylation. Expression for most (7/9) of the AD‐associated (p < 0.05) genes in this network was higher in AD cases than controls. The magenta network was enriched for genes involved in DNA binding, transcription regulation, and cilium organization. Salmon was the most significantly associated network with AD (p = 5.19x10⁻⁶, p _adj = 0.00014) and was enriched for genes related to protein complexes, oxidative phosphorylation, and metabolic processes. 

3.4. eQTLs among AD‐associated variants identified by GWAS in AA cohorts

SNP association and gene expression data were available for 177 AA brain donors. Ten of 12 SNPs with a MAF > 0.05 which were significantly associated with AD/ADRD in the large AA GWAS ¹¹ were located within 200 kb of at least one gene (range 1–15, mean = 5.5). After correction for multiple testing, significant eQTL associations were observed with five variants (Table 3). The minor alleles of two highly correlated SNPs (rs2234253 and rs73427293) in the TREM2/TREML2 region (D’ = 1, R ² = 0.99 in the 1000 Genomes AFR reference data) were associated with higher expression of ADCY10P1 (p = 0.0048 and 0.0080, respectively) and TREML2 (p = 0.0038 and 0.0044, respectively). Of note, rs2234253 is a missense mutation in TREM2 (T96M) that is relatively common (MAF = 0.046) in the AA population. Although the frequency of this variant is rare in EA individuals (MAF = 0.00083), it has been associated with frontotemporal dementia in a German cohort. ³⁷ The minor allele of CD2AP SNP rs7738720 was associated with reduced CD2AP expression (p = 0.026). The minor allele of MSRA intronic SNP rs4607615 was associated with expression of MSRA (p = 0.026), RP1L1 (p = 0.041), and most strongly with PRSS51 (p = 6.11x10⁻⁷) even though rs4607615 is 61 kb downstream of that gene. The minor allele APOE SNP rs429358, which encodes the ɛ4 isoform, was associated with reduced expression of APOE (p = 5.61x10⁻⁴) and adjacent genes APOC2 and CLPTM1 (p = 0.017 for both). More detailed comparisons of the AA ADRD GWAS findings in each region excluding APOE ¹¹ with the strength of effect of each SNP on expression of genes (i.e., eQTLs) underlying the association peaks are shown in Figures S4–S6.

TABLE 3.

Association of AA GWAS‐implicated variants and expression of nearby (< 200 kb) genes.

SNP	Gene	β a	p	p cor
rs2234253	ADCY10P1	0.11	0.0048	0.026
	TREML2	0.18	0.0038	0.026
rs73427293	ADCY10P1	0.11	0.0080	0.044
	TREML2	0.17	0.0044	0.044
rs7738720	CD2AP	−0.09	0.013	0.026
rs4607615	MSRA	−0.04	0.026	0.041
	RP1L1	0.10	0.031	0.041
	PRSS51	−0.20	1.53E‐07	6.11E‐07
rs429358	APOE	−0.13	5.61E‐04	8.42E‐03
	APOC2	−0.15	0.0033	0.017
	CLPTM1	−0.06	0.0023	0.017

^a Effect of each minor allele on gene expression.

Abbreviations: AA, African American; GWAS, genome‐wide association studies; SNP, single nucleotide polymorphism.

We examined the Genotype‐Tissue Expression (GTEx) Portal (https://www.gtexportal.org/ accessed April 02, 2025) for evidence that the associations observed in our AA sample are present in an independent largely EA cohort. We confirmed most of these variants as eSNPs in the cortex or frontal cortex at a nominal significance level (p < 0.05). Although results for TREML2 were not available, the two TREML2 variants in high LD, rs2234253 and rs73427293, are eSNPs for ADCY10P1 (p = 6.38x10⁻⁴ and 5.99x10⁻³, respectively), even though these variants are uncommon in EA populations. Rs4607615 was associated with expression of MSRA (p = 0.04) and PRSS51 (1.54x10⁻³) in the frontal cortex. Notably, the PRSS51 association was more significant in our smaller AA autopsy cohort (p = 6.11x10⁻⁷), perhaps indicating a larger effect in AAs or AD cases. The CD2AP SNP rs7738720 was nominally associated with reduced CD2AP expression (p = 0.0030) in the nucleus accumbens, but not the frontal cortex. However, this may be due to the small number of minor alleles (6) observed in the frontal cortex data.

4. DISCUSSION

We performed the largest to date transcriptome‐wide study of AD using brain tissue from AA donors. This is an important step in deciphering the genetic architecture and underlying mechanisms of AD risk in this population, in light of evidence that nearly all of the established AD risk variants are population‐specific or have divergent frequencies across populations. ⁹ , ¹¹ , ¹³ Additionally, because exposure to some AD risk factors differ between White and Black Americans including those that disproportionally affect the Black population in the United States (e.g., vascular disease, diabetes, low education, and others linked to diet, income and occupation), ⁸ , ⁹ the degree to which some biological pathways lead to AD likely varies across populations. In spite of such expected differences, we found high correspondence of differential expression of many genes in the pre‐frontal cortex from AA and EA brain donors.

ADAMTS2 was the most significant DEG in our study and among the top‐ranked DEGs assessed in PFC tissue in a very recent study of 982 EA brain donors (p = 5.1x10⁻¹³) ¹⁸ with the same effect direction in both populations. In addition, the study by Li et al. ¹⁸ found that ADAMTS2 was the most significant DEG between AD cases who had an ante mortem clinical AD diagnosis versus AD cases who were cognitively healthy prior to death (i.e., cognitively resilient). Taken together, these findings suggest that lower ADAMTS2 expression reduces the risk of cognitive impairment among those who develop hallmark AD pathology.

ADAMTS2 is a member of the disintegrin and metalloproteinase with thrombospondin motifs (ADAMTS) family of genes that perform a diverse set of functions. ³⁸ The most well studied role of ADAMTS2 relates to its proteolytically processed form resulting in a procollagen N‐proteinase. In humans, ADAMTS2 mutations cause a recessive form of Ehlers‐Danlos syndrome. ³⁹ Other studies have implicated ADAMTS2 in neurological processes related to AD risk. Yamakage et al. demonstrated in ADAMTS2 knockout mice that ADAMTS2 induces cleavage of Reelin in the PFC and hippocampus, impairing its function. ⁴⁰ Mouse studies suggest Reelin is protective and found that its level is negatively associated with tau phosphorylation and amyloid plaque formation. ⁴¹ , ⁴² A recent study of resilience in a 67‐year‐old cognitively intact individual carrying a highly penetrant PSEN1 mutation that causes autosomal dominant early‐onset AD identified a putatively protective gain of function variant (H3447R) in RELN (the gene encoding reelin), which contributed to reduced Tau phosphorylation in mouse AD models. ⁴³ Results of human studies of Reelin levels and RELN expression in brain tissue are less consistent, showing increased, ⁴¹ , ⁴⁴ reduced, ⁴² , ⁴⁵ or no difference ⁴⁶ in reelin levels and RELN expression in AD versus control tissue. In this cohort, we did not observe an association of RELN expression with AD in PFC tissue (p = 0.35). Nevertheless, our observation of higher expression of ADAMTS2 in PFC from AD cases and its role in the Reelin pathway supports the hypothesis that ADAMTS2 inhibition might be an AD therapeutic strategy. ⁴⁰

Four of the other six most significant (p < 3.0x10⁻⁷) DEGs (ITPKB, IRS4, CA12, and PDE10A) were previously implicated in AD‐related processes noting that the differential expression of IRS4 and CA12 appears to be either specific to brain tissue from AA donors or much less prominent in EA brain tissue. ¹⁸ ITPKB expression measured in cDNA microarray experiments was shown to be higher in an EA sample of 61 AD cases compared to 53 controls, ⁴⁷ which is consistent with our result in AA brain specimens. Mouse studies have linked ITPKB activation and higher ITPKB expression to AD pathology. ⁴⁸ , ⁴⁹ IRS4, which encodes insulin receptor substrate 4, is known to function in several processes relevant to AD including interacting with endosomes to control Aβ levels in neurons. ⁵⁰ In addition, insulin signaling and resistance, and diabetes more generally, have been strongly implicated in AD risk (e.g., ⁵¹ , ⁵² ). Irs4 expression was reduced in B6.APB^Tg mice in the early stages of AD pathology, ⁵⁰ which does not match our observation of higher IRS4 expression in AD cases, but this may be due to the later stage of disease in the post mortem human brain. IRS‐4 interacts with IRS‐2, ⁵³ which is also among the TWS DEGs in our AA study, and disruption of IRS‐2 impacts tau phosphorylation, although the direction of effect is not consistent across studies. ⁵⁴ , ⁵⁵

CA12 is part of the carbonic anhydrase (CA) family which has been linked to learning and memory. ⁵⁶ CA12 expression in the caudate nucleus was identified as regulated by rs117618017, a SNP in the neighboring gene APH1B, ⁵⁷ which was robustly associated with AD risk in GWAS (e.g., ¹⁰ ). Repurposing existing FDA‐approved CA inhibitors has been suggested as a potential dementia treatment. ⁵⁸ PDE10A is member of the phosphodiesterase (PDE) family of enzymes involved in the breakdown of cAMP and/or cGMP. ⁵⁹ PDE inhibitors have been shown to increase memory task performance in animals (e.g., ⁶⁰ , ⁶¹ ) and have been investigated as potential treatments for autism spectrum disorder, ⁶² schizophrenia, ⁶³ and Parkinson disease. ⁶⁴ , ⁶⁵ A clinical trial of a PDE1 inhibitor in AD cases did not find a benefit, ⁶⁶ but other PDE inhibitors may have efficacy. ⁵⁹ Here and in a recent DGE study conducted in tissue from EA brain donors, ¹⁸ PDE10A expression is lower in AD brain tissue, which is perhaps counter to the notion that PDE10A inhibition is an effective treatment strategy. However, prior expression studies of multiple PDE genes yielded conflicting results. ⁵⁹

Several other TWS DEGs have strong connections to AD. PSENEN encodes a subunit of the gamma‐secretase complex including presenilin. ⁶⁷ ROBO3 is a homolog of ROBO1 which emerged as a genome‐wide significant locus in an AA AD GWAS. ¹⁰ HDAC4 is one of several histone deacetylases previously implicated in AD and identified as potential targets for AD treatment. ⁶⁸

Analyses of pathways enriched for significant DEGs highlighted the importance of many nuclear genes involved in mitochondrial energy production, including 3 cytochrome c oxidase genes and 11 genes encoding NUDF subunits, all of which were expressed at a lower level in AD cases than controls. NDUF subunits form part of mitochondrial respiratory complex I, a multimeric enzyme composed of 44 subunits encoded by both nuclear and mtDNA genes. AD‐related impairments in mitochondrial complex I in the brain and platelets have been observed. ⁶⁹ These findings in our AA sample overlap with evidence from many studies for mitochondrial energy pathway involvement in AD ⁷⁰ and are supported by results from an AD DGE study in a large EA sample. ¹⁰ Multiple groups have identified mitochondrial dysfunction as an important cause of increased oxidative stress in brain tissue and suggest that this pathway is a potential druggable target for AD treatment and prevention. ⁷¹ , ⁷²

This study has several notable limitations. It should be acknowledged that this cohort of brain donors may not be representative of AA AD cases and controls more broadly. Although this is the largest study of post mortem brain tissue from AA AD cases and controls, the sample is much smaller than other DGE studies conducted in EA autopsy samples (e.g., ¹⁷ , ¹⁸ ). Additionally, there is substantial variability in the collection and availability of neuropathological data and ante mortem information including cognitive test data and history of relevant conditions (e.g., cardiovascular and transient ischemic attacks [TIA] events, diabetes, or other medical complications) that precluded incorporation of these factors in our analyses. Finally, because this study was performed using bulk RNA sequencing data, we could not ascribe observed gene expression association to particular cell types or detect associations evident only in underrepresented cell types. Subsequent studies could overcome this limitation through single‐cell sequencing technology ¹⁶ or deconvolution methods. ⁷³

In summary, we identified many genes that are differentially expressed in PFC tissue obtained from AA AD cases and controls. Several of the most significant DEGs were observed in EA brain donors, most notably ADAMTS2, whereas other novel DEGs and eQTLs were identified in AA individuals only. Asymmetrical findings in the two groups may be due to population differences in the frequency of genetic variants, modifying effects of other genes or other risk factors. Our findings support the hypothesis that reelin and poor bioenergetics due to mitochondrial dysfunction play a role in AD susceptibility, and may lead to new hypotheses about pathogenic mechanisms, targets for drug development, and biomarkers for risk assessment and profiling subjects for clinical trials. The inclusion of AA participants in AD research is important not only to ensure that predictions made based on genetic and ‘omic data are accurate in this population, ⁹ but also because of the potential it will lead to new and important advances in knowledge about AD risk that will benefit everyone.

CONFLICT OF INTEREST STATEMENT

Mark Logue received grants from the NIH and Department of Veterans Affairs. Marla Gearing, Lee‐Way Jin, Richard Mayeux, Richard Perrin, Shih‐Hsiu Wang and Lindsay Farrer received grants from the NIH. Melissa Murray received grants from NIH, was a paid consultant for Biogen Pharmaceuticals, and served on committees for the Alzheimer's Association and International Conference on Alzheimer's and Parkinson's Diseases. Thor Stein received grants from the NIH and Department of Veterans Affairs, and an honorarium from Brown University. Andrew Teich received grants from the NIH, a contract from Regeneron Pharmaceuticals and an honorarium from Ono Pharmaceuticals, owns stock in Ionis Pharmaceuticals and Biogen Pharmaceuticals, and served on committees for the Department of Defense and the Alzheimer's Association. The effort of Katarnut Tobunluepop and Zihan Wang was supported by NIH grants. Benjamin Wolozin received grants from the NIH, consulting fees from Aquinnah Pharmaceuticals and Abbingworth Ventures, honoraria for several lectures, and owns stock and is Co‐Founder and CSO of Aquinnah Pharmaceuticals Inc. Other authors have no competing interests to report. Author disclosures are available in the supporting information.

Supporting information

Supporting Information

Supporting Information

Logue MW, Labadorf A, O'Neill NK, et al. Novel differentially expressed genes and multiple biological pathways for Alzheimer's disease identified in brain tissue from African American donors. Alzheimer's Dement. 2025;21:e70629. 10.1002/alz.70629

DATA AVAILABILITY STATEMENT

RNA sequencing data are available at the National Institute on Aging Genetics of Alzheimer's Disease Data Storage Site (NIAGADS; https://www.niagads.org).