Reduced Expression of Brain Expressed X-linked Genes in Alzheimer’s Disease

Abstract

Background:

Altered expression of Brain Expressed X-linked (BEX) genes has been implicated in Alzheimer’s disease (AD) with inconsistencies and a lack of experimental confirmations.

Objective:

This study aims to characterize BEX expression patterns, examine the association of their expression with AD pathology, and investigate cellular changes induced by altered BEX expression.

Methods:

We integrated bulk and single-cell transcriptomics datasets to characterize cortical BEX expression changes in 5xFAD mice and AD patients, which were validated by RT-qPCR in postmortem AD cortical tissue. We manipulated the expression of BEX genes in SH-SY5Y cells and examined changes in oxidative stress.

Results:

Cortical expression of BEX genes was reduced in AD. With the dysregulation being more pronounced at a later stage of the disease, the reduction was closely related to the impairment of synaptic transmission and oxidative phosphorylation in excitatory neurons. Consistent with these, BEX3 knockdown increased oxidative stress.

Conclusions:

Our findings suggest that reduced BEX expression in neurons appears to be a contributing factor to AD pathogenesis and BEX depletion may increase oxidative stress.

Introduction

Decades of AD research has established that aggregated amyloid-β (Αβ) peptides and hyperphosphorylated Tau are two hallmarks of AD pathology^1–3. A variety of cellular processes are involved in AD pathogenesis, including lipid metabolism, immune response, metabolic stress, mitochondria function, RNA splicing, neurovascular function, etc^4–7. Brain Expressed X-linked (BEX) genes have been implicated in AD through transcriptomics analyses^8–11, although a comprehensive characterization of their expression and function is lacking. The BEX gene family is evolutionarily recent¹². Restricted in eutherian mammals, the majority of BEX genes are located on the X chromosome and have enriched expression in neural organs^12–14. This gene family contains five members in humans, namely BEX1, BEX2, BEX3 (NADE, NGFRAP1, pHGR74), BEX4 (NADE3, BEXL1), and BEX5 (NADE2, NGFRAP1L1)¹². A consensus biochemical function of human BEX proteins is its regulation of oxidative phosphorylation (OXPHOS) via binding to the E3 ubiquitin ligase adaptor Fem1 homolog B (FEM1B) and inhibiting the activity of the E3 ligase Culin2-FEM1B (CRL2^FEM1B)¹⁵. The binding of BEX proteins is mediated by residual Arg126 of FEM1B; its mutation (p.Arg126Gln) abolishes binding to BEX¹⁵. Heterozygous mutation of FEM1B at this site (p.Arg126Gln)^16,17 and the Xq22 microdeletion of BEX genes^18–21 both lead to neurodevelopmental disorders, including intellectual disability and autism-like phenotypes. Although BEX proteins have not been directly linked to amyloid, human BEX3 has been suggested to form heating reversible amyloid-like aggregates²². As an intrinsically disordered protein, BEX3 binding with nucleic acids such as small fragments of tRNA can transit its structure into a folded state, implicating its potential role in cell signaling via nucleic acid binding²².

While BEX genes generally show enriched expression in neurons¹⁴, their functions in neurons remain unclear. Mouse Bex3, the best studied BEX, has been proposed to bind to the p75 neurotrophin receptor (p75NTR) and mediate apoptosis^23,24. Bex3 knockout mice exhibit a range of abnormal behavior traits, including repetitive behavior, impaired social interaction, and memory deficits¹⁴. Conflicting evidence suggests that both mouse and rat Bex3 regulate the transcription of tropomyosin receptor kinase A (TrkA) and promote neuronal survival²⁵. Separately, mouse Bex1 has been proposed to interact with calmodulin²⁶ and facilitate axon growth after nerve injury²⁷, whilst rat Bex1 is involved in p75NTR binding that inhibits neuronal differentiation²⁸. Importantly, human BEX genes have shown sequence and functional divergence from their rodent homologues. Human BEX3 does not interact with p75NTR^29,30. Human, but not mouse, BEX2 interacts with LMO2³¹. BEX5 is not present in rodents, instead, the mouse Bex6 gene is located on an autosome¹². Thus, animal models may not be able to recapitulate the biological functions of BEX genes in humans and the pathological involvement of BEX genes in human diseases.

A consistent down-regulation of BEX1 and BEX3 has been reported from independent analyses in neurons from AD patients^8,10,11. Decreased expression of BEX2 and BEX3 is associated with cognitive decline in women in aging and AD based on transcriptomics analysis of postmortem brain tissue⁹. Here, leveraging published transcriptomics datasets, first we characterized the expression pattern of BEX genes in human tissues, then focused on the role of BEX genes in AD. Our results indicated a reduction in BEX gene expression in the cortex of AD patients and 5xFAD mice, which was not confounded by sex or age. The reduction of BEX expression was accompanied by the loss of co-expression of synaptic transmission and OXPHOS genes in excitatory neurons, whilst the change was not as evident in inhibitory neurons. Finally, we employed gene overexpression and knockdown strategies in SH-SY5Y neuroblastoma cells to show the critical role of BEX3 in regulating oxidative stress. Our results provided a comprehensive characterization of BEX expression in AD, highlighting the reduction as a contributing factor to AD pathogenesis, and implicating oxidative stress and dysregulated synaptic transmission as potential biological consequences associated with reduced expression of BEX in AD.

Materials and methods

Preprocessing of published datasets

A summary of datasets analyzed in the study is provided in Supplemental Table 1. To assess BEX gene expression in human, we retrieved cell type transcriptomics datasets from Human Protein Atlas (HPA)³² computed from multiple single cell RNA-sequencing (scRNA-seq) datasets based on healthy human tissues. Values of 557 cell types were visualized based on log2(normalized transcript per million (nTPM) + 1). To examine BEX gene expression in the human brain, we also retrieved scRNA-seq dataset on middle temporal gyrus (MTG) of neurotypical donors from the Seattle Alzheimer’s Disease Brain Cell Atlas (SEA-AD)³³. BEX gene counts of 137303 cells from 5 donors were visualized by dot plot heatmap with cell type labels provided by SEA-AD.

To examine BEX gene expression in the brain of AD patients, we retrieved microarray datasets from three independent studies. For the Zhang 2013 dataset³⁴, Brodmann area 9 (BA9) samples from 51 healthy controls and 127 late-onset AD donors (age > 60 yr) were used. Normalization of probe signals was processed by a customed procedure from the study³⁴. For the Webster 2009 dataset³⁵, frontal and temporal cortex samples from 178 neuropathologically normal and 174 late-onset AD donors (age > 60 yr) were used. To normalize the microarray signals, probes were set as missing for values less than 0 or detection scores less than 0.9. Probes detected in less than 90% samples were removed, then quantile normalization was applied to the log expression value matrix with missing values set to 0. For the Berchtold 2013 dataset³⁶, 28 cognitively normal controls and 25 AD donors (age > 60 yr) were used. Three brain cortex regions (entorhinal cortex, postcentral gyrus and superior frontal gyrus; some donors provided multiple samples) were combined for a total of 68 control and 60 AD cases. GC-RMA normalized values were provided by the study³⁶. Probes used for quantification are listed in Supplemental Table 2.

We also retrieved scRNA-seq datasets as read count matrices from three independent studies. For the Otero-Garcia 2022 dataset³⁷, cells were from BA9 of 8 cognitively normal controls and 8 AD donors (Braak stage VI). For the Xiong 2023 dataset³⁸, cells were from prefrontal cortex (PFC) of 48 control (non-AD), 29 early-stage AD and 15 late-stage AD cases from the Religious Orders Study and Rush Memory and Aging Project (ROSMAP)³⁹. The disease stages were defined previously by clinical and pathological measurements⁸. For the Gabitto 2024 dataset from SEA-AD³³, cells were from MTG and dorsal lateral prefrontal cortex (DLPFC) of 84 donors with a range of AD progression status. We stratified MTG donors into 21 not AD & low AD pathology, 21 intermediate AD pathology and 42 high AD pathology cases, and stratified DLPFC donors into 21 not AD & low AD pathology, 20 intermediate AD pathology and 39 high AD pathology cases by AD neuropathological change (ADNC) score reported in SEA-AD³³. For accessing the expression change of BEX genes in neurons, GAPDH was used as a reference for correlation comparison between gene pairs. To account for the variation from biological replicates in scRNA-seq, we aggregated cells from the same sample and quantify on the sample level⁴⁰. To accommodate for the sparsity of scRNA-seq, we filtered cells by removing cells with zero count in either of the genes in the gene pair. Given the abundant expression of all the genes in neurons (Figure 1), this step yielded enough cells for the following steps, and avoided statistical assumption on the proportion of non-biological zeros from different datasets due to technical or sampling procedures⁴¹. Next, linear regression was computed for pairwise read counts of GAPDH and the target gene. Statistical significance of each regression was determined by two-sided pairing permutation test with 2000 resampling. Regression coefficients with p < 0.05 were subjected to further analysis. Pseudobulk differential expression is a different processing procedure than above and has been evaluated in AD scRNA-seq datasets elsewhere⁴². It was not used in this study.

Figure 1. BEX expression is enriched in neurons.

(A) Chromosomal locations of BEX genes based on hg38 human genome assembly. Red lines indicate the location of corresponding genes. (B) Heatmap showing BEX expression across 557 cell subtypes of human tissues from Human Protein Atlas (HPA). Each column represents a distinct cell subtype cluster defined computationally by HPA. Hierarchical clustering on BEX expression grouped neurons (marked by yellow bar) and glial cells (marked by green bar) together. Ast. Astrocytes; Mic. Microglial cells; Olig. Oligodendrocytes; OPC Oligodendrocyte progenitor cells; Exc. Excitatory neurons; Inh. Inhibitory neurons. (C) Dot plot showing BEX expression of major cell types in MTG of neurotypical reference subjects from the Seattle Alzheimer’s Disease Brain Cell Atlas (SEA-AD). A more refined cell subtypes of BEX expression is available in Supplemental Figure 1.

To examine Bex gene expression in the brain of 5xFAD mice, the Forner 2021 dataset⁴³ was retrieved as normalized expression (logTPM) matrix from bulk RNAseq. Cortex samples were from 46 C57BL/6 wildtype and 50 5xFAD hemizygous mice at 4, 8, 12, 18 months of age.

Glycolysis score analysis on scRNA-seq datasets

To examine the expression of glycolysis genes as a response variable to covariates, we first computed glycolysis score in neurons representing the average relative expression of all glycolysis genes (KEGG: hsa00010) following a previously established method⁴⁴, then applied a linear mixed model with the design “glycolysis score ~ pathological status + sex + RBFOX3 expression + GAPDH expression + log10(total counts/UMIs) + (1 | donor)”, where excitatory neurons and inhibitory neurons were considered separately. In the design, donor identity was modeled as a random effect, while other covariates were fixed effects.

For the Otero-Garcia 2022 dataset³⁷, 8 control and 8 AD donors were used. For the Xiong 2023 dataset³⁸, 48 control and 15 late-stage AD donors were used. The coefficients were reported along with 95 % confidence intervals. For the Gabitto 2024 dataset from SEA-AD³³, MTG of 21 not AD & low AD pathology and 42 high AD pathology donors, and DLPFC of 21 not AD & low AD pathology and 39 high AD pathology donors were used. Cell number down-sampling was applied to SEA-AD dataset by randomly taking 5% of cells from MTG and 1% of cells from DLPFC for 10 times, and repeating the linear mixed modeling each time. The coefficients associated with one covariate were reported as the mean, and the 95% confidence interval for one covariate was reported as the extrema of 10 replications.

Gene co-expression analysis on scRNA-seq datasets

For the Otero-Garcia 2022 dataset³⁷, 46,070 and 50,059 excitatory neurons from 8 control and 8 AD donors were used. For the Xiong 2023 dataset³⁸, 81,770 and 26,649 excitatory neurons from 48 control and 15 late-stage AD donors were used. For the Gabitto 2024 dataset from SEA-AD³³, 183,273 and 292,810 glutamatergic neurons from MTG of 21 not AD & low AD pathology and 42 high AD pathology donors, and 173,732 and 270,589 glutamatergic neurons from DLPFC of 21 not AD & low AD pathology and 39 high AD pathology donors were used. Genes expressed in at least one third of the cells in each disease state in each dataset were retained for computing Pearson correlation matrices. Dendrogram from hierarchical clustering on correlation matrix was flattened with maximum threshold such that BEX genes were mostly retained in the same cluster. The flattened cluster containing BEX genes was denoted as BEX co-expression network.

Gene ontology (GO) enrichment analysis (biological process) was performed with PANTHER (release 2025–10-10)^45–47. GO terms with FDR < 0.05 and at least 5 hit genes were kept for further analysis. REVIGO was applied to group similar GO terms into clusters by SimRel similarity and visualize GO results as treemaps⁴⁸.

Human postmortem tissue

Frozen blocks of postmortem brain tissue from Brodmann area 10 (BA10) of normal and AD subjects were obtained from NIH NeuroBioBank. Brain samples were received frozen and stored in −80 °C freezer until RNA extraction. Subject information is listed in Supplemental Table 3.

Plasmids and lentivirus production

Plasmids for knockdown were generated by cloning a target shRNA sequence into the pLKO.1 vector (Addgene, #1864). BEX3 was PCR amplified from cDNA of SH-SY5Y cells and cloned to FUW-tetO-lox vector (Addgene, #20728) to generate FUW-tetO-lox-BEX3⁴⁹. FUW-M2rtTA (Addgene, #20342) was used in combination with FUW-tetO-lox-BEX3 for the purpose of overexpressing BEX3. Grx1-roGFP2 was subcloned from pEIGW-Grx1-roGFP2 (Addgene, #64990) to pLenti6-V5-TOPO vector (Invitrogen). Plasmids were confirmed by Sanger sequencing. Primers used for cloning are listed in Supplemental Table 4. Lentiviruses was packaged with 0.9 μg pMD2.G (Addgene, #12259), 2.7 μg psPAX (Addgene, #12260) and 3.6 μg target construct in 293FT cells in 60mm dish using Lipofectamine 2000 (Invitrogen, 11668027). Viral titer was determined by p24 ELISA (XpressBio, XB-1000).

Cell lines and virus transduction

293FT (Invitrogen, R70007) was maintained in high glucose Dulbecco's Modified Eagle Medium (DMEM) (Gibco, 11965092) supplemented with 10% FBS (Phoenix-Scientific), 2 mM L-glutamine (Gibco, 25030081), non-essential amino acids (NEAA) (Gibco, 11140050) and 1 mM sodium pyruvate (Gibco,11360070). SH-SY5Y (ATCC) was maintained in high glucose DMEM supplemented with 10% heated FBS, 2 mM L-glutamine and NEAA. SH-SY5Y cells were transduced by adding lentivirus at multiplicity of infection of 10 to the medium containing 8 μg/ml polybrene (Sigma, H9268). Medium was replaced by fresh medium after 16 h, and cells were cultured for 4 days before cultivation or live imaging. SH-SY5Y cells stably expressing Grx1-roGFP2 were obtained by first transducing cells with Grx1-roGFP2 encoding lentivirus, then manually isolating the single fluorescent colony through serial dilution in 96-well plate. All cells were maintained at 37°C with 5% CO₂.

Real-time reverse transcription quantitative PCR (RT-qPCR)

Total RNA from tissue was extracted with RNeasy kit (Qiagen, 74104). Total RNA from cells was extracted with TRIzol (Invitrogen). Total RNA was reverse transcribed to cDNA with high-capacity cDNA reverse transcription kit (Applied Biosystems). RT-qPCR was performed with iQ SYBR Green Supermix (Bio-Rad, 1708880) and gene specific primers (Supplemental Table 4) on CFX Duet Real-Time PCR system (Bio-Rad). Threshold cycle (Ct) values of target genes were normalized against GAPDH from the same sample for dCt calculation. Relative expression by ddCt were normalized against control samples.

Western blotting

Quantification of protein carbonylation by Western blotting was performed with protein carbonyl assay kit (abcam, ab-178020). Protein was collected by in-well lysis with 1% sodium dodecyl sulfate (SDS), and sonicated on ice for 3 s. 15 μg of protein was denatured with equal volume of 12% SDS, then treated with two volumes of 2,4-dinitrophenylhydrazine (DNPH) for 15 min until the reaction was terminated by two volumes of neutralization solution. For derivatization control, control solution without DNPH was used. Proteins were separated on 7.5% SDS-PAGE, and transferred to nitrocellulose membrane. Total protein on the membrane was visualized by Ponceau S (Sigma-Aldrich, P3504) staining. Primary (rabbit anti-DNP, 1:5000) and secondary antibodies (HRP-conjugated goat anti-rabbit, 1:5000) were included in the kit.

Live cell imaging

SH-SY5Y cells stably expressing Grx1-roGFP2 were plated on FluoroDish (World Precision Instruments, 35 mm;10 mm well) and incubated in imaging medium (Leibovitz’s L-5 medium without phenol red, Gibco, 21083027; supplemented with 10% FBS) for at least 3 h before imaging. Leica SP8 confocal microscope was set up to excite samples at 405 nm and 488 nm sequentially with the same emission range at 510–550 nm. Lasers were warmed up for at least 1 h. For sample oxidization or reduction, 2 mM H₂O₂ (Sigma, 386790-M) or 10 mM dithiothreitol (DTT) (Thermo Scientific, J15397.03) was freshly prepared in DPBS (Gibco, 14190144), then equal volume was added to the dish to a final concentration of 1 mM H₂O₂ or 5 mM DTT. Single images were collected without treatment for baseline, after 1 min of H₂O₂ treatment for oxidation, or after 6 min of DTT treatment for reduction. Time lapse images were continuously collected with 15 s interval where either H₂O₂ or DTT was added at 1 min after starting the experiment.

405/488 nm pseudocolor images were generated in Python by dividing each pixel of the 405 nm image by the same pixel of 488 nm image after thresholding. Colormap “twilight” from matplotlib was used to reconstitute the divided values into images. ImageJ was used for labeling ROIs for individual cells and intensity quantification. The same cell on frames of 405 nm and 488 nm excitation shares the same ROI. 405/488 nm ratio of individual cell was calculated as integrated density (IntDen) with 405 nm excitation divided by IntDen with 488 nm excitation. Degree of oxidation was calculated by the Nernst equation as described before⁵⁰ and normalized to the median of the group. For time lapse imaging, individual cell was tracked and labeled across time series, and 405/488 nm ratio was normalized against baseline value without treatment to reflect the change of ratio.

Statistical analysis

Biological replicate number (n) and statistical test appropriate for each published dataset were stated in figures and figure legends. The postmortem tissues used for RT-qPCR were balanced for age and sex. The statistical effect of age and sex in tissue level quantification were examined and shown in Supplemental Tables 5–9. We applied Shapiro-Wilk test for normality and Levene’s test for homoscedasticity before comparing the group means. For analyzing two groups, two-tailed t-test was used for normally distributed groups with equal variance. Otherwise, Wilcoxon rank-sum test was used. For analyzing multiple groups, one-way ANOVA was applied to normally distributed groups with equal variance. If significance was found, Tukey’s HSD test was used for multiple comparison. If the data was not normally distributed, Kruskal-Wallis test was used, followed by Dunn’s multiple comparison test with Bonferroni correction. For analyzing two factors, two-way ANOVA was applied to normally distributed data with equal variance. Otherwise, Scheirer-Ray-Hare test was used. Statistical tests were performed with Python. Statistical significance is denoted as * p < 0.05; ** p < 0.01; *** p < 0.001 when the exact p value was not shown.

Results

BEX expression pattern in human tissues

Human BEX gene family is located on X chromosome’s q22 region, spanning ~1.3 Mb (Figure 1A). None of the BEX genes has been reported with disease-causing mutations in Online Mendelian Inheritance in Man (OMIM). To examine the expression pattern of BEX genes in different human tissues, we clustered expression profile of BEX genes across cell subtypes from multiple tissues from Human Protein Atlas (HPA)³². BEX3 and BEX1 were the most highly expressed BEX genes, followed by BEX2 and BEX4, while BEX5 expression was much lower than other BEX genes. BEX3 and BEX4 were expressed across all tissue types, whereas other BEX genes were more tissue-specific and primarily expressed in the brain (Figure 1B). Neuronal subtypes in the brain and retina were clustered in separate groups where BEX gene expression was enriched. Glial cells were also grouped together with lower BEX expression but similar pattern (Figure 1B). Further examination in scRNA-seq of neurotypical cortical brain tissue from SEA-AD³³ revealed BEX1–4 were highly expressed in ~90% of excitatory neurons, and slightly less expressed in ~80% of inhibitory neurons. BEX5 had much lower expression in only ~30% of neurons, while all BEX genes were minimally expressed in non-neuronal (glial) (Figure 1C). Consistent with HPA (Figure 1B), analysis on refined neuronal subtypes suggested a similar BEX expression pattern across subtypes from both excitatory and inhibitory neurons (Supplemental Figure 1). Overall, BEX genes had enriched expression in neurons, suggesting that brain tissue-level expression of BEX originated largely from neurons instead of glial cells.

Decreased BEX expression in AD brains

To evaluate AD-associated BEX expression change, we examined three independent microarray gene expression datasets of various cortical regions from postmortem brains from AD patients and aged controls (age > 60 yr) for a total of 583 subjects^34–36. BEX expression was drastically decreased in AD in multiple cortex regions, including prefrontal cortex (BA9) (Figure 2A), temporal and frontal cortex (Figure 2B), entorhinal cortex, postcentral gyrus and superior frontal gyrus (Figure 2C). This effect was evident as compared to GAPDH expression change in corresponding datasets, which served as an internal reference since three datasets were reported with different normalization strategies. Next, we examined BEX expression by RT-qPCR on prefrontal cortex (BA10) of postmortem brain from controls and AD patients (subject information in Supplemental Table 3). The results support the findings from microarray datasets that BEX expression was reduced in AD brains, while expression of the neuron-specific gene RBFOX3 remained similar (Figure 2D). Sex and age had limited impact on BEX expression compared with the disease-related effect, as the overall effect size of sex was much lower (Supplemental Figure 2; Supplemental Tables 5 and 6), and no consistent and significant correlations or effects between age and BEX expression were observed across datasets (Supplemental Tables 7 and 8). Together, the examinations at tissue level indicated that BEX expression is reduced in AD brain cortex.

Figure 2. Decreased BEX expression in AD brains.

(A) Violin plots of BEX expression in control (blue) and late-onset AD (red) brains in BA9. (B) Violin plots of BEX expression in control (blue) and late-onset AD (red) brains in temporal cortex and frontal cortex. BEX2 probe was absent in the dataset. (C) Violin plots of BEX expression in control (blue) and AD (red) brains in entorhinal cortex, postcentral gyrus and superior frontal gyrus. Horizontal lines represent 25, 50 and 75th percentiles, respectively. All datasets were from microarray expression profiling. As datasets were normalized differently by the original studies, GAPDH expression were shown for reference. Statistical differences were determined by Wilcoxon rank-sum test. Microarray probes used for quantification are listed in Supplemental Table 2. (D) Bar plot showing expression of BEX genes and RBFOX3 in AD (red) brains in BA10 relative to controls (blue) measured by RT-qPCR. Data are represented as mean ± SD. Statistical differences were determined by two-tailed t-test.

Decreased Bex expression in 5xFAD mice

Next, we examined Bex expression in 5xFAD mice, which are widely used to model neuropathological and cognitive changes in AD^43,51,52. Transcriptomic profiling by RNA-seq on brain cortex of WT and 5xFAD mice from both sexes across lifespan were analyzed⁴³, which provided a longitudinal view along pathological progression (Figure 3A). Bex1 expressed lower in 5xFAD than WT as early as 8 months of age (Figure 3C), while a trend of reduced expression at age of 8 months or older in 5xFAD was observed in other Bex genes but not significant after adjustment for multiple comparison (Figure 3D–F). The effect of genotype was profound on Bex expression (Supplemental Table 9), as was also suggested by reduced Bex expression but not Gapdh in 5xFAD mice when subjects of different ages were combined (Figure 3B–F). Age-related effect was greater in Bex2 and Bex4 compared with Bex1 and Bex3, whilst sex had a smaller but significant effect on Bex1, Bex3 and Bex4 compared with genotype (Supplemental Table 9). Although expression of Bex in female was lower than in male 5xFAD mice after 12 months (Supplemental Figure 3), the interaction between genotype and sex was not significant, potentially due to the inconsistent sex differences in WT mice from different ages (Supplemental Table 9). Together, the result suggested Bex expression in the cortex of 5xFAD mice started to decrease after 4 months, later than the detection of initial amyloid plaques and memory deficits at 4 months^43,51. Thus, BEX reduction is likely a contributing, not a primary causal factor, in AD pathogenesis.

Figure 3. Decreased BEX expression in the brain of 5xFAD mice.

(A) Schematic of samples in the dataset. (B-F) Plots showing the normalized expression of (B) Gapdh (C) Bex1 (D) Bex2 (E) Bex3 (F) Bex4 in the cortex of WT (blue) and 5xFAD (red) mice across lifespan from RNA-seq. Bex6 expression was not detected. Data in line graphs are represented as mean ± SD. Statistical differences were determined by three-way ANOVA followed by Tukey’s HSD for multiple comparison. For violin plots, data was merged from all samples regardless of age. Horizontal lines represent 25, 50 and 75th percentiles, respectively. Statistical differences were determined by Wilcoxon rank-sum test.

The co-expression of BEX and GAPDH is dysregulated in AD neurons

Since BEX genes were primarily expressed in neurons, we focused our analyses on neurons from postmortem human brain cortex. Four scRNA-seq datasets from three independent studies^33,37,38 for a total of 192 controls and AD patients were analyzed (Figure 4A–D for each dataset). As the differentially expressed genes across AD scRNA-seq datasets has been evaluated elsewhere⁴², we decided to focus on the relative relationship between BEX genes and the reference gene with a linear model. To account for the statistical bias introduced by multiple cells from the same subject (pseudo biological replicates), as well as the zero-inflation issue specific to scRNA-seq^40,41, we employed a within-subject cell aggregation approach and removed zero-count genes in cells, whose proportion varies across chemistry and platforms.

Figure 4. Decreased BEX to GAPDH regression coefficient in AD excitatory neurons.

(A) Boxplots of regression coefficient using GAPDH count as predictor and target gene count as response in excitatory neurons from BA9 of control (blue) and late-stage AD (red) brains. (B) Boxplots of target gene/GAPDH regression coefficient in excitatory neurons from PFC of control (blue), early-stage AD (orange), and late-stage AD (red) brains. (C) Boxplots of target gene/GAPDH regression coefficient in glutamatergic neurons from MTG of not AD & low AD pathology (blue), intermediate AD pathology (orange), and high AD pathology (red) brains. (D) Boxplots of target gene/GAPDH regression coefficient in glutamatergic neurons from DLPFC of not AD & low AD pathology (blue), intermediate AD pathology (orange), and high AD pathology (red) brains. Each dot represents one subject. Regression coefficient was only included when permutation test p < 0.05. The center horizontal line represents median, the bounds of the box represent 25 and 75th percentiles, and the whiskers represent 1.5 interquartile range. All datasets were from scRNA-seq with cell subtype annotated from the original studies. Statistical differences were determined by two-tailed t-test (A), or Kruskal-Wallis test followed by Dunn’s multiple comparison test with Bonferroni correction (B-D).

GAPDH was used as the reference for BEX genes as it was highly expressed, exhibited consensus copy numbers across brain cortex, and reflected between-subject biological variations^53–55. Since GAPDH is involved in glycolysis, we first examined whether glycolysis score was stable across disease states and estimated how GAPDH expression in neurons correlated with glycolysis score at single cell level. A glycolysis score was defined as the average expression of genes in glycolysis pathway in a cell relative to the background gene expression. Leveraging the strength of high-resolution cell type mapping by scRNA-seq, we examined excitatory and inhibitory subgroups of neurons separately. Among pathological status, sex, RBFOX3 expression and other covariates, GAPDH expression explained glycolysis score the best with a significant and greater positive association, which was also consistent across datasets in both excitatory and inhibitory neurons (Supplemental Figure 4). In contrast, pathological status had no significant association with glycolysis score, and the trend of association varied across datasets (Supplemental Figure 4). As a reference to compare with GAPDH, RBFOX3, a canonical neuronal marker involved in neuronal RNA splicing⁵⁶ but not glycolysis, also showed limited association with glycolysis score (Supplemental Figure 4).

Next, we estimated the correlation between GAPDH and BEX genes where permuted GAPDH counts served as the predictor variable and BEX counts as response in the linear regression for each subject. To exclude subjects that were not adequately sampled, only regression coefficients with significant permutation test statistics (p < 0.05) were further analyzed. The regression coefficients between GAPDH and BEX1, BEX2, BEX3, BEX4 in controls were positive and greater than regression coefficients of GAPDH to RBFOX3, suggesting BEX function could be related to glycolysis. Significant reductions in the regression coefficient of BEX1 to GAPDH and BEX3 to GAPDH were consistently observed in excitatory neurons from multiple cortex regions at late stage of AD (Figure 4A–D), while this effect was subtle in early stage of the disease or in inhibitory neurons (Supplemental Figure 5). These results suggested that the relationship between BEX and glycolysis might be altered in excitatory neurons at late stage of AD and called for further examination on genes co-expressed with BEX.

Dysregulated BEX co-expression networks in AD excitatory neurons

The dysregulated BEX expression in AD excitatory neurons prompted us to analyze the co-expression gene networks in excitatory neurons from controls and late stage AD patients respectively. Pearson correlation matrices were first computed on universally expressed genes that were detected in at least one third of the excitatory neurons, then hierarchical clustering was applied to order the genes by linkage. Interestingly, more than twice of universally expressed genes were detected in SEA-AD (Figure 5C, D) as in other datasets (Figure 5A, B), reflecting a lower proportion of non-biological zeros and better quality of SEA-AD. Next, the BEX co-expression network was estimated by flattening the dendrogram into clusters with maximum threshold and grouping BEX genes in the same cluster, which yielded highly co-expressed gene networks containing BEX genes (Figure 5A–D). Intriguingly, BEX co-expression networks from SEA-AD contained similar number of genes as the networks from other datasets, and GAPDH was present in all BEX co-expression networks, suggesting potentially shared biological functions for the network. To investigate such possibilities, we examined the common and unique genes belonging to BEX co-expression networks in control and AD neurons. We identified 58 genes shared by both control and AD networks, 41 genes shared by control but not AD networks, whilst no gene was specifically shared by AD but not control networks (Figure 5, full list of genes in Supplemental Table 10). Moreover, control networks from SEA-AD contained fewer genes that were shared by control but not AD networks, one reason could be that control subjects from SEA-AD were not devoid of AD pathology. GO enrichment analysis showed that regulation of synapse structure and activity, nitric oxide production and glycolysis were the major functions of shared BEX co-expression networks common to both control and AD (Figure 5F). However, control networks contained a more complete gene list involved in synaptic transmission and oxidative phosphorylation (OXPHOS), which were partially lost in AD networks (Figure 5E). These results indicated that BEX genes were decoupled from specific genes participating in synaptic transmission and OXPHOS in AD excitatory neurons. Future studies are needed to understand the mechanistic details of the decoupling and its contribution to AD pathogenesis.

Figure 5. BEX gene co-expression networks revealed loss of function in AD excitatory neurons.

(A-D) Heatmap showing gene correlation matrix in excitatory neurons from (A) BA9, (B) PFC, (C) MTG and (D) DLPFC of control (left panel) and AD (right panel) brains. Genes in BEX co-expression networks were clustered by hierarchical clustering, and marked by blue bar in control (blue/total genes n = 244/3259 (A), n = 237/2869 (B), n = 362/8062 (C), n = 187/8459 (D)) and red bar in AD (red/total genes n = 215/2128 (A), n = 157/3012 (B), n = 157/7351 (C), n = 301/8151 (D)). (E) Comparison of genes in control and AD BEX co-expression networks in the four datasets (a-d, corresponding to panels A-D). Each vertical line represents one gene, with colored line indicating the presence of the gene in corresponding BEX co-expression network. Genes shared by at least 7 networks were marked by black bar (n = 58). Genes that were present in control (at least 2 control networks) but not AD networks were marked by red bar (n = 41). The genes are listed in Supplemental Table 10. (F) Treemap showing gene ontology (GO) categories of shared genes in BEX co-expression networks. Size of the block represents -log₁₀FDR. Shared genes in representative categories were labeled. (G) Treemap showing GO categories of genes that were lost in AD BEX co-expression networks. Size of the block represents -log₁₀FDR. Lost genes in representative categories were labeled.

BEX3 regulates cellular balance of reactive oxygen species (ROS)

To determine if BEX genes were OXPHOS regulators in a neuronal context, we performed in vitro experiments on the neuroblastoma cell line SH-SY5Y with controlled expression of BEX genes. Focusing on the prototypical oxidative stress-responsive pathway with transcription factor NEF2L2 (NRF2)⁵⁷, we measured the expression of HMOX1 (HO-1), NQO1, SRXN1 and ABCC4 as NEF2L2 responders coping with excessive ROS^58–60, and KEAP1 as NEF2L2 binder and inhibitor⁵⁸ in response to the knockdown of each BEX gene. BEX3 knockdown led to the most dramatic expression elevation of the stress response genes, and BEX1 and BEX2 knockdown had similar though milder effects as to BEX3 knockdown, while BEX4 knockdown had opposite effects (Figure 6A). We then overexpressed BEX3 and found it did not significantly alter the expression of stress response genes (Figure 6A).

Figure 6. BEX genes are involved in ROS regulation.

(A) Heatmap showing expression of oxidative stress related genes in SH-SY5Y with BEX genes knocked down or BEX3 overexpressed, measured by RT-qPCR. shScr, scramble control of non-targeting shRNA. Data are represented as mean quantified with three biological replicates and two technical replicates. (B) Protein carbonylation derivatized by DNPH in SH-SY5Y with BEX3 overexpression or knockdown, analyzed by Western blot. Total protein on the same blot was stained with Ponceau S. (C) Quantification of protein carbonylation with three biological replicates. Data are represented as mean ± SD. (D) Mechanism of Grx1-roGFP2 as a glutathione redox status (GSH/GSSG) sensor. (E) Fluorescence imaging of shRNA control SH-SY5Y stably expressing Grx1-roGFP2 with 405 nm or 488 nm excitation under baseline, oxidation with 1 mM H₂O₂, or reduction with 5 mM DTT. Scale bar, 100 μm. (F) Distribution of baseline 405/488 nm ratio in cells with shRNA control (n = 265), BEX3 overexpression (n = 325) or BEX3 knockdown (n = 404) based on fluorescence imaging. (G) Distribution of degree of oxidation calculated by fully oxidized (1 mM H₂O₂, median normalized to 1) and fully reduced (5 mM DTT, median normalized to 0) cells with scramble shRNA control (n = 805), BEX3 overexpression (n = 800) or BEX3 knockdown (n = 887) based on fluorescence imaging. (H) Time-lapse fluorescence changes of cellular roGFP2 treated with 1mM H₂O₂ in scramble shRNA control (n = 395), BEX3 overexpression (n = 411) or BEX3 knockdown (n = 395). (I) Time-lapse 405 nm/488 nm fluorescence changes of cellular roGFP2 treated with 5 mM DTT in scramble shRNA control (n = 287), BEX3 overexpression (n = 389) or BEX3 knockdown (n = 393). Data are represented as mean ± SD. Fluorescence imaging results were from at least three biological replicates. Statistical differences were determined by one-way ANOVA followed by Tukey’s HSD for multiple comparison in (A) and (C), and determined by Kruskal-Wallis test followed by Dunn’s test for multiple comparison with Bonferroni correction in (F-I).

To investigate the link between BEX3 and cellular ROS balance, we first quantified protein carbonylation in the cells. Western blots showed that BEX3 knockdown significantly increased protein carbonylation, whereas BEX3 overexpression had non-significant effect (Figure 6B–C and Supplemental Figure 6). To further elucidate this link in live cells, we generated SH-SY5Y cells stably expressing Grx1-roGFP2 (Supplemental Figure 7A), a fluorescent protein probe sensing cellular redox state^50,61. As a ratiometric probe, the light emission ratio of Grx1-roGFP2 at excitation of 405 nm and 488 nm reflects the redox balance of oxidized glutathione (GSSG) and reduced glutathione (GSH) (Figure 6D). Live cell imaging indicated cells were highly reduced at baseline without treatment, as further reduction with dithiothreitol (DTT) only slightly decreased 405/488 nm ratio, and they reacted to H₂O₂ oxidation with a drastic increase in 405/488 nm ratio (Figure 6E). At baseline, cells with BEX3 knockdown were in a more oxidized state, as indicated by cell ratio distribution quantified by cell segmentation in images (Figure 6F). This observation was confirmed with a higher degree of probe oxidation in BEX3 knocked down cells (Figure 6G and Supplemental Figure 7B, C), which was normalized by fully oxidized and fully reduced fluorescence with Nernst equation. In addition, BEX3 overexpression also increased degree of probe oxidation but to a lesser extent. Finally, we traced real time cellular response to oxidation or reduction, and recorded change of ratio to baseline (R/R₀) at 15 s interval. Control and BEX3 overexpressed cells had same kinetics and magnitude of response upon treatment with H₂O₂ or DTT, while BEX3 knocked down cells exhibited smaller change of ratio upon oxidation and larger change of ratio upon reduction, suggesting a lower oxidation capacity and a higher reduction capacity in these cells. Our results supported that BEX genes were involved in ROS regulation, underscoring the importance of BEX3 as its knocking down led to cellular stress due to excessive ROS, a contributing factor to AD pathogenesis.

Discussion

Despite emerging evidence suggesting the possible association of BEX genes and AD neuropathology^8–11,42, a focused characterization of their expression and function in AD is lacking. Here, we addressed this question by examining changes in BEX expression pattern in AD brains with publicly available transcriptomic datasets. We substantiated the finding by modifying BEX expression in SH-SY5Y cells and found a link between BEX3 and ROS regulation.

The rapidly evolving sequencing techniques have allowed easy access to transcriptomics data on complex diseases like AD, fueling target prediction studies yielding extensive lists of targets. With the ultimate goal of clinical translation, fundamental questions need to be addressed regarding computational and biological validations. To bridge the gap between in silico predictions and in vitro effects, we first approached computational validation of AD associated BEX expression change by cross-study comparison at brain tissue level^34–36, which showed consistent reduction of BEX in the AD brain. Being also supported by our biological validation, such reduction can be attributed to a significant decrease of neuronal proportion and/or an expression reduction in neurons, as glial cells had minimal BEX expression. Nevertheless, brain expression of neuronal marker RBFOX3 were similar in controls and AD patients, and previous studies on cell population suggested non-significant change in relative abundance of major brain cell types associated with AD patients, except for a subgroup of decreased SST⁺ inhibitory neurons and a subgroup of AD specific microglia^33,62. As a result, the reduction of BEX expression observed at tissue level was mainly driven by expression reduction in neurons. Considering the potential influence of sex or age imbalance in the datasets, we also reported sex- or age-specific effects on BEX expression. A lesser impact was observed for sex or age, indicating BEX expression difference was mainly separated by AD pathological status (Supplemental Figure 2 and Supplemental Tables 5–8). Despite BEX genes being not differentially expressed between sexes in human cortical neurons⁶³, BEX1, BEX2 and BEX3 were implicated as female-associated genes downregulated in AD brains in several cohorts^9,10. However, the sex-associated effect was not supported by cross-dataset evaluation, while AD status-associated effect was consistently observed for BEX genes⁴². Given the potential heterogeneous status of X chromosome inactivation in distinct cells and evidence of BEX2 and BEX4 biallelic expression in females^63,64, examination at transcriptomic level might not capture the consequence of sex-associated epigenetic regulation, which requires future study to clarify.

We also extended the validation to 5xFAD mouse model⁴³, which suggested Bex1 and Bex3 were reduced after the onset of AD-like plaques in the cortex, whilst sex and age had inconsistent effect compared with the effect of genotypes (Supplemental Figure 3 and Supplemental Table 9). These results suggested that BEX reduction appears to be a contributing factor, not a primary driver, in AD pathogenesis. Considering the divergence of BEX proteins between humans and mice^12,29,31, the interpretation may not be translated to humans. As such, we followed up with examinations on human datasets.

To cross-compare scRNA-seq datasets from independent studies^33,37,38 with regard to the association of BEX to AD pathogenesis beyond the known reduction in neurons at transcript level^8,10,11,42, we first evaluated the stability of glycolysis gene expression and established GAPDH as a reference for BEX genes at cell level. As a member of glycolysis pathway, GAPDH expression but not AD status exhibited positive association with the average expression of genes from glycolysis pathway in neurons (Supplemental Figure 4), indicating the impact of AD pathology on neurons was not homogeneous, and cell level GAPDH is a better representation of cell level expression of overall glycolysis genes. In addition, the result suggested that reduced glucose utilization reported in AD65 was not due to global monotonic transcriptional change in glycolysis. Instead, other mechanisms including cell specific response to the neurofibrillary tangle burden³⁷, gene specific isoform switch altering metabolic function⁶⁶, post-translational modification of enzymes⁶⁵ were potentially more involved. We next applied a linear regression-based method to evaluate the relative relationship between BEX and GAPDH expression. Our results highlighted a reduced positive correlation between BEX and GAPDH in excitatory neurons at late stage but not early stage of the disease, indicating the dysregulation of BEX was later than the onset of the disease. Further examination on BEX co-expression network identified several shared functions regardless of disease state: regulation of synapse structure and activity, nitric oxide production and glycolysis. Moreover, a consensus loss of correlation in AD excitatory neurons was observed for genes participating synaptic transmission and OXPHOS. As the primary energy consumption process in the brain, synaptic transmission is coupled with local glycolysis and mitochondrial OXPHOS to meet its high demand for adenosine triphosphate (ATP). While glycolysis provides a rapid and flexible supply of ATP, OXPHOS is critical to support synaptic vesicle endocytosis and fuel intensive synaptic transmission activities^53,67–69. Despite the caveats of studying BEX3 in mouse model, dysregulated synaptic transmission has been reported in hippocampal neurons from Bex3 knockout mice, in that evoked postsynaptic current amplitude and spontaneous postsynaptic current frequency were reduced¹⁴. Our results suggested that most glycolysis genes co-expressed with BEX in control neurons still passed the co-expression criteria in AD neurons at late stage, whereas several OXPHOS genes were impacted more severely and potentially earlier during disease progression. Such changes could be coupled with synaptic transmission, and additional experimental dissection in human neurons is required to establish a causal link between BEX genes, OXPHOS and synaptic transmission.

Nevertheless, focusing on the relationship of BEX genes and OXPHOS, we performed biological validation in neuroblastoma cells. Significant increase in the expression of oxidative stress defense genes and the amount of oxidation modified proteins, and a shift toward more oxidized glutathione redox status were detected in BEX3 knocking down cells, supporting its role in regulating cellular ROS balance. Paradoxically, BEX3 overexpression also increased oxidation. This effect could be a result of reductive stress, where short term inhibition of OXPHOS in mitochondria decreases cellular oxidation¹⁵, while chronic and persistent inhibition of mitochondria energy production could lead to mitochondria electron leakage and impairment of oxidation dependent signaling pathways, which inversely increases the cell oxidation and damage⁷⁰. More dynamic measurement of change of OXPHOS at much shorter intervals may clarify the issue. Meanwhile, the function of other BEX genes in regulating OXPHOS was less clear, as BEX1 or BEX2 knocking down increased expression of fewer oxidative stress defense genes, while BEX5 knocking down had no significant effect, and BEX4 knocking down had opposite effect. Although it was shown by previous studies that co-depletion of cellular BEX1–4 increased general ROS level¹⁵, the role of each BEX was not clear in the same context. It was also suggested that BEX genes control the cell ROS production by inhibiting CRL2^FEM1B, which mediates the mitochondria import of electron transport chain component complex IV when active⁷¹, but BEX1 had much lower affinity binding to CRL2^FEM1B, whereas BEX2 and BEX3 occupied FEM1B more than other BEX proteins in cells¹⁵. Thus, future works are necessary to understand the functional difference of different BEX genes. Additionally, provided that blocking mitochondria import is downstream of BEX activity⁷¹, either overexpression or knocking down of BEX would presumably increase oxidative stress as observed in BEX3 experiments, either due to impaired electron transport chain integrity, or due to increased OXPHOS activity. Our results suggested oxidative stress was much more pronounced in the case of BEX3 knocking down than overexpression. Future effort on dissecting the effect at organelle level will help elucidate the connection between BEX expression and mitochondria dysfunction, which is an established feature of early AD pathogenesis⁷².

Although our study presented a focused characterization of AD-associated BEX expression reduction in neurons as a contributing pathological event, and linked such reduction to ROS dysregulation, several limitations should be noted. First, our analyses on transcriptional expression does not necessarily translate into protein abundance and activity. Combining transcriptomic and proteomic analyses could help uncover post-transcriptional mechanisms, providing that antibodies are highly specific. Second, as a general term, ROS involves multiple chemical species⁷³. Our measurements focused on the overall outcome of ROS, while approaches of measuring specific chemical species can be applied in future work to dissect different ROS mechanisms. Finally, SH-SY5Y is a neuroblastoma cell line with known differences in metabolism, cell-cycle regulation and stress responses compared with primary or stem cell-differentiated neurons, and the lack of neuronal electrophysiology render it unfit for evaluating physiologically relevant neuronal and neurotransmission activities⁷⁴. Thus, biological validations in neurons and in vivo studies in humanized animal models will be the ultimate approach to refining disease-relevant neuronal functions that can translate better as in human AD pathogenesis. Finally, future works focusing on the relationship of BEX to Αβ, hyperphosphorylated Tau, and neuroinflammation could be useful to determine the primary cause of BEX reduction. With the availability of effective therapeutic antibodies against Αβ and their limited benefits in improving cognition⁷⁵, answering whether BEX reduction can be rescued by clearance of Αβ or other therapeutic relevant approaches will introduce more translational relevance for future studies.

Data availability

Publicly available datasets were listed in Supplemental Table 1. Code for reproducing the transcriptomic analyses can be accessed from GitHub at https://github.com/fenglabbuf/2026_Li_et_al.