Meta-Analysis of Transcriptomic Studies of Blood and Six Brain Regions Identifies a Consensus of 15 Cross-Tissue Mechanisms in Alzheimer's Disease and Suggests an Origin of Cross-Study Heterogeneity.

Meta-analysis of transcriptomic studies of blood and six brain regions identifies a consensus of 15 cross-tissue mechanisms in Alzheimer’s disease and suggests an origin of cross-study heterogeneity - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Am J Med Genet B Neuropsychiatr Genet . Author manuscript; available in PMC: 2026 Jun 1. Published in final edited form as: Am J Med Genet B Neuropsychiatr Genet. 2024 Dec 16;198(4):e33019. doi: 10.1002/ajmg.b.33019 Meta-analysis of transcriptomic studies of blood and six brain regions identifies a consensus of 15 cross-tissue mechanisms in Alzheimer’s disease and suggests an origin of cross-study heterogeneity Jiahui Hou 1 , Jonathan L Hess 1 , Chunling Zhang 2 , Jeroen GJ van Rooij 3 , Gentry C Hearn 4 , Chun Chieh Fan 5 , Stephen V Faraone 6, 2 , Christine Fennema-Notestine 7 , Shu-Ju Lin 8 , Valentina Escott-Price 9, 10 , Sudha Seshadri 11 ; the Alzheimer’s Disease Neuroimaging Initiative , Peter Holmans 10 , Ming T Tsuang 8 , William S Kremen 8 , Chris Gaiteri 6 , Stephen J Glatt 1, 2, 12 Author information Article notes Copyright and License information 1 Psychiatric Genetic Epidemiology & Neurobiology Laboratory (PsychGENe Lab), Department of Psychiatry and Behavioral Sciences, SUNY Upstate Medical University, Syracuse, NY, USA 2 Department of Neuroscience and Physiology, SUNY Upstate Medical University, Syracuse, NY, USA 3 Department of Internal Medicine, Erasmus MC, University Medical Center Rotterdam, Rotterdam, the Netherlands 4 Norton College of Medicine, SUNY Upstate Medical University, Syracuse, NY, USA 5 Department of Cognitive Science, University of California San Diego, La Jolla, CA, USA 6 Department of Psychiatry and Behavioral Sciences, SUNY Upstate Medical University, Syracuse, NY, USA 7 Departments of Psychiatry and Radiology, University of California San Diego, La Jolla, CA, USA 8 Department of Psychiatry, University of California San Diego, La Jolla, CA, USA 9 Dementia Research Institute, School of Medicine, Cardiff University, Cardiff, UK 10 Division of Psychological Medicine and Clinical Neurology and Medical Research Council (MRC) Centre for Neuropsychiatric Genetics and Genomics, School of Medicine, Cardiff University, Cardiff, UK 11 Department of Neurology, School of Medicine, Boston University, Boston, MA, USA 12 Department of Public Health and Preventive Medicine, SUNY Upstate Medical University, Syracuse, NY, USA Author Contributions J.H., C.G., and S.J.G. contributed to the conception and design of the study. The Alzheimer’s Disease Neuroimaging Initiative (ADNI) provided data for the ADNI study. J.G.J.V.R provided data for the Rooij et al ., 2018 study. J.H. performed the analyses for this study. C.Z. conducted quality control and processing for RNA-sequencing data. P.H. performed the gene-set analyses of genome-wide association study and provided summary statistics. J.L.H developed the algorithms for polytranscript risk scores. J.H. wrote the first draft of the manuscript. C.G. and S.J.G. provided critical edits of the manuscript. J.L.H., C.Z., G.C.H., C.C.F, S.V.F, C.F.N, S.J.L, V.E.P, S.S., P.H., M.T.T, and W.S.K. provided critical review of the manuscript. All authors gave final approval of the version of the manuscript submitted for peer-review. Issue date 2025 Jun. PMC Copyright notice PMCID: PMC12048288  NIHMSID: NIHMS2038543  PMID: 39679839 The publisher's version of this article is available at Am J Med Genet B Neuropsychiatr Genet Abstract The comprehensive genome-wide nature of transcriptome studies in Alzheimer’s disease (AD) should provide a reliable description of disease molecular states. However, the genes and molecular systems nominated by transcriptomic studies do not always overlap. Even when results do align, it is not clear if those observations represent true consensus across many studies. A couple of sources of variation have been proposed to explain this variability, including tissue-of-origin and cohort type, but its basis remains uncertain. To address this variability and extract reliable results, we utilized all publicly available blood or brain transcriptomic datasets of AD, comprised of 24 brain studies with 4,007 samples from six different brain regions, and eight blood studies with 1,566 samples. We identified a consensus of AD-associated genes across brain regions and AD-associated gene-sets across blood and brain, generalizable machine learning and linear scoring classifiers, and significant contributors to biological diversity in AD datasets. While AD-associated genes did not significantly overlap between blood and brain, our findings highlighted 15 dysregulated processes shared across blood and brain in AD. The top five most significantly dysregulated processes were DNA replication, metabolism of proteins, protein localization, cell cycle, and programmed cell death. Conversely, addressing the discord across studies, we found that large-scale gene co-regulation patterns can account for a significant fraction of variability in AD datasets. Overall, this study ranked and characterized a compilation of genes and molecular systems consistently identified across a large assembly of AD transcriptome studies in blood and brain, providing potential candidate biomarkers and therapeutic targets. Keywords: Alzheimer’s disease, blood, brain, gene expression, transcriptome, meta-analysis, classification Introduction The intention of Alzheimer’s disease (AD) transcriptome studies was to provide an actionable basis for the disease that goes beyond canonical amyloid plaques and tau tangles. Despite the genome-wide nature of these studies, it is still difficult to determine the precursors or consequences of AD in the brain or blood, in part due to inconsistent results from these studies. For example, genes that subserve mitochondrial activities were found to be dysregulated in blood in several biomarker studies of AD 1 – 5 , but not in a pair of others 6 ; 7 . In addition, inflammation- and immune-related pathways were implicated in AD participants in multiple brain transcriptomic studies 8 – 11 , but were not observed in some others 12 ; 13 . These disparate findings in prior studies could be a function of methodological differences, such as sample size, profiling platform, analytical approaches, or a function of biological differences among diverse pools of study participants or brain-tissue sites. Therefore, in the present study, we used a meta-analytic approach to identify the most reliable transcriptomic abnormalities of AD - in each of the six brain regions and blood - as well as those that are consistent between blood and brain. Because machine-learning techniques are becoming a common way to encapsulate differential expression, we applied them to blood-transcriptome studies to understand their potential for aiding in the classification of AD diagnosis. Finally, to understand the origins of variation in AD, we measured the overall molecular organization of each dataset, and evaluated that as a possible factor in the production of divergent conclusions across studies. Systematically combining gene-expression profiles by meta-analysis is a standard statistical technique to increase power and provide more precise and generalizable estimates. Gene-expression differences with small—yet reliable—effects may be undetectable in a single study due to lack of power but become significant with the increased sample size of a meta-analytic data set. Consistent and rigorous pre-processing and quality-control procedures across individual studies can also limit false-positives within the meta-analysis, and potentially overcome the variation imparted by methodological differences. Multiple meta-analyses of AD transcriptomic studies were reported previously but had some limitations that we seek to overcome. These limitations are broadly related to the lack of integration of blood and brain analysis, lack of integration with machine learning approaches, lack of consistent brain region taxonomy, and lack of direct tests on the origin of such heterogeneity. Most of these prior analyses were constrained to either peripheral blood or brain separately, and did not attempt integration of findings across tissues 14 – 26 . Paranjpe et al ., 2021 27 did compare the gene-expression signatures of AD between blood and brain, but they adopted a sex-stratified meta-analysis approach, which only included a small set of publicly available data - three blood studies and four brain studies. The accessibility of blood transcriptomes naturally raises the question of AD classifiers. Several studies have investigated this, with a wide range of reported AD diagnosis classification accuracy (0.66–0.98) 3 ; 6 ; 7 ; 17 ; 28 – 33 . Yet, with the public availability of large samples such as the AddNeuroMed study, half of all published AD blood transcriptomic biomarker studies relied solely on this study for model training and testing 17 ; 29 – 32 ; 34 . We aimed to improve upon this approach and generate accurate yet generalizable classifiers using a large dataset with more diverse characteristics. Further, we addressed the inconsistent taxonomy of brain regions within AD brain studies. Previous AD brain meta-analyses used the brain region taxonomy from the original studies, which could misalign the meta-analysis results. For instance, brain samples from Brodmann areas 8, 10, and 11, or Brodmann areas 9 and 46, were all reported as superior frontal cortex in two AD studies 13 ; 35 , while samples from Brodmann areas 9 and 46 were commonly reported as dorsolateral prefrontal cortex in multiple other AD studies 11 ; 36 – 38 . To infer which genes and pathways are consistently dysregulated in AD, we searched for common disease signals across blood and six brain regions using differential gene-expression meta-analysis, network analysis, and pathway-enrichment analysis. To do this, we collated a large sample of data from 32 transcriptomic studies, which included 1,566 samples of peripheral blood and 4,007 samples of brain. We tested the hypothesis that gene co-regulation patterns could identify and organize themes of gene-expression differences seen across AD studies. To gauge the potential of blood gene-expression as a useful source of AD biomarkers, we tested the hypothesis that transcriptomic perturbations could be exploited to construct diagnostic classifiers for AD using machine learning and our more easily explainable classification method: polytranscript risk-scoring 39 . Finally, to track why such heterogeneity exists, we utilized known and novel factors influencing gene expression to determine which are potentially driving study differences. Materials and Methods Literature Search for Blood and Brain Transcriptome Studies of AD We conducted an electronic literature search of the PubMed database up to February 2022 using combinations of keywords including “Alzheimer’s disease”, “gene expression”, “transcriptom*”, “RNA sequencing”, “microarray”, “blood” and “brain”. Meanwhile, we conducted a query of published gene-expression data sets in Gene Expression Omnibus (GEO), ArrayExpress, and Synapse platforms using the combination of terms: “Alzheimer’s disease”, “Homo sapiens”, “Expression profiling by array” or “Expression profiling by high throughput sequencing”. We retained studies that met the following inclusion criteria for further review: 1) blood studies that recruited affected cases with a clinical diagnosis of AD, and postmortem brain studies with clinically and pathologically diagnosed AD cases (diagnostic criteria of each study are shown in Supplementary Table 1 ); 2) studies that recruited unrelated, cognitively unimpaired (CU) comparison participants; and (3) studies that had available transcriptome-wide mRNA expression data (not quantitative PCR or single-cell sequencing) generated from whole blood or peripheral blood leukocytes, or bulk brain tissues from whole-tissue section or laser-capture microdissection. We retained a total of 32 studies for the primary analyses. After the initial data-freeze and completion of all analyses, we performed a second search of the literature with the same keywords (up to February 2023) to identify any newly available studies that might be useful for replicating results on the primary group of retrieved studies. This secondary search found a publicly available blood study by Nachun et al ., 2019 40 which was published as a preprint; thus, it was used as a replication dataset for evaluating AD-associated genes in the meta-analysis of blood. Supplementary Table 2 shows potentially useful studies excluded due to limitations on the available data, e.g ., they were publicly unavailable or inaccessible from corresponding authors, generated on less common platforms, missing essential phenotypic variables, or had small sample size (studies with fewer than six cases or comparison participants). Table 1 provides an overview of the sample sizes, profiling platforms, and demographics for the studies included in our analysis. For brain studies, we leveraged Brodmann areas and grouped them into six general regions ( i.e ., cerebellum [CER]; entorhinal cortex [ENT]; dorsolateral prefrontal cortex [DLPFC]; frontal cortex [FCX]; hippocampus [HIP]; lateral temporal cortex [TCX]). We combined studies that sampled different areas of the lateral temporal cortex in the analysis of TCX, which included Brodmann areas 20, 21, and 22. We combined studies that sampled Brodmann area 9 and/or 46 as the dorsolateral prefrontal cortex. We used other frontal cortex study samples in the analysis of the frontal cortex, which mostly included samples from Brodmann areas 10 and 11. Table 1. Demographics of studies included in the meta-analysis CU AD Tissue Study n Age: mean (SD) Female % n Age: mean (SD) Female % PC genes ( k ) Platform Blood AddNeuroMed batch-1 106 72.63 (6.72) 58.5 96 75.60 (6.86) 71.9 16,279 Illumina HumanHT-12 V3.0 AddNeuroMed batch-2 130 75.70 (5.99) 61.5 117 77.50 (6.70) 63.2 16,950 Illumina HumanHT-12 V4.0 ADNI 246 76.37 (6.51) 52.4 116 77.46 (7.66) 35.3 17,832 Affymetrix HG U219 Array Bai et al . (2014) 21 73.71 (4.08) 66.7 17 75.94 (6.73) 70.6 11,665 Affymetrix HG U133 Plus 2.0 Array Leandro et al . (2018) 13 77.31 (6.17) 76.9 22 79.36 (6.56) 63.6 13,254 Agilent 4×44K v2 026652 G4845A ROSMAP batch-1/2 87 85.28 (5.49) 66.7 55 86.23 (4.69) 65.5 12,207 Illumina HiSeq 2500 Samsudin et al . (2016) 90 75.23 (7.17) 47.8 90 77.68 (5.81) 50.0 17,486 Agilent-028004 SurePrint G3 8×60K Shigemizu et al . (2020) 91 71.29 (5.07) 45.1 269 79.55 (5.85) 68.4 9,564 Illumina Nova-Seq 6000 Total n 784 782 Replication Nachun et al . (2019) 247 73.72 (6.24) 55.9 203 72.97 (7.10) 51.2 7,922 Illumina HumanHT-12 V4.0 Cerebellum mayoeGWAS 177 71.74 (5.51) 35.6 197 73.59 (5.61) 51.3 15,821 Illumina Whole Genome DASL assay MAYOseq 63 82.73 (7.59) 41.3 79 82.47 (7.68) 59.5 14,072 Illumina HiSeq 2000 Patel et al . (2019) 17 77.82 (8.35) 41.2 37 81.86 (9.83) 54.1 3,602 Illumina HumanHT-12 V4.0 Semick et al . (2019) 22 67.78 (7.28) 31.8 41 80.25 (9.51) 51.2 13,794 Illumina HiSeq 2000 Zhang et al . (2013) 56 68.05 (6.63) 23.2 127 80.19 (8.89) 52.0 14,245 Rosetta/Merck Human 44k 1.1 microarray Total n 335 481 DLPFC Berchtold et al . (2014) 26 80.85 (10.34) 53.8 21 87.14 (6.19) 66.7 11,333 Affymetrix HG U133 Plus 2.0 Array Low et al . (2021) 9 84.67 (6.24) 55.6 9 85.67 (6.75) 55.6 18,841 Affymetrix Human Transcriptome Array 2.0 MSBB_array 8 79.75 (12.09) 50.0 16 87.88 (8.65) 81.2 5,279 Affymetrix HG U133 A Narayanan et al . (2014) 97 68.68 (6.79) 24.7 304 80.96 (8.38) 56.2 14,100 Rosetta/Merck Human 44k 1.1 microarray ROSMAP 68 83.42 (6.26) 45.6 321 88.23 (3.24) 73.2 14,085 Illumina HiSeq/Illumina NovaSeq 6000 Scheckel et al . (2016) 7 78.14 (10.57) 71.4 9 90.67 (8.77) 88.9 14,944 Illumina HiSeq 2500 Semick et al . (2019) 23 68.24 (7.01) 30.4 38 79.74 (9.66) 44.7 14,230 Illumina HiSeq 2000 Zhang et al . (2013) 57 67.95 (6.62) 22.8 127 80.19 (8.89) 52.0 14,245 Rosetta/Merck Human 44k 1.1 microarray Total n 295 845 Entorhinal Berchtold et al . (2014) 18 80.67 (10.34) 50.0 15 86.47 (5.46) 53.3 11,363 Affymetrix HG U133 Plus 2.0 Array Liang et al . (2007) 13 80.31 (9.20) 23.1 8 84.12 (5.54) 75.0 7,115 Affymetrix HG U133 Plus 2.0 Array Patel et al . (2019) 18 80.89 (8.14) 38.9 36 83.44 (9.25) 63.9 2,608 Illumina HumanHT-12 V4.0 Semick et al . (2019) 23 67.69 (7.05) 34.8 36 79.75 (8.97) 47.2 14,416 Illumina HiSeq 2000 Total n 72 95 Frontal Hokama et al . (2013) 12 80.83 (7.38) 50.0 14 93.29 (6.43) 50.0 17,960 Affymetrix Human Gene 1.0 ST Array Liang et al . (2007) 10 79.30 (10.73) 40.0 18 79.27 (8.00) 50.0 8,786 Affymetrix HG U133 Plus 2.0 Array McKay et al . (2019) 11 78.82 (8.84) 54.5 10 80.80 (8.15) 70.0 17,528 Agilent-039494 SurePrint G3 Human GE v2 8×60K MSBB_seq 17 78.65 (7.57) 52.9 120 83.91 (6.95) 66.7 14,861 Illumina HiSeq 2500 Patel et al . (2019) 15 78.20 (7.90) 33.3 39 82.38 (9.42) 66.7 3,134 Illumina HumanHT-12 V4.0 Total n 65 201 Hippocampus ACT 31 87.65 (6.92) 38.7 17 92.71 (7.15) 64.7 14,423 Illumina HiSeq 2500 Berchtold et al . (2014) 25 81.96 (9.99) 48.0 19 83.05 (8.48) 52.6 11,463 Affymetrix HG U133 Plus 2.0 Array Blalock et al . (2004) 9 85.33 (7.98) 22.2 15 83.67 (7.30) 73.3 5,846 Affymetrix HG U133 A Hokama et al . (2013) 6 76.83 (3.25) 50.0 6 93.67 (6.31) 50.0 17,960 Affymetrix Human Gene 1.0 ST Array Miller et al . (2013) 15 82.43 (6.36) 26.7 15 76.23 (8.94) 53.3 7,968 Illumina HumanHT-12 V3.0 Rooij et al . (2018) 9 78.89 (9.85) 44.4 18 75.44 (7.22) 72.2 14,227 Illumina HiSeq 2000 Semick et al . (2019) 22 68.41 (7.13) 31.8 36 80.20 (9.21) 50.0 14,470 Illumina HiSeq 2000 Total n 117 126 Temporal ACT 30 87.43 (6.26) 40.0 18 91.67 (7.21) 55.6 14,287 Illumina HiSeq 2500 Hokama et al . (2013) 13 77.38 (6.36) 61.5 9 90.00 (4.56) 44.4 17,960 Affymetrix Human Gene 1.0 ST Array Huntley et al . (2019) 70 87.00 (6.32) 47.1 218 83.38 (8.13) 44.5 14,092 Illumina HiSeq 2500 Lai et al . (2014) 8 80.12 (7.06) 37.5 8 72.62 (9.15) 37.5 18,667 Affymetrix Human Exon 1.0 ST Array mayoeGWAS 194 71.59 (5.59) 39.7 202 73.57 (5.53) 53.5 15,821 Illumina Whole Genome DASL assay MAYOseq 66 83.06 (7.59) 48.5 80 82.61 (7.65) 61.3 14,290 Illumina HiSeq 2000 McKay et al . (2019) 11 78.82 (8.84) 54.5 10 80.80 (8.15) 70.0 17,528 Agilent-039494 SurePrint G3 Human GE v2 8×60K MSBB_seq 14 77.07 (9.10) 57.1 106 83.19 (7.25) 63.2 14,890 Illumina HiSeq 2500 Nativio et al . (2018) 10 68.20 (5.12) 10.0 12 68.33 (6.04) 8.3 14,451 Illumina NextSeq 500 Nitsche et al . (2021) 14 72.14 (9.39) 42.9 16 81.56 (6.19) 62.5 13,412 Agilent-034879 ADchip_1.0 033934 Patel et al . (2019) 21 78.81 (8.87) 33.3 50 81.52 (9.91) 58.0 2,613 Illumina HumanHT-12 V4.0 Piras et al . (2019) 98 84.98 (6.90) 49.0 97 85.02 (6.75) 49.5 15,732 Illumina HumanHT-12 V4.0 Total n 549 826 Open in a new tab Abbreviations: AD: Alzheimer’s disease cases; CU: cognitively unimpaired comparisons; DLPFC: dorsolateral prefrontal cortex; SD: standard deviation; PC genes: protein-coding genes. Data Import and Quality Control As all the datasets in our analysis were from previous studies with various sampling and profiling procedures, we reprocessed data using a consistent set of tools and a consistent pipeline to mitigate unwanted between-study variance that could obscure biologically meaningful signals. Overall, we strove to derive a QC-adjusted and normalized matrix of continuously distributed expression values, allowing the combinability of data from disparate sources while retaining relative rank and distance measures. We applied quantile normalization to all microarray studies when non-normalized data were available. We normalized RNA-sequencing data to effective library size using a scaling factor called trimmed mean of M-values 41 . Due to the negative gene-expression values in some studies, we used the inverse hyperbolic sine function to transform all studies with accessible non-normalized gene expression values, given its robustness to the influence of outliers in right-skewed data. Five studies included in our analysis provided pre-normalized data, which were log 2 -transformed. When batch information was available, we adjusted batch effects using the ComBat function 42 of the R package SVA . We applied a data-driven de-convolution approach in the R package dtangle 43 to estimate the relative abundance of plasma and seven blood cell-types ( i.e ., B cell, T cell, dendritic cell, monocytes, mast cell, natural killer cell, and granulocytes), or brain cell-types (neuron, astrocyte, oligodendrocytes, microglia). To account for hidden confounding effects, we performed a surrogate variable analysis (SVA) 44 using the sva function of the R package SVA on each study to obtain surrogate variables (SVs). We detected the SVs using AD diagnosis as the variable of interest and other known covariates as the adjustment variables ( i.e ., age, sex, estimated cell proportions, race, RNA integrity number, postmortem interval, brain pH). We identified SVs that explained unmodeled variation of gene expression at the default p < 0.1 as covariates. We defined and excluded sample outliers if their component scores on each of the first three gene-expression principal components (PCs) were greater than four standard deviations from the mean. More details of the data-processing procedures can be found in Supplementary Methods . Transcriptome-Wide Meta-Analysis For each study, we fit a multivariate linear regression model that specified expression of each gene as a dependent variable, and diagnosis as an independent variable, using the R package limma 45 , outputting empiric Bayes moderated t -statistics. All subjects had age and sex information, which we included as covariates. When self-reported race/ethnicity, education years, RNA integrity number, postmortem interval, and brain pH were available, we added these as covariates. We included any identified SVs as covariates as described above. In addition, we used seven estimated blood-cell proportions or three estimated brain-cell proportions as covariates per study (the estimated proportions of plasma or one brain cell-type were excluded to avoid collinearity). Supplementary Table 3 documents the covariates included for each study in the analysis of differential gene expression. We applied this model to every gene per study to compute the adjusted mean difference in gene expression between AD cases relative to CU comparisons. To perform the meta-analysis in blood and six brain regions separately, we retained protein-coding genes that were present in at least three studies. Then we used a random-effects inverse-variance weighted meta-analysis model to pool differential-expression coefficient estimates and sampling variances of the coefficient estimates per gene across studies per tissue. We chose this model because it accounts for between-study variances and produces conservative estimates of significance in the possible presence of heterogeneity of coefficient estimates. We used the R package metafor 46 to perform the meta-analysis with the DerSimonian-Laird estimator. To decouple the effects of sample size on the number of detected differentially expressed genes (DEGs) across tissues, we calculated a DEG index by normalizing the number of DEGs by the effective total sample size ( n DEG s 1 1 n A D cases + 1 n CU s ). For DEGs exhibiting significant heterogeneity between studies (Cochran’s Q-test, p > 0.05), we conducted a leave-one-study-out sensitivity analysis to assess the stability of the meta-analysis results. For each examined gene that was present in n studies, we excluded one study from the primary meta-analysis and re-ran the meta-analysis using the remaining n -1 studies. We repeated this process n times, ensuring that each study was excluded once. Differential Expression of Weighted Gene Co-Expression Networks In biological systems, a group of genes can be correlated and exhibit similar expression patterns across different samples and conditions, which could be due to shared regulatory mechanisms or biological pathways. The analysis of gene co-expression networks offers several advantages over individual gene-expression analysis, including the ability to identify clusters of functionally correlated genes without reference to ontologies and increase the power of analysis by reducing the multiple-testing burden. For each tissue, we performed weighted gene co-expression network analysis (WGCNA) 47 on a single data matrix, which was generated from multiple normalized and scaled studies batch-adjusted with non-missing data. More details of the procedure are described in Supplementary Methods . We used the one-step automatic network detection method blockwiseModules with the default values except for the data-specific power value and these parameters: TOMType = “signed”, minModuleSize = 30, and corType = “bicor”. We chose an outlier-robust correlation method (bi-weight mid-correlation) to measure gene-gene correlations, which were raised to a soft-threshold power value ( β ) chosen by the scale-free fit index so that the correlation network approximated a scale-free topology. We identified clusters of highly correlated genes (called ‘modules’) using K -means clustering. PCs were obtained for each cluster using singular value decomposition. The first PC was selected as the module eigengene, which can be interpreted as a representation of the expression profiles of genes in a module. In each module, hub genes were defined as the genes with the highest connectivity with other genes within the module based on the following formula: ( k i = ∑ u ≠ i a ui where a ij = |(1 + cor( x i , x j )) / 2| β ). Here, we chose the top-ranked gene as the hub-gene for each module. In each tissue, we fit a linear regression model to estimate the effect of diagnostic group on module eigengene expression levels while simultaneously adjusting for potential effects of age, sex, estimated cell proportions, and SVs. We examined the enrichment of AD-associated brain modules for transcriptomic markers of eight major brain cell-types ( i.e ., endothelial cells, pericytes, oligodendrocytes, oligodendrocyte progenitor cells, microglia, astrocytes, excitatory neurons, and inhibitory neurons) using one-tailed Fisher’s exact tests. We downloaded the cell-type marker genes from PsychENCODE 48 , which were generated from single-cell gene-expression data in unique molecular identifier (UMI) units. Similarity of Module Partitions and Gene Effect Sizes Across Studies As the data used in the WGCNA analysis were combined from multiple studies of the same tissue-of-origin, we investigated the preservation of gene-module partitions across studies (a partition is the full gene set divided into a set of clusters). We retained 41 datasets that had at least 8,000 protein-coding genes for this analysis, including eight blood datasets and 33 brain datasets (from 20 independent brain studies) covering six brain regions. We kept a total of 3,101 protein-coding genes that were present in all 41 datasets in each study/sub-study to identify gene-modules using WGCNA. We performed a gene-set over-representation analysis to investigate whether the 3,101 common genes were significantly enriched in processes given the whole transcriptome as the background. The results of this analysis are described in the Supplementary Results . We used normalized mutual information (NMI) to measure the similarity of gene-module partitions between any two of the 41 datasets. To further investigate whether module partitions can account for the differences in AD-associated genes across datasets, we measured the DEG similarity between all pairs in the 41 datasets, by calculating the Spearman’s r using effect sizes of the 3,101 common genes from the differential gene-expression analysis. We assessed the association between module partition ( i.e ., NMI values) and DEG similarity ( i.e ., correlation coefficients for effect sizes between studies) using Spearman’s correlation. To determine if the heterogenous signals across studies were influenced by technical variability, we examined the association between study cluster assignments (based on gene-partition similarity or DEG effect-size similarity) and the profiling platforms ( i.e ., array vs . sequencing) or profiling library types ( i.e ., Affymetrix array, Illumina array, Agilent array, Rosetta/Merck array, and Illumina sequencing). We examined the associations across eight blood datasets and across 33 brain datasets of six brain regions. We used Chi-squared tests to evaluate the associations between profiling platforms (array vs . sequencing) and cluster assignments ( K = 2). We used an extension of Fisher’s exact test to evaluate the associations between profiling library types and cluster assignments ( K = 4 in blood or 5 in brain). We fit linear regression models to assess the interaction between AD diagnosis and study clusters, as defined by their NMI similarity. We combined scaled gene-expression data of the eight blood datasets to generate a single data matrix. Similarly, we merged scaled gene-expression data of the 33 brain datasets for the downstream analysis. We estimated the missing gene-expression values for any genes with a missing rate of no more than 20% across the samples using the SVDimpute algorithm 49 of the R package pcaMethods , which utilizes the linear combination of top principal components. Based on the inflection point of the scree plot identified by the find_curve_elbow function of the R package pathviewr 50 , we used the top seven PCs to impute the missing gene-expression values for both blood and brain. After imputation, we had 1,566 individuals and 8,518 protein-coding genes for blood, and 3,673 individuals and 9,982 protein-coding genes for brain. In the model using blood data, we used age, sex, estimated cell proportions, AD diagnosis, study cluster, and the interaction between AD diagnosis and study cluster as the independent variables and the gene-expression level of each gene as the dependent variable. In the analysis of brain, we added brain region as an additional independent variable. Gene-set Over-Representation and Enrichment Analysis We used the R package piano 51 for gene-set over-representation analysis using summary statistics obtained from our differential gene-expression meta-analysis of protein-coding genes. Effect sizes and p -values of differential expression were supplied to the method runGSA , which computed the gene-set-level statistics across Gene Ontology (GO) terms and Reactome pathways. We performed three enrichment tests: 1) to test whether genes differentially expressed between AD cases and CU comparisons were significantly enriched in gene-sets, regardless of the direction of the differential-expression changes observed in the genes mapped in the gene-set; 2) to test whether genes up-regulated in AD cases were significantly enriched in gene-sets; and 3) to test whether genes down-regulated in AD cases were significantly enriched in gene-sets. The gene-set-level statistic in the non-directional enrichment test was the mean raw p -values of the genes mapped to the gene-set. In the directional enrichment test, we derived the gene-set-level statistic from the mean of transformed p -values. For the upregulation (downregulation) test, the method transformed the raw p -values of the up (down)-regulated genes ( i.e ., p /2) to range from 0 to 0.5, and the raw p -values of the down (up)-regulated genes were transformed ( i.e ., 1- p /2) to range between 0.5 and 1. To reduce the number of broad biological categories tested and to provide better insight into specific underlying biological processes and molecular functions, we limited gene-sets to those with a range of 10 to 100 genes. To determine the significance of the enrichment test, gene labels were permuted 100,000 times, and mean differential-expression statistics were recomputed for each permuted gene-set. We performed a separate gene-set over-representation test for WGCNA modules whose eigengene expression was significantly associated with AD diagnosis, while adjusting for covariates. For each AD-associated module within each tissue, we used a hypergeometric test to evaluate the enrichment of GO terms and Reactome pathways among genes in the module. The p -value of enrichment was calculated as follows: = 1 - ∑ i = 0 k - 1 M i N - M n - i N n , where k is the number of genes from a module that overlapped with a GO or Reactome gene-set, M is the number of genes in a targeted gene-set, n is the total number of genes in a tested module, and N is the total number of genes in gene-sets. For Reactome pathways significantly over-represented by differentially expressed genes or gene modules, we classified these pathways into a set of the highest-order umbrella pathways, called root-pathways, using the hierarchical relationship between pathways in the Reactome database. Collapsing these detailed sub-pathways into the broader root-pathways can provide a clearer view of the dysregulated biological processes in AD across tissues. As different root-pathways contain various numbers of sub-pathways, we used a normalized count to account for the sub-pathway size of root-pathways in the enrichment analysis. For each root pathway in each tissue, we calculated the normalized count of significantly enriched sub-pathways as Number of significantly enriched sub - pathways Total number of tested sub - pathways × 100 % , which indicates the number of significantly enriched sub-pathways per 100 tested sub-pathways within the root pathway. To provide a synopsis of the dysregulated pathways in AD across blood and brain, we performed a gene-set enrichment analysis (GSEA) 52 using the R package piano . For genes that were present in at least three of the seven tissue types ( i.e ., blood and six brain regions), we calculated the median of effect sizes from the meta-analysis. We performed two distinct directional enrichment analyses using 29 Reactome root-pathways as the reference gene-sets. Similarly, to provide an overview of dysfunctional cell types in AD brain, we performed a GSEA using the gene markers of eight major brain cell-types ( i.e ., endothelial cells, pericytes, oligodendrocytes, oligodendrocyte progenitor cells, microglia, astrocytes, excitatory neurons, and inhibitory neurons) as the reference gene-sets. The gene markers of brain cell-types were downloaded from PsychENCODE as described above. For the cell-type enrichment analysis, we computed the median of effect sizes from the meta-analysis for genes that were present in at least three of the six brain regions. We performed two distinct directional enrichment analyses to identify brain cell-types enriched by up- or down-regulated genes. Gene-Set Analysis of Genome-Wide Association Study (GWAS) data We examined whether AD-associated genes identified from the meta-analysis or genes comprising AD-associated modules were enriched with AD GWAS signals using MAGMA 53 . We used AD GWAS meta-analysis summary statistics from Bellenguez C. et al ., 2022 54 for the GWAS enrichment analysis, which included both clinically diagnosed AD cases and controls and “proxy” AD cases and controls from the UK Biobank. We used a window of 35kb upstream/10kb downstream around the longest transcript to assign SNPs to genes, which was intended to capture potential regulatory regions. Polytranscript Risk-Scoring We used our polytranscript risk score (PTRS) methodology, which was originally developed and evaluated in the setting of major psychosis 39 , to generate a de novo standardized score that measures the AD-like status of an individual’s transcriptome profile. PTRSs were derived using blood and lateral temporal cortex (TCX) data, which are the two tissue types that had the largest sample sizes and relatively balanced case/control ratios. Our method is analogous to polygenic risk scoring 55 in that it yields a single score per individual that represents a weighted sum of disease-related effects over multiple risk genes; however, our method applies to transcriptome-wide gene-expression data as opposed to polygenic risk scores that use genome-wide genotypes. We retained normalized, scaled, and batch-adjusted transcriptome-wide gene-expression data (excluding missing values) for our classification analyses. Prior to downstream analyses ( i.e ., gene feature-selection), we split the entire data set into independent training and held-out test sets, which roughly approximated an 80:20 or 70:30 split ( Supplementary Table 4 ). In each analysis (blood or TCX), we set aside a study with a balanced case/control ratio as test data to ensure that the final model could be evaluated in an out-of-bag sample. This poses a stringent test for the classification models and is critical for assessing classification performance in real-world conditions. The training set in blood included four studies with a total of 881 subjects ( n AD cases = 353; n CU comparisons = 528), and the test set (AddNeuroMed batch-2) comprised 247 subjects ( n AD cases = 117; n CU comparisons = 130). In TCX, the training set included 871 subjects from nine studies ( n AD cases = 554; n CU comparisons = 317), and the test set (mayoeGWAS) included 395 subjects ( n AD cases = 201; n CU comparisons = 194). In addition to binary AD diagnosis, we derived polytranscript risk-scoring models to predict Braak scores 56 in TCX. We retained normalized, scaled, and batch-adjusted transcriptome-wide gene-expression data (excluding missing values) across six TCX studies that have Braak stage scores, including ACT, Hokama et al ., 2013, Huntley et al ., 2019, MAYOseq, MSBB_seq, and Nativio et al ., 2018 ( Table 1 ; Supplementary Table 1 ). The combined TCX dataset included 598 individuals and 13,210 protein-coding genes. We set aside the MSBB_seq study as the held-out test data, which included 120 individuals. In the training set only, we identified the scoring weights of genes via linear regression with AD diagnosis or Braak scores as an independent variable, and covariates including the measured variables age and sex, as well as estimated abundances of cell types. To compute AD PTRSs or Braak PTRSs, we then summed the observed expression of genes weighted by their t -standardized difference ( i.e ., t -score) in mean expression between AD cases and CU comparisons or with one unit increase of Braak scores. To keep the most informative predictors for the model, we generated PTRSs using a range of ranked discriminant genes ( i.e ., 2, 5, 10, 15, 25, 50, 100, 125, 150, 200) as selected features. We ranked these genes by p -values and selected top-ranked genes only using the training datasets to prevent data leakage. We supplied each AD PTRS to a logistic regression model in the independent held-out test set with AD status specified as the outcome variable to evaluate generalizability. We measured AD diagnostic classification performance using the area under the curve (AUC) of the receiver operating characteristic (ROC) plot. For each Braak PTRS, we evaluated its predictive performance on Braak scores in the held-out test data using a linear regression model, with Braak scores as the outcome variable and each Braak PTRS as the predictor. To directly compare model performance with Beebe-Wang et al ., 2021 57 , we used the same predictive performance error metric for Braak PTRSs, which is 1- R 2 . This was calculated by dividing the mean squared error (MSE) by the variance of Braak scores. In the held-out test set, we examined the differences of best-performing AD PTRSs between apolipoprotein E ( APOE ) ε4 carriers and non-carriers using logistic regression models with APOE -ε4 carrier status as the dependent variable. We used a linear regression model to test the association between Mini–Mental State Examination (MMSE) scores (dependent variable) and the best-performing AD PTRSs of blood (independent variable) in the test set, with full-time education years, sex, and age as covariates. In the test data, we examined whether the best-performing Braak PTRS ( i.e ., with the lowest 1- R 2 ) was predictive of AD diagnosis using a logistic regression model with AD diagnosis as the dependent variable. We next examined the hypothesis that the AD PTRS scoring weights derived in TCX could be transferred to blood data and achieve significant AD diagnostic classification. We also derived Braak PTRSs in blood data using Braak PTRS scoring weights trained in TCX. In the scoring formula for TCX, we retained the genes that were present in the blood held-out test set (AddNeuroMed batch-2). We generated AD or Braak PTRSs using the scoring weights from TCX over a range of ranked discriminant genes ( i.e ., 10, 50, 100, 150, 200, 300, 400, 500, 600, 700) for the individuals in AddNeuroMed batch-2. Each AD PTRS was supplied to a logistic regression model with AD status specified as the outcome variable to evaluate the classification performance. We evaluated the association between each transferred Braak PTRS and AD diagnosis using a logistic regression model with AD status specified as the outcome variable. Diagnostic Classification Using Machine Learning To further assess the biomarker utility of blood-based gene expression for diagnosing AD, we evaluated the performance of three supervised machine learning algorithms (MLAs) in discerning AD cases from CU comparisons. The algorithms of choice for our classification analysis were random forest (RF), linear-kernel support vector machine (SVM), and XGBoost 58 . We evaluated the generalizability of the MLAs using the same training/testing framework and blood datasets as described in the PTRSs subsection above. We adjusted the blood gene-expression data in the training set by regressing out the effects of age, sex, study, and estimated abundances of cell types. After splitting the training/testing datasets, we ranked genes by differences in mean expression between AD cases and CU comparisons according to a two-tailed t -test that was performed in the training set only. To compare the predictive performances between MLAs and blood PTRSs, we evaluated the same sets of top-ranked discriminant genes ( i.e ., 2, 5, 10, 15, 25, 50, 100, 125, 150, 200). For each MLA, we performed a 5-fold cross-validation approach within the training set to identify a model that best separated the two diagnostic groups. Model-specific hyperparameters included the number of decision trees (100, 200, 500, 1000) for RF, cost (0.001, 1.001, 2.001) for SVM, and learning rate (0.01, 0.0825, 0.155, 0.2275, 0.3) for XGBoost. We selected the model for each MLA that yielded the highest cross-validation classification accuracy to deploy in the test set. Significance Criteria and Multiple-Testing Corrections For each tissue, we used a false-discovery rate (FDR) of 0.05 for the multiple-testing corrections for differentially expressed genes and enriched gene-sets. For each tissue, we used Bonferroni-correction in identifying AD-associated WGCNA modules, and cell-type enrichment analysis of AD-associated modules. In the GWAS gene-set analysis, we used a family-wise Bonferroni-correction to adjust for the number of tissues or the number of AD-associated modules. We reported FDR-adjusted p -values as “FDR- p ” and Bonferroni-adjusted p -values as “Bonf- p ”. We measured classification accuracy using the area under the curve (AUC) of the receiver operating characteristic (ROC) plot. We calculated p -values for AUCs via one-tailed z -tests with 0.5 as the null. Results Genes Differentially Expressed in AD Our meta-analysis uncovered a total of 8,575 significant DEGs in AD cases relative to CU comparisons (FDR- p < 0.05) combined across blood and six brain regions, which included 6,634 unique DEGs found in at least one tissue. We identified a total of 6,618 unique genes that were differentially expressed in at least one of the six brain regions, and 23 DEGs in blood. In the brain, 309 unique genes were differentially expressed in at least half of the six brain regions. The summary statistics for all DEGs are provided in Supplementary Table 5 . DLPFC accounted for the highest share of DEGs (45.5%); the tissue with the lowest share of DEGs (0.2%) was FCX, which comprises other regions of the frontal cortex outside of DLPFC. Figure 1 shows the number of DEGs detected in each tissue type and the number of DEGs shared between tissues. The DEG index (the number of DEGs normalized by sample size) ranged from 0.06 to 17.84, with the smallest value in blood and the largest value in DLPFC. This implied that, under the condition of infinite sample size, the number of identifiable genes with true differential expression in AD is the smallest in blood and the largest in DLPFC. The effect size and p -value of the top 10 most significant DEGs in each tissue are shown as a volcano plot in Figure 2 . Across seven tissue types (i.e., blood and six brain regions), we did not identify significant enrichment (Bonf- p < 0.05) of GWAS signals in genes significantly differentially expressed in AD. Figure 1. The number of DEGs in each tissue and the number of DEGs shared between tissues. Open in a new tab The histogram on the left shows the total number of genes that were significantly dysregulated in AD cases compared with CU comparisons (FDR- p < 0.05) in each of the seven tissues. Beside each tissue label, the number inside the parentheses indicates the total sample size in the meta-analysis. The histogram on the top shows the number of DEGs that are unique in a tissue or shared across tissues. When the top histogram column matches with a stand-alone dot that is not connected by lines, the column represents the number of unique DEGs in a tissue. When the top histogram column matches with linked dots, the column represents the number of DEGs that were identified across two or more tissues. The red asterisk indicates a significant overlap of DEGs across tissues (one-tailed Fisher’s exact test, Bonf- p < 0.05). Abbreviations: DEGs: differentially expressed genes; AD: Alzheimer’s disease; CU: cognitively unimpaired; CER: cerebellum; DLPFC: dorsolateral prefrontal cortex; ENT: entorhinal cortex; FCX: frontal cortex; HIP: hippocampus; TCX: temporal cortex; FDR- p : false discovery rate-adjusted p -value; Bonf- p : Bonferroni-corrected p -value. Figure 2. Volcano plot of the top 10 most significant DEGs in each meta-analysis. Open in a new tab The horizontal axis shows the effect size of gene-expression differences between AD cases and CU comparisons in each meta-analysis. The vertical axis indicates the transformed FDR-corrected p -value of the effect-size estimate. A blue horizontal dotted line denotes FDR- p = 0.05. DEGs in different tissues are coded with different colors according to the key. Abbreviations: DEGs: differentially expressed genes; AD: Alzheimer’s disease; CU: cognitively unimpaired; FDR- p : false discovery rate-adjusted p -value. During the second literature search of available studies, we found only one new study of blood that was available; thus, we used it to examine whether the DEGs identified in the meta-analysis of blood were replicable (no new available brain studies were identified). Among the 23 DEGs identified in blood, 22 genes were present in the replication study of Nachun et al ., 2019 ( n AD cases = 203, n CU comparisons = 247; Table 1 ). While none of the 22 genes was significantly replicated (FDR- p < 0.05), the gene ELL (Elongation Factor For RNA Polymerase II) showed the strongest evidence of dysregulation ( β = 0.024, CI = [−1.5e-03, 0.049], p = 0.066), and the effect sizes of the 22 genes from meta-analysis were significantly positively correlated with the ones from the replication study (Pearson’s r = 0.52, p = 0.014). Overall, the lack of significant replication suggests limited generalization and substantial heterogeneity in gene differential effects across blood transcriptomic studies of AD. Across blood and six brain regions, TCX and DLPFC shared the most DEGs in common (count = 676). We further examined Spearman’s correlation of per-gene effect sizes for the whole transcriptome between pairs of tissues ( Supplementary Figure 1 ; Supplementary Table 6 ). All pairs of tissues were significantly correlated except for blood-DLPFC ( p = 0.097). We observed larger correlations between all pairs of brain regions than between blood and any one brain region. The Spearman’s correlation coefficients between brain regions and blood ranged from 0.013 to 0.150, while the correlations were between 0.202 and 0.514 for pairs of brain regions. We identified seven DEGs that were shared between blood and brain. However, the overlap of DEGs between blood and brain was not significant (Fisher’s exact test, one-tailed p = 0.74). Across blood and the six brain regions, the differential expression of most AD-associated genes did not show significant evidence of heterogeneity between studies (Cochran’s Q-test, p > 0.05), except for 7.4% (289/3901) of genes in DLPFC, 33.7% (999/2967) of genes in TCX, and 8.2% (124/1521) of genes in CER ( Supplementary Figure 2 ). The leave-one-study-out sensitivity analysis results for all the DEGs with significant study-level heterogeneity are shown in Supplementary Table 8 . Leave-one-study-out sensitivity analysis for the 999 heterogenous DEGs in TCX indicated that pooled effects were relatively stable for most genes, as only seven DEGs failed to remain significant ( p < 0.05) in these analyses. For DLPFC, the association between 122 DEGs and AD did not remain significant ( p < 0.05), and 121 of those genes lost significance after removing one of the largest DLPFC studies, which may indicate the lost significance in the sensitivity analysis was due to reduced power. Specifically, 106 out of the 121 DEGs lost significance ( p < 0.05) when Narayanan et al ., 2014 ( n AD = 304; n CU = 97) was excluded, 71 out of the 121 DEGs lost significance when Zhang et al ., 2013 ( n AD = 127; n CU = 57) was excluded, and 23 out of the 121 DEGs lost significance when ROSMAP ( n AD = 321; n CU = 68) was excluded. For CER, the association between 18 DEGs and AD no longer remained significant in sensitivity analyses ( p < 0.05). Across the three brain regions, the fraction of DEGs that had significant heterogeneity and lost significance in the sensitivity analysis was small ( i.e ., 1.2% in CER; 3.1% in DLPFC; 0.24% in TCX; Supplementary Figure 2 ), and the direction of pooled effect sizes of these genes remained stable after removal of any given study ( Supplementary Table 8 ). Gene Modules Associated with AD The WGCNA approach generated a variable number of co-expression modules across blood and the six brain regions: 20 in blood; 46 in CER; 18 in DLPFC; 6 in ENT; 21 in FCX; 18 in HIP; 12 in TCX). Among all the gene modules, we detected 57 whose eigengene values were significantly different (Bonf- p < 0.05) in AD cases compared with CU comparisons while adjusting for covariates ( i.e ., age, sex, estimated cell proportions, and any detected SVs). Of the differentially expressed modules, 34 had up-regulated eigengene expression in AD cases (5 in blood; 10 in CER; 8 in DLPFC; 5 in HIP; 6 in TCX; no up-regulated module was identified in ENT), while the other 23 showed down-regulation in AD cases (3 in blood; 9 in CER; 5 in DLPFC; 1 in ENT; 4 in HIP; 1 in TCX). The results of the association between WGCNA module eigengenes and AD diagnosis are shown in Supplementary Table 10 . The number of genes comprising the AD-associated modules ranged from 1,095 to 3,621 across blood and the five brain regions ( Supplementary Figure 3 ). Similar to the meta-analysis, we observed the smallest number of AD-associated genes ( i.e ., genes comprising the AD-associated modules) in blood and the largest number in DLPFC. With the exception of DLPFC, there were generally more genes mapped to AD-associated modules than DEGs from individual gene differential-expression meta-analyses in blood and four brain regions ( i.e ., CER, ENT, HIP, TCX); the ratio ranged from 1.1 in TCX to 108.7 in HIP. We compared the biological similarities of AD-associated modules based on the top five significantly over-represented Reactome pathways ranked by p -values ( Supplementary Figure 5 ). Across blood and brain modules that were up-regulated in AD, we found nine modules that were characterized by immune system pathways, including blood Module 14, CER Modules 18, 24 and 25, DLPFC Modules 12 and 15, HIP Module 2, and TCX Modules 8 and 11. Across modules that were down-regulated in AD, we identified two modules ( i.e ., DLPFC Module 5 and TCX Module 1) that were featured with dysregulation of genes involved in NMDA signaling and cardiac conduction. In addition, we identified five down-regulated modules associated with protein metabolism pathways ( i.e ., blood Modules 3 and 5, CER Module 26, DLPFC Modules 1 and 18) and seven down-regulated modules related to metabolic pathways ( i.e ., CER Modules 35, 38 and 42, DLPFC Module 14, ENT Module 1, HIP Modules 1 and 17). Out of the 57 AD-associated modules in the brain, 28 modules (6 in CER; 10 in DLPFC; 1 in ENT; 5 in HIP; 6 in TCX) were enriched for gene markers of at least one of eight major brain cell-types (Bonf- p < 0.05). Across the five brain regions, the number of unique cell types enriched by AD-associated modules ranged from two to eight. Specifically, two cell types ( i.e ., excitatory neurons and inhibitory neurons) were enriched by the one and only AD-associated module in ENT. Ten modules in DLPFC were enriched across eight unique brain cell-types, with each cell type being enriched by at least one module. Among the eight brain cell-types, astrocyte markers were enriched by the greatest number of AD-associated brain modules (count = 9). The cell types enriched by each module and the p -values are shown in Supplementary Table 11 . Hub genes can be defined as the gene nodes that have the highest connectivity with other genes inside a gene co-expression module, which could be important drivers or regulators of disease-associated biological processes. We identified 10 hub genes from AD-associated WGCNA modules in CER, DLPFC, and TCX that were also DEGs identified in the same region ( Supplementary Table 12 ). Among them, prior studies suggested some genes may play a role in the etiology of AD. For example, we found TCF7L1 (Transcription Factor 7 Like 1), the hub gene of Module 3 in DLPFC, was up-regulated in AD. Increased expression of TCF7L1 was implicated in impaired WNT signaling that contributes to AD pathogenesis 59 , such as the production and clearance of A β , neuroinflammation, and synapse vulnerability 60 . In line with this, we found the Reactome pathway “Repression of WNT target genes” was over-represented by the genes comprising Module 3 in DLPFC ( Figure 3A ; Supplementary Table 13 ). Additionally, FPR1 (Formyl Peptide Receptor 1) was the hub gene for Module 8 in TCX, and it was downregulated in AD patients in our meta-analysis. FPR1 is known to modulate phagocytosis and A β uptake 61 . Similarly, we found seven phagocytosis-related Reactome pathways that were significantly over-represented by the genes comprising Module 8 in TCX ( Figure 3A ; Supplementary Table 13 ). Interestingly, Module 8 in TCX was the only AD-associated module significantly enriched with AD GWAS signals (Bonf- p = 0.011, Supplementary Table 14 ). Fifteen gene members in TCX Module 8 were significantly enriched with AD GWAS signals (Bonf- p < 0.05, Supplementary Table 15 ). Among them, four genes are known candidate risk genes in AD that play critical roles in microglia efferocytosis ( i.e ., the phagocytic clearance of apoptotic cells by microglia) 62 , including TREM2 (Triggering Receptor Expressed On Myeloid Cells 2, Bonf- p = 1.7e-09), PLCG2 (Phospholipase C Gamma 2, Bonf- p = 8.1e-08), SPI1 (Spi-1 Proto-Oncogene, Bonf- p = 9.5e-08) and BLNK (B Cell Linker, Bonf- p = 3.3e-04) ( Figure 3B ). Figure 3. Over-represented Reactome pathways and GWAS gene-set analysis highlight phagocytosis-related Module 8 in TCX. Open in a new tab In line with the biological roles of hub genes in AD, panel A shows the WNT signaling-related Reactome pathway that was significantly (FDR- p < 0.05) over-represented by genes comprising Module 3 in DLPFC, and phagocytosis-related Reactome pathways that were significantly (FDR- p < 0.05) over-represented by genes comprising Module 8 in TCX. Bar colors indicate different Modules. The dotted line annotates FDR- p -value 0.05. In panel B, the x-axis indicates the transformed Bonferroni-corrected p -values of gene-level GWAS enrichment for Module 8 in TCX. A grey vertical dotted line denotes Bonf- p = 0.05. The dots in red color are known microglia-related candidate risk genes in AD. Abbreviations: AD: Alzheimer’s disease; GWAS: genome-wide association study; DLPFC: dorsolateral prefrontal cortex; TCX: temporal cortex; FDR- p : false discovery rate-adjusted p -value; Bonf- p : Bonferroni-corrected p -value. Gene-Sets Associated with AD Gene-set over-representation analysis using meta-analysis summary statistics of DEGs across seven tissues identified a combined total of 199 gene-sets over-represented in AD cases (FDR- p < 0.05) that were distinctly up-regulated ( Supplementary Tables 16 and 17 ). A total of 1,323 distinctly down-regulated gene-sets were significantly over-represented in AD cases relative to CU comparisons (FDR- p < 0.05) ( Supplementary Tables 18 and 19 ). The Reactome pathways and GO terms that were significantly enriched by DEGs without incorporating the directionality of gene-expression changes are shown in Supplementary Tables 20 and 21 , respectively. Gene-set over-representation analysis across blood and five brain regions uncovered a total of 1,443 gene-sets significantly enriched by up-regulated WGCNA co-expression modules, and 2,423 gene sets significantly enriched by down-regulated WGCNA co-expression modules (FDR- p < 0.05). Among them, four unique GO terms and 191 unique Reactome pathways were shared between up-regulated and down-regulated co-expression modules, as some pathways were over-represented by both up- and down-regulated modules of the same or different tissues. In the case of the same tissue, the pathways could be enriched by two different sets of genes that were dysregulated in opposite directions. For example, in DLPFC, “ER-phagosome pathway”, which participates in antigen processing and cross-presentation, was enriched by a set of down-regulated proteasome subunit genes in a down-regulated module, as well as by a set of up-regulated innate immunity-related genes in an up-regulated module ( Supplementary Table 22 ). The complete set of Reactome pathways and GO terms significantly over-represented by genes comprising AD-associated modules across six tissue types (except for FCX, which has no AD-associated modules) are provided in Supplementary Tables 22 and 23 , respectively. In gene-set over-representation analyses of both DEGs and AD-associated co-expression modules, there is substantial redundancy between significantly over-represented GO terms and Reactome pathways. Here we describe the enriched Reactome pathways in detail. Using the Reactome pathway-level hierarchy, we summarized Reactome sub-pathways into root-pathways for each tissue. For each root-pathway in each tissue, the normalized counts of distinctly up- or down-regulated sub-pathways enriched by DEGs are shown in Figures 4A and 4B , and the raw counts are provided in Supplementary Figures 4A and 4B . Overall, we identified more downregulated pathways than upregulated pathways in each tissue. There was substantial overlap of the dysregulated pathways across different brain tissues. Seven downregulated root-pathways were shared between blood and brain, where most of the distinctly downregulated pathways in blood were involved in the “metabolism of proteins” ( Figure 4B ). Figure 4. The number of over-represented Reactome sub-pathways within each root-pathway. Open in a new tab This figure summarizes the Reactome pathways that were significantly over-represented by differentially expressed genes in the meta-analysis. Panel A and B present the Reactome pathways that were distinctly upregulated and downregulated, respectively. For each tissue, each over-represented pathway was traced down to a root-pathway term through the Reactome parent-child pathway hierarchy. In each tissue, the color of heatmap cells indicates the raw counts of over-represented sub-pathways within the root-pathway. Cells with a larger raw count are highlighted with a darker-color shading. The number within cells indicates the normalized counts of over-represented sub-pathways within the root-pathway to adjust for the total number of sub-pathways tested within each root-pathway for a tissue type. The normalized count was calculated using the formula Number of significantly enriched sub - pathways Total number of tested sub - pathways × 100 % . Abbreviations: CER: cerebellum; DLPFC: dorsolateral prefrontal cortex; ENT: entorhinal cortex; FCX: frontal cortex; HIP: hippocampus; TCX: temporal cortex. Note: we did not identify significantly over-represented pathways that were distinctly upregulated in blood and hippocampus. All root-pathways from analysis using DEGs (count = 25) were the subset of all observed root-pathways from co-expression modules (count = 27). The two root-pathways uncovered only by co-expression modules were “muscle contraction” and “sensory perception”. Notably, the pathway analysis of co-expression modules also revealed many sub-pathways that were not over-represented when using DEG summary statistics. Taken together results from both analyses, we observed 15 root-pathways dysregulated across all seven tissue types. A rank of these processes based on the median FDR- p -values of the sub-pathways within the root-pathways was “DNA replication”, “metabolism of proteins”, “protein localization”, “cell cycle”, “programmed cell death”, “cellular responses to stimuli”, “vesicle-mediated transport”, “metabolism”, “immune system”, “metabolism of RNA”, “signal transduction”, “transport of small molecules”, “disease”, “gene expression (transcription)”, and “DNA repair”. In addition, we found downregulated “neuronal system” in all brain regions except for CER, and downregulated “autophagy” in all tissue types except for FCX. We identified some root-pathways in brain regions but not blood, such as “cell-cell communication”, “chromatin organization”, “developmental biology”, “extracellular matrix organization”, “muscle contraction”, “neuronal system”, “organelle biogenesis and maintenance”, “reproduction”, and “sensory perception”. The GSEA using median meta-analysis effect sizes of genes across blood and six brain regions corroborated that the “immune system” was among the most up-regulated biological processes in AD, and several metabolism-related pathways were the most down-regulated processes in AD ( Figure 5 ). Cell-type enrichment using median effect sizes of genes across six brain regions found that down-regulated genes were significantly over-represented among known markers for excitatory neurons (Bonf- p = 7.8e-05) and inhibitory neurons (Bonf- p = 7.9e-05). Although substantial evidence from prior research 63 indicated that microglia play a major role in the pathogenesis of AD, we didn’t identify the enrichment of DEGs in microglia in this analysis. A possible explanation could be that microglia profiles change dynamically with AD progression 64 and may vary between brain regions 65 ; 66 . Figure 5. Reactome root-pathways dysregulated in AD across blood and/or brain. Open in a new tab The figure shows the Reactome root-pathways that were significantly (FDR- p < 0.05) over-represented by genes that showed relatively robust association with AD in blood and/or six brain regions. We used the median effect sizes of genes that were present in at least three tissue types for gene-set enrichment analysis in 29 Reactome root-pathways. Pathways in red color were distinctly up-regulated in AD. Pathways in blue color were distinctly down-regulated in AD. The dotted line annotates FDR- p -value 0.05. Abbreviations: AD: Alzheimer’s disease; FDR- p : false discovery rate-adjusted p -value. Immune- and inflammatory-response-related pathways were the most significant pathways enriched in the upregulated co-expression modules of three brain tissues (CER, DLPFC, and TCX), including the “immunoregulatory interactions between a lymphoid and a non-lymphoid cell” and “interleukin-4 and interleukin-13 signaling” ( Figure 6A ). Notably, DLPFC demonstrated a stronger pattern of up-regulated immune system pathways than other tissues. Multiple toll-like receptor (TLR-2, 3, 4, 5, 6, 7, 8, 9, 10) cascades and “MAP kinase activation” were among the significantly upregulated pathways in DLPFC. In addition, several adaptive immune response-related pathways were distinctly downregulated in AD across multiple brain regions, including “activation of NF- κ B in B cells”, “downstream signaling events of B cell receptor (BCR)”, “antigen processing-cross presentation”, and “downstream T cell receptor (TCR) signaling”. Moreover, “ER-phagosome pathway” was distinctly downregulated in all six brain tissues of AD cases, which may suggest impaired phagocytosis. Figure 6. The most significant over-represented Reactome sub-pathway of the top five root-pathways in each tissue. Open in a new tab This figure shows the pathways that were significantly over-represented by genes comprising AD-associated WGCNA modules. Panel A shows the Reactome pathways that were enriched by gene members of WGCNA modules that were significantly upregulated in AD cases compared with CU comparisons. Panel B shows the Reactome pathways that were enriched by gene members of WGCNA modules that were significantly downregulated in AD cases. In each tissue, each triangle indicates the most significant Reactome pathway within the root-pathway term, and its transformed FDR-corrected p -value is shown in the vertical axis. The size of the triangle is relative to the number of sub-pathways mapped into the root-pathway term. For each tissue, only the top five root-pathway categories are shown in the figure, which were ranked by the smallest p -value within the root-pathway category. The most significant pathway among all the pathways over-represented in a tissue is labeled with the pathway name. Abbreviations: AD: Alzheimer’s disease; CU: cognitively unimpaired; WGCNA: weighted gene co-expression network analysis; FDR- p : false discovery rate-adjusted p -value; CER: cerebellum; DLPFC: dorsolateral prefrontal cortex; ENT: entorhinal cortex; HIP: hippocampus; TCX: temporal cortex. Note: No AD-associated co-expression modules were identified in the frontal cortex. No significantly over-represented Reactome pathways were identified in the upregulated modules of the entorhinal cortex. Several protein translation-related pathways were among the most significantly over-represented pathways in the downregulated co-expression modules of blood, CER, and DLPFC ( Figure 6B ). Most enriched pathways within the “metabolism of proteins” root-pathway were downregulated. In addition to downregulated eukaryotic and mitochondrial translation, several protein quality-control-related pathways were also downregulated in AD, such as “chaperonin-mediated protein folding”, “protein ubiquitination”, and “endoplasmic reticulum (ER) quality control compartment”. In addition, genes mediating multiple subtypes of autophagy, another protein degradation mechanism, were downregulated in AD cases, including aggrephagy, chaperone-mediated autophagy, mitophagy, and macroautophagy. The respiratory electron transport, ATP synthesis, and heat production-related pathway was the most significantly over-represented in the downregulated co-expression modules of HIP ( Figure 6B ). The “respiratory electron transport” pathway was distinctly downregulated in blood and all six brain regions, implying the impairment of mitochondria in AD. We also found other downregulated mitochondrial-related processes, such as the TCA cycle, cholesterol biosynthesis, and the metabolism of numerous cellular components ( i.e ., amino acid, glucose, fatty acid, and nucleotides). Some heat-shock and stress-related pathways were distinctly upregulated in AD, which may indicate mitochondrial dysfunction-related oxidative stress, including “heme signaling” and “attenuation phase”. Concordantly, multiple cellular anti-stress activities were among the significantly downregulated pathways in AD, such as “cellular response to hypoxia”, “detoxification of reactive oxygen species”, and “cryoprotection by HMOX1”. The “activation of N-methyl-D-aspartate (NMDA) receptors and postsynaptic events” was the most significant pathway enriched in the downregulated co-expression modules of both ENT and TCX ( Figure 6B ), which was also distinctly downregulated in HIP. Additionally, “long-term potentiation” was enriched in the downregulated co-expression modules of four brain regions ( i.e ., ENT, DLPFC, HIP, and TCX). The “GABA synthesis, release, reuptake and degradation” pathway was distinctly downregulated in multiple brain regions ( i.e ., DLPFC, FCX, HIP, and TCX), which may imply imbalanced excitatory and inhibitory synaptic signals. Notably, the only distinctly upregulated neuronal system-related pathway was the “voltage-gated potassium channels” in TCX. Gene Co-Regulation Variation Drives Heterogeneous AD Signals across Studies We hypothesized that differences in gene regulation might result in distinct sets of DEGs. To test this, we examined Spearman’s association between module preservation measured by the medianRank statistics and DEG similarity measured by mean p -value rank. We found no significant associations for each of the five tissue types ( i.e ., blood, DLPFC, ENT, HIP, and TCX) that had AD-associated modules in the reference study. The Supplementary Methods section provides a description of this analysis in detail. We then investigated if larger-scale changes in gene co-expression patterns could account for various sets of DEGs between AD studies using NMI as the module preservation statistic. We confirmed that tissue-of-origin ( i.e ., blood and six brain regions) was a primary factor driving the differences in DEGs across datasets, which was significantly associated with the assignment of seven clusters (equivalent to blood and six brain regions) in DEG effect-size similarity (Fisher’s exact test, p = 0.019; Supplementary Figure 7 ). However, the regional origin of brain datasets was not a key determinant for DEG similarity (Fisher’s exact test, p = 0.082), which motivates us to investigate what other factors might be generating subgroups of studies. We found that the partition similarities ( i.e ., NMI values) were significantly correlated with DEG similarity ( i.e ., Spearman’s r of per-gene effect size correlation) across brain datasets (Spearman’s r = 0.19, p = 9.5e-06), and across blood datasets (Spearman’s r = 0.61, p = 5.7e-04). We replicated these results with an independent clustering method (see Supplementary Results ) which assigned all genes to modules, as WGCNA typically places thousands in the unassigned “gray” module. Our results suggest that, to a significant extent, gene-regulatory patterns are contributing to different DEG results between AD studies for both blood and brain. We further examined whether technical variability could contribute to the distinct gene-regulatory patterns. Across eight blood datasets, we found a significant association between cluster assignments ( K = 4) based on gene-partition similarity and four library types ( i.e ., Affymetrix array, Illumina array, Agilent array, and Illumina sequencing) (Fisher’s exact test, p = 0.029). Across 33 brain datasets from six brain regions, we found that cluster assignments ( K = 2) based on gene-partition similarity were associated with profiling platforms (array vs . sequencing) (Chi-squared test, p = 0.010). In addition, in brain studies, we found the five library types ( i.e ., Affymetrix array, Illumina array, Agilent array, Rosetta/Merck array, and Illumina sequencing) could explain the differences in cluster assignments ( K = 5) based on DEG effect-size similarity (Fisher’s exact test, p = 0.001) and the cluster assignments ( K = 5) based on gene-partition similarity (Fisher’s exact test, p = 0.016). Therefore, variability from profiling library types could drive gene-partition differences and result in heterogeneous DEG signals across brain studies. However, the gene expression variance explained by different library types in brain studies was very small ( R 2 = 1.9e-06). We then tested the correlation between gene-partition similarity and DEG similarity in brain studies profiled by Illumina sequencing platforms, but we did not identify a significant correlation (Spearman’s r = 0.009, p = 0.93). This further supports that technical variability, such as the differences in profiling library types, is one of the factors driving the heterogeneity of gene-regulatory patterns and gene differential effects across AD studies. We observed the presence of two major sub-groups of AD studies in both blood and brain based on the study partition similarities ( Figures 7A and 7B ). A granular cluster definition could be used to identify more study sub-groups, but we adopted the most parsimonious number of clusters to simplify the analysis and showcase the differential biological processes between AD studies for the first time. We found 483 genes whose AD-associated gene-expression changes significantly differed between blood datasets in cluster-1 and cluster-2 (FDR- p < 0.05, Supplementary Table 25 ) after adjusting for the effects of age, sex, and estimated cell proportions. For brain datasets, 3,854 genes had differential AD-associated effects between cluster-1 and cluster-2 (FDR- p < 0.05, Supplementary Table 26 ) after adjusting for the effects of age, sex, brain regions, and estimated cell proportions. We found seven GO terms and 132 Reactome pathways in blood, and 65 GO terms and 163 Reactome pathways in the brain, that were significantly over-represented by genes with differential AD-associated effects between the two study clusters (FDR- p < 0.05, Supplementary Tables 27 and 28 ). There were 94 Reactome gene-sets shared between blood and brain that showed differential AD association between study sub-groups, including genes involved in metabolism and mitochondrial activities, degradation of cell cycle proteins, DNA damage response, DNA replication, DNA repair, Hedgehog signaling, WNT signaling, noncanonical NF- κ B signaling, B cell signaling, and apoptosis. We found six Reactome root-pathway categories that were only revealed in the brain (autophagy, for instance). Interestingly, “cell surface interactions at the vascular wall” was among the brain-specific Reactome pathways, which may suggest heterogenous expression changes of leukocyte extravasation-related genes between groups of brain studies. In addition, genes involved in GO gene-sets “amyloid-beta clearance” and “response to unfolded protein” exhibited significantly different AD-associated effects between the two study clusters in brain, implicating heterogeneous gene expression changes involved in A β -related biological responses across AD brain studies, with stronger or more positive AD-associated effects in cluster-2 studies. Because technical variability (such as profiling library types) might drive the differences in gene-partition patterns and lead to heterogeneous DEG signals across studies, these results suggest that technical variables might obfuscate AD-associated signals and contribute to inconsistent results across AD studies. Figure 7. Heatmap of module partition similarity between any two of the blood or brain studies. Open in a new tab The color of heatmap cells indicates the partition similarity between any two studies, which was measured by normalized mutual information (NMI). The NMI color legend shows the magnitude of NMI values. Panel A presents the NMI values between any two of the eight blood studies. Panel B shows the NMI values between any two of the 33 brain datasets of six brain regions. The color bar on the side of the heatmaps indicates the assignment of two study clusters in blood and brain. Abbreviations: CER: cerebellum; DLPFC: dorsolateral prefrontal cortex; ENT: entorhinal cortex; FCX: frontal cortex; HIP: hippocampus; TCX: temporal cortex. Note: the row annotations are a combination of tissue-of-origin and study IDs. For study IDs in the format of author name and publication year, the last name of the first author and the last two digits of the publication year were used. To examine whether the disease-associated biological variation between blood study sub-groups could explain the lack of replication of DEGs in the meta-analysis, we conducted two separate meta-analyses using the two clusters of blood studies. We found that 12 out of the 117 DEGs (FDR- p < 0.05) in cluster-1 studies were nominally significant ( p < 0.05) in the replication study of Nachun et al ., 2019, while none of the nine DEGs (FDR- p < 0.05) in cluster-2 studies were nominally replicated. This is consistent with the fact that Nachun et al ., 2019 showed the strongest positive association with two cluster-1 studies (AddNeuroMed batch-1: Spearman’s r = 0.34, p = 3.2e-193; AddNeuroMed batch-2: Spearman’s r = 0.26, p = 1.4e-118; Supplementary Table 29 ) based on the per-gene effect size correlation across the whole transcriptome. AD Diagnostic Classification and Braak Score Prediction AD PTRSs in both blood and TCX, and MLAs ( i.e ., RF, linear SVM, and XGBoost) distinguished held-out AD cases from CU comparisons significantly better than expected by chance when summing across top genes ranked by p -value ( Figure 8 ). The best-performing AD PTRS model in blood, which used just the top 2 ranked DEGs, yielded an AUC of 0.727 (95% CI = [0.665, 0.789], p = 7.3 e -09) ( Figure 8A ). The AD PTRS model in TCX comprised of the top 10 ranked DEGs yielded the highest AUC of 0.835 (95% CI = [0.796, 0.875], p = 1.4e-21) ( Figure 8A ). For each top-ranked gene in the best-performing AD PTRS model, Supplementary Figure 8 shows the gene-expression values in AD cases and CU comparisons, adjusting for age, sex, and estimated cell-type proportions. AD PTRSs in blood derived from a range of top-ranked DEGs showed similar performance to MLAs in terms of AUC for the AD diagnostic classification ( Figure 8B ). The highest held-out set classification performance of all ML methods was from linear SVM using the top 25 differentially expressed genes in AD (AUC = 0.763, 95% CI = [0.704, 0.821], p = 7.6e-19) ( Figure 8B ). Figure 8. Classification performance of AD PTRSs in blood and TCX, and of machine learning models (RF, SVM-linear, and XGBoost) in blood. Open in a new tab Each method was trained on gene-expression profiles to distinguish AD cases from CU comparisons. Vertical axes show AUCs in a fully independent held-out test sample of AD cases and CU comparisons. Horizontal axes indicate classification methods. For each classification method, the color of each bar indicates the number of top-ranked AD-associated genes included as features in the model. The p -values for AUCs were computed using a one-tailed t -test with a population mean of 0.5 ( i.e ., chance accuracy, designated by a dotted line). All the classification models had AUCs significantly higher than 0.5. For each classification method, we used an asterisk (*) to indicate the model with the highest AUC. Abbreviations: AD: Alzheimer’s disease; PTRSs: polytranscript risk scores; TCX: temporal cortex; RF: random forests; SVM-linear: linear-kernel support vector machines; CU: cognitively unimpaired; AUC: Receiver Operating Characteristics Area Under the Curve. APOE -ε4 carriers had higher AD PTRSs in blood ( k genes = 2) than non-carriers ( β = 0.32, 95% CI = [0.068, 0.575], p = 0.013) ( Figure 9A ). We observed a similar positive association with APOE -ε4 with the best-performing AD PTRSs in TCX ( k genes = 10) ( β = 0.37, 95% CI = [0.178, 0.571], p = 2.2e-04) ( Figure 9B ). In addition, the best-performing scores in blood ( k genes = 2) were negatively associated with MMSE scores ( β = −1.40, 95% CI = [−2.170, −0.621], p = 5.6e-04) after adjusting for age, sex, and education years ( Figure 9C ). Figure 9. The correlation between the best-performing PTRSs and phenotypic variables in the held-out test study. Open in a new tab Panels A and B demonstrate the differences between APOE -ε4 carriers and non-carriers of the best-performing AD PTRSs in blood and temporal cortex, respectively. Panel C shows the linear association of the best-performing AD PTRS in blood with MMSE scores, adjusting for age, sex, and years of education. Panel D shows the differences in Braak PTRSs between AD cases and CU comparisons. Abbreviations: PTRSs: polytranscript risk scores; APOE : Apolipoprotein E; MMSE: Mini Mental State Examination; AD: Alzheimer’s disease; CU: cognitively unimpaired. Note: one asterisk (*) symbol indicates p <0.05; three asterisks (***) symbol indicates p <0.001. The prediction error 1- R 2 (MSE divided by the variance of Braak scores) of Braak PTRSs in TCX ranged from 0.785 to 0.872. The Braak PTRS model comprised of the top five ranked genes achieved the lowest prediction error ( Supplementary Table 30 ). Supplementary Figure 9 shows how gene expression levels change with Braak scores for each of the top five ranked genes in the held-out testing study, including ADAMTS2 (ADAM Metallopeptidase With Thrombospondin Type 1 Motif 2), CIC (Capicua Transcriptional Repressor), NPNT (Nephronectin), VGF (VGF Nerve Growth Factor Inducible), and RIDA (Reactive Intermediate Imine Deaminase A Homolog). For the best-performing Braak PTRSs, higher PTRSs were associated with a greater likelihood of being an AD case than a CU comparison in the held-out testing data (OR = 4.56, 95% CI = [2.04, 10.22], p = 2.2e-04) ( Figure 9D ). In addition, we found that the AD PTRS and Braak PTRS models in TCX could be significantly transferred to blood data. We applied the TCX scoring formula in the blood held-out study ( i.e ., AddNeuroMed batch-2) to derive a series of AD or Braak PTRSs of top-ranked genes ranging from 10 to 700. The best-performing AD PTRS model included the top 500 ranked DEGs in TCX and achieved an AUC of 0.629 in blood (95% CI = [0.559, 0.698], p = 1.4 e -04) ( Supplementary Table 31 ). For all 10 Braak PTRSs in blood, higher PTRSs were associated with an increased likelihood of being AD cases compared to CU comparisons ( Supplementary Table 32 ). Discussion Here, we performed a meta-analysis using the largest collection of extant peripheral blood and brain transcriptomic data of AD cases and CU comparison participants. We reproduced signals at both the gene and pathway levels that had been observed in some prior studies but not observed in others. These meta-analytic results, thus, lend credibility to those findings and may help settle the ambiguity surrounding inconsistent reports. We also revealed findings that had not been reported in any prior gene-expression study, underscoring the power of meta-analytic data-pooling. We found evidence to support that the divergent findings of individual AD gene-expression studies could be explained by the heterogeneous coregulation patterns of transcripts in different AD samples. These results shed light on new candidate biomarkers and pathogenic mechanisms for AD. In our study, several classification models built on blood gene-expression data showed significant diagnostic classification performance for AD. Using a simple additive scoring algorithm (polytranscript risk-scoring, PTRS), we achieved equivalent performance in AD classification as more complex machine-learning algorithms. More importantly, we showed that PTRSs in blood and brain can be useful tools in quantifying the overall disease risk attributable to transcriptomic perturbations, which could be leveraged to identify dimensions or subgroups of individuals carrying the highest load of transcriptomic risk, and to better understand the mechanisms that protect resilient individuals from AD pathology or symptomatology. Differentially Expressed Genes Provide Candidate Targets in AD As expected, the brain regions with the largest sample sizes ( i.e ., DLPFC, TCX, and CER) had the highest power to detect significant diagnostic-group differences in gene expression, and this was borne out in the results. In fact, these three brain regions accounted for 97.9% of the total number of DEGs identified in our meta-analysis. Interestingly, we detected only 23 DEGs in peripheral blood, which had the largest sample size among all seven tissue types. This could be explained by the greater variability of gene-expression levels in peripheral blood than in brain regions ( Supplementary Figure 6 ). Although seven out of the 23 blood DEGs were also significantly dysregulated in the brain in our analysis, the overlap of DEGs between blood and brain was not significant, suggesting a limited capacity of blood gene-expression changes to parallel biological alterations in the brains of AD cases at the level of individual genes. In line with the amyloid cascade hypothesis, multiple genes directly involved in A β production and tau hyperphosphorylation were significantly dysregulated in AD brain tissues. Those include downregulated PSEN2 , which encodes presenilin 2 and is one of the genes linked with early-onset autosomal dominant AD, and numerous upregulated tau kinases, such as the two genes that have drawn substantial interest in the field: GSK3β (glycogen synthase kinase 3β) 67 and FYN (FYN Proto-Oncogene) 68 . The complete list of autosomal-dominant AD-associated genes ( i.e ., APP , PSEN1 , PSEN2 , and SORL1 ) and tau kinase-encoding genes, curated from Guo et al ., 2020 69 , that were identified as significant DEGs in the current meta-analysis is presented in Supplementary Table 7 . In addition, we cross-referenced the top 5 most significant DEGs in each tissue with the existing literature to determine if our strongest findings were concordant with prior studies. The evidence from prior gene-expression studies, proteomic studies, genome-wide association studies, and cell/animal experimental studies is summarized in Table 2 . Below, we describe in detail the top DEGs in blood and brain with the most compelling support from prior literature. Table 2. Literature evidence for the top five differentially expressed genes in the meta-analysis of each tissue. Tissue Gene symbol Effect size SE p FDR- p Het- p Genetic/GWAS Animal/cell model Human transcript Human protein Blood CRYZL1 −0.05 0.009 2.1E-09 3.8E-05 0.692 PMID: 31385390 CHD2 −0.03 0.006 5.2E-08 4.8E-04 0.798 PMID: 34011927 UBR7 −0.04 0.008 8.2E-08 0.001 0.581 PMID: 26125932 ELL 0.06 0.012 2.5E-07 0.001 0.456 CD226 −0.07 0.014 1.3E-06 0.005 0.827 PMID: 25114072 Cerebellum SLCO4A1 0.24 0.028 1.1E-17 1.4E-13 0.457 PMID: 34600516; PMID: 29523845 HSD17B8 −0.16 0.021 1.2E-13 7.0E-10 0.723 PMID: 31019427 PCK2 −0.16 0.022 5.7E-13 1.9E-09 0.855 PMID: 31019427 M6PR −0.05 0.007 6.3E-13 1.9E-09 0.691 PMID: 11551970; PMID: 30704515; PMID: 15330337 RNF24 0.13 0.018 1.0E-12 2.5E-09 0.594 DLPFC GNRH1 0.11 0.010 1.7E-30 2.9E-26 0.420 PMID: 25089901 GEM 0.23 0.023 4.8E-24 3.2E-20 0.912 PMID: 15087445 PMID: 34589742 PNRC1 0.10 0.010 5.5E-24 3.2E-20 0.726 ATF7IP2 −0.09 0.009 1.6E-22 6.8E-19 0.626 PMID: 31346329 TPD52L1 0.13 0.014 1.4E-21 4.8E-18 0.442 PMID: 34589742 Entorhinal NAPG −0.08 0.010 6.8E-15 5.1E-11 0.540 PMID: 29490708 TMEM158 0.16 0.021 4.0E-14 1.5E-10 0.928 PMID: 23705665 MCL1 0.09 0.012 6.5E-14 1.6E-10 0.637 PMID: 28751497; PMID: 33184293; PMID: 33342330 PMID: 26258756; PMID: 28648936 HOPX −0.21 0.029 8.1E-13 1.5E-09 0.952 https://doi.org/10.1101/2022.06.08.495344 PFDN2 −0.09 0.013 2.1E-12 3.1E-09 0.567 PMID: 21403392 PMID: 11788046 Frontal VAMP1 −0.09 0.017 2.6E-07 0.003 0.723 PMID: 25881291 PMID: 34305524; PMID: 27701410 PMID: 23273601 PMID: 33326751 RASSF9 0.23 0.046 4.5E-07 0.003 0.906 PDE10A −0.17 0.035 9.4E-07 0.004 0.791 PCID2 0.06 0.013 2.0E-06 0.007 0.453 NFIB 0.05 0.011 4.2E-06 0.012 0.632 PMID: 35341343 Hippocampus SYNM 0.08 0.014 1.7E-08 2.4E-04 0.545 PMID: 17356066 PMID: 24269023 ANKRD20A1 −0.11 0.021 4.8E-08 3.5E-04 0.967 PMID: 23705665 WIF1 −0.26 0.051 4.9E-07 0.002 0.757 PMID: 31980586 PMID: 25380588; PMID: 29426890 FKBP11 −0.07 0.015 1.0E-06 0.004 0.950 PMID: 16945109 TPST1 0.08 0.017 2.0E-06 0.006 0.454 PMID: 23705665 Temporal POLR2G −0.08 0.009 3.8E-18 7.2E-14 0.738 PMID: 24348898 CCDC3 0.12 0.015 2.0E-16 1.5E-12 0.485 AEBP1 0.23 0.028 2.4E-16 1.5E-12 0.577 PMID: 35177665 PMID: 31176712; PMID: 31256118; PMID: 31768052 PMID: 27997051 SST −0.48 0.063 9.0E-15 3.5E-11 0.090 PMID: 15778722 PMID: 20164562 PMID: 6107862 TBC1D1 0.08 0.010 9.3E-15 3.5E-11 0.623 PMID: 34873149 PMID: 33446646; PMID: 35585924 Open in a new tab Abbreviations: SE: standard error; FDR- p : false discovery rate-adjusted p -value; Het- p : heterogeneity p -value; GWAS: genome-wide association study; DLPFC: dorsolateral prefrontal cortex. Note: Genes in red were highlighted in the Discussion section. Our analysis in blood highlighted CHD2 (Chromodomain helicase DNA binding protein 2), which was found to be downregulated in AD patients. This gene plays an essential role in neural circuit development and long-term memory 70 . One genetic variant in CHD2 has been associated with AD risk among APOE ε4-carriers 71 . The top dysregulated genes in the brain highlighted several biological processes that may underly distinct pathophysiological mechanisms in AD, including hormonal regulation ( i.e ., GNRH1 , Gonadotropin Releasing Hormone 1 ; SST , Somatostatin), protein quality control ( i.e ., PFDN2 , Prefoldin Subunit 2; MCL1 , Myeloid cell leukemia 1), and immune and inflammatory processes ( i.e ., AEBP1 , Adipocyte enhancer binding protein 1; WIF1 , WNT Inhibitory Factor 1). Some of the most significant DEGs in the AD brain encode hormones of the hypothalamic–pituitary–gonadal axis, suggesting that the role of hormonal dysregulation in the development of AD deserves heightened attention in future investigations. The most significantly dysregulated gene in DLPFC was GNRH1 (Gonadotropin Releasing Hormone 1), which was upregulated in AD cases. A previous study also observed increased GNRH mRNA levels in AD mice models of both sexes 72 . Multiple prior studies reported an association between gonadotropin signaling and AD pathophysiology 73 – 75 . GNRH -analog, which causes the suppression of gonadotropins, was shown to be a promising treatment for AD 76 . SST (Somatostatin) is another neuronal hormone whose insufficiency was shown to reduce neprilysin activity, which could lead to increased levels of a hydrophobic 42-mer form of A β (A β 42 ) 77 ; 78 . In line with this, we found the expression of SST was downregulated in TCX of AD cases. Notably, SST was also shown to decrease the GNRH- induced gonadotropins ( i.e ., luteinizing hormone) that were linked with AD pathophysiology 79 ; 80 . Together, the elevated GNRH1 and reduced SST supported the involvement of the hypothalamic–pituitary–gonadal axis in AD pathobiology. Dysregulated Pathways Implicated in AD Underscore Several Biological Mechanisms We identified thousands of pathways dysregulated in AD cases across blood and six brain regions. Some of the dysregulated pathways have been implicated in genome-wide association studies of AD, such as immunity, inflammation, lipid metabolism, and endocytosis 54 ; 81 ; 82 . Here, we describe several AD-associated major biological processes that have been consistently observed in blood and/or multiple brain regions. The inferred relationship between these processes based on literature evidence was summarized as a schematic diagram ( Figure 10 ). Table 3 presents the summary statistics of the top five significantly enriched Reactome pathways across tissues for each biological process highlighted below. Figure 10. A schematic diagram of dysregulated biological processes in Alzheimer’s disease. Open in a new tab Biological processes in mint green were found to be dysregulated in peripheral blood and brain regions of AD cases in our analysis. The biological process in purple ( i.e ., neuronal function) was found to be dysregulated only in the brain of AD cases. Abbreviations: AD: Alzheimer’s disease; A β : amyloid beta; GABA: gamma-aminobutyric acid; LTP: long-term potentiation; LTD: long-term depression; NMDAR: N-methyl-D-aspartate receptor. Table 3. The summary statistics of the top five significantly enriched Reactome pathways across tissues for each highlighted biological process. Highlighted process Pathway ID Pathway name Root pathway ID Root pathway name FDR- p Tissue Direction Protein folding R-HSA-391251 Protein folding R-HSA-392499 Metabolism of proteins 0.001 TCX Down Protein folding R-HSA-390466 Chaperonin-mediated protein folding R-HSA-392499 Metabolism of proteins 0.002 TCX Down Protein folding R-HSA-391251 Protein folding R-HSA-392499 Metabolism of proteins 0.020 DLPFC Down Protein folding R-HSA-391251 Protein folding R-HSA-392499 Metabolism of proteins 0.024 HIP Down Protein folding R-HSA-390466 Chaperonin-mediated protein folding R-HSA-392499 Metabolism of proteins 0.033 DLPFC Down Protein ubiquitination R-HSA-8852135 Protein ubiquitination R-HSA-392499 Metabolism of proteins 3.5E-04 TCX Down Protein ubiquitination R-HSA-8866654 E3 ubiquitin ligases ubiquitinate target proteins R-HSA-392499 Metabolism of proteins 0.001 TCX Down Protein ubiquitination R-HSA-8852135 Protein ubiquitination R-HSA-392499 Metabolism of proteins 0.014 FCX Down Protein ubiquitination R-HSA-8866652 Synthesis of active ubiquitin: roles of E1 and E2 enzymes R-HSA-392499 Metabolism of proteins 0.015 TCX Down Protein ubiquitination R-HSA-8866652 Synthesis of active ubiquitin: roles of E1 and E2 enzymes R-HSA-392499 Metabolism of proteins 0.018 FCX Down Autophagy R-HSA-9663891 Selective autophagy R-HSA-9612973 Autophagy 1.2E-04 TCX Down Autophagy R-HSA-5205685 PINK1-PRKN Mediated Mitophagy R-HSA-9612973 Autophagy 1.2E-04 TCX Down Autophagy R-HSA-5205647 Mitophagy R-HSA-9612973 Autophagy 1.2E-04 TCX Down Autophagy R-HSA-5205685 PINK1-PRKN Mediated Mitophagy R-HSA-9612973 Autophagy 0.001 ENT Down Autophagy R-HSA-5205647 Mitophagy R-HSA-9612973 Autophagy 0.002 ENT Down Mitocondrial metabolism R-HSA-71403 Citric acid cycle (TCA cycle) R-HSA-1430728 Metabolism 1.2E-04 TCX Down Mitocondrial metabolism R-HSA-611105 Respiratory electron transport R-HSA-1430728 Metabolism 1.2E-04 TCX Down Mitocondrial metabolism R-HSA-6799198 Complex I biogenesis R-HSA-1430728 Metabolism 1.2E-04 TCX Down Mitocondrial metabolism R-HSA-70895 Branched-chain amino acid catabolism R-HSA-1430728 Metabolism 1.2E-04 TCX Down Mitocondrial metabolism R-HSA-163210 Formation of ATP by chemiosmotic coupling R-HSA-1430728 Metabolism 1.2E-04 TCX Down Stress response R-HSA-9633012 Response of EIF2AK4 (GCN2) to amino acid deficiency R-HSA-8953897 Cellular responses to stimuli 1.2E-04 TCX Down Stress response R-HSA-1234176 Oxygen-dependent proline hydroxylation of Hypoxia-inducible Factor Alpha R-HSA-8953897 Cellular responses to stimuli 1.2E-04 TCX Down Stress response R-HSA-8932339 ROS sensing by NFE2L2 R-HSA-8953897 Cellular responses to stimuli 1.2E-04 TCX Down Stress response R-HSA-1234174 Cellular response to hypoxia R-HSA-8953897 Cellular responses to stimuli 1.2E-04 TCX Down Stress response R-HSA-9707587 Regulation of HMOX1 expression and activity R-HSA-8953897 Cellular responses to stimuli 1.2E-04 TCX Down DNA damage response R-HSA-5696399 Global Genome Nucleotide Excision Repair (GG-NER) R-HSA-73894 DNA repair 1.2E-04 TCX Down DNA damage response R-HSA-6781823 Formation of TC-NER Pre-Incision Complex R-HSA-73894 DNA repair 1.2E-04 TCX Down DNA damage response R-HSA-6782135 Dual incision in TC-NER R-HSA-73894 DNA repair 1.2E-04 TCX Down DNA damage response R-HSA-6782210 Gap-filling DNA repair synthesis and ligation in TC-NER R-HSA-73894 DNA repair 1.2E-04 TCX Down DNA damage response R-HSA-6781827 Transcription-Coupled Nucleotide Excision Repair (TC-NER) R-HSA-73894 DNA repair 1.2E-04 TCX Down Degradation of cyclins R-HSA-174184 Cdc20:Phospho-APC/C mediated degradation of Cyclin A R-HSA-1640170 Cell cycle 1.2E-04 TCX Down Degradation of cyclins R-HSA-75815 Ubiquitin-dependent degradation of Cyclin D R-HSA-1640170 Cell cycle 1.2E-04 TCX Down Degradation of cyclins R-HSA-75815 Ubiquitin-dependent degradation of Cyclin D R-HSA-1640170 Cell cycle 1.6E-04 FCX Down Degradation of cyclins R-HSA-174184 Cdc20:Phospho-APC/C mediated degradation of Cyclin A R-HSA-1640170 Cell cycle 1.6E-04 FCX Down Degradation of cyclins R-HSA-174184 Cdc20:Phospho-APC/C mediated degradation of Cyclin A R-HSA-1640170 Cell cycle 2.6E-04 DLPFC Down Apoptosis R-HSA-169911 Regulation of Apoptosis R-HSA-5357801 Programmed cell death 1.2E-04 TCX Down Apoptosis R-HSA-211733 Regulation of activated PAK-2p34 by proteasome mediated degradation R-HSA-5357801 Programmed cell death 1.2E-04 TCX Down Apoptosis R-HSA-211733 Regulation of activated PAK-2p34 by proteasome mediated degradation R-HSA-5357801 Programmed cell death 1.6E-04 FCX Down Apoptosis R-HSA-169911 Regulation of Apoptosis R-HSA-5357801 Programmed cell death 1.6E-04 FCX Down Apoptosis R-HSA-169911 Regulation of Apoptosis R-HSA-5357801 Programmed cell death 2.6E-04 DLPFC Down Innate immunity and inflammation R-HSA-168138 Toll Like Receptor 9 (TLR9) Cascade R-HSA-168256 Immune system 0.001 DLPFC Up Innate immunity and inflammation R-HSA-166166 MyD88-independent TLR4 cascade R-HSA-168256 Immune system 0.001 DLPFC Up Innate immunity and inflammation R-HSA-937061 TRIF(TICAM1)-mediated TLR4 signaling R-HSA-168256 Immune system 0.001 DLPFC Up Innate immunity and inflammation R-HSA-975138 TRAF6 mediated induction of NFkB and MAP kinases upon TLR7/8 or 9 activation R-HSA-168256 Immune system 0.001 DLPFC Up Innate immunity and inflammation R-HSA-975155 MyD88 dependent cascade initiated on endosome R-HSA-168256 Immune system 0.001 DLPFC Up Adaptive immunity R-HSA-1236978 Cross-presentation of soluble exogenous antigens (endosomes) R-HSA-168256 Immune system 1.2E-04 TCX Down Adaptive immunity R-HSA-5676590 NIK-->noncanonical NF-kB signaling R-HSA-168256 Immune system 1.2E-04 TCX Down Adaptive immunity R-HSA-5607761 Dectin-1 mediated noncanonical NF-kB signaling R-HSA-168256 Immune system 1.2E-04 TCX Down Adaptive immunity R-HSA-1168372 Downstream signaling events of B Cell Receptor (BCR) R-HSA-168256 Immune system 1.2E-04 TCX Down Adaptive immunity R-HSA-2871837 FCERI mediated NF-kB activation R-HSA-168256 Immune system 1.2E-04 TCX Down Neuronal system R-HSA-888590 GABA synthesis, release, reuptake and degradation R-HSA-112316 Neuronal system 0.003 HIP Down Neuronal system R-HSA-888590 GABA synthesis, release, reuptake and degradation R-HSA-112316 Neuronal system 0.003 TCX Down Neuronal system R-HSA-112310 Neurotransmitter release cycle R-HSA-112316 Neuronal system 0.007 HIP Down Neuronal system R-HSA-888590 GABA synthesis, release, reuptake and degradation R-HSA-112316 Neuronal system 0.010 DLPFC Down Neuronal system R-HSA-442755 Activation of NMDA receptors and postsynaptic events R-HSA-112316 Neuronal system 0.024 HIP Down Open in a new tab Abbreviations: DLPFC: dorsolateral prefrontal cortex; ENT: entorhinal cortex; FCX: frontal cortex; HIP: hippocampus; TCX: temporal cortex; FDR- p : false discovery rate-adjusted p -value. As a neurodegenerative disorder marked by the aggregation of A β and tau proteins, the impaired metabolism of protein plays an essential role in the early pathogenic stage of AD. In our analysis, we observed a pattern of downregulated protein translation in the blood and brain of AD cases, which may underlie AD-associated neuronal atrophy and cognitive deficits 83 . In line with prior research 84 , multiple protein quality-control processes were impaired in AD, such as chaperonin-mediated protein folding, the ubiquitin proteasomal system, and autophagy, whose disturbance may lead to the accumulation of A β and tau. Specifically, accumulating evidence supports that the malfunction of autophagy occurs in both AD cases and animal models, which may precede synaptic and neuronal loss 85 . Downregulated transcription and translation in mitochondria, and disrupted mitophagy, imply that mitochondrial functions were impaired in AD cases. Mitochondria play a critical role in maintaining normal neuronal activities 86 . Thus, mitochondrial dysfunction has been thought to be a predisposing factor for AD 87 . Here, we observed multiple pathways that implicate the impairment of mitochondrial energy metabolism in AD, such as downregulated respiratory electron transport, ATP synthesis, TCA cycle, and the metabolism of glucose, amino acids, and fatty acids. Additionally, we observed the downregulation of genes that participate in the metabolism of cholesterol and multiple phospholipids (such as phosphatidic acid, phosphatidylcholine, and phosphatidylinositol) in AD cases. Mitochondria-associated ER membranes play a key role in phospholipid and cholesterol metabolism, and are the predominant place where APP is processed by presenilin 1 and 2 88 . Therefore, our results may provide additional evidence to support the link between mitochondrial dysfunction and A β pathogenesis. In addition to hypometabolism, dysfunctional mitochondria could also lead to increased reactive oxygen species generation, oxidative stress, and DNA damage. The association between AD and DNA damage, specifically double-strand breaks and single-strand breaks, is well established, and it is believed that endogenous ROS is the major source of DNA damage in the brain in AD 89 . We revealed downregulated pathways in cellular responses to chemical stress, heat stress, ER stress, hypoxia, and starvation in AD cases in our meta-analysis. In addition, we found impaired endogenous defense mechanisms in AD, such as the detoxification of ROS, and the antioxidant activities of HMOX1 and NFE2L2 . Our results suggest that the DNA damage in AD could be further exacerbated by reductions of the DNA damage-repair mechanisms. Base-excision repair, transcription-coupled nucleotide excision repair, and global-genome nucleotide excision repair were downregulated, suggesting defects in the single-strand break-repairing system. Additionally, we observed downregulated homologous recombination and non-homologous end-joining, which are the DNA double-strand-break repair pathways 90 that mainly activate during the cell cycle. Although normal neurons are terminally differentiated and non-dividing, previous studies implied that neurons could reenter the cell cycle when induced by AD-associated mitotic stimuli 91 ; 92 . The compromised ubiquitin–proteasome pathway in AD could also disrupt the precise regulation of cell-cycle-checking proteins and potentially contribute to neuronal cell-cycle reentry. In line with the elevations of cyclins and cyclin-dependent kinases observed in AD neurons 93 , we found the downregulated degradation of cyclins in all four cell-cycle phases. In particular, we identified the reduction of ubiquitin-dependent degradation of cyclin D, whose abundance signifies when a cell exits from G0 and enters the G1 phase of the cell cycle 91 . We found that the p53-dependent DNA damage-response pathway was impaired and the stabilization of p53 was compromised in AD. The compromised p53 function would also affect p53-mediated controlled cell death through apoptosis. Although abortive cell-cycle reentry and subsequent apoptosis-like cell death have been reported in AD 94 – 96 , emerging evidence showed that apoptosis may not always be the outcome of cell-cycle reentry. It has been observed that cell-cycle-reactivated AD neurons could remain in the brain for months or even years until the trigger to cell death 97 . Additionally, cell-cycle reentry has been proposed as a protective mechanism to prevent apoptosis-mediated neuronal death in AD 98 . In light of the debate over the link between cell-cycle reentry and apoptosis in AD, our study showed evidence to support that the regulation of apoptosis was impaired in AD. Potentially, the mitochondrial-initiated intrinsic pathway of apoptotic death 99 might be reduced in AD. In addition, necrosis and pyroptosis may be upregulated in AD, both of which are cell-death mechanisms associated with inflammation. Extensive research has demonstrated that innate immunity and inflammation play a pivotal role in the pathogenesis of AD. This is consolidated by the association of multiple innate immune-related genes with AD susceptibility in genome-wide association studies 100 . In line with these results, we detected transcriptomic evidence of increased innate immunity-mediated inflammatory responses in blood and multiple brain regions of AD cases, especially in the DLPFC. In addition, we observed that AD was associated with impaired noncanonical NF- κ B signaling. Due to the important role of non‐canonical NF‐ κ B signaling in the survival and maturation of B cells, this may imply aberrant B cell functions in AD. Interestingly, our results showed that activation of NF- κ B in B cells was disrupted and the pathway of downstream signaling events of B-cell receptors was reduced in the blood and brain of AD participants. Several other pathways involved in adaptive immunity were aberrant, including a reduction in downstream signaling events of T-cell receptors, and the decreased antigen presentation by MHC class I and II. Taken together, our analysis implies that adaptive immunity was compromised in AD. In support of our results, a prior study using immune-deficient mouse models reported that loss of B and T cells impairs microglial phagocytosis which could lead to exacerbated amyloid deposition 101 . In contrast to well-studied innate immunity, very few studies have investigated the role of adaptive immunity in the pathogenesis of AD 102 . Addressing how B and T cells impact the development of AD in further studies is warranted. In the AD cases, multiple neuronal system-related pathways were significantly dysregulated in five brain regions that are commonly altered in AD ( i.e ., DLPFC, ENT, FCX, HIP, and TCX). NMDA receptors (NMDARs) are glutamate-activated channels that play a pivotal role in excitatory synaptic transmission and plasticity in the nervous system 103 . It is widely accepted that A β may elevate glutamate levels by impairing glutamate uptake/recycling mechanisms, or directly stimulate NMDAR, leading to impaired NMDAR-dependent synaptic long-term potentiation, excitotoxicity, and neurodegeneration 104 . Although long-term potentiation was found to be downregulated in AD cases, we identified NMDA-related pathways that were contrary to previous studies, such as the downregulated glutamate release cycle, and reduced NMDAR activation and its postsynaptic events in AD cases. A possible explanation is that one of the few FDA-approved prescription drugs for the treatment of AD is memantine, which is an NMDA antagonist 105 . Due to the lack of uniformly available medication data in the included studies, we could not account for possible confounding effects related to medications in our analyses. This limitation further underscores an urgent need for future postmortem brain studies to collect and report antemortem medication use by study participants. Whether and how AD treatments affect postmortem gene-expression changes are important questions to investigate in future studies. NMDARs can be classified into two groups, namely synaptic and extrasynaptic. The activation of the synaptic NMDARs induces plasticity and cell survival, while the activation of extrasynaptic NMDARs leads to cell death, which may contribute to the pathogenesis of AD 106 . Thus, our observation of downregulated NMDAR signaling may result from the suppression of synaptic NMDARs. Synaptic and extrasynaptic NMDARs exert opposite effects on cyclic-AMP response element binding protein (CREB)-dependent and forkhead box O (FOXO)-dependent gene expression 107 , and both proteins have important roles in long-term memory consolidation 108 ; 109 . We found multiple CREB phosphorylation-related Reactome pathways enriched in downregulated WGCNA modules, and FOXO-mediated transcription enriched in upregulated WGCNA modules, which collectively suggested reduced synaptic NMDAR activity and/or increased extrasynaptic NMDAR activity. Further experiments are warranted to elucidate the alterations of NMDAR-mediated neuronal transmission and its role in AD etiology. Our current study also found that AD was linked with impaired synthesis, release, uptake, and degradation processes of gamma-aminobutyric acid (GABA), the major inhibitory neurotransmitter, which has not been extensively studied in AD. Recent studies indicated that dysfunction of GABA inhibitory interneurons may contribute to the over-activity of neural circuits observed in AD patients and AD mouse models 110 . Interestingly, we found the upregulated activity of voltage-gated potassium channels in TCX of AD participants. In line with our finding, increased voltage-gated potassium channel subtype 3.4 (Kv3.4) protein expression may exacerbate A β -mediated neurotoxicity and synaptotoxicity 111 , and elevated protein expression of voltage-gated potassium channel Kv1.3 may be implicated in microglial-mediated proinflammatory response 112 . Polytranscript Risk Scores and Machine Learning Algorithms Discriminate AD Cases from CU Comparisons Our study sought to evaluate the classification performance of a linear additive method—polytranscript risk scoring (PTRS)—which provides more intuitive results than typical machine learning methods. AD PTRSs of both blood and TCX achieved an accuracy better than chance, and the performance of scores in blood was comparable with machine learning models for classifying AD diagnostic groups. Additional validation work is warranted to determine whether the performance can be generalized to other data sets. Nonetheless, the apparent advantage of PTRS over some machine learning methods is that it provides an easier means of discerning variable importance. A recent study 57 built deep learning models to predict AD neuropathological phenotypes using gene expression. Similarly, we built linear additive scoring models to predict Braak scores. The prediction error in the out-of-sample data for the best-performing Braak PTRSs (1- R 2 = 0.785) was approximately 0.05 higher than that of the deep learning neuropathology model. Similar to the diagnosis-differential effects of the neuropathology scores predicted by the deep learning model, our linear Braak PTRSs also demonstrated a significant difference between AD cases and CU comparisons. The prior study demonstrated that the non-linear deep learning models trained in brain data could be transferred to blood data ( i.e ., AddNeuroMed study) to predict the neuropathology of AD, which the linear models failed to achieve. However, we demonstrated that linear Braak PTRSs trained in TCX could be transferred to the blood test study ( i.e ., AddNeuroMed batch-2), and the transferred scores in blood were significantly associated with AD diagnosis. In addition, we showed that the scoring weights of AD PTRSs derived from TCX can be applied in blood data (AddNeuroMed batch-2), and the generalized AD PTRSs can significantly distinguish AD cases from CU comparisons, with a best AUC of 0.629. Our results suggest that straightforward linear models, such as PTRSs, might be beneficial for both the task of AD diagnosis classification and AD neuropathology prediction. The approach of integrating AD risk over multiple genes at the gene-expression level was also utilized in a study 113 in which the risk score was derived by summing up the expression levels of genetically regulated genes. In comparison, our risk-scoring approach could potentially account for more variance in AD risk because it is not constrained to only genes whose transcript levels are heritable; instead, transcripts included in our PTRSs included both genetically regulated and non-regulated genes. It is worth noting that the diagnostic classification performance of the AD PTRSs in TCX was comparable to the polygenic risk scores in a pathologically confirmed AD cohort (AUC = 0.84, 95% CI = [0.82, 0.86]) 114 . Although the primary objective of AD classification in this study was to build a model using gene-expression levels, the performance could potentially be improved by integrating other biological and demographic features ( i.e ., age, sex, behavioral traits, brain imaging data, other omics data, and environmental characteristics), as factors such as brain structural MRI outcomes also are strong predictors of AD classification 115 . Diagnostic classification with the AD PTRSs in blood resulted in a somewhat lower AUC of 0.727 (95% CI = [0.665, 0.789]), yet this classification accuracy is similar to that of some individual plasma A β and tau biomarkers 116 while exposing a wider range of biological involvement. Studies of AD-related plasma biomarkers suggest that single biomarkers may be insufficient, but combining plasma biomarkers results in meaningfully increased AUC values 116 . When combined with other biomarkers, the PTRSs in blood might thus add to predictive accuracy. It is remarkable that higher AD PTRSs in both blood ( k gene = 2) and TCX ( k gene = 10) were associated with higher odds of being an APOE -ε4 carrier. As APOE -ε4 is the strongest genetic risk factor for late-onset AD, our results demonstrate that PTRSs may capture the gene-expression perturbations contributed by the pathogenic effects of APOE -ε4. Furthermore, we showed that higher PTRSs in blood ( k gene = 2) were associated with poorer MMSE performance, even after adjusting for the effects of age, sex, and education. This further implies that AD PTRSs in blood could be an indicator of biological disturbances underlying poor cognitive outcomes. Altogether, we provided evidence to support that PTRSs in blood could be useful biomarkers for AD, and PTRSs in TCX could be leveraged in future study designs to better understand the risk- or resilience-associated biological mechanisms in AD. The performance of PTRSs reached a maximum using only a small number of top-ranked gene features, which underscores the potential importance of these genes in AD risk. The best-performing AD PTRSs in blood included just two top-ranked genes, including RPA3 (Replication Protein A3) and COX17 (Cytochrome C Oxidase Copper Chaperone COX17), both of which were downregulated in AD cases compared with CU individuals. Prior research indicates that RPA3 is critical in DNA repair and DNA damage response 117 . COX17 has an important role in the mitochondrial respiratory chain and copper homeostasis, both processes that have been implicated in AD pathogenesis 118 . The best-performing AD PTRSs in TCX comprised 10 top-ranked genes. Among them, numerous genes were implicated in AD pathophysiology in prior studies. Specifically, multiple lines of evidence suggest that the overexpression of ELK1 (ETS Transcription Factor ELK1) contributes to AD pathogenesis. ELK1 could repress the transcription of PS1 119 , which can further lead to increased amyloidogenic A β 42/40 polypeptide 120 . We found the expression levels of ADCYAP1 (Adenylate Cyclase Activating Polypeptide 1) were downregulated in AD patients. Consistently, prior work reported that the protein product of ADCYAP1 could play a neuroprotective role in AD by enhancing the non-amyloidogenic pathway of APP processing 121 . Increased mRNA levels of ITPKB (Inositol-Trisphosphate 3-Kinase B) were observed in AD cases in our study. In line with this, increased protein levels of ITPKB were observed in human AD brain tissue and the overexpression was associated with increased neuronal cell apoptosis, APP processing, and tau phosphorylation in AD 122 . Limitations, and Future Directions to Overcome Them Like all meta-analyses, we inherit all the limitations of the constituent studies. For example, our results could be affected by confounding factors, including inaccurate diagnosis, lifetime medications, postmortem brain changes, disease severity, and study designs. How these confounders impact our outcomes remains to be addressed in future studies when the data in demand are available. Additional avenues for future work have also been suggested by this study. First, since 1984 123 , the diagnosis of AD in living individuals has been largely dependent on cognitive symptoms, and its definite diagnosis relies on pathological confirmation from postmortem brain tissues. The lower accuracy of AD diagnosis in peripheral-blood studies possibly introduced false-positives and caused inconsistency across the studies included in our analysis. Recent revisions of the AD diagnostic criteria recommend using pathological biomarkers to enhance the specificity of AD diagnosis 124 . Therefore, it is important to validate our findings in blood studies using this more accurate and recent diagnostic scheme. Second, lifetime and current medication use was not reported in most of the individuals included in our study; therefore, we used surrogate variable analysis 44 to indirectly control for potentially confounding effects of medication and other factors associated with chronic disease on gene expression. Future analyses would be better served by careful documentation of medication and inclusion as a covariate. Third, the biological changes detected in postmortem brain tissues may be different from living brain tissues due to several factors, such as the cause of death and postmortem interval. Our findings showed that peripheral blood may serve as a meaningful—though imperfect— surrogate when investigating biological changes associated with brain disorders. By leveraging the known cross-tissue similarity of blood and brain, we developed a new transcriptome imputation method called BrainGENIE 39 , which predicts gene-expression levels in 10 distinct brain regions using orthogonal components of blood-based gene expression. We are currently undertaking analyses using this computational tool for modeling brain-specific ‘omics profiles from living individuals to help advance our understanding of AD; comparison of the results of that effort with the blood- and brain-identified patterns of differential expression observed here could be illuminating, and is forthcoming. Fourth, it is notable that the sex hormone gene GNRH1 was found to be the most significant DEG in DLPFC after accounting for sex effects. As sex plays an important role in the development of AD, continued study is needed to further investigate the molecular mechanisms that underlie the impact of sex in AD, either by using sex as an interaction term or by conducting a sex-stratified analysis. Fifth, here we demonstrated that the differences in AD-associated genes across studies of the same tissue-of-origin ( i.e ., blood or brain) could be partly attributed to heterogenous gene co-regulation patterns, however, we are uncertain whether the inconsistent findings in prior AD studies may be due—to some degree—to the biological variations of this complex disease. We demonstrated that technical variability (such as differences in profiling library types) could drive distinct gene-coregulatory patterns and explain the variability of DEGs across studies. Given the small variance in gene expression explained by profiling library types, other technical variables may contribute to the heterogeneous AD-associated gene expression changes observed between studies. In our meta-analysis, we sought to examine AD-associated genes and pathways that were robust across multiple studies. In future research, an in-depth investigation of the sources of heterogeneity among AD studies could help achieve a better understanding of potential confounding effects and their impact on association signals with AD. In addition, our results highlighted the importance of designing and selecting homogeneous AD samples to replicate the findings of prior studies. Sixth, due to sample-size limitations, only one independent study was held out to test the performance of polytranscript risk scores in our study. The best-performing polytranscript risk score model in blood only included two top-ranked genes, suggesting that a parsimonious selection of biomarker signature may be favored for AD classification. However, it is notable that the model was not significantly different from other models comprising more top-ranked genes. Further independent studies are required to validate the performance of AD classification. In particular, it is important to examine whether polytranscript risk scores in peripheral blood could be useful biomarkers in longitudinal studies, which could help understand the role of altered transcriptome profiles at the early stages of AD. Moreover, several top genes in TCX with the best classification performance are involved in A β processes. Investigating the correlations of polytranscript risk scores with A β pathological measures could help unravel whether the aggregated transcriptomic effects were related to A β plaques or other AD-related manifestations in the brain. Aside from monitoring disease progression and treatment effects, the relative risk associated with polytranscript risk scores may help identify at-risk individuals and allow for the discovery of resilience factors 125 that buffer against disease risk, which would eventually inspire preventive strategies for AD. Last but not least, our results support the existing evidence of an array of etiological pathways, but they cannot be used to directly implicate causal mechanisms. Summary In conclusion, we conducted a systematic analysis of the AD blood and brain transcriptomes with the largest sample size currently available. Our meta-analyses and network analyses identified a set of genes and pathways that may play important roles in the pathophysiology or etiology of AD. The gene-expression abnormalities in AD cases observed in blood reflected some similar biological patterns as seen in the brain. The 15 dysregulated Reactome root-pathways shared across blood and six brain regions in AD suggest an integrated mechanism related to impaired protein quality control, damaged mitochondrial processes, dysregulated immune and inflammatory responses, and potentially their downstream effects. Additional in-depth research on the pathophysiological roles of these dysregulated genes and pathways is warranted. In addition, our investigation of heterogeneity implicated that different gene co-regulation patterns may drive biological variability across AD studies, highlighting the importance of homogeneous samples in AD cohorts. Furthermore, our derived AD classifiers from blood transcriptome data yielded promising results under heterogeneous analytic circumstances, suggesting that further work to refine AD classifiers will be fruitful. Lastly, our polytranscript risk scoring method showed potential as a useful and intuitive tool to aid research in risk and resilience. Supplementary Material Supinfo2 NIHMS2038543-supplement-Supinfo2.docx (1.7MB, docx) Supinfo1 NIHMS2038543-supplement-Supinfo1.xlsx (3.8MB, xlsx) Acknowledgements We thank all the participants of this study for their contributions. Our effort on this project was supported by the U.S. National Institute on Aging [grant numbers R01AG064955 and R01AG054002]. Additional acknowledgements and detailed acknowledgements of funding sources for the study are provided in Supplementary acknowledgements . Conflicts of Interest SVF, in the past year, received income, potential income, travel expenses, continuing education support, and/or research support from Aardvark, Akili, Genomind, Ironshore, KemPharm/Corium, Noven, Ondosis, Otsuka, Rhodes, Supernus, Takeda, Tris and Vallon. In previous years, he received support from Alcobra, Arbor, Aveksham, CogCubed, Eli Lilly, Enzymotec, Impact, Janssen, Lundbeck/Takeda, McNeil, NeuroLifeSciences, Neurovance, Novartis, Pfizer, Shire, and Sunovion. He is also the Program Director of www.adhdinadults.com . He is supported by the European Union’s Horizon 2020 research and innovation program under grant agreement No 965381; NIMH grants U01AR076092-01A1, 1R21MH1264940, R01MH116037; Oregon Health and Science University, Otsuka Pharmaceuticals, Noven Pharmaceuticals Incorporated, and Supernus Pharmaceutical Company. The rest of the authors have no conflicts of interest to declare. Data and Code Availability The non-normalized gene expression data of the studies included in our analysis can be accessed publicly or requested from the original authors. The access approaches and IDs can be found in Supplementary Table 1 . The R function for computing polytranscript risk scores can be found on GitHub ( https://github.com/hessJ/ptrs ). References 1. Chen KD, Chang PT, Ping YH, Lee HC, Yeh CW, and Wang PN (2011). Gene expression profiling of peripheral blood leukocytes identifies and validates ABCB1 as a novel biomarker for Alzheimer’s disease. Neurobiol Dis 43, 698–705. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Lunnon K, Ibrahim Z, Proitsi P, Lourdusamy A, Newhouse S, Sattlecker M, Furney S, Saleem M, Soininen H, Kloszewska I, et al. (2012). Mitochondrial dysfunction and immune activation are detectable in early Alzheimer’s disease blood. J Alzheimers Dis 30, 685–710. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Bai Z, Stamova B, Xu H, Ander BP, Wang J, Jickling GC, Zhan X, Liu D, Han G, Jin LW, et al. (2014). Distinctive RNA expression profiles in blood associated with Alzheimer disease after accounting for white matter hyperintensities. Alzheimer Dis Assoc Disord 28, 226–233. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Naughton BJ, Duncan FJ, Murrey DA, Meadows AS, Newsom DE, Stoicea N, White P, Scharre DW, McCarty DM, and Fu H (2015). Blood genome-wide transcriptional profiles reflect broad molecular impairments and strong blood-brain links in Alzheimer’s disease. J Alzheimers Dis 43, 93–108. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Leandro GS, Evangelista AF, Lobo RR, Xavier DJ, Moriguti JC, and Sakamoto-Hojo ET (2018). Changes in Expression Profiles Revealed by Transcriptomic Analysis in Peripheral Blood Mononuclear Cells of Alzheimer’s Disease Patients. J Alzheimers Dis 66, 1483–1495. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Booij BB, Lindahl T, Wetterberg P, Skaane NV, Saebo S, Feten G, Rye PD, Kristiansen LI, Hagen N, Jensen M, et al. (2011). A gene expression pattern in blood for the early detection of Alzheimer’s disease. J Alzheimers Dis 23, 109–119. [ DOI ] [ PubMed ] [ Google Scholar ] 7. Shigemizu D, Mori T, Akiyama S, Higaki S, Watanabe H, Sakurai T, Niida S, and Ozaki K (2020). Identification of potential blood biomarkers for early diagnosis of Alzheimer’s disease through RNA sequencing analysis. Alzheimers Res Ther 12, 87. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. van Rooij JGJ, Meeter LHH, Melhem S, Nijholt DAT, Wong TH, Netherlands Brain B, Rozemuller A, Uitterlinden AG, van Meurs JG, and van Swieten JC (2019). Hippocampal transcriptome profiling combined with protein-protein interaction analysis elucidates Alzheimer’s disease pathways and genes. Neurobiol Aging 74, 225–233. [ DOI ] [ PubMed ] [ Google Scholar ] 9. Patel H, Hodges AK, Curtis C, Lee SH, Troakes C, Dobson RJB, and Newhouse SJ (2019). Transcriptomic analysis of probable asymptomatic and symptomatic alzheimer brains. Brain Behav Immun 80, 644–656. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Wang M, Roussos P, McKenzie A, Zhou X, Kajiwara Y, Brennand KJ, De Luca GC, Crary JF, Casaccia P, Buxbaum JD, et al. (2016). Integrative network analysis of nineteen brain regions identifies molecular signatures and networks underlying selective regional vulnerability to Alzheimer’s disease. Genome Med 8, 104. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Zhang B, Gaiteri C, Bodea LG, Wang Z, McElwee J, Podtelezhnikov AA, Zhang C, Xie T, Tran L, Dobrin R, et al. (2013). Integrated systems approach identifies genetic nodes and networks in late-onset Alzheimer’s disease. Cell 153, 707–720. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Piras IS, Krate J, Delvaux E, Nolz J, Mastroeni DF, Persico AM, Jepsen WM, Beach TG, Huentelman MJ, and Coleman PD (2019). Transcriptome Changes in the Alzheimer’s Disease Middle Temporal Gyrus: Importance of RNA Metabolism and Mitochondria-Associated Membrane Genes. J Alzheimers Dis 70, 691–713. [ DOI ] [ PubMed ] [ Google Scholar ] 13. Liang WS, Dunckley T, Beach TG, Grover A, Mastroeni D, Ramsey K, Caselli RJ, Kukull WA, McKeel D, Morris JC, et al. (2008). Altered neuronal gene expression in brain regions differentially affected by Alzheimer’s disease: a reference data set. Physiol Genomics 33, 240–256. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Han G, Wang J, Zeng F, Feng X, Yu J, Cao HY, Yi X, Zhou H, Jin LW, Duan Y, et al. (2013). Characteristic transformation of blood transcriptome in Alzheimer’s disease. J Alzheimers Dis 35, 373–386. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Bottero V, and Potashkin JA (2019). Meta-Analysis of Gene Expression Changes in the Blood of Patients with Mild Cognitive Impairment and Alzheimer’s Disease Dementia. Int J Mol Sci 20. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Wang Z, Shen L, Wang Y, and Huang S (2020). Integrated analysis of miRNA and mRNA expression in the blood of patients with Alzheimer’s disease. Mol Med Rep 22, 1053–1062. [ DOI ] [ PubMed ] [ Google Scholar ] 17. Lee T, and Lee H (2020). Prediction of Alzheimer’s disease using blood gene expression data. Sci Rep 10, 3485. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Nho K, Nudelman K, Allen M, Hodges A, Kim S, Risacher SL, Apostolova LG, Lin K, Lunnon K, Wang X, et al. (2020). Genome-wide transcriptome analysis identifies novel dysregulated genes implicated in Alzheimer’s pathology. Alzheimers Dement 16, 1213–1223. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Wang Q, Li WX, Dai SX, Guo YC, Han FF, Zheng JJ, Li GH, and Huang JF (2017). Meta-Analysis of Parkinson’s Disease and Alzheimer’s Disease Revealed Commonly Impaired Pathways and Dysregulation of NRF2-Dependent Genes. J Alzheimers Dis 56, 1525–1539. [ DOI ] [ PubMed ] [ Google Scholar ] 20. Moradifard S, Hoseinbeyki M, Ganji SM, and Minuchehr Z (2018). Analysis of microRNA and Gene Expression Profiles in Alzheimer’s Disease: A Meta-Analysis Approach. Sci Rep 8, 4767. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Su L, Chen S, Zheng C, Wei H, and Song X (2019). Meta-Analysis of Gene Expression and Identification of Biological Regulatory Mechanisms in Alzheimer’s Disease. Front Neurosci 13, 633. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Patel H, Dobson RJB, and Newhouse SJ (2019). A Meta-Analysis of Alzheimer’s Disease Brain Transcriptomic Data. J Alzheimers Dis 68, 1635–1656. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Brooks LRK, and Mias GI (2019). Data-Driven Analysis of Age, Sex, and Tissue Effects on Gene Expression Variability in Alzheimer’s Disease. Front Neurosci 13, 392. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Wan YW, Al-Ouran R, Mangleburg CG, Perumal TM, Lee TV, Allison K, Swarup V, Funk CC, Gaiteri C, Allen M, et al. (2020). Meta-Analysis of the Alzheimer’s Disease Human Brain Transcriptome and Functional Dissection in Mouse Models. Cell Rep 32, 107908. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Yuen SC, Zhu H, and Leung SW (2020). A Systematic Bioinformatics Workflow With Meta-Analytics Identified Potential Pathogenic Factors of Alzheimer’s Disease. Front Neurosci 14, 209. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Noori A, Mezlini AM, Hyman BT, Serrano-Pozo A, and Das S (2021). Systematic review and meta-analysis of human transcriptomics reveals neuroinflammation, deficient energy metabolism, and proteostasis failure across neurodegeneration. Neurobiol Dis 149, 105225. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Paranjpe MD, Belonwu S, Wang JK, Oskotsky T, Gupta A, Taubes A, Zalocusky KA, Paranjpe I, Glicksberg BS, Huang Y, et al. (2021). Sex-Specific Cross Tissue Meta-Analysis Identifies Immune Dysregulation in Women With Alzheimer’s Disease. Front Aging Neurosci 13, 735611. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Fehlbaum-Beurdeley P, Sol O, Desire L, Touchon J, Dantoine T, Vercelletto M, Gabelle A, Jarrige AC, Haddad R, Lemarie JC, et al. (2012). Validation of AclarusDx, a blood-based transcriptomic signature for the diagnosis of Alzheimer’s disease. J Alzheimers Dis 32, 169–181. [ DOI ] [ PubMed ] [ Google Scholar ] 29. Lunnon K, Sattlecker M, Furney SJ, Coppola G, Simmons A, Proitsi P, Lupton MK, Lourdusamy A, Johnston C, Soininen H, et al. (2013). A blood gene expression marker of early Alzheimer’s disease. J Alzheimers Dis 33, 737–753. [ DOI ] [ PubMed ] [ Google Scholar ] 30. Voyle N, Keohane A, Newhouse S, Lunnon K, Johnston C, Soininen H, Kloszewska I, Mecocci P, Tsolaki M, Vellas B, et al. (2016). A Pathway Based Classification Method for Analyzing Gene Expression for Alzheimer’s Disease Diagnosis. J Alzheimers Dis 49, 659–669. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Li H, Hong G, Lin M, Shi Y, Wang L, Jiang F, Zhang F, Wang Y, and Guo Z (2017). Identification of molecular alterations in leukocytes from gene expression profiles of peripheral whole blood of Alzheimer’s disease. Sci Rep 7, 14027. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Li X, Wang H, Long J, Pan G, He T, Anichtchik O, Belshaw R, Albani D, Edison P, Green EK, et al. (2018). Systematic Analysis and Biomarker Study for Alzheimer’s Disease. Sci Rep 8, 17394. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Moradi E, Marttinen M, Hakkinen T, Hiltunen M, and Nykter M (2019). Supervised pathway analysis of blood gene expression profiles in Alzheimer’s disease. Neurobiol Aging 84, 98–108. [ DOI ] [ PubMed ] [ Google Scholar ] 34. Sood S, Gallagher IJ, Lunnon K, Rullman E, Keohane A, Crossland H, Phillips BE, Cederholm T, Jensen T, van Loon LJ, et al. (2015). A novel multi-tissue RNA diagnostic of healthy ageing relates to cognitive health status. Genome Biol 16, 185. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Berchtold NC, Coleman PD, Cribbs DH, Rogers J, Gillen DL, and Cotman CW (2013). Synaptic genes are extensively downregulated across multiple brain regions in normal human aging and Alzheimer’s disease. Neurobiol Aging 34, 1653–1661. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Narayanan M, Huynh JL, Wang K, Yang X, Yoo S, McElwee J, Zhang B, Zhang C, Lamb JR, Xie T, et al. (2014). Common dysregulation network in the human prefrontal cortex underlies two neurodegenerative diseases. Mol Syst Biol 10, 743. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Semick SA, Bharadwaj RA, Collado-Torres L, Tao R, Shin JH, Deep-Soboslay A, Weiss JR, Weinberger DR, Hyde TM, Kleinman JE, et al. (2019). Integrated DNA methylation and gene expression profiling across multiple brain regions implicate novel genes in Alzheimer’s disease. Acta Neuropathol 137, 557–569. [ DOI ] [ PubMed ] [ Google Scholar ] 38. Scheckel C, Drapeau E, Frias MA, Park CY, Fak J, Zucker-Scharff I, Kou Y, Haroutunian V, Ma’ayan A, Buxbaum JD, et al. (2016). Regulatory consequences of neuronal ELAV-like protein binding to coding and non-coding RNAs in human brain. Elife 5. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Hess JL, Tylee DS, Barve R, de Jong S, Ophoff RA, Kumarasinghe N, Tooney P, Schall U, Gardiner E, Beveridge NJ, et al. (2020). Transcriptomic abnormalities in peripheral blood in bipolar disorder, and discrimination of the major psychoses. Schizophr Res 217, 124–135. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Nachun D, Ramos E, Karydas A, Dokuru D, Gao F, Yang Z, Van Berlo V, Sears R, Kramer J, Boxer A, et al. (2019). Systems-level analysis of peripheral blood gene expression in dementia patients reveals an innate immune response shared across multiple disorders. bioRxiv. [ Google Scholar ] 41. Robinson MD, McCarthy DJ, and Smyth GK (2010). edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics 26, 139–140. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 42. Johnson WE, Li C, and Rabinovic A (2007). Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics 8, 118–127. [ DOI ] [ PubMed ] [ Google Scholar ] 43. Hunt GJ, Freytag S, Bahlo M, and Gagnon-Bartsch JA (2019). dtangle: accurate and robust cell type deconvolution. Bioinformatics 35, 2093–2099. [ DOI ] [ PubMed ] [ Google Scholar ] 44. Leek JT, Johnson WE, Parker HS, Jaffe AE, and Storey JD (2012). The sva package for removing batch effects and other unwanted variation in high-throughput experiments. Bioinformatics 28, 882–883. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, and Smyth GK (2015). limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res 43, e47. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Viechtbauer W (2010). Conducting Meta-Analyses in R with the metafor Package. Journal of Statistical Software 36, 1–48. [ Google Scholar ] 47. Langfelder P, and Horvath S (2008). WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics 9, 559. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 48. Wang D, Liu S, Warrell J, Won H, Shi X, Navarro FCP, Clarke D, Gu M, Emani P, Yang YT, et al. (2018). Comprehensive functional genomic resource and integrative model for the human brain. Science 362. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Troyanskaya O, Cantor M, Sherlock G, Brown P, Hastie T, Tibshirani R, Botstein D, and Altman RB (2001). Missing value estimation methods for DNA microarrays. Bioinformatics 17, 520–525. [ DOI ] [ PubMed ] [ Google Scholar ] 50. Luo W, and Brouwer C (2013). Pathview: an R/Bioconductor package for pathway-based data integration and visualization. Bioinformatics 29, 1830–1831. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 51. Varemo L, Nielsen J, and Nookaew I (2013). Enriching the gene set analysis of genome-wide data by incorporating directionality of gene expression and combining statistical hypotheses and methods. Nucleic Acids Res 41, 4378–4391. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, et al. (2005). Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci U S A 102, 15545–15550. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. de Leeuw CA, Mooij JM, Heskes T, and Posthuma D (2015). MAGMA: generalized gene-set analysis of GWAS data. PLoS Comput Biol 11, e1004219. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Bellenguez C, Kucukali F, Jansen IE, Kleineidam L, Moreno-Grau S, Amin N, Naj AC, Campos-Martin R, Grenier-Boley B, Andrade V, et al. (2022). New insights into the genetic etiology of Alzheimer’s disease and related dementias. Nat Genet 54, 412–436. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. Schizophrenia Working Group of the Psychiatric Genomics, C. (2014). Biological insights from 108 schizophrenia-associated genetic loci. Nature 511, 421–427. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 56. Braak H, and Braak E (1997). Frequency of stages of Alzheimer-related lesions in different age categories. Neurobiol Aging 18, 351–357. [ DOI ] [ PubMed ] [ Google Scholar ] 57. Beebe-Wang N, Celik S, Weinberger E, Sturmfels P, De Jager PL, Mostafavi S, and Lee SI (2021). Unified AI framework to uncover deep interrelationships between gene expression and Alzheimer’s disease neuropathologies. Nat Commun 12, 5369. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 58. Chen T, and Guestrin C (2016). XGBoost: A Scalable Tree Boosting System. arXiv. [ Google Scholar ] 59. Riise J, Plath N, Pakkenberg B, and Parachikova A (2015). Aberrant Wnt signaling pathway in medial temporal lobe structures of Alzheimer’s disease. J Neural Transm (Vienna) 122, 1303–1318. [ DOI ] [ PubMed ] [ Google Scholar ] 60. Palomer E, Buechler J, and Salinas PC (2019). Wnt Signaling Deregulation in the Aging and Alzheimer’s Brain. Front Cell Neurosci 13, 227. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 61. Schroder N, Schaffrath A, Welter JA, Putzka T, Griep A, Ziegler P, Brandt E, Samer S, Heneka MT, Kaddatz H, et al. (2020). Inhibition of formyl peptide receptors improves the outcome in a mouse model of Alzheimer disease. J Neuroinflammation 17, 131. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. Andrews SJ, Renton AE, Fulton-Howard B, Podlesny-Drabiniok A, Marcora E, and Goate AM (2023). The complex genetic architecture of Alzheimer’s disease: novel insights and future directions. EBioMedicine 90, 104511. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. Hansen DV, Hanson JE, and Sheng M (2018). Microglia in Alzheimer’s disease. J Cell Biol 217, 459–472. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 64. Mathys H, Adaikkan C, Gao F, Young JZ, Manet E, Hemberg M, De Jager PL, Ransohoff RM, Regev A, and Tsai LH (2017). Temporal Tracking of Microglia Activation in Neurodegeneration at Single-Cell Resolution. Cell Rep 21, 366–380. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 65. Mastroeni D, Nolz J, Sekar S, Delvaux E, Serrano G, Cuyugan L, Liang WS, Beach TG, Rogers J, and Coleman PD (2018). Laser-captured microglia in the Alzheimer’s and Parkinson’s brain reveal unique regional expression profiles and suggest a potential role for hepatitis B in the Alzheimer’s brain. Neurobiol Aging 63, 12–21. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 66. Masuda T, Sankowski R, Staszewski O, Bottcher C, Amann L, Sagar, Scheiwe C, Nessler S, Kunz P, van Loo G, et al. (2019). Spatial and temporal heterogeneity of mouse and human microglia at single-cell resolution. Nature 566, 388–392. [ DOI ] [ PubMed ] [ Google Scholar ] 67. Lauretti E, Dincer O, and Pratico D (2020). Glycogen synthase kinase-3 signaling in Alzheimer’s disease. Biochim Biophys Acta Mol Cell Res 1867, 118664. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 68. Nygaard HB (2018). Targeting Fyn Kinase in Alzheimer’s Disease. Biol Psychiatry 83, 369–376. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 69. Guo T, Zhang D, Zeng Y, Huang TY, Xu H, and Zhao Y (2020). Molecular and cellular mechanisms underlying the pathogenesis of Alzheimer’s disease. Mol Neurodegener 15, 40. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 70. Kim YJ, Khoshkhoo S, Frankowski JC, Zhu B, Abbasi S, Lee S, Wu YE, and Hunt RF (2018). Chd2 Is Necessary for Neural Circuit Development and Long-Term Memory. Neuron 100, 1180–1193 e1186. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 71. Park JH, Park I, Youm EM, Lee S, Park JH, Lee J, Lee DY, Byun MS, Lee JH, Yi D, et al. (2021). Novel Alzheimer’s disease risk variants identified based on whole-genome sequencing of APOE epsilon4 carriers. Transl Psychiatry 11, 296. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 72. Nuruddin S, Syverstad GH, Lillehaug S, Leergaard TB, Nilsson LN, Ropstad E, Krogenaes A, Haraldsen IR, and Torp R (2014). Elevated mRNA-levels of gonadotropin-releasing hormone and its receptor in plaque-bearing Alzheimer’s disease transgenic mice. PLoS One 9, e103607. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 73. Meethal SV, Smith MA, Bowen RL, and Atwood CS (2005). The gonadotropin connection in Alzheimer’s disease. Endocrine 26, 317–326. [ DOI ] [ PubMed ] [ Google Scholar ] 74. Smith MA, Bowen RL, Nguyen RQ, Perry G, Atwood CS, and Rimm AA (2018). Putative Gonadotropin-Releasing Hormone Agonist Therapy and Dementia: An Application of Medicare Hospitalization Claims Data. J Alzheimers Dis 63, 1269–1277. [ DOI ] [ PubMed ] [ Google Scholar ] 75. Caceres A, Vargas JE, and Gonzalez JR (2017). APOE and MS4A6A interact with GnRH signaling in Alzheimer’s disease: Enrichment of epistatic effects. Alzheimers Dement 13, 493–497. [ DOI ] [ PubMed ] [ Google Scholar ] 76. Bowen RL, Verdile G, Liu T, Parlow AF, Perry G, Smith MA, Martins RN, and Atwood CS (2004). Luteinizing hormone, a reproductive regulator that modulates the processing of amyloid-beta precursor protein and amyloid-beta deposition. J Biol Chem 279, 20539–20545. [ DOI ] [ PubMed ] [ Google Scholar ] 77. Saito T, Iwata N, Tsubuki S, Takaki Y, Takano J, Huang SM, Suemoto T, Higuchi M, and Saido TC (2005). Somatostatin regulates brain amyloid beta peptide Abeta42 through modulation of proteolytic degradation. Nat Med 11, 434–439. [ DOI ] [ PubMed ] [ Google Scholar ] 78. Watamura N, Kakiya N, Nilsson P, Tsubuki S, Kamano N, Takahashi M, Hashimoto S, Sasaguri H, Saito T, and Saido TC (2022). Somatostatin-evoked Abeta catabolism in the brain: Mechanistic involvement of alpha-endosulfine-K(ATP) channel pathway. Mol Psychiatry 27, 1816–1828. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 79. Yu WH, Kimura M, and McCann SM (1997). Effect of somatostatin on the release of gonadotropins in male rats. Proc Soc Exp Biol Med 214, 83–86. [ DOI ] [ PubMed ] [ Google Scholar ] 80. Starcevic V, Milosevic V, Brkic B, and Severs WB (2002). Somatostatin affects morphology and secretion of pituitary luteinizing hormone (LH) cells in male rats. Life Sci 70, 3019–3027. [ DOI ] [ PubMed ] [ Google Scholar ] 81. Kunkle BW, Grenier-Boley B, Sims R, Bis JC, Damotte V, Naj AC, Boland A, Vronskaya M, van der Lee SJ, Amlie-Wolf A, et al. (2019). Genetic meta-analysis of diagnosed Alzheimer’s disease identifies new risk loci and implicates Abeta, tau, immunity and lipid processing. Nat Genet 51, 414–430. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 82. Jansen IE, Savage JE, Watanabe K, Bryois J, Williams DM, Steinberg S, Sealock J, Karlsson IK, Hagg S, Athanasiu L, et al. (2019). Genome-wide meta-analysis identifies new loci and functional pathways influencing Alzheimer’s disease risk. Nat Genet 51, 404–413. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 83. Hernandez-Ortega K, Garcia-Esparcia P, Gil L, Lucas JJ, and Ferrer I (2016). Altered Machinery of Protein Synthesis in Alzheimer’s: From the Nucleolus to the Ribosome. Brain Pathol 26, 593–605. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 84. Liu Y, Ding R, Xu Z, Xue Y, Zhang D, Zhang Y, Li W, and Li X (2021). Roles and Mechanisms of the Protein Quality Control System in Alzheimer’s Disease. Int J Mol Sci 23. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 85. Liu J, and Li L (2019). Targeting Autophagy for the Treatment of Alzheimer’s Disease: Challenges and Opportunities. Front Mol Neurosci 12, 203. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 86. Spinelli JB, and Haigis MC (2018). The multifaceted contributions of mitochondria to cellular metabolism. Nat Cell Biol 20, 745–754. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 87. Wang W, Zhao F, Ma X, Perry G, and Zhu X (2020). Mitochondria dysfunction in the pathogenesis of Alzheimer’s disease: recent advances. Mol Neurodegener 15, 30. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 88. Fisar Z (2022). Linking the Amyloid, Tau, and Mitochondrial Hypotheses of Alzheimer’s Disease and Identifying Promising Drug Targets. Biomolecules 12. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 89. Canugovi C, Misiak M, Ferrarelli LK, Croteau DL, and Bohr VA (2013). The role of DNA repair in brain related disease pathology. DNA Repair (Amst) 12, 578–587. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 90. Lin X, Kapoor A, Gu Y, Chow MJ, Peng J, Zhao K, and Tang D (2020). Contributions of DNA Damage to Alzheimer’s Disease. Int J Mol Sci 21. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 91. Moh C, Kubiak JZ, Bajic VP, Zhu X, Smith MA, and Lee HG (2011). Cell cycle deregulation in the neurons of Alzheimer’s disease. Results Probl Cell Differ 53, 565–576. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 92. Seward ME, Swanson E, Norambuena A, Reimann A, Cochran JN, Li R, Roberson ED, and Bloom GS (2013). Amyloid-beta signals through tau to drive ectopic neuronal cell cycle re-entry in Alzheimer’s disease. J Cell Sci 126, 1278–1286. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 93. McShea A, Harris PL, Webster KR, Wahl AF, and Smith MA (1997). Abnormal expression of the cell cycle regulators P16 and CDK4 in Alzheimer’s disease. Am J Pathol 150, 1933–1939. [ PMC free article ] [ PubMed ] [ Google Scholar ] 94. Raina AK, Hochman A, Zhu X, Rottkamp CA, Nunomura A, Siedlak SL, Boux H, Castellani RJ, Perry G, and Smith MA (2001). Abortive apoptosis in Alzheimer’s disease. Acta Neuropathol 101, 305–310. [ DOI ] [ PubMed ] [ Google Scholar ] 95. Folch J, Junyent F, Verdaguer E, Auladell C, Pizarro JG, Beas-Zarate C, Pallas M, and Camins A (2012). Role of cell cycle re-entry in neurons: a common apoptotic mechanism of neuronal cell death. Neurotox Res 22, 195–207. [ DOI ] [ PubMed ] [ Google Scholar ] 96. Chauhan M, Modi PK, and Sharma P (2020). Aberrant activation of neuronal cell cycle caused by dysregulation of ubiquitin ligase Itch results in neurodegeneration. Cell Death Dis 11, 441. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 97. Herrup K, and Yang Y (2007). Cell cycle regulation in the postmitotic neuron: oxymoron or new biology? Nat Rev Neurosci 8, 368–378. [ DOI ] [ PubMed ] [ Google Scholar ] 98. Ippati S, Deng Y, van der Hoven J, Heu C, van Hummel A, Chua SW, Paric E, Chan G, Feiten A, Fath T, et al. (2021). Rapid initiation of cell cycle reentry processes protects neurons from amyloid-beta toxicity. Proc Natl Acad Sci U S A 118. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 99. Goel P, Chakrabarti S, Goel K, Bhutani K, Chopra T, and Bali S (2022). Neuronal cell death mechanisms in Alzheimer’s disease: An insight. Front Mol Neurosci 15, 937133. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 100. Wightman DP, Jansen IE, Savage JE, Shadrin AA, Bahrami S, Holland D, Rongve A, Borte S, Winsvold BS, Drange OK, et al. (2021). A genome-wide association study with 1,126,563 individuals identifies new risk loci for Alzheimer’s disease. Nat Genet 53, 1276–1282. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 101. Marsh SE, Abud EM, Lakatos A, Karimzadeh A, Yeung ST, Davtyan H, Fote GM, Lau L, Weinger JG, Lane TE, et al. (2016). The adaptive immune system restrains Alzheimer’s disease pathogenesis by modulating microglial function. Proc Natl Acad Sci U S A 113, E1316–1325. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 102. Chen X, and Holtzman DM (2022). Emerging roles of innate and adaptive immunity in Alzheimer’s disease. Immunity 55, 2236–2254. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 103. Cull-Candy S, Brickley S, and Farrant M (2001). NMDA receptor subunits: diversity, development and disease. Curr Opin Neurobiol 11, 327–335. [ DOI ] [ PubMed ] [ Google Scholar ] 104. Wang R, and Reddy PH (2017). Role of Glutamate and NMDA Receptors in Alzheimer’s Disease. J Alzheimers Dis 57, 1041–1048. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 105. Danysz W, Parsons CG, Mobius HJ, Stoffler A, and Quack G (2000). Neuroprotective and symptomatological action of memantine relevant for Alzheimer’s disease--a unified glutamatergic hypothesis on the mechanism of action. Neurotox Res 2, 85–97. [ DOI ] [ PubMed ] [ Google Scholar ] 106. Sattler R, Xiong Z, Lu WY, MacDonald JF, and Tymianski M (2000). Distinct roles of synaptic and extrasynaptic NMDA receptors in excitotoxicity. J Neurosci 20, 22–33. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 107. Hardingham GE, and Bading H (2010). Synaptic versus extrasynaptic NMDA receptor signalling: implications for neurodegenerative disorders. Nat Rev Neurosci 11, 682–696. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 108. Kandel ER (2012). The molecular biology of memory: cAMP, PKA, CRE, CREB-1, CREB-2, and CPEB. Mol Brain 5, 14. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 109. Salih DA, Rashid AJ, Colas D, de la Torre-Ubieta L, Zhu RP, Morgan AA, Santo EE, Ucar D, Devarajan K, Cole CJ, et al. (2012). FoxO6 regulates memory consolidation and synaptic function. Genes Dev 26, 2780–2801. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 110. Xu Y, Zhao M, Han Y, and Zhang H (2020). GABAergic Inhibitory Interneuron Deficits in Alzheimer’s Disease: Implications for Treatment. Front Neurosci 14, 660. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 111. Yeap J, Sathyaprakash C, Toombs J, Tulloch J, Scutariu C, Rose J, Burr K, Davies C, Colom-Cadena M, Chandran S, et al. (2022). Reducing voltage-dependent potassium channel Kv3.4 levels ameliorates synapse loss in a mouse model of Alzheimer’s disease. Brain Neurosci Adv 6, 23982128221086464. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 112. Rangaraju S, Gearing M, Jin LW, and Levey A (2015). Potassium channel Kv1.3 is highly expressed by microglia in human Alzheimer’s disease. J Alzheimers Dis 44, 797–808. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 113. Park YH, Hodges A, Simmons A, Lovestone S, Weiner MW, Kim S, Saykin AJ, Nho K, AddNeuroMed, c., and the Alzheimer’s Disease Neuroimaging, I. (2020). Association of blood-based transcriptional risk scores with biomarkers for Alzheimer disease. Neurol Genet 6, e517. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 114. Escott-Price V, Myers AJ, Huentelman M, and Hardy J (2017). Polygenic risk score analysis of pathologically confirmed Alzheimer disease. Ann Neurol 82, 311–314. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 115. Qiu S, Joshi PS, Miller MI, Xue C, Zhou X, Karjadi C, Chang GH, Joshi AS, Dwyer B, Zhu S, et al. (2020). Development and validation of an interpretable deep learning framework for Alzheimer’s disease classification. Brain 143, 1920–1933. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 116. Palmqvist S, Stomrud E, Cullen N, Janelidze S, Manuilova E, Jethwa A, Bittner T, Eichenlaub U, Suridjan I, Kollmorgen G, et al. (2023). An accurate fully automated panel of plasma biomarkers for Alzheimer’s disease. Alzheimers Dement 19, 1204–1215. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 117. Cavero S, Limbo O, and Russell P (2010). Critical functions of Rpa3/Ssb3 in S-phase DNA damage responses in fission yeast. PLoS Genet 6, e1001138. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 118. Ejaz HW, Wang W, and Lang M (2020). Copper Toxicity Links to Pathogenesis of Alzheimer’s Disease and Therapeutics Approaches. Int J Mol Sci 21. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 119. Pastorcic M, and Das HK (2003). Ets transcription factors ER81 and Elk1 regulate the transcription of the human presenilin 1 gene promoter. Brain Res Mol Brain Res 113, 57–66. [ DOI ] [ PubMed ] [ Google Scholar ] 120. Wolfe MS (2007). When loss is gain: reduced presenilin proteolytic function leads to increased Abeta42/Abeta40. Talking Point on the role of presenilin mutations in Alzheimer disease. EMBO Rep 8, 136–140. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 121. Rat D, Schmitt U, Tippmann F, Dewachter I, Theunis C, Wieczerzak E, Postina R, van Leuven F, Fahrenholz F, and Kojro E (2011). Neuropeptide pituitary adenylate cyclase-activating polypeptide (PACAP) slows down Alzheimer’s disease-like pathology in amyloid precursor protein-transgenic mice. FASEB J 25, 3208–3218. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 122. Stygelbout V, Leroy K, Pouillon V, Ando K, D’Amico E, Jia Y, Luo HR, Duyckaerts C, Erneux C, Schurmans S, et al. (2014). Inositol trisphosphate 3-kinase B is increased in human Alzheimer brain and exacerbates mouse Alzheimer pathology. Brain 137, 537–552. [ DOI ] [ PubMed ] [ Google Scholar ] 123. McKhann G, Drachman D, Folstein M, Katzman R, Price D, and Stadlan EM (1984). Clinical diagnosis of Alzheimer’s disease: report of the NINCDS-ADRDA Work Group under the auspices of Department of Health and Human Services Task Force on Alzheimer’s Disease. Neurology 34, 939–944. [ DOI ] [ PubMed ] [ Google Scholar ] 124. McKhann GM, Knopman DS, Chertkow H, Hyman BT, Jack CR Jr., Kawas CH, Klunk WE, Koroshetz WJ, Manly JJ, Mayeux R, et al. (2011). The diagnosis of dementia due to Alzheimer’s disease: recommendations from the National Institute on Aging-Alzheimer’s Association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimers Dement 7, 263–269. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 125. Hou J, Hess JL, Armstrong N, Bis JC, Grenier-Boley B, Karlsson IK, Leonenko G, Numbers K, O’Brien EK, Shadrin A, et al. (2022). Polygenic resilience scores capture protective genetic effects for Alzheimer’s disease. Transl Psychiatry 12, 296. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supinfo2 NIHMS2038543-supplement-Supinfo2.docx (1.7MB, docx) Supinfo1 NIHMS2038543-supplement-Supinfo1.xlsx (3.8MB, xlsx) Data Availability Statement The non-normalized gene expression data of the studies included in our analysis can be accessed publicly or requested from the original authors. The access approaches and IDs can be found in Supplementary Table 1 . The R function for computing polytranscript risk scores can be found on GitHub ( https://github.com/hessJ/ptrs ). ACTIONS View on publisher site PDF (3.0 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• MRI detects blood-brain barrier alterations in a rat model of Alzheimer’s disease and lung infection#274370
• Overrepresentation of glutamate signaling in Alzheimer's disease: Network-based pathway enrichment using meta-analysis of genome-wide association studies#1005236
• A Cross-Disease Microglial Transcriptional Program Characterizes Neurodegeneration and Highlights SPP1 as a Biomarker.#122738
• Characterization and pharmacological modulation of Alzheimers disease-associated human microglial states#1136700
• GBoost-CTL: A novel method in multi-tissue transcriptome-wide associations studies in cross-tissue learner incorporating GWAS information#179763