Multi-region transcriptomics and unsupervised clustering reveal recurrent and region-specific gene signatures in Alzheimer’s disease

Abstract

Alzheimer’s disease (AD) exhibits anatomically heterogeneous transcriptomic alterations, complicating the identification of disease-associated genes with consistent external support. We compared four gene-prioritization strategies, absolute log₂ fold change, false discovery rate (FDR), CombinedScore, and RankSum, using laser-capture-microdissected neuronal expression profiles from six brain regions in GSE5281. Differential expression was assessed separately within each region using limma, followed by within-region gene prioritization and cross-region comparison to identify recurrent and anatomically specific candidates. Robustness was evaluated through resampling stability, sensitivity to candidate-list size and scoring formulation, and external benchmarking in three anatomically matched regions of GSE48350. STAT4, SST, and NVL emerged as recurrent candidates, whereas ZNRD2, ANKRD12, and NME7 were prominent regionally prioritized signals. External directional agreement across Top-100 lists ranged from 68.3% to 72.0%; RankSum achieved 71.7% agreement and the highest or joint-highest directional agreement at Top-50 and Top-200. A fixed expression score comprising the three recurrent genes showed higher values in AD in the external cohort. Exploratory six-gene clustering revealed diagnosis-associated structure with modest separation. 14 of 18 gene–region comparisons showed concordant AD–control directions, although cross-dataset effect-size correlations were weak. This study establishes a comparative framework linking regional gene prioritization with stability assessment and independent-cohort benchmarking, identifying recurrent AD-associated candidates and quantifying their external support.

1. Introduction

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by neuronal dysfunction, synaptic loss, and pronounced anatomical heterogeneity across the human brain. Neurofibrillary tau pathology follows a stereotypical anatomical progression, beginning in transentorhinal and entorhinal regions and extending through hippocampal and other limbic structures before involving broader neocortical territories, whereas primary sensory cortices are generally affected later (Braak and Braak, 1991, Jagust, 2018, Seeley et al., 2009, Shi et al., 2026). This spatial gradient is accompanied by region-dependent molecular and transcriptomic alterations and reflects differences in the vulnerability of specific neuronal populations, circuits, and molecular programs (Mathys et al., 2024, Yao et al., 2024). Anatomical vulnerability therefore encompasses differences among neuronal populations and cellular environments that cannot be adequately captured by assigning each sampled region to a single neuropathological stage. Accordingly, disease-associated molecular alterations should be examined within individual brain regions rather than assuming a uniform transcriptional response across the brain.

The multifactorial nature of AD reflects the interplay of genetic susceptibility, aging-related processes, and environmental influences (Jackson et al., 2024). Among the disease-associated processes, cholinergic dysfunction and neuroinflammation have motivated investigation of complementary therapeutic targets. In particular, acetylcholinesterase inhibition remains an important pharmacological strategy, and several chemically diverse compounds, including thiosemicarbazone derivatives, have been investigated for their inhibitory activity against acetylcholinesterase and related enzyme targets using biochemical and molecular-docking approaches (Demir et al., 2023, Yakan et al., 2023). Recent experimental studies have also evaluated diaryl ether–phenolic Mannich base derivatives and pyrimidine-based compounds as acetylcholinesterase and butyrylcholinesterase inhibitors (Alım and Demir, 2026, Şenol et al., 2026), while thienopyrimidine hybrids have been investigated for combined cholinesterase inhibition and anti-neuroinflammatory activity (Tokalı et al., 2026). Beyond therapeutic-target studies, multi-region single-nucleus transcriptomic analyses have revealed differences in disease-associated cellular responses across brain regions, including distinct patterns of neuronal vulnerability and glial activation (Mathys et al., 2024). These findings further support examining AD-associated expression changes within their anatomical and cellular contexts.

High-throughput transcriptomic profiling has transformed the study of neurodegeneration by enabling systematic evaluation of gene-expression changes in affected brain regions and neural circuits. Major AD-focused consortia and large-scale resources, including AMP-AD cohorts such as the Religious Orders Study and Memory and Aging Project (ROSMAP), the Mount Sinai Brain Bank (MSBB), and the Mayo Clinic RNA-seq study, have generated rich transcriptomic and multi-omic datasets that highlight dysregulation of synaptic function, neuroinflammation, metabolic stress, mitochondrial dysfunction, RNA processing, and chromatin remodeling (Allen et al., 2016, Bennett et al., 2018, Hodes and Buckholtz, 2016, Wang et al., 2018). Integrative analyses across multiple cohorts, brain regions, and molecular layers have further refined these insights by identifying convergent AD-associated transcriptional modules, molecular subtypes, and disease-associated cellular states (Mathys et al., 2024, Neff et al., 2021). Together, these studies underscore the need to integrate region-specific transcriptomic evidence with robust computational approaches.

Within this context, the GSE5281 microarray dataset remains an influential early transcriptomic resource for investigating regional gene-expression alterations in AD (Liang et al., 2008). It comprises expression profiles from laser-captured neurons obtained from six postmortem brain regions: the hippocampus (HIP), entorhinal cortex (EC), middle temporal gyrus (MTG), posterior cingulate cortex (PC), superior frontal gyrus (SFG), and primary visual cortex (VCX). These regions encompass differing patterns of AD involvement, from early-affected medial temporal structures to association cortices and comparatively later-affected primary sensory cortex. The original study identified prominent alterations in metabolic and mitochondrial gene expression (Liang et al., 2008), providing a foundation for subsequent investigations of regional molecular dysfunction. Later RNA-seq, systems-level, and cell-type-resolved studies have further characterized AD-associated transcriptional changes (Allen et al., 2016, Mathys et al., 2024, Wang et al., 2018). A recent GSE5281 analysis combined multi-region differential expression, protein-interaction networks, and compound–target information to prioritize candidates associated with mitochondrial function and proteostasis, illustrating the application of this dataset to integrated molecular candidate prioritization (Isıyel et al., 2026).

Despite the value of GSE5281 and related datasets, identifying biologically meaningful and reproducible candidate genes remains challenging. Gene prioritization commonly relies on fold-change magnitude or false discovery rate (FDR)-adjusted statistical significance, which capture different aspects of differential expression. Effect-size ranking emphasizes the magnitude of expression changes, whereas FDR-based ranking emphasizes statistical evidence after multiple-testing correction. Consequently, the two criteria can yield different candidate lists, particularly in heterogeneous AD datasets, where anatomical region, cellular composition, disease stage, technical variability, and inter-individual differences influence expression estimates and their uncertainty. Integrative approaches, including rank aggregation, provide a framework for combining complementary evidence (Kolde et al., 2012). However, whether integrated approaches provide more stable or externally reproducible prioritization than individual criteria requires empirical evaluation. Comparing individual and integrated ranking criteria within a consistent regional framework can therefore distinguish candidates that recur across prioritization approaches from those whose selection depends on a particular ranking rule.

Machine learning has expanded AD transcriptomic analysis through feature selection, classification, clustering, and biclustering. Network- and SVM-based approaches, as well as random forest and LASSO methods, have been used to identify candidate genes and disease-associated expression patterns across multiple brain regions (Crist et al., 2021, Sharma and Dey, 2021, Wang and Liu, 2019, Wu et al., 2021). Fuzzy clustering and type-2 biclustering have further been applied to model uncertainty and heterogeneous gene-expression patterns in microarray data, including AD datasets (Moattar Husseini et al., 2024, Moattar Husseini et al., 2023). Cross-dataset feature-selection and classification have also been explored in blood-expression studies (Lee and Lee, 2020). These studies motivate the present combination of gene prioritization, exploratory clustering, resampling-based stability assessment, and external expression evaluation.

A central challenge in AD transcriptomics is the pronounced anatomical heterogeneity of disease-associated molecular alterations. Consistent with regional differences in neuropathological vulnerability, transcriptomic and spatial transcriptomic studies have identified both shared molecular responses and region-specific transcriptional programs across the AD brain (Bayaraa et al., 2026, Niu et al., 2025). The six regions examined in GSE5281 span this anatomical spectrum, from highly vulnerable medial temporal regions such as the entorhinal cortex (EC) and hippocampus (HIP), through association cortices including the middle temporal gyrus (MTG), posterior cingulate cortex (PC), and superior frontal gyrus (SFG), to the comparatively later-affected primary visual cortex (VCX). As summarized in Table 1, these regions also differ in their reported synaptic, metabolic, mitochondrial, glial, and regulatory alterations. This regional diversity provides the biological rationale for analyzing differential expression and gene prioritization separately within each brain region before assessing cross-regional recurrence.

Table 1.

Regional transcriptomic signatures in Alzheimer’s disease based on the literature.

Brain Region	Main Dysregulated Pathways	Notable Genes / Markers	Selected Key References
Hippocampus (HIP)	Synaptic transmission, calcium signaling, mitochondrial metabolism, inhibitory neurotransmission	SST, CAMK1G, mitochondrial ribosomal genes	(Almeida and Higa, 2026, Liang et al., 2008, Liu et al., 2025, Miller et al., 2013)
Entorhinal Cortex (EC)	Synaptic organization, RNA processing, ECM, copper metabolism, Reelin signaling	ELAVL3, TMEM106A, RELN, SEMA3C, FGF2, SPP1	(Liang et al., 2008, Mathys et al., 2024, Santa-Maria et al., 2010; Y. Zhang et al., 2023)
Middle Temporal Gyrus (MTG)	Synaptic vesicle cycling, glutamatergic signaling, MAM, plasticity-to-stress shift	AEBP1, DNALI1, NPTX2, BDNF, VGF, SCG2	(Gabitto et al., 2024, Lao and Xiao, 2025, Piras et al., 2019)
Posterior Cingulate Cortex (PC)	Mitochondrial OXPHOS, synaptic maintenance, energy metabolism	Presynaptic/postsynaptic & ATP-related genes	(Kelley et al., 2022, Liang et al., 2008, Soreq, 2024)
Superior Frontal Gyrus (SFG)	Immune/glial activation, synaptic regulation, epigenetic (TOMM40-APOE)	APOE, TOMM40 (methylation)	(Bezuch et al., 2021, Pereira et al., 2026, Wang et al., 2024)
Primary Visual Cortex (VCX)	Immune activation, synaptic vesicle secretion, neuronal energetics	TREM2, MS4A6A; RORB, CUX2, EYA4, KCNIP4	(Dharshini et al., 2026, Guennewig et al., 2021, Patel et al., 2023)

An important analytical consideration is how the choice of ranking criterion influences the identification of regionally prioritized and recurrent AD-associated candidates and whether patterns observed within the discovery dataset are supported in an independent cohort. Agreement among results derived from the same differential-expression estimates does not constitute independent biological support, just as cross-regional recurrence does not necessarily imply reproducibility across datasets. Distinguishing these dimensions is therefore essential for evaluating the strength and robustness of candidate prioritization.

The present study systematically compares four complementary gene-prioritization strategies, absolute log₂ fold change, FDR, CombinedScore, and RankSum, across the six brain regions of GSE5281. Differential expression and gene prioritization are first performed separately within each region, after which the resulting ranked lists are compared across regions to identify recurrent and anatomically specific candidates. Cross-regional overlap, recurrence across ranking criteria, resampling stability, and sensitivity to candidate-list size and scoring formulation are evaluated. External support is then assessed in the independent GSE48350 cohort, restricted to the anatomically matched EC, HIP, and SFG regions, through directional concordance and cross-dataset comparison of gene-level effect sizes. These analyses are complemented by a fixed three-gene expression score comprising STAT4, SST, and NVL and exploratory clustering of a six-gene panel. Because gene selection and sample-level clustering were performed using the discovery dataset, the clustering analysis is interpreted as exploratory characterization of internal molecular structure rather than independent validation. Collectively, the study distinguishes region-wise prioritization, cross-regional recurrence, ranking robustness, discovery-set molecular organization, and independent external expression support.

The remainder of this paper is organized as follows. Section 2 describes dataset acquisition and preprocessing, region-wise differential-expression analysis, gene-prioritization strategies, cross-regional recurrence and overlap analyses, robustness assessments, three-gene expression-score analysis, exploratory six-gene clustering, and external evaluation. Section 3 presents the main findings, including recurrent and regionally prioritized candidate genes, ranking-stability and external-benchmarking results, the three-gene expression score, and exploratory molecular stratification based on the six-gene panel. Section 4 discusses the biological implications, methodological contributions, study limitations, and directions for future independent validation.

2. Methods

2.1. Dataset acquisition, preprocessing, and quality control

The discovery dataset GSE5281 was obtained from the Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo) and was generated using the Affymetrix Human Genome U133 Plus 2.0 Array platform (GPL570) (Liang et al., 2008). It comprises 161 postmortem brain samples from 87 Alzheimer’s disease (AD) cases and 74 non-demented controls across six brain regions: entorhinal cortex (EC), hippocampus (HIP), middle temporal gyrus (MTG), posterior cingulate cortex (PC), superior frontal gyrus (SFG), and primary visual cortex (VCX). Region-specific sample numbers are reported in Table 2.

Table 2.

Distribution of samples in GSE5281 and GSE48350.

Brain Region	GSE5281		GSE48350
	Controls	AD Cases	Controls ≥ 60	AD Cases
Middle Temporal Gyrus (MTG)	12	16	-	-
Entorhinal Cortex (EC)	13	10	18	15
Hippocampus (HIP)	13	10	25	19
Posterior Cingulate Cortex (PC)	13	9	-	-
Superior Frontal Gyrus (SFG)	11	23	26	21
Primary Visual Cortex (VCX)	12	19	-	-
Post-Central Gyrus (PCG)	-	-	24	25
Total	74	87	93	80

Table Note: For GSE48350, an age criterion of ≥ 60 years was applied to controls, whereas all available AD cases were retained. External evaluation was restricted to EC, HIP, and SFG. PCG samples were not considered anatomically equivalent to the posterior cingulate cortex (PC) represented in GSE5281 and were therefore not included in the external-evaluation analysis.

GEO Series Matrix expression values were retrieved in R using the GEOquery package (Davis and Meltzer, 2007). The analysis was based on the processed expression values deposited in GEO rather than on reprocessing raw CEL files. Inspection of the downloaded GSE5281 matrix confirmed that the deposited values were not logarithmically transformed: the median probe intensity was 54.04 and the 99th percentile was 3566.57. Therefore, a log2 transformation was applied before probe-to-gene summarization. Following log2 transformation, the probe-level median and 99th-percentile expression values were approximately 5.76 and 11.80, respectively. After annotation filtering and probe-to-gene summarization, the corresponding median and 99th-percentile values of the final gene-level matrix were approximately 6.69 and 12.44.

GPL570 probe identifiers were mapped to gene symbols using the hgu133plus2.db annotation package through AnnotationDbi (Carlson, 2021). Probes without valid gene-symbol annotations and probes mapping to more than one gene symbol were excluded. Probes uniquely mapping to the same gene were summarized using WGCNA::collapseRows with the MaxMean method (connectivityBasedCollapsing = FALSE, methodFunction = NULL, connectivityPower = 1, selectFewestMissing = TRUE, and thresholdCombine = NA) (Langfelder and Horvath, 2008). This workflow retained 43,121 uniquely mapped probes and excluded 1528 ambiguous and 10,026 unmapped probes, yielding 20,825 gene-level features for downstream analysis. A probe-level annotation audit was retained to document probe mapping and retention status.

Sample-level quality control included assessment of missing values, expression-density and sample-wise expression distributions, and principal component analysis (PCA) based on the 5000 most variable genes. The post-transformation expression-density profiles showed broadly comparable distributions across samples (Supplementary Figure S1A), and no missing expression values were detected across the 161 GSE5281 samples. PCA was used to assess global sample structure and identify potential outliers (Supplementary Figure S1B). Potential PCA outliers were screened using an absolute standardized PC1 or PC2 score greater than 3 and were subjected to additional QC inspection rather than automatically excluded. This screening identified one potential outlier, GSM119676, an SFG control sample, which was retained following additional QC inspection. Available GEO metadata were also reviewed for potential technical and batch-related variables; no usable sample-specific batch covariate was identified. The metadata documented the Affymetrix hybridization and GCOS scanning procedures. Because the present analysis used the deposited Series Matrix rather than reprocessing raw CEL files, raw-array QC metrics were not recalculated; QC was performed on the downloaded expression matrix and corresponding sample metadata.

For external evaluation, GSE48350, generated on the same GPL570 platform, was used as an independent dataset (Berchtold et al., 2014). According to the GEO metadata, the deposited GSE48350 expression data had already undergone preprocessing by the original investigators, including GC-RMA/PLIER processing and GeneSpring normalization. Inspection of the downloaded probe-level values indicated a positive, right-skewed expression scale; therefore, a log2 transformation was applied before downstream analysis. The same probe-annotation, ambiguous-probe exclusion, and gene-level summarization strategy used for GSE5281 was then applied to GSE48350. Control samples were required to be aged ≥ 60 years. External evaluation was restricted to the three anatomically shared regions, EC, HIP, and SFG. The resulting cohort comprised 124 samples: 33 EC samples (18 controls and 15 AD), 44 HIP samples (25 controls and 19 AD), and 47 SFG samples (26 controls and 21 AD), corresponding to 69 controls and 55 AD cases overall. GSE48350 also contained postcentral gyrus (PCG) samples; however, these were excluded because PCG was not considered anatomically equivalent to the posterior cingulate cortex represented in GSE5281.

Age information was available for both datasets. In GSE5281, the mean age was 79.79 ± 6.91 years for AD samples (n = 87) and 79.51 ± 8.92 years for control samples (n = 74). In the GSE48350 external-evaluation cohort restricted to EC, HIP, and SFG, the corresponding mean ages were 85.55 ± 7.02 years for AD samples (n = 55) and 81.20 ± 10.08 years for control samples (n = 69). These values summarize regional tissue samples rather than unique donors.

To provide an objective external comparison of the four prioritization strategies, genes selected from GSE5281 were evaluated in the anatomically matched EC, HIP, and SFG regions of GSE48350. For each strategy, Top-50, Top-100, Top-200, and Top-500 ranked genes were examined. External agreement was assessed using directional concordance of AD-versus-control log2 fold changes and strict replication, defined as concordant direction together with FDR < 0.05 in both datasets. Cross-dataset effect-size agreement was additionally assessed using Pearson and Spearman correlations. These comparisons were interpreted descriptively because the ranked gene sets overlap and therefore do not represent independent observations.

Sample-level expression diagnostics and PCA based on the 5000 most variable genes were used for quality-control assessment of the final GSE48350 external-evaluation cohort. PCA-based screening using (|z| > 3) for PC1 or PC2 identified four potential outliers (GSM300168, GSM300170, GSM300171, and GSM300172) all of which were hippocampal control samples. Because additional inspection showed no missing expression values or extreme shifts in median expression for these samples, they were retained in the primary analysis rather than being excluded solely on the basis of PCA.

2.2. Analytical workflow

This study followed a multi-stage analytical workflow to identify and evaluate Alzheimer’s disease-associated transcriptomic signals in the GSE5281 postmortem brain microarray dataset. First, the processed expression data were obtained from the Gene Expression Omnibus and subjected to probe annotation and gene-level summarization. Second, differential expression analysis was performed separately within each of the six brain regions to preserve anatomical specificity. Third, genes were prioritized within each region using four complementary strategies based on effect magnitude, statistical evidence, and integrated ranking criteria: absolute log₂ fold change, FDR, CombinedScore, and RankSum. The resulting region-specific ranked lists were then compared across regions to identify recurrent and anatomically specific candidates. Robustness was further assessed through candidate-list sensitivity, resampling-based stability, and sensitivity to scoring formulation. Finally, selected candidates and ranking strategies were evaluated in the independent GSE48350 cohort using anatomically matched regions, including external directional concordance, strict replication, and cross-dataset effect-size comparison. These analyses were complemented by evaluation of a fixed three-gene expression score and exploratory clustering of a six-gene panel. The complete computational workflow is summarized in Algorithm 1.

Algorithm 1

Multi-metric gene-prioritization and external evaluation framework.

3. Results

3.1. Region-wise differential expression and comparative gene-prioritization analysis

Differential expression analysis was performed independently for each brain region in the discovery dataset GSE5281 (Table 2) to account for the anatomical heterogeneity of Alzheimer’s disease. Six regions were analyzed: the entorhinal cortex (EC), hippocampus (HIP), middle temporal gyrus (MTG), posterior cingulate cortex (PC), superior frontal gyrus (SFG), and primary visual cortex (VCX).

Within each region, gene expression was compared between AD and control samples using the limma framework with empirical Bayes moderation (Ritchie et al., 2015). Diagnosis was modeled as a categorical variable with control as the reference group, and the contrast was defined as AD minus control. Accordingly, positive log2 fold-change (log2FC) values indicated higher expression in AD, whereas negative values indicated lower expression in AD. Benjamini–Hochberg correction was applied to control the false discovery rate (FDR). Genes with FDR < 0.05 were considered statistically significant differentially expressed genes, whereas the subsequent ranking analysis was performed across the complete gene set to identify top-ranked candidate genes irrespective of the significance threshold.

To examine the influence of alternative prioritization criteria, four gene-ranking strategies were evaluated. The first strategy ranked genes according to decreasing absolute log2FC, whereas the second ranked genes according to increasing FDR. The third strategy combined effect size and statistical evidence using: CombinedScore = |log2FC| × min [−log10 (FDR), 10], where the −log10 (FDR) contribution was capped at 10 to limit the influence of extremely small FDR values. The fourth strategy used a rank-based integration: RankSum = Rank (|log2FC|) + Rank (FDR), where absolute log2FC was ranked in descending order and FDR in ascending order; therefore, a smaller RankSum indicated higher priority. CombinedScore and RankSum were used as complementary prioritization measures rather than as additional statistical significance tests. The top-ranked genes obtained using the four strategies for each of the six brain regions are summarized in Table 3-Table 8.

Table 3.

Top-ranked genes in the MTG using four ranking strategies.

Strategy (MTG)	Gene	logFC	P-value	FDR	Rank |logFC|	Rank FDR	Combined Score	RankSum
logFC	PCSK1	-4.1005	1.92E-08	7.00E-06	1	56	21.1376	57
	STAT4	-3.7367	5.32E-08	1.26E-05	2	83	18.3049	85
	SIMC1P1	3.6292	3.15E-08	9.50E-06	3	68	18.2267	71
	MAFF	3.5914	5.48E-08	1.26E-05	4	83	17.5929	87
	ENC1	-3.5783	4.69E-07	3.67E-05	5	266	15.8702	271
FDR	FIG4	-2.4435	8.74E-12	1.82E-07	222	1	16.4686	223
	ACTB	-1.8928	2.16E-10	1.20E-06	1024	2	11.2056	1026
	ATP5F1C	-2.8428	2.31E-10	1.20E-06	72	2	16.8301	74
	NME7	-3.3928	1.80E-10	1.20E-06	12	2	20.0859	14
	NCALD	-2.6069	4.10E-10	1.71E-06	139	5	15.0344	144
CombinedScore	PCSK1	-4.1005	1.92E-08	7.00E-06	1	56	21.1376	57
	NME7	-3.3928	1.80E-10	1.20E-06	12	2	20.0859	14
	CALM1	-3.5626	3.61E-09	4.19E-06	7	14	19.1588	21
	NSUN6	3.4971	1.06E-08	5.20E-06	9	42	18.4792	51
	STAT4	-3.7367	5.32E-08	1.26E-05	2	83	18.3049	85
RankSum	NME7	-3.3928	1.80E-10	1.20E-06	12	2	20.0859	14
	CALM1	-3.5626	3.61E-09	4.19E-06	7	14	19.1588	21
	STMN2	-3.3815	4.81E-09	4.21E-06	14	19	18.1778	33
	NFIA	3.1866	3.31E-09	4.19E-06	20	14	17.1368	34
	HSPB3	-3.2774	5.58E-09	4.30E-06	17	27	17.5890	44

Table 8.

Top-ranked genes in the VCX using four ranking strategies.

Strategy (VCX)	Gene	logFC	P-value	FDR	Rank |logFC|	Rank FDR	Combined Score	RankSum
logFC	LOC101927709	3.4669	1.51E-05	0.052424	1	1	4.4393	2
	ABCA13	3.0452	1.13E-04	0.059524	2	33	3.7314	35
	SST	-2.7553	4.15E-05	0.052757	3	7	3.5205	10
	HTR1B	2.7464	1.41E-05	0.052424	4	1	3.5167	5
	LOC105377776	2.7165	1.08E-04	0.059524	5	33	3.3285	38
FDR	ADAM12	1.4248	7.67E-06	0.052424	672	1	1.8245	673
	COLCA1	1.0494	1.29E-05	0.052424	2328	1	1.3438	2329
	DOK3	1.3748	5.70E-06	0.052424	784	1	1.7604	785
	FAM167A-AS1	2.5439	6.94E-06	0.052424	9	1	3.2574	10
	HTR1B	2.7464	1.41E-05	0.052424	4	1	3.5167	5
CombinedScore	LOC101927709	3.4669	1.51E-05	0.052424	1	1	4.4393	2
	ABCA13	3.0452	1.13E-04	0.059524	2	33	3.7314	35
	SST	-2.7553	4.15E-05	0.052757	3	7	3.5205	10
	HTR1B	2.7464	1.41E-05	0.052424	4	1	3.5167	5
	LOC105377776	2.7165	1.08E-04	0.059524	5	33	3.3285	38
RankSum	LOC101927709	3.4669	1.51E-05	0.052424	1	1	4.4393	2
	HTR1B	2.7464	1.41E-05	0.052424	4	1	3.5167	5
	FAM167A-AS1	2.5439	6.94E-06	0.052424	9	1	3.2574	10
	SST	-2.7553	4.15E-05	0.052757	3	7	3.5205	10
	TOP1P2	2.5656	6.29E-05	0.052757	8	7	3.2781	15

The four ranking strategies identified partially overlapping but region-dependent gene-prioritization patterns. In the middle temporal gyrus (MTG; Table 3), ranking by absolute log2 fold change prioritized PCSK1, STAT4, SIMC1P1, MAFF, and ENC1, whereas FDR-based ranking emphasized FIG4, ACTB, ATP5F1C, NME7, and NCALD. The integrated strategies identified genes supported jointly by effect magnitude and statistical evidence. PCSK1 achieved the highest CombinedScore, whereas NME7 was highly ranked by both CombinedScore and RankSum; CALM1 was also retained among the top-ranked genes by both integrated approaches. These differences illustrate that effect-size- and significance-based ranking can prioritize distinct subsets of genes.

In the entorhinal cortex (Table 4), the four ranking strategies emphasized partly overlapping but distinct subsets of genes. Effect-size ranking identified C1QTNF4, ACTL6B, GABRD, CAMK1G, and PSMD8 as the genes with the largest absolute log2 fold changes, whereas FDR-based ranking prioritized MRPL41, PRXL2B, RTL8C, SDF4, and ZNRD2. The integrated strategies showed greater convergence: ZNRD2 achieved both the highest CombinedScore and the lowest RankSum, while CAMK1G and YJEFN3 were highly prioritized by both integrated approaches. ACTL6B and C1QTNF4 were additionally among the leading CombinedScore candidates, whereas MRPL41 and SMYD5 appeared among the highest-ranked genes according to RankSum. These results illustrate how integrating effect magnitude and statistical evidence can prioritize candidates that differ from those identified by either criterion alone.

Table 4.

Top-ranked genes in the EC using four ranking strategies.

Strategy (EC)	Gene	logFC	P-value	FDR	Rank |logFC|	Rank FDR	Combined Score	RankSum
logFC	C1QTNF4	-4.98014	1.94E-07	3.16E-05	1	128	22.41233	129
	ACTL6B	-4.45676	1.30E-08	6.04E-06	2	44	23.25851	46
	GABRD	-4.45323	4.02E-06	0.000234	3	357	16.1665	360
	CAMK1G	-4.41442	3.88E-09	3.37E-06	4	23	24.16008	27
	PSMD8	-4.12198	6.03E-06	0.000307	5	408	14.48187	413
FDR	MRPL41	-3.42807	6.27E-11	5.68E-07	25	1	21.41104	26
	PRXL2B	-3.15809	1.25E-10	5.68E-07	39	1	19.72481	40
	RTL8C	-2.8865	1.13E-10	5.68E-07	68	1	18.02847	69
	SDF4	-2.70446	5.44E-11	5.68E-07	98	1	16.89153	99
	ZNRD2	-4.09964	1.36E-10	5.68E-07	6	1	25.60553	7
CombinedScore	ZNRD2	-4.09964	1.36E-10	5.68E-07	6	1	25.60553	7
	CAMK1G	-4.41442	3.88E-09	3.37E-06	4	23	24.16008	27
	ACTL6B	-4.45676	1.30E-08	6.04E-06	2	44	23.25851	46
	YJEFN3	-3.89532	9.69E-10	1.55E-06	8	12	22.62745	20
	C1QTNF4	-4.98014	1.94E-07	3.16E-05	1	128	22.41233	129
RankSum	ZNRD2	-4.09964	1.36E-10	5.68E-07	6	1	25.60553	7
	YJEFN3	-3.89532	9.69E-10	1.55E-06	8	12	22.62745	20
	MRPL41	-3.42807	6.27E-11	5.68E-07	25	1	21.41104	26
	CAMK1G	-4.41442	3.88E-09	3.37E-06	4	23	24.16008	27
	SMYD5	-3.62841	1.28E-09	1.77E-06	15	15	20.86629	30

In the hippocampus (Table 5), the integrated rankings showed greater concordance. ANKRD12 and ELAVL3 were among the highest-ranked genes according to FDR, CombinedScore, and RankSum, while GPR155 and THRA were also consistently prioritized across these approaches. In contrast, ARPC1A exhibited the largest absolute log2 fold change, illustrating the distinction between ranking based primarily on expression magnitude and ranking based on joint statistical and effect-size evidence.

Table 5.

Top-ranked genes in the HIP using four ranking strategies.

Strategy (HIP)	Gene	logFC	P-value	FDR	Rank |logFC|	Rank FDR	Combined Score	RankSum
logFC	ARPC1A	-3.886	1.48E-05	0.0009	1	358	11.9122	359
	SIMC1P1	3.8248	3.13E-05	0.0014	2	464	10.9109	466
	TMEM106A	3.6151	6.64E-05	0.0023	3	609	9.5601	612
	LINC01949	3.5493	2.60E-06	0.0003	4	182	12.5402	186
	CD9	-3.4089	3.21E-05	0.0014	5	468	9.6969	473
FDR	ANKRD12	3.1109	4.41E-12	9.19E-08	8	1	21.8901	9
	ELAVL3	2.8961	5.24E-11	4.89E-07	16	2	18.277	18
	YLPM1	2.5206	7.04E-11	4.89E-07	45	2	15.9075	47
	GPR155	2.8416	1.21E-10	6.30E-07	20	4	17.6203	24
	THRA	2.87	1.60E-10	6.66E-07	19	5	17.7255	24
CombinedScore	ANKRD12	3.1109	4.41E-12	9.19E-08	8	1	21.8901	9
	ELAVL3	2.8961	5.24E-11	4.89E-07	16	2	18.277	18
	THRA	2.87	1.60E-10	6.66E-07	19	5	17.7255	24
	GPR155	2.8416	1.21E-10	6.30E-07	20	4	17.6203	24
	RTN3	-2.6583	1.96E-10	6.81E-07	28	6	16.3939	34
RankSum	ANKRD12	3.1109	4.41E-12	9.19E-08	8	1	21.8901	9
	ELAVL3	2.8961	5.24E-11	4.89E-07	16	2	18.277	18
	GPR155	2.8416	1.21E-10	6.30E-07	20	4	17.6203	24
	THRA	2.87	1.60E-10	6.66E-07	19	5	17.7255	24
	FXR1	2.9041	6.74E-09	7.80E-06	14	18	14.8343	32

In the posterior cingulate cortex (Table 6), NDUFA7, TCAP, and TMEM106A were consistently prioritized by both CombinedScore and RankSum, with SLC35E1 also shared between these approaches. The FDR-based ranking identified a largely distinct set, including RABEP1, CCSER2, CCT2, RPL15, and CENPBD2P. In the superior frontal gyrus (Table 7), SST showed the largest absolute log2 fold change and the highest CombinedScore, while LOC729870 and TYRP1 were repeatedly prioritized across several strategies. ADCYAP1 ranked highly by effect size but did not remain among the five highest-ranked genes under the integrated strategies.

Table 6.

Top-ranked genes in the PC using four ranking strategies.

Strategy (PC)	Gene	logFC	P-value	FDR	Rank |logFC|	Rank FDR	Combined Score	RankSum
logFC	PSMA3	-3.6004	2.14E-05	0.001032	1	432	10.752	433
	CACNB2	3.5578	1.26E-05	0.000777	2	338	11.0635	340
	SNX31	3.4699	1.28E-04	0.002997	3	888	8.7554	891
	ZNF705G	3.4695	1.47E-05	0.000833	4	365	10.6843	369
	ITGA9-AS1	3.3964	2.39E-06	0.000326	5	153	11.8446	158
FDR	RABEP1	-2.3763	1.66E-09	3.46E-05	268	1	10.5992	269
	CCSER2	-1.7669	7.17E-09	3.97E-05	1370	2	7.7759	1372
	CCT2	-2.0979	7.63E-09	3.97E-05	574	2	9.2324	576
	RPL15	-1.845	5.56E-09	3.97E-05	1126	2	8.1196	1128
	CENPBD2P	1.7178	2.49E-08	5.84E-05	1575	5	7.2723	1580
CombinedScore	NDUFA7	-3.2661	3.21E-07	0.000125	9	53	12.7427	62
	TCAP	2.999	1.97E-08	5.84E-05	32	5	12.6963	37
	TMEM106A	3.2385	3.12E-07	0.000125	10	52	12.6388	62
	ITGA9-AS1	3.3964	2.39E-06	0.000326	5	153	11.8446	158
	SLC35E1	2.7979	4.73E-08	6.09E-05	62	11	11.7934	73
RankSum	TCAP	2.999	1.97E-08	5.84E-05	32	5	12.6963	37
	NDUFA7	-3.2661	3.21E-07	0.000125	9	53	12.7427	62
	TMEM106A	3.2385	3.12E-07	0.000125	10	52	12.6388	62
	SLC35E1	2.7979	4.73E-08	6.09E-05	62	11	11.7934	73
	PGF	2.7781	4.51E-08	6.09E-05	67	11	11.7098	78

Table 7.

Top-ranked genes in the SFG using four ranking strategies.

Strategy (SFG)	Gene	LogFC	P-value	FDR	Rank |logFC|	Rank FDR	Combined Score	RankSum
logFC	SST	-3.6703	4.13E-07	0.001228	1	7	10.6838	8
	FOXC1	3.3659	4.45E-06	0.002808	2	33	8.5882	35
	ADCYAP1	-3.1041	1.56E-05	0.003965	3	81	7.4552	84
	TYRP1	-3.0126	2.77E-07	0.001155	4	4	8.8496	8
	LOC729870	-2.9895	4.94E-08	0.000642	5	1	9.5433	6
FDR	AICDA	-2.5425	8.16E-08	0.000642	16	1	8.1166	17
	LOC729870	-2.9895	4.94E-08	0.000642	5	1	9.5433	6
	POP4	-1.94	9.25E-08	0.000642	220	1	6.1931	221
	LOC105370549	-2.7984	2.41E-07	0.001155	8	4	8.2204	12
	TYRP1	-3.0126	2.77E-07	0.001155	4	4	8.8496	8
CombinedScore	SST	-3.6703	4.13E-07	0.001228	1	7	10.6838	8
	LOC729870	-2.9895	4.94E-08	0.000642	5	1	9.5433	6
	TYRP1	-3.0126	2.77E-07	0.001155	4	4	8.8496	8
	FOXC1	3.3659	4.45E-06	0.002808	2	33	8.5882	35
	LOC105370549	-2.7984	2.41E-07	0.001155	8	4	8.2204	12
RankSum	LOC729870	-2.9895	4.94E-08	0.000642	5	1	9.5433	6
	SST	-3.6703	4.13E-07	0.001228	1	7	10.6838	8
	TYRP1	-3.0126	2.77E-07	0.001155	4	4	8.8496	8
	LOC105370549	-2.7984	2.41E-07	0.001155	8	4	8.2204	12
	SIMC1P1	2.7894	3.51E-07	0.001217	9	6	8.1303	15

The primary visual cortex showed a distinct pattern (Table 8). LOC101927709, HTR1B, and SST recurred among the highest-ranked candidates, with LOC101927709 attaining the highest CombinedScore and lowest RankSum. However, no VCX gene reached an FDR threshold of 0.05; accordingly, these genes were considered ranking-based candidates rather than statistically significant differentially expressed genes.

Fig. 1 and Fig. 2 provide a visual comparison of the two integrated prioritization strategies, CombinedScore and RankSum, across the six brain regions. Absolute log2 fold change is shown on the x-axis and −log10 (FDR) on the y-axis, while point size and color represent CombinedScore and RankSum, respectively. Fig. 1 displays the five highest-ranked genes according to CombinedScore, whereas Fig. 2 displays the five highest-ranked genes according to RankSum. The two integrated approaches showed varying degrees of agreement across brain regions, indicating that they provide related but non-identical gene-prioritization profiles.

Fig. 1.

Top-ranked genes across brain regions based on the CombinedScore prioritization strategy.

Fig. 2.

Top-ranked genes across brain regions based on the RankSum prioritization strategy.

Agreement among the four ranking strategies was further examined using the Top-100 ranked gene sets (Fig. 3). The number of genes shared by all four strategies was 48 in EC, 24 in MTG, 14 in HIP, 12 in PC, 32 in SFG, and 27 in VCX. Thus, the extent of agreement among the ranking strategies varied across brain regions, while each strategy also retained method-specific genes.

Fig. 3.

Overlap of the Top-100 ranked genes across four prioritization strategies in six brain regions. Venn diagrams show the overlap among genes prioritized by absolute log2FC, FDR, CombinedScore, and RankSum in the (A) entorhinal cortex (EC), (B) middle temporal gyrus (MTG), (C) hippocampus (HIP), (D) posterior cingulate cortex (PC), (E) superior frontal gyrus (SFG), and (F) primary visual cortex (VCX). Numbers indicate the counts of genes unique to individual strategies or shared among multiple strategies. Each ranking strategy contributed 100 genes per region.

Regionally prioritized candidates were selected from the top-ranked genes shown in Fig. 2 based on their combined performance across multiple criteria. NME7 in MTG, ZNRD2 in EC, and ANKRD12 in HIP showed high priority according to RankSum while also exhibiting relatively large absolute log₂ fold changes, strong FDR-based evidence, and high CombinedScore values. These genes were therefore selected as representative regionally prioritized candidates for subsequent analyses.

3.2. Cross-regional recurrence of top-ranked genes

To characterize the extent of cross-regional consistency in gene prioritization, the Top-100 ranked genes from each of the six brain regions were compared separately for the logFC, FDR, CombinedScore, and RankSum strategies. The multi-set overlap diagrams (Fig. 4A–D) revealed substantial regional heterogeneity, with no gene occurring among the Top-100 genes of all six regions under any of the four ranking strategies. Nevertheless, recurrent genes were observed across subsets of regions, indicating that the regional prioritization profiles contained both shared and region-specific components.

Fig. 4.

Cross-regional overlap of the Top-100 ranked genes across six brain regions in GSE5281. The top row shows multi-set overlap diagrams for four gene-prioritization strategies: (A) logFC, (B) FDR, (C) CombinedScore, and (D) RankSum. The bottom row shows the corresponding pairwise overlap-count heatmaps: (E) logFC, (F) FDR, (G) CombinedScore, and (H) RankSum. The six regions analyzed were the entorhinal cortex (EC), hippocampus (HIP), middle temporal gyrus (MTG), posterior cingulate cortex (PC), superior frontal gyrus (SFG), and primary visual cortex (VCX).

Pairwise overlap analysis further resolved these relationships (Fig. 4E–H). Under the logFC strategy, the largest overlap was observed between MTG and SFG, with 13 shared genes. This relationship remained prominent under both CombinedScore and RankSum, with 12 shared genes under each strategy. HIP and MTG also showed comparatively high overlap under the integrated approaches, sharing 11 genes under both CombinedScore and RankSum. In contrast, EC generally exhibited limited overlap with the other regions, and VCX similarly showed relatively low cross-regional recurrence. Thus, cross-regional agreement among highly ranked genes was concentrated in specific regional pairs rather than being uniformly distributed across the six brain regions.

Supplementary Table S1–S4 report the detailed pairwise overlap counts and corresponding gene identities underlying Fig. 4E–H, enabling the pairwise overlap patterns to be traced to the specific genes shared between regions under each prioritization strategy. To evaluate cross-strategy robustness, genes occurring among the Top-100 ranked genes in at least two brain regions were compared across the four prioritization strategies. Five genes (ADCYAP1, PCSK1, SIMC1P1, SST, STAT4) were recurrent under all four strategies, whereas 18 additional genes were recurrent under three of the four strategies (Table 9). This analysis identified a subset of candidates whose recurrence was not dependent on a single ranking criterion.

Table 9.

Cross-strategy robustness of genes recurrent among the Top-100 ranked genes in at least two of the six brain regions and supported by at least two of the four prioritization strategies in GSE5281.

Cross-strategy recurrence	Number of genes	Genes
Recurrent under all 4 strategies	5	ADCYAP1, PCSK1, SIMC1P1, SST, STAT4
Recurrent under 3 of 4 strategies	18	ARPC1A, CACNB2, CALM1, CDC37, CDK13, CRYM, DHRS7B, EP400, GABRA1, GARIN5A, HSPB3, NME1, NSUN6, NVL, RTN3, SLC35E1, SRRM2, ZFP36L1

STAT4, SST, and NVL were selected for exploratory expression-level characterization based on their cross-strategy recurrence and predominantly negative AD-associated expression differences. STAT4 and SST met the recurrence criterion under all four prioritization strategies, whereas NVL met this criterion under three. Regional differential-expression estimates indicated negative AD–control log₂FC values for STAT4 and NVL in all six regions and for SST in five regions; SST showed a small, non-significant positive estimate in HIP. The expression distributions are presented in Fig. 5. Together, these findings indicate predominantly reduced expression of the selected genes in AD, with variation in effect magnitude across regions.

Fig. 5.

Region-wise expression of three recurrent genes in GSE5281. Violin-box-jitter plots show normalized expression levels of (A) STAT4, (B) SST, and (C) NVL in AD and control samples across EC, HIP, MTG, PC, SFG, and VCX. Region-wise differences between AD and control groups were evaluated using Wilcoxon rank-sum tests. Significance annotations indicate P < 0.05, P < 0.01, P < 0.001, and P < 0.0001; non-significant comparisons are not shown.

Sensitivity analyses examined Top-50, Top-100, Top-200, and Top-500 lists and recurrence requirements ranging from at least two to all six regions. Candidate counts varied with both settings. Nevertheless, SST remained recurrent in at least two regions under all four strategies at every tested list size. STAT4 met this recurrence criterion under all four strategies at Top-100 and larger list sizes, whereas NVL showed greater dependence on the ranking criterion. An alternative significance-based analysis identified SST and STAT4 as differentially expressed at FDR < 0.05 in four regions and NVL in three. Under both integrated strategies, MTG–SFG and HIP–MTG shared 12 and 11 Top-100 genes, respectively, with Jaccard similarities of 0.0638 and 0.0582 and BH-adjusted hypergeometric P-values of 3.87 × 10⁻¹ ³ and 1.09 × 10⁻¹ ¹. These findings indicate persistent recurrence for selected candidates despite threshold-dependent list membership, together with modest regional overlaps that were enriched relative to the specified gene-universe null model.

3.3. Discovery-set expression score and exploratory six-gene clustering

The three-gene AD-like expression score, −z(STAT4) − z(SST) − z(NVL), showed higher median values in AD than in controls across all six GSE5281 brain regions (Fig. 6 and Table 10). Regional Wilcoxon rank-sum comparisons remained significant after Benjamini–Hochberg adjustment across the six tests. The largest difference in median scores occurred in MTG, with values of 2.595 in AD and −2.829 in controls (adjusted P = 3.94 × 10⁻⁷). Median differences were also pronounced in SFG, PC, and VCX, whereas EC and HIP showed smaller differences in the same direction.

Fig. 6.

Three-gene AD-like expression signature across six brain regions in GSE5281. The signature was calculated as −z(STAT4) − z(SST) − z(NVL), with gene directions oriented according to their predominant AD-associated expression changes. Violin-box-jitter plots show the distribution of the composite score in AD and control samples across EC, HIP, MTG, PC, SFG, and VCX. Region-wise differences were assessed using Wilcoxon rank-sum tests with Benjamini–Hochberg correction across the six regional comparisons. Significance annotations indicate adjusted p-values: P_adj < 0.05, P_adj < 0.01, P_adj < 0.001, and P_adj < 0.0001.

Table 10.

Region-wise comparison of the three-gene AD-like expression signature score between AD and control samples in GSE5281.

Region	Median AD	Median control	BH-adjusted p
EC	1.170	-0.597	0.0147
HIP	0.278	-0.975	0.0145
MTG	2.595	-2.829	3.94 × 10⁻⁷
PC	1.855	-1.509	9.64 × 10⁻⁵
SFG	1.362	-2.217	4.08 × 10⁻⁶
VCX	1.770	-1.586	2.21 × 10⁻⁴

The three-gene score therefore summarized a recurrent AD-associated downregulation pattern across multiple brain regions in the GSE5281 discovery dataset. Because gene selection, score orientation, and score evaluation were all performed using GSE5281, these results were considered discovery-set evidence and were subsequently evaluated in the independent GSE48350 cohort.

To examine whether the selected genes captured sample-level molecular structure within GSE5281, exploratory unsupervised analysis was performed using a combined six-gene panel comprising the regional candidates ZNRD2, ANKRD12, and NME7 and the recurrent candidates STAT4, SST, and NVL. The analysis included 161 samples, comprising 87 AD and 74 control samples across six brain regions. Expression values for each gene were z-standardized across all samples so that genes contributed comparably to the distance-based analyses.

A single global UMAP embedding was constructed from the standardized six-gene expression matrix and displayed separately by brain region. UMAP used Euclidean distance, 10 nearest neighbors, and a minimum distance of 0.3. Complementary k-means clustering was performed in the original standardized six-dimensional expression space. Candidate solutions from (k = 2) to (k = 6) were evaluated using average silhouette width, with 100 random initializations for each solution. The highest average silhouette width was obtained for (k = 2) (0.258), compared with 0.221, 0.194, 0.200, and 0.198 for (k = 3), 4, 5, and 6, respectively. The two-cluster solution was therefore retained for subsequent analysis.

Diagnosis labels were not supplied to UMAP or k-means. However, the input genes had been selected through preceding analyses of the discovery dataset that incorporated AD–control information. Consequently, these analyses constituted an internal evaluation of exploratory molecular stratification conditional on the selected features. Cluster–diagnosis associations were interpreted descriptively within the discovery cohort.

The UMAP visualization showed regional variation in AD–control separation (Fig. 7). HIP and MTG exhibited relatively distinct diagnostic group distributions, while EC also showed appreciable separation. PC displayed partial separation with overlap, whereas more extensive overlap was apparent in SFG and VCX. These visual patterns suggested that the selected panel captured diagnosis-associated expression structure to varying degrees across anatomical regions.

Fig. 7.

UMAP visualization of the selected six-gene panel in GSE5281. A single global UMAP embedding was constructed using the z-standardized expression values of ZNRD2, ANKRD12, NME7, STAT4, SST, and NVL across 161 samples and displayed by brain region. Points represent samples colored by diagnosis. Ellipses show diagnosis-specific dispersion at the 68% level under a normal approximation. UMAP used Euclidean distance, 10 nearest neighbors, and a minimum distance of 0.3. Diagnosis labels were used for annotation only. Because gene selection incorporated diagnosis information, the visualization represents exploratory stratification within the discovery cohort.

The global k-means analysis identified two molecular clusters, whose labels were assigned post hoc according to diagnostic enrichment. Cluster 1 contained 91 samples, including 78 AD and 13 control samples, whereas Cluster 2 contained 70 samples, including 9 AD and 61 control samples. The AD proportions were therefore 85.7% and 12.9%, respectively. Assigning Cluster 1 to AD and Cluster 2 to control yielded a post hoc cluster–diagnosis concordance of 86.3% (139/161 samples). The between-cluster sum of squares represented 29.7% of the total sum of squares in the standardized expression space. These statistics describe discovery-set association and do not constitute estimates of independently validated diagnostic performance.

Regional cluster composition further demonstrated variation in diagnostic enrichment (Fig. 8A–B; Table 11). MTG showed the largest contrast in AD proportions between clusters: all 15 samples in Cluster 1 were AD, compared with 1 of 13 samples in Cluster 2 (7.7%). In EC, AD samples comprised 83.3% of Cluster 1 and none of Cluster 2; the corresponding proportions in PC were 75.0% and 0.0%. Cluster 1 was also AD-enriched in HIP, SFG, and VCX, with AD proportions of 85.7%, 90.9%, and 78.3%, compared with 25.0%, 25.0%, and 12.5% in Cluster 2, respectively.

Fig. 8.

Internal evaluation of molecular stratification using the selected six-gene panel. (A) Percentage of AD samples within each molecular cluster by brain region; annotations indicate AD counts relative to the total number of samples in each region–cluster subgroup. (B) Percentage of each region’s samples assigned to each cluster; annotations show cluster percentages and sample counts, with percentages summing to 100% within each region. (C) Individual silhouette widths from the global (k = 2) solution, displayed by region and ordered by silhouette width within each panel. Clustering used the z-standardized expression values of ZNRD2, ANKRD12, NME7, STAT4, SST, and NVL in the original six-dimensional expression space. The two-cluster solution maximized average silhouette width among (k = 2) to (k = 6), with 100 random initializations per candidate solution. Diagnosis labels were excluded from clustering; cluster labels were assigned post hoc according to diagnostic enrichment.

Table 11.

Distribution of global k-means clusters by brain region using the selected six-gene panel in GSE5281.

Region	Cluster	Total	AD	Control	% AD
EC	Cluster 1	12	10	2	83.3
EC	Cluster 2	11	0	11	0.0
HIP	Cluster 1	7	6	1	85.7
HIP	Cluster 2	16	4	12	25.0
MTG	Cluster 1	15	15	0	100.0
MTG	Cluster 2	13	1	12	7.7
PC	Cluster 1	12	9	3	75.0
PC	Cluster 2	10	0	10	0.0
SFG	Cluster 1	22	20	2	90.9
SFG	Cluster 2	12	3	9	25.0
VCX	Cluster 1	23	18	5	78.3
VCX	Cluster 2	8	1	7	12.5

Table Note: Clustering was performed globally using the standardized six-gene expression matrix. Cluster 1 was designated post hoc as the cluster with the higher overall AD proportion. % AD denotes the percentage of AD samples within each region–cluster subgroup.

Silhouette analysis provided an internal assessment of cluster cohesion and separation (Fig. 8C). The global mean silhouette width was 0.258, with a median of 0.282; four samples (2.48%) had negative silhouette values. These results indicated modest overall separation between the molecular clusters. Regional summaries of the global silhouette values were highest in MTG (0.317), followed by VCX (0.273), SFG (0.247), EC (0.246), PC (0.245), and HIP (0.207). These values summarize how samples from each region fitted the global clustering solution; they were not derived from separate regional clustering analyses. Because multiple regional samples in GSE5281 may originate from the same donor and explicit donor identifiers were not available in the metadata used for the present analysis, cluster–diagnosis relationships were interpreted descriptively rather than subjected to sample-level inferential testing.

Together, these analyses identified an AD-associated expression pattern in the selected three-gene score and diagnosis-enriched sample groups in the six-gene expression space. The two-cluster solution showed 86.3% post hoc concordance with diagnosis, alongside modest overall separation. Subsequent analyses assessed external support for the selected gene-expression patterns and three-gene score in GSE48350

3.4. External evaluation of ranking strategies, selected genes, and the three-gene score

External evaluation used GSE48350 and was restricted to the anatomically shared EC, HIP, and SFG regions. Postcentral gyrus samples were excluded because this region is anatomically distinct from the posterior cingulate cortex represented in GSE5281. External benchmarking compared the four prioritization strategies across Top-50, Top-100, Top-200, and Top-500 candidate lists. At Top-100, each strategy contributed 300 gene–region comparisons across the three shared regions. Directional agreement was 68.3% for absolute-logFC ranking, 72.0% for FDR, 69.0% for CombinedScore, and 71.7% for RankSum. CombinedScore and RankSum each recovered three gene–region associations with concordant direction and FDR < 0.05 in both datasets, compared with two for absolute-logFC ranking and none for FDR ranking. Across all evaluated list sizes, directional agreement ranged from 66.7% to 72.0%. Relative performance varied with list size and evaluation criterion, with no consistently superior strategy (Supplementary Table S7).

For the selected six-gene panel, regional effect estimates showed directional agreement alongside differences in magnitude (Fig. 9A). ZNRD2, STAT4, and NVL had negative AD-associated effects in all three shared regions in both datasets. ANKRD12 showed concordant directions in HIP and SFG but discordance in EC. NME7 showed concordant negative directions in EC and HIP, with a small positive external estimate in SFG. SST showed concordant downregulation in SFG but discordant directions in EC and HIP. Complete effect estimates, moderated standard errors, 95% confidence intervals, nominal P-values, FDR-adjusted P-values, and sample counts for both datasets are provided in Supplementary Tables S8 and S9.

Fig. 9.

External evaluation of directionality of the selected six-gene panel in GSE48350.

Directional agreement was observed in four of six comparisons in EC and five of six comparisons in both HIP and SFG, corresponding to 66.7%, 83.3%, and 83.3%, respectively (Fig. 9B; Table 13). Overall, 14 of 18 gene–region comparisons were concordant (77.8%). Cross-dataset effect-size correspondence was weak (Pearson r = 0.158; Spearman ρ = 0.216; Fig. 9C). The corresponding nominal P-values were 0.532 and 0.390; dependence among gene–region comparisons was not modeled.

Table 13.

Summary of directional concordance for the selected six-gene panel between GSE5281 and GSE48350 across shared brain regions.

Region	Matched comparisons	Total comparisons	Match rate
EC	4	6	66.7%
HIP	5	6	83.3%
SFG	5	6	83.3%
Overall	14	18	77.8%

Table Note: Directional concordance was defined as agreement in the sign of the AD-versus-control log₂ fold change between GSE5281 and GSE48350. Six genes were evaluated across three shared brain regions, yielding 18 gene–region comparisons. Directional agreement does not imply replication of statistical significance or effect-size magnitude

The discovery-selected STAT4–SST–NVL expression score was additionally evaluated in 124 external tissue samples without gene reselection or sign reorientation. Expression was standardized within each external region before score calculation. An exploratory joint model adjusting for brain region estimated higher scores in AD than in controls (mean difference = 1.201 score units; 95% CI, 0.469–1.932; nominal P = 0.00149). This sample-level model did not account for donor dependence.

External evaluation therefore identified directional agreement in 14 of 18 selected gene–region comparisons, including consistent directions for ZNRD2, STAT4, and NVL across all three shared regions. The benchmarking results further quantified how external support varied with prioritization strategy and selection threshold.

(A) AD-versus-control log₂ fold changes for ZNRD2, ANKRD12, NME7, STAT4, SST, and NVL in GSE5281 and GSE48350 across the shared brain regions EC, HIP, and SFG. (B) Direction-concordance heatmap showing agreement in the sign of the regional log₂ fold changes between datasets; green indicates matched direction and salmon indicates discordance. Values within cells show the corresponding log₂ fold changes in GSE5281 and GSE48350. (C) Cross-dataset comparison of regional effect sizes, with circles indicating directionally concordant gene–region comparisons and triangles indicating discordant comparisons. Across all 18 comparisons, Pearson r = 0.158 (P = 0.532) and Spearman ρ = 0.216 (P = 0.390). PCG in GSE48350 was not considered equivalent to posterior cingulate cortex (PC) in GSE5281 and was therefore excluded from the cross-dataset comparison.

4. Discussion

This study presents a region-stratified, multi-metric framework for prioritizing Alzheimer’s disease (AD)-associated transcriptomic signals across anatomically heterogeneous brain regions. Four complementary ranking strategies (absolute log₂ fold change, FDR, CombinedScore, and RankSum) were systematically compared. Rather than assuming that integrated criteria are inherently superior to individual measures, the analysis evaluated cross-regional overlap, recurrence across ranking strategies, resampling stability, formulation sensitivity, and independent external support. This multi-level design helps distinguish signals that persist across analytical choices from those that depend more strongly on a particular ranking criterion or anatomical context.

External benchmarking showed that relative performance varied according to candidate-list size and evaluation criterion, although RankSum showed generally strong directional agreement across the evaluated thresholds. RankSum achieved the highest or joint-highest directional concordance at Top-50 and Top-200 and remained close to the highest value at Top-100, whereas CombinedScore showed the highest directional concordance at Top-500. At Top-100, directional agreement across the four strategies ranged from 68.3% to 72.0%, with RankSum reaching 71.7%; CombinedScore and RankSum each recovered three gene–region associations meeting the prespecified strict-replication criterion. Cross-regional recurrence was likewise dependent on candidate-list size and the required degree of regional representation, indicating that recurrence is better regarded as a graded property than as a fixed characteristic of a gene. Nevertheless, several signals persisted across analytical settings: SST remained recurrent under all four strategies across all tested list sizes, STAT4 showed broad recurrence from the Top-100 level onward, and NVL was recovered under most configurations. At the same time, the modest Jaccard similarity observed among regional ranked-gene lists indicated substantial anatomical specificity, consistent with previous multi-region and cell-type-resolved studies of AD (Liang et al., 2008, Mathys et al., 2024).

A predefined three-gene expression score comprising STAT4, SST, and NVL provided a complementary sample-level representation of the recurrent transcriptional pattern. Although higher AD-like scores were observed across the GSE5281 regions, these discovery-set comparisons are not independent because gene selection, score orientation, and score evaluation were performed using the same cohort. When the predefined score was evaluated without gene reselection or reorientation in GSE48350, it remained higher in AD in an exploratory model accounting for brain region, with an adjusted mean difference of 1.201 score units (95% CI, 0.469–1.932). This finding provides external support for the composite expression pattern but should not be interpreted as validation of diagnostic performance, particularly because only three regions were anatomically matched and complete donor-level information was not available for all external analyses.

Exploratory clustering of a six-gene panel comprising ZNRD2, ANKRD12, NME7, STAT4, SST, and NVL further examined whether the prioritized candidates captured sample-level molecular structure. The resulting clustering showed diagnosis-associated organization, but geometric separation was modest, indicating that the selected genes captured only part of the heterogeneity present in the discovery cohort. Because the six genes were themselves selected using AD–control information from GSE5281, the observed post hoc diagnosis enrichment should be interpreted as exploratory characterization of internal molecular structure rather than classification performance. External evaluation of the same six-gene panel showed concordant AD–control directions in 14 of 18 gene–region comparisons, corresponding to 77.8% directional agreement. In contrast, cross-dataset correlations of effect-size magnitude were weak and nonsignificant, indicating that reproducibility was stronger for expression direction than for quantitative effect magnitude.

SST was among the recurrent candidates identified in the present analysis and has also been implicated in previous AD transcriptomic and single-nucleus studies. Reduced SST expression and vulnerability of somatostatin-positive interneurons have been reported across several affected brain regions (Almeida and Higa, 2026, Mathys et al., 2024, Mathys et al., 2019). Its persistence across ranking strategies and candidate-list sizes in the present analysis is therefore consistent with an established disease-associated signal. However, SST showed less consistent directional reproducibility in GSE48350 than STAT4 or NVL, illustrating that strong prior biological support does not necessarily imply uniform expression behavior across cohorts and anatomical regions.

STAT4 emerged as a second recurrent candidate with strong external directional support. STAT4 is a cytokine-responsive transcription factor involved in immune and inflammatory signaling. Although less extensively studied in AD than SST, experimental work has linked STAT4-dependent signaling to inflammatory regulation and modulation of synaptic plasticity (X. Zhang et al., 2023). Its broad recurrence in GSE5281 together with concordant AD-associated direction in all three anatomically matched regions of GSE48350 therefore identifies STAT4 as a biologically plausible but still under-explored candidate. Nevertheless, tissue-level STAT4 expression should not be interpreted directly as evidence of activation or suppression of the corresponding signaling pathway, and cell-type-resolved studies will be required to clarify the biological origin of this signal.

NVL represents a comparatively under-characterized finding. The gene encodes a nuclear AAA-type ATPase involved in pre-ribosomal RNA processing and maturation of the 60S ribosomal subunit. Alterations in ribosomal function and protein synthesis have previously been implicated in AD-associated cellular dysfunction (Ding et al., 2005, Prattes et al., 2019). The recurrent reduction of NVL observed here therefore provides a plausible transcriptomic connection to this broader biological process. However, a direct mechanistic role for NVL in AD has not been established, and its identification should be regarded as hypothesis-generating rather than evidence of a demonstrated disease mechanism.

The regionally prioritized candidates provide a complementary view of anatomical heterogeneity. ZNRD2 was strongly prioritized in the entorhinal cortex, ANKRD12 in the hippocampus, and NME7 in the middle temporal gyrus. NME7 has recently been implicated in primary-cilium assembly and microtubule stability (Ji et al., 2025), whereas ZNRD2 and ANKRD12 remain comparatively poorly characterized in neurodegeneration. Their recovery through multiple complementary ranking criteria, without reliance on prior AD annotation, illustrates a practical advantage of the region-wise prioritization framework: less extensively studied candidates can be retained when their statistical and effect-size characteristics support their relevance within a specific anatomical context. At the same time, these genes should remain exploratory until their expression patterns and biological functions are confirmed in additional cohorts and cell-type-resolved systems.

The regional distribution of the prioritized signals should also be interpreted in relation to the selective anatomical vulnerability characteristic of AD. The entorhinal cortex and hippocampus are among the earliest and most severely affected regions, whereas association cortices such as the middle temporal, posterior cingulate, and superior frontal regions generally become involved as disease progresses, and the primary visual cortex is comparatively affected later. The observation that different genes were preferentially prioritized across these regions is therefore consistent with the possibility that AD-associated transcriptional responses vary with regional cellular composition, vulnerability, and pathological burden. Conversely, recurrent genes such as STAT4, SST, and NVL may represent molecular alterations that extend across anatomically distinct regions rather than being restricted to a single regional context. The coexistence of recurrent and regionally restricted signals therefore supports a model in which AD transcriptomic alterations include both broadly shared disease-associated processes and region-dependent molecular responses. Because GSE5281 is cross-sectional and does not provide a longitudinal representation of pathological spread, however, regional differences in expression should not be interpreted as direct temporal stages of disease progression.

These findings should be considered within the broader landscape of AD transcriptomics. Large multi-region, single-cell, and single-nucleus studies have demonstrated that AD-associated transcriptional alterations reflect interactions among anatomical vulnerability, cell type, pathological stage, and individual heterogeneity rather than a uniform pan-brain molecular program (Allen et al., 2016, Mathys et al., 2024). The present study integrates region-wise transcriptomic prioritization with an exploratory machine-learning layer. The transcriptomic component systematically compares alternative gene-prioritization criteria and evaluates their robustness across regions, resampling conditions, and an independent cohort, whereas the machine-learning component examines whether the selected multigene panel captures coherent diagnosis-associated structure at the sample level. UMAP and k-means clustering were therefore used as complementary unsupervised tools to characterize the multivariate organization represented by the prioritized genes rather than to perform independent gene selection or estimate predictive classification performance. This integration of transcriptomic prioritization and exploratory machine learning provides two complementary levels of evidence: gene-level robustness and recurrence, and sample-level molecular structure.

Several limitations should be acknowledged. GSE5281 is a microarray dataset with modest region-specific sample sizes, and complete donor-level information was not available for all analyses. Although laser-capture microdissection enriches neuronal populations, it does not fully resolve neuronal subtype or cell-state heterogeneity. Differences in tissue sampling, preprocessing, donor composition, and disease heterogeneity may also contribute to variation between GSE5281 and GSE48350. External evaluation was restricted to EC, HIP, and SFG because these were the anatomically matched regions available in both datasets, and the weak cross-dataset effect-size correlations indicate that external support remains partial rather than quantitative replication of effect magnitude. Larger RNA-seq cohorts with detailed donor-level, clinical, and neuropathological information, together with single-nucleus and spatial transcriptomic studies, will be required to determine whether STAT4, SST, NVL, ZNRD2, ANKRD12, and NME7 represent cell-type-specific disease responses, markers of selective regional vulnerability, or downstream consequences of neurodegeneration.

Overall, AD-associated gene prioritization was both criterion-dependent and anatomically heterogeneous, yet a subset of transcriptional signals persisted across ranking strategies, candidate-list sizes, resampling, and an independent cohort. RankSum showed generally strong external directional agreement, although its relative performance depended on list size and evaluation criterion. The principal strength of the framework therefore lies not in selecting a single “best” ranking strategy, but in identifying which candidate signals remain credible when ranking criterion, anatomical region, recurrence definition, sampling variation, scoring formulation, and external cohort are varied. The convergence of recurrent signals such as SST, STAT4, and NVL with regionally prioritized candidates such as ZNRD2, ANKRD12, and NME7 provides a focused set of hypotheses for subsequent cell-type-resolved and mechanistic investigation.

5. Conclusion

This study applied a region-stratified, multi-metric gene-prioritization framework to the GSE5281 Alzheimer’s disease transcriptomic dataset. Differential expression and gene prioritization were performed separately across six brain regions using absolute log₂ fold change, FDR, CombinedScore, and RankSum, followed by cross-regional comparison, robustness assessment, exploratory molecular characterization, and external evaluation.

STAT4, SST, and NVL emerged as recurrent cross-regional candidates, whereas ZNRD2, ANKRD12, and NME7 were preferentially prioritized within specific brain regions, highlighting the coexistence of broadly shared and region-dependent transcriptional alterations in AD. Recurrence varied with candidate-list size and ranking criterion, while resampling and formulation-sensitivity analyses indicated that the main findings were reasonably robust to sampling variation and score specification.

A predefined three-gene expression score comprising STAT4, SST, and NVL remained higher in AD in an exploratory region-adjusted analysis of GSE48350. Exploratory clustering of the six-gene panel revealed diagnosis-associated expression structure, although cluster separation was modest. External evaluation showed concordant AD–control directions in 14 of 18 gene–region comparisons, while cross-dataset effect-size correlations were weak, indicating stronger reproducibility of direction than of effect magnitude. RankSum showed generally strong external directional agreement, although relative performance varied across thresholds and evaluation criteria.

Overall, the framework provides a transparent approach for identifying transcriptomic signals that remain credible across ranking criteria, anatomical regions, sampling variation, and an independent cohort. The recurrent and regionally prioritized candidates identified here provide focused hypotheses for future validation in larger RNA-seq, single-nucleus, spatial transcriptomic, and functional studies.

CRediT authorship contribution statement

Zohre Moattar-Husseini: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Abbas Ahmadi: Writing – review & editing, Validation, Supervision, Project administration, Methodology, Investigation, Conceptualization. Nafiseh Mohebi: Writing – review & editing, Validation, Investigation, Conceptualization.

Ethics declaration

The authors have NOT obtained informed consent from participants or their legal representatives. The present study involved secondary analysis of publicly available GEO datasets (GSE5281 and GSE48350). The authors did not recruit participants or collect new biological specimens. Accordingly, new informed consent was not required for the present secondary analysis. Ethics approval and consent requirements related to the original collection of the samples were addressed by the original investigators.

This study was performed in compliance with relevant laws, regulatory frameworks and guidelines where the research took place. Ethics committee approval was not required under relevant laws and institutional guidelines. This study involved a secondary analysis of publicly available gene-expression datasets obtained from the Gene Expression Omnibus (GEO; GSE5281 and GSE48350). No new participants were recruited and no new human specimens were collected for the present study. Therefore, additional ethics committee approval was not required for this secondary analysis.

Declaration of Generative AI and AI-Assisted Technologies in the Manuscript Preparation Process

During the preparation of this work, the author(s) used ChatGPT in order to improve the language and readability of the manuscript. After using this tool, the author(s) reviewed and edited the content as needed and take full responsibility for the content of the article.

Funding

This work was supported by the Iran National Science Foundation (INSF), Grant No. 4039476.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary material

Supplementary material