Integrating polygenic and transcriptional risk scores for detecting Alzheimer's disease Authors: Jiyun Hwang, Jung‐Min Pyun, Joo‐Yeon Lee, Jeong Su Park, Paula J. Bice, Andrew J. Saykin, Joohon Sung, SangYun Kim, Young Ho Park, Kwangsik Nho, for the Alzheimer's Disease Neuroimaging Initiative Published: 2026 ABSTRACT AbstractINTRODUCTIONEarly detection of Alzheimer's disease (AD) is essential, yet existing biomarkers are invasive or costly. Polygenic risk scores (PRS) and transcriptional risk scores (TRS) may offer accessible alternatives, but their combined predictive performance remains understudied.METHODSWe calculated PRS and TRS using genome‐wide genotype and blood transcriptome data from two ancestrally distinct cohorts: Alzheimer's Disease Neuroimaging Initiative (ADNI, N = 313) and Seoul National University Bundang Hospital (SNUBH, N = 173). Logistic regression and machine learning models assessed associations of PRS and TRS with AD and AD classification performance.RESULTSIndividuals with high PRS and TRS values showed larger odds ratios for AD, 2.5‐fold in ADNI and 3.4‐fold in SNUBH, compared to those with low PRS and TRS values. The integrated PRS–TRS model achieved better classification performance (area under the curve [AUC] 0.705) than the PRS model (AUC 0.635).DISCUSSIONIntegrating static genetic and dynamic transcriptomic information from blood improves early detection of AD across diverse populations. BACKGROUND Alzheimer's disease (AD) is the most common cause of dementia, affecting approximately 50 million individuals worldwide. 1 Its pathological process begins long before clinical symptoms, marked by amyloid‐β (Aβ) plaque accumulation and neurofibrillary tangles of hyperphosphorylated tau. 2 Despite therapeutic advances, no definitive cure exists, underscoring the need for early detection strategies to enable timely intervention. 3 , 4 To support early diagnosis, various biomarkers have been developed, including cerebrospinal fluid (CSF) Aβ and tau measures, and neuroimaging modalities such as positron emission tomography (PET). 5 , 6 , 7 However, the invasiveness of lumbar puncture and the high cost and limited availability of PET imaging restrict their clinical utility, particularly in population‐level screening. 6 , 8 Additionally, these markers primarily detect AD‐related pathology at prodromal or symptomatic stages, highlighting the need for complementary biomarkers that are less invasive, more scalable, and suitable for earlier, population‐level risk assessment. This has prompted interest in blood‐based biomarkers as accessible and non‐invasive alternatives. Among blood‐based biomarker approaches, polygenic risk scores (PRS) 9 and transcriptional risk scores (TRS) 10 have each shown promise. PRS aggregates the effects of numerous common genetic variants identified through genome‐wide association studies (GWAS), capturing lifelong inherited susceptibility to AD. 11 , 12 However, as a static measure based on germline information, PRS cannot reflect dynamic disease processes or stage‐specific changes. TRS, derived from the polarized expression of disease‐relevant genes, captures molecular alterations influenced by both genetic and environmental factors and may reflect dynamic pathological processes not represented by static genetic risk profiles such as PRS. 10 , 13 , 14 While PRS and TRS each provide value, using them separately may not fully capture AD's complex etiology. We hypothesized that integrating blood‐based PRS and TRS would improve AD risk stratification by leveraging both static genomic susceptibility and dynamic transcriptomic dysregulation. This study developed and validated a combined PRS–TRS model across two independent cohorts, the Alzheimer's Disease Neuroimaging Initiative (ADNI) and a Korean cohort from Seoul National University Bundang Hospital (SNUBH), aiming to enhance diagnostic performance beyond single‐omics approaches. METHODS Study participants This study utilized data from two independent cohorts, the ADNI as the discovery cohort, and the SNUBH cohort as the replication cohort. ADNI, launched in 2003 as a public‐private partnership, is a multicenter longitudinal study designed to evaluate neuroimaging, biomarker, and cognitive measures for tracking the progression of mild cognitive impairment (MCI) and early‐stage AD. For the present analysis, we included ADNI participants classified as cognitively normal (CN) or with dementia due to AD, for whom both genotype and blood gene expression data were available. Additional information is described in previous reports 15 , 16 and summarized in the Supplementary Methods. The SNUBH cohort included individuals aged 65 years or older who were evaluated for cognitive symptoms in the Department of Neurology, SNUBH, between April 2020 and February 2023. All participants underwent comprehensive clinical and neuropsychological assessments. CN and probable AD diagnoses were assigned based on standardized criteria, including Mini‐Mental State Examination (MMSE) performance, informant‐based cognitive screening, and clinical guidelines. 17 Detailed inclusion criteria and diagnostic definitions are provided in prior publications 18 and outlined in the Supplementary Methods. RESEARCH IN CONTEXT Systematic review: We searched PubMed and Google Scholar for previous studies integrating polygenic risk scores (PRS) and transcriptional risk scores (TRS) in Alzheimer's disease (AD). While several studies evaluated each risk score separately, few investigated their combined classification performance for AD using independent and ancestrally distinct cohorts. Interpretation: Our findings show that integrating PRS and TRS substantially improves the risk prediction and classification performance for AD, suggesting that static genetic predisposition and dynamic gene expression dysregulation may provide complementary information on early detection of AD. Future directions: Future studies should validate our integrative framework in larger, independent, and multiethnic cohorts, and assess whether our PRS–TRS integration predicts preclinical or cognitive decline trajectories. In addition, expanding our omics‐based risk score approach to other omics layers may further improve early detection and personalized prevention. Systematic review: We searched PubMed and Google Scholar for previous studies integrating polygenic risk scores (PRS) and transcriptional risk scores (TRS) in Alzheimer's disease (AD). While several studies evaluated each risk score separately, few investigated their combined classification performance for AD using independent and ancestrally distinct cohorts. Systematic review: We searched PubMed and Google Scholar for previous studies integrating polygenic risk scores (PRS) and transcriptional risk scores (TRS) in Alzheimer's disease (AD). While several studies evaluated each risk score separately, few investigated their combined classification performance for AD using independent and ancestrally distinct cohorts. Interpretation: Our findings show that integrating PRS and TRS substantially improves the risk prediction and classification performance for AD, suggesting that static genetic predisposition and dynamic gene expression dysregulation may provide complementary information on early detection of AD. Interpretation: Our findings show that integrating PRS and TRS substantially improves the risk prediction and classification performance for AD, suggesting that static genetic predisposition and dynamic gene expression dysregulation may provide complementary information on early detection of AD. Future directions: Future studies should validate our integrative framework in larger, independent, and multiethnic cohorts, and assess whether our PRS–TRS integration predicts preclinical or cognitive decline trajectories. In addition, expanding our omics‐based risk score approach to other omics layers may further improve early detection and personalized prevention. Future directions: Future studies should validate our integrative framework in larger, independent, and multiethnic cohorts, and assess whether our PRS–TRS integration predicts preclinical or cognitive decline trajectories. In addition, expanding our omics‐based risk score approach to other omics layers may further improve early detection and personalized prevention. This study was approved by the institutional review board (IRB) of the SNUBH (IRB No. B‐2004‐604‐305) in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to enrollment. Genotyping and imputation In the ADNI, genome‐wide genotyping was performed on DNA extracted from peripheral blood using the Human610‐Quad BeadChip, HumanOmni Express BeadChip, and HumanOmni 2.5 M BeadChip, and apolipoprotein E (APOE) genotyping was conducted separately to determine ε4 allele status. 15 , 16 Genotyping in the SNUBH was performed using the Illumina Infinium Asian Screening Array, with APOE genotyping completed using a TaqMan assay. Quality control (QC) was conducted using PLINK version 1.9 software (www.cog‐genomics.org/plink). 19 Genotype imputation was conducted separately for each cohort using population‐appropriate reference panels: the Haplotype Reference Consortium (HRC) panel for ADNI 20 and the 1000 Genomes Project Phase 3 version 5 for SNUBH. 21 Details on quality control process and imputation procedures are provided in the Supplementary Methods. Blood RNA gene expression profiling In the ADNI, blood‐based gene expression profiling was conducted using the Affymetrix Human Genome U219 microarray platform (Affymetrix, Santa Clara, CA). In the SNUBH, RNA sequencing was performed using the NovaSeq 6000 platform (Macrogen Inc., South Korea) following standardized library preparation protocols. Both cohorts underwent data preprocessing and quality control prior to downstream analyses. Detailed procedures for RNA extraction, sequencing, normalization, and quality control are provided in the Supplementary Methods. Polygenic risk score calculation Individual PRSs were calculated using PRSice software version 2.3.3. 22 Summary statistics from a large‐scale GWAS of AD by Jansen et al. were used as the base dataset, comprising approximately 13,367,300 variants derived from 71,880 AD cases and 383,378 controls. 23 All single nucleotide polymorphisms (SNPs) located within a 1 Mb window surrounding the APOE gene were excluded from PRS computation. The number of SNPs included in the final PRS for each cohort is reported in Table S1. In ADNI, PRS distributions in the study subset were compared with those of the full ADNI cohort with GWAS genotyping data (Figure S1). Transcriptional risk score calculation Before calculating individual TRSs, we identified AD‐associated target genes by integrating GWAS summary data 23 and publicly available blood expression quantitative trait loci (eQTL) data. 24 This integration was performed using summary data‐based Mendelian randomization (SMR), a method designed to prioritize genes whose expression levels are likely to mediate genetic risk. 25 Genes reaching significance after false discovery rate (FDR) correction were retained for downstream analyses. Using the identified target genes, TRSs for each individual were then computed as follows. 10 Gene expression levels were first standardized to z‐scores with a mean of 0 and a standard deviation of 1 across individuals. Subsequently, we determined the directionality of risk for each gene based on eQTL‐derived associations between AD risk alleles and gene expression levels. Genes for which risk alleles were associated with upregulated expression were classified as high expression, whereas those for which risk alleles were linked to reduced expression were categorized as low expression. In rare instances, certain genes were linked to both increased and decreased expression levels due to their association with distinct SNPs exerting opposing eQTL effects. Such genes were excluded from the TRS calculation to avoid ambiguity. To harmonize the direction of risk across genes, we applied a polarity correction to the standardized expression values. For genes identified as low expression, the sign of z‐scores was changed so that higher values uniformly reflected greater inferred AD risk. Finally, an individual's TRS was calculated as the sum of these polarized z‐scores across all selected target genes. The number of genes and the specific gene names included in the final TRS for each cohort are reported in Table S1. Statistical analysis To evaluate the independent associations of PRS and TRS with AD diagnosis, multivariable logistic regression models were fitted for each score separately in the ADNI and SNUBH cohorts. Odds ratios (ORs) and 95% confidence intervals (CIs) were estimated, with adjustments for age, sex, and APOE ε4 carrier status. To explore whether combining PRS and TRS improved stratification of AD diagnosis rates compared with either score alone, participants were categorized into three groups based on PRS, TRS, or both. For PRS and TRS analyses, individuals were classified into tertiles according to the distribution of each score. For the combined PRS–TRS analysis, individuals with both PRS and TRS in the first tertile and those with both scores in the third tertile were designated as “PRS 1st tertile & TRS 1st tertile” and “PRS 3rd tertile & TRS 3rd tertile” groups, respectively. All other combinations were classified as “Intermediate”. AD diagnosis rates were descriptively compared across the three groups. Visualizations were generated using the ggplot2 package in R. Next, we assessed the association between jointly stratified PRS and TRS risk groups and AD diagnosis. Both PRS and TRS were dichotomized at the cohort‐specific median and combined to define four groups: PRS low–TRS low (reference), PRS low–TRS high, PRS high–TRS low, and PRS high–TRS high. Multivariable logistic regression models were used to estimate ORs and 95% CIs for AD diagnosis in each group compared with the reference group. All models were adjusted for age, sex, and APOE ε4 carrier status. Forest plots were generated using ggplot2 in R. Lastly, multiple machine learning approaches including elastic net, logistic regression, random forest, support vector machine, and XGBoost were applied to evaluate the classification performance of PRS, TRS, and their combination in distinguishing AD from CN participants. Models were trained using data from the ADNI cohort and externally validated in the SNUBH cohort. To address the 1:2 case–control ratio in ADNI, the CN group was downsampled via propensity score matching based on age and sex using the MatchIt package in R. Model performance was assessed using five‐fold cross‐validation in the training set and external validation in the test set. Evaluation metrics included area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, specificity, and F1‐score. AUCs were compared across models using the DeLong test, implemented in the MLstatkit Python package. All statistical analyses and visualizations were conducted using R version 4.1.2 and Python version 3.8. Statistical significance was defined as two‐sided p < 0.05. RESULTS Participant characteristics The study design is illustrated in Figure 1. Baseline characteristics of the discovery (ADNI, N = 313) and replication (SNUBH, N = 173) cohorts are presented in Table 1. AD participants were older than CN individuals in both cohorts. APOE ε4 carrier frequency was higher in AD groups across both datasets. Standardized PRS and TRS values were elevated among AD participants. When comparing the two cohorts, no statistically significant differences were observed in age or APOE ε4 carrier frequency, whereas the proportion of female participants was higher in SNUBH than in ADNI. Flow chart. AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; eQTL, expression quantitative trait loci; PRS, polygenic risk scores; SNP, single‐nucleotide polymorphism; SNUBH, Seoul National University Bundang Hospital; TRS, transcriptional risk scores. TABLE 1Participants characteristics in the discovery and replication cohortsADNI (Discovery, N = 313)SNUBH (Replication, N = 173) p‐ValueCharacteristics CN (N = 211) AD (N = 102) p‐Value CN (N = 88) AD (N = 85) p‐ValueAge, mean (SD), y76.40 (6.42)77.47 (7.80)0.23272.53 (5.39)79.42 (4.96)1.95 × 10−15 0.176Female, No. (%)105 (49.76)38 (37.25)0.05065 (73.86)54 (63.53)0.1931.62 × 10−6 APOE ε4 carrier, No. (%)57 (27.01)69 (67.65)1.50 × 10−11 14 (15.91)44 (51.76)1.34 × 10−6 0.172Standardized PRS, mean (SD)−0.13 (0.97)0.27 (1.02)0.001−0.24 (1.01)0.25 (0.93)0.0015.56 × 10−6 Standardized TRS, mean (SD)−0.11 (1.02)0.22 (0.92)0.005−0.27 (1.00)0.27 (0.93)3.10 × 10−4 1.05 × 10−5 Note: Standardized PRS and TRS values represent z‐scores, calculated within each cohort by subtracting the mean and dividing by the standard deviation of the respective score across all participants. p‐Values within each cohort compare AD and CN participants, while the p‐value in the rightmost column corresponds to between‐cohort comparisons. p‐Values were calculated using two‐sample t‐tests for continuous variables and chi‐squared tests for categorical variables.Abbreviations: AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; CN, cognitively normal; PRS, polygenic risk scores; SD, standard deviation; SNUBH, Seoul National University Bundang Hospital; TRS, transcriptional risk scores.John Wiley & Sons, Ltd. Participants characteristics in the discovery and replication cohorts CN (N = 211) AD (N = 102) CN (N = 88) AD (N = 85) Note: Standardized PRS and TRS values represent z‐scores, calculated within each cohort by subtracting the mean and dividing by the standard deviation of the respective score across all participants. p‐Values within each cohort compare AD and CN participants, while the p‐value in the rightmost column corresponds to between‐cohort comparisons. p‐Values were calculated using two‐sample t‐tests for continuous variables and chi‐squared tests for categorical variables. Abbreviations: AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; CN, cognitively normal; PRS, polygenic risk scores; SD, standard deviation; SNUBH, Seoul National University Bundang Hospital; TRS, transcriptional risk scores. Association of PRS and TRS with AD In the ADNI cohort, both PRS and TRS were independently associated with increased AD risk in a model adjusted for age and sex, as well as in a model additionally adjusted for APOE ε4 carrier status (Table S2). In the SNUBH cohort, the association with PRS remained significantly associated with AD risk, while the association for TRS was attenuated after further adjustment for APOE ε4 carrier status (Table S2). Stratified AD risk by PRS, TRS, and combined scores AD diagnosis rates increased progressively across PRS and TRS tertiles in both cohorts (Figure 2). In ADNI, rates rose from 20% to 45% for PRS, and from 25% to 41% for TRS. In SNUBH, rates increased from 29% to 63% for PRS, and from 33% to 67% for TRS. When stratified by combined PRS–TRS scores, individuals in the highest tertile for both PRS and TRS showed substantially higher AD diagnosis rates (ADNI, 56%; SNUBH, 80%) compared to those in the lowest tertiles (ADNI, 17%; SNUBH: 14%). Alzheimer's disease diagnosis rates by PRS, TRS, and the combined risk score. Diagnosis rates were assessed according to tertile groupings of PRS, TRS, and their combination (PRS+TRS). For PRS and TRS, individuals were classified into tertiles based on each respective score in (A) ADNI and (C) SNUBH. For the combined PRS+TRS group, individuals who fell into the bottom tertile (first tertile) or top tertile (third tertile) of both PRS and TRS distributions were assigned to each group, respectively in (B) ADNI and (D) SNUBH. All others were categorized as the intermediate group. AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; PRS, polygenic risk scores; SNUBH, Seoul National University Bundang Hospital; TRS, transcriptional risk scores. AD risk by combined risk profiles Compared to the reference group (PRS low–TRS low), the PRS high–TRS high group showed significantly increased AD risk, with adjusted ORs of 2.53 (95% CI: 1.21‐5.40) in ADNI and 3.39 (95% CI: 1.05‐11.65) in SNUBH (Figure 3). Other groups were not associated with significant risk elevations. Odds ratios for Alzheimer's disease across combined PRS and TRS risk groups in each cohort. Participants were stratified into four groups based on median splits of PRS and TRS: (1) PRS low–TRS low (reference), (2) PRS low–TRS high, (3) PRS high–TRS low, (4) PRS high–TRS high. Logistic regression models were used to estimate the ORs and 95% CIs for AD diagnosis in each group, compared to the reference group, adjusting for age, sex, and APOE ε4 carrier status. (A) ADNI, (B) SNUBH. ADNI, Alzheimer's Disease Neuroimaging Initiative; APOE, apolipoprotein E; CI, confidence interval; OR, odds ratio; PRS, polygenic risk scores; SNUBH, Seoul National University Bundang Hospital; TRS, transcriptional risk scores. Classification performance of PRS and TRS for AD The integration of PRS and TRS in elastic net models, which showed the best overall performance among the five machine learning approaches, improved AD classification performance, achieving an AUC of 0.705 (95% CI: 0.621‐0.776), which outperformed 0.635 (95% CI: 0.553‐0.716) for the PRS model and 0.653 (95% CI: 0.559‐0.730) for the TRS model (Figure 4A). The difference in AUC between the combined model and PRS model was statistically significant (p = 0.046). In addition to improved discrimination, the combined model showed consistently superior classification performance across all evaluated metrics, including accuracy (0.676), precision (0.679), recall (0.647), specificity (0.705), and F1 score (0.663), compared with either PRS or TRS models (Figure 4B). These findings were replicated across the other machine learning models, including logistic regression, random forest, support vector machine, and XGBoost, where adding TRS to PRS consistently improved overall performance (including AUC) and significantly enhanced classification compared with PRS alone (Figure S2 and Table S3). Classification performance of elastic net models for AD diagnosis using PRS and TRS features. (A) ROC curves for AD classification evaluated in the independent validation cohort (SNUBH) using elastic net models trained in the discovery cohort (ADNI) with PRS, TRS, and their combination (PRS+TRS). Shaded areas represent 95% CIs derived from bootstrap resampling. (B) Performance metrics, including accuracy, precision, recall, specificity, F1 score, and AUC, obtained from the independent validation cohort (SNUBH) for each feature set. p‐Values for AUC comparisons were calculated using the DeLong test, with the PRS only model as the reference. AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; AUC, area under the curve; CI, confidence interval; PRS, polygenic risk scores; ROC, receiver operating characteristic; SNUBH, Seoul National University Bundang Hospital; TRS, transcriptional risk scores. DISCUSSION In this study, we demonstrated that integrating PRS and TRS, both derived from blood‐based biomarkers, enhances the stratification and classification of AD beyond either score alone. Using two independent and ancestrally distinct cohorts, ADNI and SNUBH, we found that individuals with simultaneously elevated PRS and TRS had significantly increased odds of AD diagnosis. Moreover, the combined PRS–TRS model consistently outperformed single score models in classification tasks, underscoring the clinical relevance of a multi‐omics approach in blood‐based AD risk assessment. The complementary value of genomic and transcriptomic risk profiling underpins this improved performance. While PRS reflects inherited susceptibility based on the cumulative effects of common genetic variants, TRS captures gene expression alterations by both genetic and environmental influences. The lack of significant risk elevation in PRS high–TRS low or PRS low–TRS high further supports the hypothesis that both static genetic predisposition and dynamic transcriptomic dysregulation must co‐occur to confer substantial disease risk. This synergistic effect was observed in external validation cohorts, strengthening the generalizability of our findings. Unlike many studies that rely on internal cross‐validation, our use of an independent test cohort provides robust evidence for the applicability of the combined model across diverse populations. These results extend prior work in several key respects. Our earlier research demonstrated that blood‐based TRS was significantly associated with AD diagnosis and neuroimaging biomarkers, 13 and that a brain‐based TRS model provided additive predictive power beyond PRS, improving AD classification performance by up to 16%. 14 The current study confirms these findings using peripheral blood‐derived transcriptomic data, which are more clinically accessible and provide critical validation for the broader applicability of the PRS–TRS integration across tissues. This validation is crucial as it establishes the consistency of PRS–TRS integration across different tissue types and supports the generalizability of combined risk scoring approaches in distinct ethnic populations. Our findings are also in line with emerging literature on multi‐omics risk scoring. Recent studies have shown that module‐based PRS derived from brain transcriptomic profiles can reveal molecular heterogeneity in AD and enable more precise patient stratification. 26 Furthermore, several efforts have aimed to enhance PRS models by incorporating genetically imputed transcriptomic data. 27 , 28 , 29 , 30 Going a step further, our study integrates directly measured gene expression data, rather than imputed values, providing stronger empirical support that transcriptomic information contributes complementary value to traditional polygenic approaches. The strength of our integrative strategy also aligns with the growing recognition that comprehensive multi‐omics frameworks are essential for understanding complex neurodegenerative diseases. Beyond the genome and transcriptome, integrative analyses now include proteomic, metabolomic, and epigenomic layers, both in brain and peripheral compartments. 31 , 32 , 33 , 34 , 35 , 36 In AD specifically, brain‐based multi‐omics approaches have revealed mechanistic insights, 32 while plasma proteomics, now an especially active area of investigation, has identified promising diagnostic biomarkers with translational potential. 33 , 34 Additional efforts integrating brain and blood data 31 , 35 and CSF profiling 36 have further contributed to identify diagnostic biomarkers and dissect the molecular mechanisms of AD pathogenesis. Our findings support the continued expansion of blood‐based multi‐omics approaches, which offer a pragmatic path toward scalable, population‐level screening and personalized risk prediction, considering their accessibility and potential to reflect both central and systemic pathophysiology. Nonetheless, several limitations should be acknowledged. Because of differences in population background and RNA expression profiling platforms between cohorts, PRS and TRS were constructed using cohort‐specific SNP and gene sets, precluding the direct transfer of identical PRS and TRS models across datasets. Accordingly, cross‐cohort evaluation in this study was designed to assess the robustness of PRS–TRS integration as a risk stratification strategy across heterogeneous settings, rather than the direct portability of individual PRS or TRS features. In the main analyses, PRS and TRS were each constructed using a single representative method and eQTL reference, respectively. However, sensitivity analyses using an alternative state‐of‐the‐art PRS approach, SBayesRC, 37 and an alternative large‐scale blood eQTL resource from eQTLGen 38 yielded similar overall patterns (Figure S3 and S4), supporting the robustness of our findings. Notably, TRS performance was less consistent in ADNI when alternative eQTL references were applied, potentially reflecting differences in gene expression profiling platforms and gene coverage between cohorts and eQTL resources. In addition, the cross‐sectional design of this study limits causal inference, particularly with respect to the interpretation of blood‐based transcriptomic signals. The observed associations captured by the TRS may reflect pleiotropic genetic effects or systemic, reactive transcriptional responses rather than direct causal mechanisms in AD. Longitudinal studies with serial measurements of peripheral transcriptomic markers will be necessary to clarify temporal relationships and disentangle causal versus pleiotropic effects. Furthermore, although we adjusted for major confounders such as age, sex, and APOE ε4 status, residual confounding from unmeasured variables cannot be ruled out. AD status was defined based on clinical diagnosis rather than amyloid‐based biological criteria, which may introduce pathological heterogeneity and limit biological interpretability. While amyloid‐based definitions are increasingly emphasized in contemporary AD research, the lack of sufficient biomarker‐defined samples precluded robust analyses using amyloid PET‐based diagnostic labels in the present study. Future studies incorporating larger biomarker‐characterized cohorts will be essential to more precisely evaluate the biological specificity of PRS–TRS integration. Finally, as our study focused on clinically diagnosed AD cases from memory clinic‐based cohorts, the generalizability of our findings to preclinical or population‐based screening settings remains to be established. This study provides evidence that integrating blood‐based PRS and TRS improves AD risk prediction compared to either score alone. By combining inherited genetic predisposition with dynamic transcriptomic profiles, this multi‐omics approach offers a more comprehensive and clinically accessible assessment of disease susceptibility. Its consistent performance across two ancestrally distinct cohorts supports its generalizability. Future studies should expand this framework by incorporating additional omics layers such as proteomics and metabolomic and by validating its utility in preclinical and community settings. Ultimately, blood‐based multi‐omics models may facilitate personalized strategies for early detection and risk prediction, as well as approaches for population‐level risk stratification. AUTHOR CONTRIBUTIONS Drs J.‐M. Pyun and Y. H. Park had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Concept and design: J. Hwang, J.‐M. Pyun, Y. H. Park, K. Nho. Acquisition, analysis, or interpretation of data: Y. H. Park, A. J. Saykin, K. Nho. Drafting of the manuscript: J. Hwang, J.‐M. Pyun, Y. H. Park, K. Nho. Critical revision of the manuscript for important intellectual content: J.‐Y. Lee, P. J. Bice, A. J. Saykin, S. Kim, J. Sung, Y. H. Park, K. Nho. Statistical analysis: J. Hwang, J.‐M. Pyun, Y. H. Park. Obtained funding: J.‐M. Pyun, Y. H. Park, K. Nho, A. J. Saykin. Administrative, technical, or material support: J. S. Park. Supervision: J.‐Y. Lee, J. Sung, S. Kim, Y. H. Park, K. Nho. CONFLICT OF INTEREST STATEMENT Dr. A. J. Saykin reported receiving in‐kind support from Avid Radiopharmaceuticals, a subsidiary of Eli Lilly (PET tracer precursor) and Gates Ventures, LLC and Sanofi (Proteomics panel assays on IADRC and KBASE participants as part of the Global Neurodegeneration Proteomics Consortium) and gift funds (GV) supporting technical contributions to the GRIP platform, and he has participated in Scientific Advisory Boards (Bayer Oncology, Eisai, Novo Nordisk, and Siemens Medical Solutions USA, Inc) and an Observational Study Monitoring Board (MESA, NIH NHLBI), as well as External Advisory Committees for multiple NIA grants. He also serves as Editor‐in‐Chief of Brain Imaging and Behavior, a Springer‐Nature Journal. No other disclosures were reported. Author disclosures are available in the supporting information. ETHICS APPROVAL AND CONSENT TO PARTICIPATE This study was approved by the IRB of the SNUBH (IRB No. B‐2004‐604‐305), in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to enrollment. Supporting information Supporting Information Supporting Information