Scholarlink

Background <p>Alzheimer’s disease and related dementias (ADRD) are complex, polygenic conditions with substantial public health impact. Accurate genetic risk prediction may enable earlier identification and stratification of individuals at elevated risk.</p> Objective <p>To evaluate the predictive performance of polygenic risk scores (PRS) for ADRD using a Bayesian variational autoencoders approach and to assess the modifying effects of age and <i>APOE</i> genotype on model performance.</p> Methods <p>We analyzed data from 276,566 unrelated individuals of European ancestry in the UK Biobank, with a median follow-up of 9.19 years. PRS and polygenic hazard scores (PHS) were constructed using genome-wide association study summary statistics, with PHS incorporating age-at-onset information. Three PRS methods were compared: DDML (Bayesian variational autoencoders), SBayesR (Bayesian multiple regression), and clumping and thresholding (C + T). Models were stratified by age and <i>APOE</i> genotype. Predictive performance was evaluated using time-dependent AUC, C-index, and hazard ratios (HRs), using a prespecified 2:1 training/testing split with identical ADRD case proportion across splits. All primary results are based on covariate-adjusted models, incorporating PRS together with age, sex, and 10 genetic principal components, and <i>APOE</i> genotype where indicated. Classification performance was compared between individuals in the top and bottom PRS quartiles to assess stratified risk.</p> Results <p>Among the participants (mean age 56.8 ± 8.0 years; 46.7% male), 1,328 (0.48%) developed ADRD. In covariate-adjusted models, DDML_PRS achieved the highest predictive accuracy (AUC = 0.847) in individuals aged 65–70 years. PHS models showed peak performance at 7 years of follow-up. DDML_PRS significantly improved classification in <i>APOE-ε4</i> carriers aged ≥ 65 years and outperformed other models across ADRD subtypes. Significant interactions were observed between PRS performance, age, and <i>APOE</i> genotype.</p> Conclusion <p>The DDML_PRS framework showed consistently higher discrimination than standard PRS baselines in this UK Biobank setting, particularly in older adults and <i>APOE-ε4</i> carriers, supporting its potential for individualized ADRD risk stratification. However, the observed classification accuracy remains modest, limiting immediate clinical utility and underscoring the need for external replication and multi-modal validation (e.g., biomarkers and clinical adjudication) to translate these predictive gains into practical early detection strategies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improved polygenic risk prediction for alzheimer’s disease and related dementias using deep learning: age and APOE-stratified analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月12日

查看全文
  • Shayan Mostafaei,
  • Daniel Wikström Shemer,
  • Jonathan K. L. Mak,
  • Ida K. Karlsson,
  • Sara Hägg

摘要

Background

Alzheimer’s disease and related dementias (ADRD) are complex, polygenic conditions with substantial public health impact. Accurate genetic risk prediction may enable earlier identification and stratification of individuals at elevated risk.

Objective

To evaluate the predictive performance of polygenic risk scores (PRS) for ADRD using a Bayesian variational autoencoders approach and to assess the modifying effects of age and APOE genotype on model performance.

Methods

We analyzed data from 276,566 unrelated individuals of European ancestry in the UK Biobank, with a median follow-up of 9.19 years. PRS and polygenic hazard scores (PHS) were constructed using genome-wide association study summary statistics, with PHS incorporating age-at-onset information. Three PRS methods were compared: DDML (Bayesian variational autoencoders), SBayesR (Bayesian multiple regression), and clumping and thresholding (C + T). Models were stratified by age and APOE genotype. Predictive performance was evaluated using time-dependent AUC, C-index, and hazard ratios (HRs), using a prespecified 2:1 training/testing split with identical ADRD case proportion across splits. All primary results are based on covariate-adjusted models, incorporating PRS together with age, sex, and 10 genetic principal components, and APOE genotype where indicated. Classification performance was compared between individuals in the top and bottom PRS quartiles to assess stratified risk.

Results

Among the participants (mean age 56.8 ± 8.0 years; 46.7% male), 1,328 (0.48%) developed ADRD. In covariate-adjusted models, DDML_PRS achieved the highest predictive accuracy (AUC = 0.847) in individuals aged 65–70 years. PHS models showed peak performance at 7 years of follow-up. DDML_PRS significantly improved classification in APOE-ε4 carriers aged ≥ 65 years and outperformed other models across ADRD subtypes. Significant interactions were observed between PRS performance, age, and APOE genotype.

Conclusion

The DDML_PRS framework showed consistently higher discrimination than standard PRS baselines in this UK Biobank setting, particularly in older adults and APOE-ε4 carriers, supporting its potential for individualized ADRD risk stratification. However, the observed classification accuracy remains modest, limiting immediate clinical utility and underscoring the need for external replication and multi-modal validation (e.g., biomarkers and clinical adjudication) to translate these predictive gains into practical early detection strategies.