Alzheimer's Imaging Consortium.

Background: For large genomic studies of middle-aged individuals, the prevalence of Alzheimer's disease (AD) is extremely low, making it difficult to conduct genomic analysis of the condition. To enable genome-wide association studies of AD in such datasets, an approach called Genome-wide association by proxy (GWAX) uses family history of disease as a proxy for disease status. Borrowing from the machine learning (ML) literature, we treat the development of proxy phenotypes as a pseudo-labelling task, where an ideal proxy label accurately predicts the lifetime risk of AD. Given this, this work develops ML-derived proxy phenotypes of AD for downstream GWAX based on survival models of conversion to AD and demonstrates their predictive performance and impact on downstream GWAX.

Method: Using cognitively normal (CN) and mild cognitive impaired (MCI) individuals with at least 2 timepoints in AIBL (n = 426), ADNI (n = 283) and UK Biobank (n = 56,105), we derived two proxy phenotypes of AD: i) a ML-derived boosting model with age, sex, years of education and AD family history, i.e. number of recorded parents with a history of Alzheimer's disease and ii) family history alone. The prognostic ability for future AD conversion was measured using concordance index (C-index) in 10-fold cross-validation. The two models were applied to the subset of UK Biobank with white British ancestry (n = 408,165) and a GWAX was conducted on the predicted labels. From this GWAX, PRS was derived and evaluated using area under the ROC curve (AUC).

Results: In 10-fold cross-validation, the ML-derived phenotypes were more predictive of conversion to AD than family history alone (median C-index: 0.7 vs 0.5, Mann Whitney p-value <0.001). GWAX of UKB from these proxy phenotypes indicates that the multivariate model detects a higher number of significant regions compared to family history alone. PRS constructed from these ML-derived phenotypes GWAX show greater power to separate true AD case/control labels when externally validated on ADNI.

Conclusion: Preliminary analysis of ML-derived proxy phenotypes of AD indicates may be a promising approach to improve AD genomic studies in middle age cohorts. Further analysis of the results and potential biases of this approach are needed to fully characterise this method.