Machine Learning Methods for the Genetics of Alzheimer’s Disease

Our goal is to use machine learning (ML) to understand and predict genetic cause and risk to Alzheimer's disease (AD).

Our study is purely data informatic in nature and follows common ML-practices, including data cleaning and quality control of both samples and genotypes, pre-processing (such as correcting for covariates like age, sex and general population structure), training and parameter optimization, and evaluation using hold-out sets. Our primary goal is to increase the predictive power for disease status from only the genetic information. We use several different ML-algorithms to build predictors, such as compressed sensing (also known as LASSO), neural networks, and horseshoe Bayesian regression. The underlying genetic architectures are studied through dissecting and analyzing the trained predictors — which loci are important? how important are they (in particular the polygenetic importance beyond APOE)? what are the genetic correlations for these loci within and across different ancestries or other population groups? etc. — informing both fundamental disease research and future predictor algorithm designs.

A priori, we will at least use the phenotypic characteristics sex, age/age of onset, race, ethnicity, AD status, and family history. We will also investigate whether a more informative case variable can be constructed as a function of the mentioned variables in conjunction with AD status comments. Other phenotypic characteristics may also be used in the continuous improvement of our predictor algorithms.

All analysis will be performed on high-performance computing clusters at Michigan State University (MSU), where they will be stored under strict security, accessible only to PI and three other MSU staff (who also sign the Data Transfer Agreement), in accord with regulations.

We will publish all scientific results in peer-reviewed journals and make developed general algorithms public. Published predictors will be made available, both to the public and returned to the NIAGADS.

Alzheimer’s disease (AD) is affecting more than 6 million Americans with an enormous impact on their lives and surrounding families. We aim to improve existing and develop new methods to predict the AD risk of an individual from his/her genetic information. Such a predictor can inform both life and treatment decisions, and its inner workings shed light on the genetic causes of AD. As such, it furthers the basic disease understanding and both the development and employment of preventive medications.

Our methods use modern machine learning techniques (e.g. LASSO and neural networks) which are trained on carefully processed trait and genetic data. The amount of data is a crucial factor for success and the more than 30,000 participating individuals in the NIAGADS database constitute a state-of-the-art resource for this type of research.

All data is handled under strict security policies while all scientific results will be made publicly available through publications and downloadable files.