F011 A benchmark of machine learning algorithms for Huntington’s disease age of onset prediction

<h3>Background</h3> Although most Huntington’s disease (HD) patients share the same genetic origin, there is a large variability in the onset of symptoms and their severity. Only about 50% of this variability can be accounted for by the length of CAG repeats. Importantly, 40% of the remaining variability is heritable, indicating that there is a great deal of unexplained variability due to unknown genetic factors. Recent genome-wide association studies (GWAS) have led to the discovery of a few genetic modifiers (GeMs). Nevertheless, state-of-the-art methods only consider linear marginal effects, limiting the search space to linear associations and ignoring interactions. <h3>Methods</h3> We present a comprehensive benchmark of different machine learning algorithms trained on a large Enroll-HD GWAS dataset to predict the age of onset of motor symptoms in HD. We consider linear models (lasso, elastic net) and non-linear models (random forest, XGBoost). Importantly, the latter implicitly account for interactions between SNPs. For each model considered we can systematically determine the most informative SNPs, providing alternative ways of discovering GeMs. <h3>Results</h3> We find that non-linear models perform the best in the prediction task. In addition, we rediscover well-known GeMs and propose novel GeM candidates. Lastly, we explore the potential of graph neural networks in identifying important gene subnetworks and epistatic interactions that drive the predictive model. <h3>Conclusions</h3> Our results provide new candidate modifier genes and show the potential of graph neural networks in predicting AO and discovering important gene subnetworks.