Leveraging protein-protein interactions in phenotype prediction through graph neural networks
Current genotype-to-phenotype models, such as poly-genic risk scores, only account for linear relationships between genotype and phenotype and ignore epistatic interactions, limiting the complexity of the diseases that can be properly characterized. Protein-protein interaction networks have the potential to improve the performance of the models. Moreover, interactions at the protein level can have profound implications in understanding the genetic etiology of diseases and, in turn, for drug development. In this article, we propose a novel approach for phenotype prediction based on graph neural networks (GNNs) that naturally incorporates existing protein interaction networks into the model. As a result, our approach can naturally discover relevant epistatic interactions. We assess the potential of this approach using simulations and comparing it to linear and other non-linear approaches. We also study the performance of the proposed GNN-based methods in predicting Alzheimers disease, one of the most complex neurodegenerative diseases, where our GNN approach outperform state of the art methods. In addition, we show that our proposal is able to discover critical interactions in the Alzheimers disease. Our findings highlight the potential of GNNs in predicting phenotypes and discovering the underlying mechanisms of complex diseases.