Phenotype Prediction Using Graph Neural Networks

Phenotype Prediction using Graph Neural Networks This project aims to study the potential of Graph Neural Network (GNN) models for phenotype prediction from genetic data. In particular, we seek to computationally diagnose Alzheimer’s disease, one of the most complex neurodegenerative diseases. To this purpose, we explore the potential of GNNs approaches and consider different architectures, including Convolutional Graph Neural Networks (ConvGNN) and Graph Attention Networks (GAT) and show that they provide a robust tool for phenotype prediction from genetic data. Traditional phenotype prediction methods are based on logistic regression models and, in turn, are linear and do not adequately incorporate important biological information. In contrast, GNNs facilitate the usage of prior biological knowledge in the form of graphs, such as protein- protein interactions (PPI). As a result, this approach can naturally account for higher-order effects and interactions than the state of the art. We hypothesized and proved in this thesis that this fact leads to improved performance in phenotype prediction. The proposed GNN models are benchmarked against state-of-the-art Machine Learning (ML) models for phenotype prediction, including logistic regression (LR) models with L1 and L2 reg- ularization, as well as a fully connected artificial neural network (ANN). To study the behavior of each approach under different circumstances, we perform two sets of simulations covering various amounts of training data and complexity. In addition, we compare all methods using real Alzheimer’s disease data. Remarkably, all GNN models outperformed state-of-the-art methods in phenotype prediction. In conclusion, this study proves the reliability of GNNs for Alzheimer phenotype prediction and establishes a promising foundation for future research and advancements in the field of GNN-based phenotype prediction. Moreover, these findings highlight the potential of GNNs in predicting phenotypes and discovering the underlying mechanisms in complex diseases.