Virus-host prediction tools in the era of machine learning

Abstract

[eng] Master’s Thesis. Graph neural networks (GNNs) have recently gained popularity in detecting prokaryote-phage interactions. This task is crucial as there is a demand for precise and computationally efficient models due to the exponential increase in the number of sequenced phages. Its importance is further highlighted by phage therapy becoming more popular in the West as a solution for antimicrobial resistance (AMR). However, the usage of Viral Protein Families (VPF) is lacking in the existing work for creating knowledge graphs to train GNN models. Hence, this work constructed a knowledge graph based on VPF connections and trained models based on various embedding architectures like Graph Convolutional Layer (GCN), Graph Attentional Layer (GAT), and Graph Sample and Aggregation Layer (GraphSage). The GraphSage model trained on an initial dataset of 206 prokaryotes and 1718 viruses shows a promising performance of 83% recall with a false positive rate of 11% on the species-level prediction compared to the selected StateOf-The-Art (SOTA) model’s 70% recall and 16% false positives. On the genus-level prediction, the model also outperformed the selected SOTA model and VPF-Class in these metrics. Finally, this work hypothesised the reason behind the importance of additional prokaryote nodes in training machine-learning models for higher precision