RGPA-GCN: Graph convolutional networks for rice gene-phenotype association prediction

Rice plays a pivotal role as a vital food source for human consumption. Identifying gene-phenotype associations (GPAs) can significantly enhance the tolerance of rice to environmental stress and its overall yield. Nevertheless, the experimental process of discovering GPAs is not only consume a lot of resources but also time-consuming. The computational screening for GPAs has emerged as an essential tool to complement and expedite biological experiments. In this study, we tackle the prediction of GPAs by framing it as a node classification task, and introduce RGPA-GCN, an innovative computational approach leveraging graph convolutional networks. RGPA-GCN constructs a topology graph through the application of the k-nearest neighbor method for effective information aggregation. The nodes within this graph encapsulate both gene functional similarity and phenotype semantic similarity, enhancing the accuracy of our predictions. Notably, the RGPA-GCN approach demonstrates its ability to predict both unknown GPAs and previously unseen genes or phenotypes. Leveraging 5-fold cross-validation, RGPA-GCN exhibits commendable performance, outperforming six classical machine learning methods, and three state-of-the-art models. Additionally, the ablation studies on the sampler and the case studies involving five different phenotypes yields promising results, underscoring the effectiveness of this approach.