GEGE: predicting gene essentiality with graph embeddings

Official URL: http://dx.doi.org/10.29130/dubited.1028387

A gene is considered essential if its function is indispensable for the viability or reproductive success of a cell or an organism. Distinguishing essential genes from non-essential ones is a fundamental question in genetics, and it is key to understanding the minimal set of functional requirements of an organism. Knowledge of the set of essential genes is also crucial in drug discovery. Several reports in the literature show that the gene location in a protein-protein interaction network is correlated with the target gene’s essentiality. Here, we ask whether the node embeddings of a protein-protein interaction (PPI) network can help predict gene essentiality. Our results on predicting human gene essentiality show that node embeddings alone can achieve up to 88% AUC score, which is better than using topological features to characterize gene properties and other previous work’s results. We also show that, when combined with homology information across species, this performance reaches 89% AUC. Our work shows that node embeddings of a protein in the PPI network capture the network connectivity patterns of the proteins and improve the gene essentiality predictions.