Leveraging Graph Machine Learning for Moonlighting Protein Prediction: A Local PPI Network and Physiochemical Feature Approach

Abstract

Moonlighting proteins are those proteins that perform more than one distinct function in the body. They are pivotal in various metabolic pathways and disease mechanisms. Identifying moonlighting proteins remains a challenge in Computational Biology. In this work, we propose the first graph neural network based models to identify moonlighting proteins. Our models work on large protein-protein interaction (PPI) networks with sparse labels of moonlighting and non-moonlighting proteins. In addition to PPI network, the models make use of features derived from the amino acid sequences of the proteins. We propose two frameworks: one as graph classification based on the local neighborhood of the query protein; and the other node classification based on the entire graph. These GNN-based methods outperform traditional machine learning methods that have previously been used for moonlighting prediction. The global full network-based model, operating on Homo sapiens data achieves accuracy of 88.4% and F1 score of 88.8%. The local neighborhood method is more lightweight and can be applied to larger protein sets with multiple species. CCS CONCEPTS* Applied computing [->] Computational proteomics.