LASIGE researchers Laura Balbi, Rita Sousa, Pedro Cotovio, and Cátia Pesquita published a new study, titled “The role of protein embeddings for protein-protein interaction prediction with graph neural networks”, in Briefings in Bioinformatics (Top 10% journal according to Scimago, Scopus and Web of Science rankings).

The paper investigates how different protein representations can improve protein-protein interaction (PPI) prediction with graph neural networks (GNNs). The researchers systematically compare approaches that rely on PPI network representations with those that incorporate protein-level features derived from sequence, structure and function, as well as approaches that integrate both types of representations. The results show that combining biologically meaningful protein information with network-level features generally improves predictive performance, with functional information proving particularly informative.

Overall, the study provides a taxonomic map of the field, a reproducible framework for evaluating representative methods across paradigms, and practical guidance for future PPI prediction efforts. It is available here.