Enhancing Protein Complex Identification via GO Function Annotation-Augmented Graph Embedding
Abstract
Protein complexes are necessary for cell function, and thus, in order to identify them from protein-protein interaction (PPI) networks, a problem in computational biology has emerged. Graph embedding methods can learn representations of the network structure in low-dimensional space. However, these methods only consider the structure and fail to incorporate the rich functional information provided by Gene Ontology (GO) annotations. This paper proposes a GO-augmented graph embedding framework that integrates topological node embeddings with GO functional annotation vectors as composite node features for protein complex identification. Experiments on the Saccharomyces cerevisiae PPI network from BioGRID, assessed using an extended, non-standard 150-complex evaluation set built around a CYC2008-based core (35 manually matched CYC2008 complexes plus 115 GO-coherence-filtered candidates) under both K-means and Markov Clustering (MCL), show that the benefit of GO annotations is conditional on the clustering algorithm: under MCL, adding GO annotations raises the F1 score from 0.4329 to 0.6004, the best result among all method-feature combinations tested, while under K-means, GO annotations do not improve and slightly reduce F1 from 0.5782 to 0.5011. A subset analysis confirms that the MCL improvement also holds on a GO-independent set of manually curated CYC2008 complexes, indicating that it is not merely an artifact of GO-based benchmark construction. Three classical complex-detection baselines (MCODE, standard MCL, and ClusterONE) are additionally compared using their official implementations; the best GO-augmented configuration exceeds all three on the full evaluation set. Therefore, the results indicate that functional annotations can complement network topology to improve protein complex identification, but the magnitude and direction of this benefit depend on how the downstream clustering algorithm uses the embedding space.