BackgroundProtein function prediction has evolved from sequence-based methods to deep learning-based multimodal approaches incorporating diverse data streams to better reflect biological complexity. Graph Neural Networks (GNNs) are increasingly used to integrate heterogeneous data due to their ability to handle different dimensionalities and scales. However, the contributions of various data modalities to prediction accuracy remain unclear.ResultsWe use a simplified deep learning framework with a single-layer attention-based GNN on the BeProf dataset to investigate the impact of different data modalities on protein function prediction. The results confirm GNNs as effective integrators of diverse data, including protein interaction networks, textual, sequence and structural information. Our findings highlight the central role of language modelembeddings and the benefits of integrating InterPro and Gene Ontology annotations. We show that structural information, while useful at the protein scale[1], is redundant with other data types in large-scale networks. Despite using a simple architecture, we reach first place on the BeProf benchmark’s Cellular Component subset, third and fifth respectively on the Biological Processes and Molecular Function subsets, outperforming popular methods such as DeepGoZero, DeepGraphGO or TALE.ConclusionThis study provides insights into the relative contributions of different data modalities, showing that sequence information remains foundational while interactome, structural, and ontology-based features offer avenues for improvement. We highlight the trade-offs and benefits of multimodal frameworks, providing guidelines for future innovations in computational biology and machinelearning for enhanced protein functional annotation.