FusionProt: Fusing Sequence and Structural Information for Unified Protein Representation Learning

Accurate protein representation is vital for diverse biological and biomedical applications. While three-dimensional (3D) structural context is central to protein function, most computational approaches either ignore it or fuse it with sequence information in a single late step, yielding limited benefits. We present FusionProt, a unified representation learning framework that iteratively exchanges information between a protein language model and a graph-based structure encoder via a single learnable fusion token. This early, bidirectional conditioning preserves structural cues across layers while maintaining near-constant complexity. Across EC and GO benchmarks, FusionProt achieves state-of-the-art results, improving F _max by up to 3% over strong joint baselines; on mutation stability prediction it boosts AUROC by 5.1% versus the best structure model, with only 2–5% runtime overhead. We further demonstrate how analysing the specific gains in predictive capability can help focus attention and generate hypotheses about the underlying biological mechanisms.