Summary
Structure is not used by hybrid sequence-structure models when many homologs are available (1).
Details
Quote from (1):
“the contribution of the structural information processed by the GNN-architectures to the performance of the model was more significant for simple node features and became less significant as more complex language models were used”
1.
Gamouh H, Novotný M, Hoksza D. Hybrid protein–ligand binding residue prediction with protein language models: does the structure matter? Bioinformatics. 2025;41(8). Available from: https://doi.org/10.1093/bioinformatics/btaf431