A Unified Protein Embedding Model with Local and Global Structural Sensitivity

Abstract

Structural comparison between proteins is key to many research tasks, including evolutionary analysis, peptidomimetics, and functional annotation. Traditional structure alignment tools based on three-dimensional protein structures, such as TM-Align, DALI, or ProBiS, are accurate, but they are computationally expensive and impractical at scale. Existing protein language models (PLMs), such as TM-Vec, improve computational efficiency but only capture global structural similarity, overlooking important motif-level structural details. In this paper, we propose a novel PLM consisting of a Siamese neural network, enabling efficient embedding-based structural comparison while also capturing both global and local structural similarity. Our model was trained on a dual loss function combining TM-score, a global similarity metric, and a variation of lDDT scores, a per-residue similarity metric. We tested against two datasets: a varied TM-score dataset from TM-Vec, and a high TM-score mutant dataset from VIPUR. Against these sets, our model achieved a TM-score MAE of 0.0741 and 0.0583, respectively, and a lDDT-score MAE of 0.0788 and 0.0038, respectively. Our model fulfills two key roles: first, it rapidly detects global structural differences. Second, it supports fine-grained structural assessments, improving sensitivity to subtle but functionally important structural changes.