Abstract
<title>Abstract</title> <p>Background: Protein structure comparison underlies functional annotation, protein engineering, therapeutic design, and drug discovery. Most existing approaches emphasize geometric superposition, structural alphabets, or learned latent representations. Although effective, these methods are typically optimized for global fold similarity rather than direct interpretation of residue-level biochemical interactions. In many biological settings, functional relatedness depends on conserved local interaction environments, including hydrogen bonding, ionic contacts, hydrophobic packing, aromatic stacking, van der Waals contacts, and disulfide bonding. We introduce Interaction Token Structural Alignment (ITSA), an interpretable framework that converts protein structures into residue-ordered sequences of biochemical interaction tokens and compares them using Smith–Waterman local alignment. Each residue is assigned a compact token representing its dominant local interaction environment, while secondary-structure context contributes to alignment scoring through a chemistry-aware substitution scheme. Results: ITSA was evaluated on two SCOP family-level benchmarks. On a curated benchmark comprising 4,950 pairwise comparisons across 10 structurally diverse protein families, the best blended ITSA score achieved an AUC of 0.8148 and an AUPRC of 0.4529. On a larger and more heterogeneous benchmark containing 44,253 pairwise comparisons across 30 families, ITSA achieved an AUC of 0.7271 and an AUPRC of 0.1549, corresponding to approximately 5.15 times the random AUPRC baseline. Family-level entropy analysis showed that token diversity alone does not explain performance variation across families, suggesting that performance depends less on raw interaction diversity than on the organization of conserved and alignable interaction motifs. Conclusions: ITSA provides an interpretable interaction-centered framework for protein structure comparison that captures discriminative family-level similarity while preserving residue-level biochemical context. The method complements geometry-centered structural comparison approaches and may be particularly useful in settings where local interaction environments and mechanistic interpretability are important.</p>