TINA: a geometry-aware post-hoc scoring framework for protein-protein interaction

Accurate identification of the most native-like conformation remains a critical bottleneck in protein complex structure prediction, particularly when multiple globally plausible candidate models are generated by deep learning methods such as AlphaFold3 (AF3). While AF3 provides internal confidence metrics such as ipTM and pLDDT, these signals often lack sufficient sensitivity to distinguish subtle but structurally meaningful differences in local interface geometry. Here, we present TINA, a geometry-aware post-hoc scoring framework designed for fine-grained ranking and selection among high-quality protein complex candidates. TINA constructs atom-level interface graphs and explicitly models two-body, three-body, and four-body geometric interactions through a many-body tensor modeling strategy, enabling a more comprehensive characterization of local contact patterns, angular arrangements, and torsional constraints. The model is initially trained using a joint ranking and regression objective on a large decoy set generated by ZDOCK, with additional task-specific fine-tuning for protein-peptide complexes, and is evaluated exclusively on AF3-generated candidate pools across four independent benchmarks covering general protein-protein, antibody-antigen, protein-peptide, and antibody-peptide complexes. TINA consistently achieves superior or competitive performance in terms of top-1 success rate, enrichment factor, and top-1 DockQ quality. Notably, its benefit is most pronounced in low-confidence and antibody-related scenarios, where AF3 internal ranking signals are less effective in discriminating among candidate conformations. Representative case studies demonstrate that TINA successfully identifies near-native conformations that are present in the AF3 decoy pool but overlooked by ipTM. Ablation experiments further confirm the critical contributions of atom-level graphs, many-body geometric modules, and the joint loss design. Overall, TINA demonstrates that independent, interface-focused post-hoc scoring remains a valuable complement to AF3, particularly for challenging complexes where local geometric discrimination is critical for accurate model selection.