DTC: A deep learning framework for ternary contact prediction in C3-symmetric homotrimers.

Protein complexes play crucial roles in cellular functions. The work focuses on the prediction of interchain ternary residue contacts in protein complexes. We first constructed a homotrimeric dataset from the PDB and developed DTC (Deep Ternary Contact predictor), a hybrid deep learning framework for ternary contact prediction. DTC does not require experimentally resolved structures as input, and integrates PLM embeddings with MSA-dependent representations extracted from the Evoformer module of AlphaFold2. DTC breaks through the challenge faced by existing residue-residue contact prediction methods, namely the inability to identify ternary contacts. On an independent test set, DTC achieves a Top-1 accuracy of 0.6. Furthermore, we propose the Top-k coverage metric, which provides a more reasonable standard for evaluating model performance under extreme class imbalance. Application results demonstrate that the sites predicted by DTC can be used as constraints to significantly improve the success rate of protein docking and show good generalization capability in viral trimers. The study provides a new computational tool and insights for understanding the cooperative assembly mechanisms of multimeric proteins.