An Artificial Intelligence-based framework for protein interaction design with accelerated KAN-based Positive-Unlabeled learning

Protein design seeks optimal amino acid sequences for target structures, but designing stable protein complexes remains challenging. We introduce a protein interaction design pipeline combining Monte-Carlo simulation with Metropolis-criteria (MCM) and Deep-Learning. It uses Protein-Protein-Interaction(PPI) scores from Deep-Learning-based MaTPIP model to form a PPI-score-based-MCM (PMCM). The workflow integrates PMCM-driven sequence generation, HDBSCAN-clustering-based selection, and validation via AlphaFold2 and Molecular-Dynamics (MD) simulation. Incorporating learned PPI scores enhances efficiency and feature fusion. A Positive-Unlabeled (PU) learning classifier accelerates sequence validation, while the Kolmogorov-Arnold Network (KAN) improves PU learning over Multi-Layer-Perceptron (MLP). AlphaFold2 predictions yield median Root-Mean-Square-Deviation (RMSD) 1.17 $\mathring{A}$, predicted-Template-Modelling (pTM) 0.72, and interface pTM 0.88; 73\% of complexes remain within 2 $\mathring{A}$ RMSD after 100 ns MD, confirming stability. Interface mutations reveal altered interactions. The KAN-based PU model improves F1-score, precision, and AUC by 5\%, 11\%, and 2\% over MLP. Overall, our method outperforms traditional and simulation-based methods while remaining competitive with modern Deep-Learning design frameworks.