Enhancing Protein-Protein Interaction Site Prediction through Hybrid Transformer-OGF-AGNN Architecture: A Study on Yeast Proteome

Predicting correct proteins interact (PPI) is key for understanding how cells function and for creating new drugs. Standard techniques are usually not able to model the many links between the sequence, structure, and network traits present in protein interactions. A Hybrid Transformer-OGF-AGNN architecture that merges three distinct approaches: (1) the Transformer network to detect long range patterns, (2) Object-Graph Fusion (OGF) to represent both structural and functional info, and (3) Attention-Guided Graph Neural Networks (AGNN) to make use of the interaction between proteins. The model was built and assessed using data from the Saccharomyces cerevisiae proteome, which includes 6,049 proteins, each with verified interaction sites. The test data demonstrated superior performance of the proposed architecture, achieving an F ₁ -score of 0.847 (precision: 0.823, recall: 0.872), compared to baseline methods: DeepPPISP (F ₁ : 0.782), SPRINGS (F ₁ : 0.756), and PPI-Pred (F ₁ : 0.729). Cross-validation results indicated broad applicability with minimal variance (σ ² = 0.003). Notably, 93.2% of predicted interaction sites (1,247 out of 1,323) were experimentally validated. The Hybrid Transformer-OGF-AGNN architecture makes PPI site prediction better by combining different biological data. The better coverage in yeast proteome suggests its capabilities may be applied more widely in fields like structural biology and finding targets for drugs.