A Bayesian Optimization-Based Hybrid Deep Prediction Method for Zinc-Binding Protein Interaction Sites

The binding of zinc ions to proteins plays a crucial role in normal physiological functions and life activities of organisms. To enhance the prediction accuracy of zinc-binding protein interaction sites, the paper proposes a novel hybrid deep prediction method, ZnSite_HDPM_Bayes, which is based on Bayesian weighted optimization. The method utilizes a protein sequence to construct the LSTM deep prediction submodel integrated with a self-attention mechanism and machine learning component learners. Then, the Bayesian weighted optimization algorithm is used to dynamically set weights and adjust parameters to achieve the optimal combined model. The results of the experiment have demonstrated that the performance metrics of the proposed method outperform its component model by nearly 4.42% on average; compared to existing state-of-the-art methods, ZnSite_HDPM_Bayes has accomplished an increase of 4-20% in MCC, F1-score, and AUPRC, thereby showing a better predictive power. The findings help to identify the high-throughput zinc-binding protein interaction sites and the study of metalloprotein functions, to better promote the research and development of new drugs and biotechnology.