MBPBERT: A Large Language Model for Metal-Binding Peptide Discovery

Metal-binding peptides (MBPs) are a class of peptides capable of selectively coordinating metal ions and play critical roles in various biological processes, such as metal ion transport, storage, catalysis, and signal transduction. However, traditional experimental approaches, such as mass spectrometry, chromatography, and nuclear magnetic resonance spectroscopy, are labor-intensive, time-consuming, and poorly suited for the large-scale screening of MBPs. To overcome these limitations, we introduce MBPBERT, a deep learning-based predictive framework for accurate identification of MBPs and discrimination of metal-specific binding subtypes. MBPBERT builds upon the ProteinBERT architecture and was developed using two curated datasets: a peptide pretraining corpus comprising 33,095 sequences and a labeled dataset containing 909 MBPs and non-MBPs. For rigorous external validation, an independent test dataset of 101 peptides was constructed. MBPBERT attained areas under the receiver operating characteristic curve (AUROCs) of 0.9414 and 0.8882 on the test dataset for the prediction of MBPs and the classification of their specific binding subtypes, respectively. These results underscore the framework's strong predictive capability and generalization performance. Thus, MBPBERT provides a scalable and efficient in silico solution for high-throughput discovery of novel MBPs and screening of peptides with metal-specific binding preferences, potentially reducing the reliance on resource-intensive experimental validation.