A Deep Learning Framework for Protein-to-Metal Binding Prediction Using Protein Language Models

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Manual curation of metal binding sites is labor-intensive and time-consuming.
• Accurate prediction of protein-metal ion interactions is crucial for understanding protein function and mechanisms.
• Existing computational methods often fail to capture long-range residue dependencies and positional information.

Purpose of the Study:

• To develop an end-to-end deep learning framework for predicting protein-metal ion binding sites.
• To evaluate the performance of state-of-the-art protein language models (pLMs) for this task.
• To assess the impact of positional encoding and compare with classical machine learning techniques.

Main Methods:

• Utilized a large language model (LLM) for metal-ion binding prediction.
• Compared five different protein language models (pLMs).
• Incorporated positional encoding for binding sites and evaluated against classical machine learning approaches.

Main Results:

• Achieved a Matthews Correlation Coefficient (MCC) of 0.89 using 10-fold cross-validation.
• Demonstrated precision, recall, and F1 scores exceeding 95% for six common metal ions.
• The proposed deep learning framework effectively captures residue dependencies and positional information.

Conclusions:

• The developed deep learning framework provides a highly accurate and efficient method for predicting protein-metal ion binding sites.
• The study highlights the importance of positional encoding and advanced language models in improving prediction accuracy.
• This computational pipeline offers a valuable tool for annotating uncharacterized proteins and advancing research in metalloprotein function.