Alignment-free metal ion-binding site prediction from protein sequence through pretrained language model and...

Area of Science:

• Biochemistry
• Computational Biology
• Bioinformatics

Background:

• Over one-third of proteins in the Protein Data Bank contain metal ions.
• Accurate identification of metal ion-binding residues is crucial for understanding protein function and drug design.
• Predicting metal ion-binding sites from protein sequence alone is challenging due to the small size and versatility of metal ions.

Purpose of the Study:

• To develop LMetalSite, an alignment-free, sequence-based predictor for metal ion-binding sites.
• To accurately identify binding sites for zinc (Zn2+), calcium (Ca2+), magnesium (Mg2+), and manganese (Mn2+).
• To overcome the limitations of existing sequence-based and structure-based prediction methods.

Main Methods:

• Leveraging a pretrained language model for informative sequence representations.
• Employing transformer architecture to capture long-range dependencies in protein sequences.
• Utilizing multi-task learning to address data scarcity and exploit similarities between metal ions.

Main Results:

• LMetalSite significantly outperformed state-of-the-art structure-based methods in predicting metal ion-binding sites.
• Achieved superior performance across four independent tests for Zn2+, Ca2+, Mg2+, and Mn2+.
• Demonstrated the effectiveness of self-attention modules in learning residue structural contexts from sequence data.

Conclusions:

• LMetalSite offers a highly accurate and efficient method for predicting metal ion-binding sites from protein sequences.
• The approach provides valuable insights for protein function studies and novel drug development.
• The developed tool, datasets, and models are publicly available for research use.