Accurate prediction of nucleic acid binding proteins using protein language model

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Nucleic acid binding proteins (NABPs) are essential for numerous biological processes.
• Existing machine learning methods for NABP prediction have limitations in dataset scope, feature selection, and predicting proteins with unknown functions.
• Single-stranded DNA binding proteins (SSBs) remain underexplored in functional prediction.

Purpose of the Study:

• To enhance the accuracy and scope of predicting various NABP types for proteins with unknown functions.
• To develop novel predictive models leveraging advanced features from protein language models.
• To specifically improve the identification of single-stranded DNA binding proteins (SSBs).

Main Methods:

• Developed hierarchical and multi-class machine learning models.
• Utilized features extracted from the protein language model ESM2.
• Trained and tested models on comprehensive datasets for NABP classification.

Main Results:

• Achieved high prediction accuracy, up to 95% at each stage of the hierarchical model.
• Attained an overall prediction accuracy of 85% using the multi-class approach.
• Successfully improved the prediction of various NABP types, with a notable enhancement in identifying SSBs.

Conclusions:

• The integration of ESM2 features with machine learning significantly boosts NABP prediction accuracy.
• The developed models offer a robust tool for functional annotation of proteins, especially for underexplored categories like SSBs.
• The study provides valuable resources (datasets and code) for further research in NABP prediction.