iProtDNA-SMOTE: Enhancing protein-DNA binding sites prediction through imbalanced graph neural networks.

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• Protein-DNA interactions are fundamental to cellular processes, including gene regulation and DNA repair.
• Accurate prediction of DNA binding residues is crucial for understanding protein function and disease mechanisms.
• Existing methods often struggle with class imbalance in predicting protein-DNA binding sites.

Purpose of the Study:

• To develop and validate a novel computational method, iProtDNA-SMOTE, for predicting DNA binding residues in proteins.
• To address the challenge of class imbalance in protein-DNA binding site prediction datasets.
• To enhance the accuracy, generalization, and specificity of DNA binding site predictions.

Main Methods:

• Utilized non-equilibrium graph neural networks (GNNs) integrated with pre-trained protein language models.
• Implemented the SMOTE (Synthetic Minority Over-sampling Technique) algorithm to handle unbalanced graph data.
• Trained and evaluated the model on established datasets (TR646, TR573) and independent test sets (TE46, TE129, TE181).

Main Results:

• Achieved high Area Under the Curve (AUC) values: 0.850 (TE46), 0.896 (TE129), and 0.858 (TE181).
• Demonstrated superior performance compared to existing methods in predicting DNA binding sites.
• Showcased enhanced model generalization and specificity due to effective handling of imbalanced data.

Conclusions:

• iProtDNA-SMOTE provides a reliable and accurate method for predicting protein-DNA binding sites.
• The approach effectively overcomes class imbalance issues, leading to improved predictive performance.
• The publicly available code and datasets facilitate further research in protein-DNA interaction prediction.