Improved datasets and evaluation methods for the automatic prediction of DNA-binding proteins

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate protein function annotation is crucial for understanding biological processes.
• Identifying DNA-binding proteins from sequence is a key challenge in bioinformatics.
• Existing datasets for DNA-binding protein prediction have significant limitations.

Purpose of the Study:

• To address flaws in existing datasets for DNA-binding protein prediction.
• To introduce new, improved datasets and benchmark tasks for model evaluation.
• To develop and assess novel models for predicting DNA-binding proteins.

Main Methods:

• Development of new datasets addressing weaknesses in previous literature.
• Creation of new benchmark tasks for realistic performance assessment.
• Implementation and comparison of a gradient boosting model and a nearest neighbor model against existing methods.

Main Results:

• The proposed gradient boosting model, utilizing protein language model features, outperforms previous models.
• A baseline nearest neighbor model also shows strong performance, highlighting the importance of sequence identity.
• Models demonstrate sensitivity to DNA-binding regions and maintain accuracy across taxa within kingdoms.

Conclusions:

• Improved datasets and benchmarks are essential for advancing DNA-binding protein prediction.
• Data-driven models effectively learn to identify DNA-binding regions.
• The developed models show promise for cross-kingdom taxonomic predictions, albeit with varying accuracy.