Comparative Analysis on Alignment-Based and Pretrained Feature Representations for the Identification of DNA-Binding...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• DNA-protein interactions are fundamental to biological processes.
• Traditional in silico DNA-binding protein (DBP) identification relies on computationally intensive position-specific scoring matrices (PSSMs).
• Pretrained language models, like evolutionary scale modeling (ESM), offer a novel approach to biological sequence analysis.

Purpose of the Study:

• To compare the effectiveness of alignment-based PSSM features versus pretrained ESM representations for DBP classification.
• To evaluate the impact of feature extraction and selection methods on DBP identification performance.
• To assess the potential of ESM for DBP identification and other protein function annotation tasks.

Main Methods:

• Extracted features from both PSSM and ESM representations using four standardized averaging operations.
• Applied various feature selection (FS) techniques to optimize feature sets.
• Compared classification performance using a unified evaluation framework.

Main Results:

• Pretrained ESM representations demonstrated superior performance compared to PSSM-derived features in DBP classification.
• Feature selection methods were crucial for enhancing classification accuracy.
• An ensemble scheme combining multiple feature selection models significantly improved DBP classification.

Conclusions:

• Pretrained ESM representations provide a more efficient and effective alternative to PSSM for in silico DBP identification.
• ESM holds significant promise for broader applications in protein function annotation.
• Ensemble methods can further boost the performance of DBP identification models.