Explainable Deep Multilevel Attention Learning for Predicting Protein Carbonylation Sites.

Area of Science:

• Biochemistry
• Computational Biology
• Proteomics

Background:

• Protein carbonylation, a marker of oxidative stress, alters protein function and cellular processes.
• Accurate identification of carbonylation sites is vital for disease mechanism research.
• Existing computational methods often misidentify ligand interaction sites as carbonylation sites.

Purpose of the Study:

• To develop a novel deep learning framework for accurate prediction of protein carbonylation sites.
• To address the challenge of cross-prediction between carbonylation and ligand interaction sites.
• To enhance the specificity and performance of computational prediction tools.

Main Methods:

• Introduction of Selective Carbonylation Sites (SCANS), a deep learning framework.
• Utilization of a multilevel attention strategy for local and global feature extraction.
• Application of a tailored loss function and transfer learning to improve prediction accuracy and specificity.

Main Results:

• SCANS demonstrates superior predictive performance compared to existing methods.
• The framework achieves consistently low false positive rates, including reduced cross-predictions.
• Motif analyses provide novel insights into the characteristics of protein carbonylation sites.

Conclusions:

• SCANS offers a significant advancement in predicting protein carbonylation sites.
• The framework effectively distinguishes carbonylation sites from ligand interaction sites.
• This tool provides valuable insights into oxidative stress-related diseases and protein function.