iPro2L-Kresidual: A High-Performance Promoter Identification Model for Sequence Nonlinearity and Context Mining

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Promoters are crucial non-coding DNA sequences regulating gene expression.
• Abnormalities in promoters are linked to diseases like cancer, diabetes, and heart disease.
• Existing models struggle with nonlinear feature extraction and context relationship capture, limiting promoter identification performance.

Purpose of the Study:

• To develop an advanced model for accurate promoter identification.
• To enhance the extraction of nonlinear features and sequence context relationships.
• To improve the classification performance of promoter identification models.

Main Methods:

• Proposed iPro2L-Kresidual model integrating residual structure and KAN network (Kresidual module).
• Enhanced Transformer encoder using gated recurrent units for local and global context feature extraction.
• Implemented a regularized label smoothing cross-entropy loss function for training stability.

Main Results:

• Achieved 94.28% accuracy for promoter identification and 90.55% for promoter strength identification via 5-fold cross-validation.
• Demonstrated strong generalization ability with 93.13% prediction accuracy on an independent dataset.
• The Kresidual module improved nonlinear sequence feature expression, and enhanced Transformer captured sequence context effectively.

Conclusions:

• iPro2L-Kresidual offers a novel and effective approach for promoter site prediction.
• The model overcomes limitations of existing methods in nonlinear feature extraction and context understanding.
• This work provides a significant advancement in computational genomics for disease-related gene regulation studies.