SME-MFP: A novel spatiotemporal neural network with multiangle initialization embedding toward multifunctional...

Area of Science:

• Biomedical Science
• Computational Biology
• Machine Learning

Background:

• Functional peptides offer a low-toxicity, high-absorption alternative to conventional antibiotics.
• Machine learning aids functional peptide prediction, but struggles with multifunctional identification and imbalanced data.

Purpose of the Study:

• To develop SME-MFP, a novel predictor for imbalanced multi-label functional peptide datasets.
• To improve the accuracy and feature extraction capabilities for functional peptide identification.

Main Methods:

• Utilized physicochemical and evolutionary information for peptide sequence representation.
• Employed fused features with spatial (residual connection, multiscale CNN) and temporal (AFT) feature extractors.
• Developed a novel loss function to address class imbalance issues in multi-label datasets.

Main Results:

• The SME-MFP framework effectively captures sequence features from multiple perspectives.
• Achieved a 3.89% accuracy improvement over existing methods on public peptide datasets.
• Demonstrated enhanced model performance in identifying functional peptides, particularly in imbalanced scenarios.

Conclusions:

• SME-MFP significantly enhances the ability to capture peptide sequence features.
• The proposed predictor offers improved accuracy for functional peptide identification, especially in challenging imbalanced datasets.
• This work advances machine learning applications in peptide-based therapeutics.