Structure-Aware Multimodal Learning Improves Minor-Class Signal Peptide Prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Signal peptides are crucial for protein secretion and localization.
• Class imbalance in datasets hinders accurate prediction of minor signal peptide classes.
• Existing methods struggle with performance disparities between major and minor classes.

Purpose of the Study:

• To develop a structure-aware multimodal network (SaSPNet) for improved signal peptide prediction.
• To address the challenge of class imbalance in signal peptide identification.
• To enhance the prediction accuracy for minor signal peptide classes.

Main Methods:

• Proposed SaSPNet, a network incorporating structural modality information.
• Utilized a graph convolutional network (GCN)-based structure encoder for signal peptide representation.
• Incorporated structure prediction models and created a dedicated minor signal peptide test set (SP-MinorEval).

Main Results:

• SaSPNet significantly improved prediction performance for minor signal peptide classes, exceeding existing methods by over 10% on key metrics.
• Feature visualization confirmed that the structure encoder learns discriminative patterns for minor signal peptides.
• SaSPNet demonstrated robustness to variations in structural data quality.

Conclusions:

• SaSPNet effectively enhances minor signal peptide prediction by integrating structural information.
• The structure encoder provides insights into the mechanism of improved performance for rare classes.
• SaSPNet offers a robust tool for minor-class signal peptide prediction, aiding protein secretion studies and discovery.