Dual-Modal Fusion Based on Peptide Sequence and Molecular Graph with Bidirectional Cross-Attention for Precise...

Area of Science:

• Pharmacology and Cheminformatics
• Computational Biology
• Drug Discovery

Background:

• Peptide drugs offer high specificity and biocompatibility but face toxicity challenges.
• Accurate toxicity assessment is crucial for translating peptide leads into clinical applications.
• Existing toxicity prediction tools require improvement in precision and scalability.

Purpose of the Study:

• To develop a high-precision, scalable framework for peptide toxicity prediction.
• To address the bottleneck of toxicity risk in peptide drug development.
• To improve the accuracy of discriminating toxic peptides.

Main Methods:

• Proposed a dual-modal feature fusion framework, PTP-SMGCA.
• Captured local motifs and long-range dependencies via sequence pathways.
• Characterized atomic topology and bonding semantics via molecular graph pathways.
• Utilized bidirectional cross-attention for dynamic alignment of sequence and graph features, suppressing noise.

Main Results:

• PTP-SMGCA achieved high performance in peptide toxicity prediction, with an AUROC of 0.9289 and AUPRC of 0.9397.
• Outperformed several advanced peptide toxicity prediction tools.
• Interpretability analysis identified key amino acids and functional groups, highlighting the role of aligned features.

Conclusions:

• PTP-SMGCA provides an accurate and effective method for peptide toxicity prediction.
• The framework facilitates rapid screening of toxic peptides, aiding therapeutic peptide drug development.
• Dual-modal feature fusion and cross-attention mechanisms are key to the model's predictive power.