Multi-Peptide: Multimodality Leveraged Language-Graph Learning of Peptide Properties.

Area of Science:

• Bioinformatics
• Computational Biology
• Drug Discovery

Background:

• Peptides play vital roles in biological systems and are key therapeutic agents.
• Accurate prediction of peptide properties is crucial for advancing their applications.
• Current methods may not fully capture complex peptide characteristics.

Purpose of the Study:

• To develop an advanced computational model for predicting peptide properties.
• To integrate sequence and structural information for enhanced prediction accuracy.
• To establish a new benchmark for peptide property prediction.

Main Methods:

• Introduction of Multi-Peptide, a multimodal approach integrating transformer language models and graph neural networks (GNNs).
• Utilized PeptideBERT, a transformer model, alongside a GNN encoder to process sequence and structural data.
• Employed a contrastive loss framework to align embeddings in a shared latent space.

Main Results:

• Multi-Peptide demonstrated robust performance in predicting peptide properties.
• Achieved state-of-the-art 88.057% accuracy in hemolysis prediction.
• Successfully integrated sequence-based and structural features for improved predictions.

Conclusions:

• Multimodal learning offers significant potential for bioinformatics and peptide research.
• Multi-Peptide provides a powerful tool for accurate and reliable peptide property prediction.
• This approach paves the way for novel peptide-based therapeutics and applications.