Integrating Protein Language Models and Geometric Deep Learning for Peptide Toxicity Prediction

Area of Science:

• Biomedical research
• Drug discovery
• Computational biology

Background:

• Peptide toxicity prediction is vital for drug safety.
• Existing methods struggle with complex structure-toxicity relationships.

Purpose of the Study:

• Introduce PeptiTox, a deep learning framework for enhanced peptide toxicity prediction.
• Integrate sequence and structural data for improved accuracy.

Main Methods:

• Utilize ESM2 for sequence embeddings and ESMFold for 3D structure prediction.
• Employ graph neural networks (GNNs) on graph representations of peptide structures.
• Train the GNN to classify peptide toxicity based on learned representations.

Main Results:

• PeptiTox significantly outperforms existing state-of-the-art models.
• Demonstrated superior performance across multiple evaluation metrics.
• Validated the importance of integrating sequence and structural data.

Conclusions:

• PeptiTox offers a powerful new approach for peptide toxicity prediction.
• Highlights the value of combining sequence and structural information.
• Facilitates the development of safer peptide-based therapeutics.