PLPTP: A Motif-based Interpretable Deep Learning Framework Based on Protein Language Models for Peptide Toxicity...

Area of Science:

• Computational biology and bioinformatics
• Drug discovery and development
• Machine learning in biotechnology

Background:

• Accurate peptide toxicity prediction is vital for developing safe peptide-based therapeutics.
• Traditional methods often struggle with the complexity and imbalanced nature of peptide toxicity data.

Purpose of the Study:

• To develop and validate a novel deep learning model for enhanced peptide toxicity prediction.
• To improve the accuracy and reliability of identifying toxic peptide sequences in drug development.

Main Methods:

• Integration of Evolutionary Scale Modeling (ESM2) for sequence context, Bidirectional Long Short-Term Memory (BiLSTM) for dependency extraction, and Deep Neural Network (DNN) for classification.
• Utilized motif analysis for model interpretability and transparency.
• Employed Focal Loss to effectively address class imbalance in the dataset.

Main Results:

• The proposed deep learning model demonstrated superior performance across multiple evaluation metrics.
• Significant improvements were observed in handling imbalanced datasets compared to traditional approaches.
• Motif analysis provided insights into the model's attention mechanisms and classification decisions.

Conclusions:

• The developed model offers a powerful tool for accurate peptide toxicity prediction, aiding in the design of safer peptide drugs.
• This approach has significant implications for advancing drug development and biotechnology research.
• The PLPTP web server is publicly available for broader application.