PeptideNet: An Integrative Deep Learning Framework for Predicting Diverse Bioactive Peptides Using Protein Language...

Area of Science:

• Biochemistry and Bioinformatics
• Computational Biology and Drug Discovery

Background:

• Bioactive peptides are crucial biomolecules with diverse therapeutic activities.
• Accurate computational prediction of peptide bioactivity is vital for drug development.
• Existing methods require enhancement for comprehensive bioactivity prediction.

Purpose of the Study:

• To develop and validate a deep learning model, PeptideNet, for predicting multiple bioactive peptide functions.
• To evaluate the efficacy of large protein language model embeddings and physicochemical descriptors for bioactivity prediction.
• To establish a generalized and interpretable framework for multi-bioactivity prediction.

Main Methods:

• Investigated five categories of bioactive peptides: antioxidative, antihemolytic, anticell-penetrating, antiviral, and antimicrobial.
• Utilized four feature representations: ESM1, ESM2, ProtBert embeddings, and physicochemical descriptors.
• Developed 20 hybrid deep learning models integrating Convolutional Neural Networks (CNNs) and Bidirectional Gated Recurrent Units (BiGRUs).

Main Results:

• PeptideNet achieved high predictive accuracies: 0.83 (antioxidative), 0.87 (antihemolytic), 0.89 (anticell-penetrating), 0.92 (antiviral), and 0.94 (antimicrobial).
• ESM-2 embeddings consistently outperformed other feature sets, offering rich contextual and evolutionary information.
• t-SNE visualization and sequence logo analysis confirmed effective generalization and identified key residue patterns.

Conclusions:

• The PeptideNet model provides a robust and accurate framework for predicting multiple bioactive peptide functions.
• Large protein language model embeddings, particularly ESM-2, significantly enhance prediction performance.
• The integrated approach offers a generalized and interpretable tool for accelerating peptide-based therapeutic discovery.