METFAN: Multisource Enhanced Therapeutic Peptide Function Prediction via Adapter Network.

Area of Science:

• Computational biology and bioinformatics
• Drug discovery and precision medicine

Background:

• Multifunctional therapeutic peptides (MTPs) show promise for precision medicine due to diverse biological activities.
• Accurate computational prediction of MTP functions is challenging due to multilabel characteristics and severe class imbalance.

Purpose of the Study:

• To develop an advanced deep learning model for accurate prediction of MTP functions.
• To address the challenges of multilabel classification and class imbalance in MTP prediction.

Main Methods:

• Proposed a deep learning model, METFAN (Multisource Enhanced Therapeutic Peptide Function Prediction via Adapter Network).
• Integrated local sequence features (TextCNN) with global semantic embeddings (ESM2, ProtT5) from pretrained protein language models.
• Incorporated a feature optimization module to refine embeddings and a feature aggregation network to integrate heterogeneous features.

Main Results:

• METFAN achieved superior performance compared to state-of-the-art methods, with sample-level accuracy of 0.623 and label-level F1-score of 0.522.
• Demonstrated enhanced robustness and generalizability, particularly under severe label imbalance conditions.
• The model effectively integrates multisource features, enhancing prediction sensitivity and discriminative capacity.

Conclusions:

• METFAN provides a novel and effective framework for MTP function prediction.
• The model offers a solid foundation for peptide function screening and mechanistic studies in drug discovery.
• Publicly available data and code facilitate further research and application of the METFAN model.