pLM-HP: Peptide hormone prediction using pre-trained protein language model representations.

INTRODUCTION Peptide hormones play key roles in metabolic regulation, growth and development, and the maintenance of homeostasis and are important targets for drug discovery and design. However, traditional experimental methods for screening and identifying peptide hormones are costly and inefficient. Although existing machine learning methods have made some progress, they pay limited attention to class imbalance, and there is still room to improve prediction performance.

Objectives

This study aimed to develop an effective deep learning framework for imbalanced peptide hormone prediction and to improve the recognition of hormone peptides as the minority class.

Methods

We developed pLM-HP, a deep learning framework based on a pre-trained protein language model. The framework combines the ESM2 protein language model with a bidirectional long short-term memory network. ESM2 was first used to obtain high-dimensional global contextual embeddings from peptide sequences, and BiLSTM was then applied to model sequential dependencies between residues. Class weights were incorporated into the loss function to reduce bias caused by class imbalance.

Results

Compared with traditional feature-based methods, pLM-HP achieved a BACC of 95.61% and an MCC of 0.802 in five-fold cross-validation, and a BACC of 95.64% and an MCC of 0.824 on the independent test set. The model also maintained stable and strong performance under imbalanced data conditions.

Conclusion

These results indicate that combining protein language models with sequence modeling and class-imbalance learning strategies is an effective way to improve peptide hormone prediction. pLM-HP may provide a useful tool for high-throughput functional peptide screening and prioritization of candidate peptide hormones. The source code of pLM-HP is freely available at https://github.com/aochunyan123/pLM-HP.git.