MetaMBP: Few-Shot Multilabel Prediction of Bioactive Peptides Based on Deep Metric Meta-Learning.

Area of Science:

• Computational Biology
• Bioinformatics
• Drug Discovery

Background:

• Bioactive peptides offer targeted, low-toxicity therapeutic potential.
• Many bioactive peptide types have limited sample sizes, hindering predictive model development.
• Existing methods struggle with multilabel prediction tasks involving few-sample categories.

Purpose of the Study:

• To develop a novel multilabel model, MetaMBP, for predicting bioactive peptide functions.
• To enhance the predictive performance for bioactive peptide categories with limited samples.
• To leverage meta-learning to improve predictions in few-shot scenarios.

Main Methods:

• Proposed MetaMBP, a deep metric meta-learning model for bioactive peptide function prediction.
• Utilized meta-knowledge from a meta-learning stage to aid fine-tuning on limited sample categories.
• Employed feature visualization and attention score analysis to understand model behavior.

Main Results:

• MetaMBP outperformed existing methods on benchmark datasets, especially for limited sample categories.
• Few-shot experiments demonstrated MetaMBP's adaptability and effectiveness.
• Analysis revealed insights into category relationships and model's learned features.

Conclusions:

• MetaMBP provides an accurate and adaptive approach for screening multilabel bioactive peptides.
• The model effectively addresses the challenge of limited sample sizes in bioactive peptide prediction.
• MetaMBP shows promise for accelerating the discovery of novel peptide therapeutics.