Accelerating bioactive peptide discovery via mutual information-based meta-learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Machine learning for peptide bioactivity prediction suffers from poor generalizability due to limited experimental data.
• Existing computational frameworks lack a generic approach for predicting diverse peptide bioactivities.
• A key challenge is developing effective predictive models with limited samples for various peptide functions.

Purpose of the Study:

• To propose a novel meta-learning framework, Mutual Information Maximization Meta-Learning (MIMML), for bioactive peptide discovery.
• To address the challenge of few-sample learning in biological sequence analysis for functional peptide mining.
• To develop a generic computational solution for predicting bioactivities of different peptides.

Main Methods:

• Developed Mutual Information Maximization Meta-Learning (MIMML), a novel meta-learning-based predictive model.
• Utilized few samples from various functional peptides to train the meta-learning model.
• Focused on learning discriminative information and characterizing functional differences among peptides.

Main Results:

• MIMML demonstrated excellent performance in predicting peptide bioactivities.
• The model achieved high accuracy using significantly fewer training samples compared to state-of-the-art methods.
• Latent relationships among different peptide functions were deciphered to understand model learning.

Conclusions:

• MIMML offers a pioneering solution for few-sample learning problems in biological sequence analysis.
• This approach accelerates the discovery of new functional peptides.
• The study provides the first computational framework for generic bioactive peptide prediction with limited data.