Learning peptide properties with positive examples only.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Deep learning models require extensive labeled data, including negative examples, which are scarce for peptide properties.
• High-throughput screening methods struggle to efficiently generate negative examples for peptide databases.

Purpose of the Study:

• To develop a semi-supervised deep learning framework for predicting peptide properties using limited positive examples.
• To address the challenge of missing negative data in peptide sequence analysis.

Main Methods:

• Implemented positive-unlabeled (PU) learning strategies: adapting base classifiers and reliable negative identification.
• Built deep learning models to infer peptide solubility, hemolysis, SHP-2 binding, and non-fouling activity from sequences.
• Evaluated predictive performance against traditional positive-negative (PN) classification.

Main Results:

• The PU learning method achieved competitive predictive performance using only positive peptide data.
• Demonstrated the efficacy of PU learning in scenarios with limited negative examples.
• Successfully predicted multiple antimicrobial peptide properties.

Conclusions:

• Semi-supervised PU learning is a viable alternative to PN learning when negative data is scarce.
• This approach enables effective peptide property prediction and guides molecule design.
• Advances the application of deep learning in antimicrobial peptide research.