MvAl-MFP: A Multi-Label Classification Method on the Functions of Peptides with Multi-View Active Learning

Area of Science:

• Bioinformatics
• Computational Biology
• Peptide Science

Background:

• Peptide libraries are expanding, increasing the need for predicting multifunctional peptide properties.
• Supervised learning methods require extensive labeled data for accurate peptide prediction.
• Existing methods face challenges in efficiently predicting diverse peptide functions.

Purpose of the Study:

• To introduce MvAl-MFP, a novel multi-label active learning approach for peptide prediction.
• To leverage multi-view representations and active learning to reduce the need for labeled data.
• To develop a high-performing model for predicting multifunctional peptides efficiently.

Main Methods:

• Generated nine distinct feature views from labeled peptide sequences based on various characteristics.
• Trained multi-label classifiers on each feature view using limited labeled samples.
• Employed a query-by-committee (QBC) strategy with average entropy to select informative unlabeled samples for wet-lab validation.
• Iteratively refined classifiers with an expanding labeled dataset.

Main Results:

• MvAl-MFP significantly reduced the requirement for labeled samples in training predictive models.
• The method demonstrated superior performance in predicting multifunctional peptides.
• Experimental results confirmed the rapid improvement in prediction accuracy with minimal labeling.

Conclusions:

• MvAl-MFP offers an effective solution for precise multifunctional peptide prediction.
• The approach substantially lowers the cost and effort associated with wet-lab experiments.
• This method advances bioinformatics research by enabling efficient prediction of peptide properties.