MultiPep: a hierarchical deep learning approach for multi-label classification of peptide bioactivities

Area of Science:

• Computational chemistry and bioinformatics
• Drug discovery and development
• Machine learning in pharmacology

Background:

• Peptide-based therapeutics are a growing field, necessitating efficient methods for identifying and characterizing novel peptide drug candidates.
• Current peptidomics screening generates vast numbers of peptides, requiring robust tools for prioritizing candidates for functional studies.
• Existing bioactivity classifiers often rely on multiple binary classifiers, limiting their ability to handle multi-label classification effectively.

Purpose of the Study:

• To develop an advanced deep learning multi-label classifier for predicting peptide bioactivities.
• To create a tool that can assign peptides to zero or more of 20 distinct bioactivity classes.
• To improve upon existing state-of-the-art methods for peptide bioactivity prediction.

Main Methods:

• Development of MultiPep, a deep learning multi-label classifier.
• Training and testing MultiPep using data from public databases, with architecture informed by hierarchical clustering.
• Evaluation of a novel loss function combining Matthews correlation coefficient and binary cross entropy (BCE).

Main Results:

• MultiPep successfully assigns peptides to multiple bioactivity classes.
• The novel loss function demonstrated superior performance compared to class-weighted BCE.
• MultiPep outperformed existing state-of-the-art peptide bioactivity classifiers.
• The tool accurately predicted known and novel bioactivities for FDA-approved therapeutic peptides.

Conclusions:

• Innovative machine learning techniques have been applied to create MultiPep, a powerful peptide prediction tool.
• MultiPep aids in the development of peptide-based therapies by facilitating candidate selection and hypothesis generation.
• This tool represents a significant advancement in computational approaches for peptide drug discovery.