Function Prediction of Peptide Toxins with Sequence-Based Multi-Tasking PU Learning Method.

Area of Science:

• Pharmacology
• Bioinformatics
• Computational Biology

Background:

• Peptide toxins possess significant pharmacological activities, making them valuable for drug lead discovery.
• Identifying optimal activities of novel peptide toxins is often a resource-intensive and time-consuming endeavor.

Purpose of the Study:

• To develop and validate machine learning models for predicting the biological functions of peptide toxins.
• To accelerate the identification of potential peptide toxin drug leads.

Main Methods:

• Retrieved peptide toxins from UniProt and employed three positive-unlabeled (PU) learning schemes: adaptive basis classifier, two-step method, and PU bagging.
• Integrated 14 machine learning classifiers within the PU learning frameworks.
• Optimized top-performing models using feature selection and hyperparameter tuning.
• Validated models on three-finger toxins and the HemoPI dataset.

Main Results:

• Adaptive basis classifier and two-step method yielded highly consistent prediction results.
• Models accurately predicted functions like cardiotoxicity, neurotoxicity, and hemolysis.
• Predictions for hemostasis and presynaptic neurotoxicity were relatively weaker.
• Identified biological functions include cardiotoxicity, vasoactivity, lipid binding, hemolysis, neurotoxicity, postsynaptic neurotoxicity, hypotension, and cytolysis.

Conclusions:

• Developed predictive models serve as valuable tools for discovering active peptide toxins.
• The models are expected to expedite the development of peptide toxins into therapeutic agents.
• The study highlights the potential of machine learning in accelerating drug discovery from natural peptide sources.