ToxTeller: Predicting Peptide Toxicity Using Four Different Machine Learning Approaches.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Peptide toxicity assessment is critical for developing safe and effective peptide-based therapeutics.
• Machine learning (ML) models are increasingly employed for accurate peptide toxicity prediction.

Purpose of the Study:

• To develop and validate ToxTeller, a novel tool providing four distinct ML predictors for peptide toxicity.
• To enhance the identification of toxic peptides by improving prediction accuracy and minimizing false negatives.

Main Methods:

• Constructed a comprehensive dataset of toxic and non-toxic peptides from SwissProt and ConoServer, verifying evidence levels.
• Utilized extensive SwissProt annotations to identify additional toxic peptides beyond keyword searches.
• Developed four ML models (logistic regression, SVM, random forests, XGBoost) and optimized feature combinations via 10-fold cross-validation.
• Created an independent test set with limited sequence similarity to training data for robust evaluation.

Main Results:

• ToxTeller demonstrated robust performance in predicting peptide toxicity on an independent test set.
• Evaluated and compared the performance of ToxTeller's four predictors against existing methods.
• Identified optimized feature combinations through rigorous cross-validation.

Conclusions:

• ToxTeller provides reliable peptide toxicity prediction, supporting safer drug design.
• Prioritizing sensitivity over Matthews correlation coefficient is recommended for minimizing false-negative toxic peptide predictions.
• A meta-predictor approach combining multiple models is suggested for enhanced accuracy in therapeutic peptide development.