How to build machine learning models able to extrapolate from standard to modified peptides.

Area of Science:

• Biochemistry and Cheminformatics
• Computational Chemistry
• Pharmacology

Background:

• Bioactive peptides are versatile natural products with therapeutic potential.
• Chemical modifications enhance peptide pharmacology but pose computational modeling challenges.
• Abundant data exists for standard peptides, but limited data is available for modified peptides.

Purpose of the Study:

• To assess the reliability of predictive models trained on standard peptide data when applied to modified peptides.
• To explore critical aspects of computational modeling for bioactive peptides, including similarity functions and molecular representations.
• To evaluate similarity-based dataset partitioning for model validation.

Main Methods:

• Investigated the impact of similarity functions on dataset partitioning for model evaluation.
• Examined different molecular representations for modeling peptides.
• Utilized similarity-based dataset partitioning to create distinct training and testing sets.
• Assessed the performance of models trained on standard peptide data against modified peptide data.

Main Results:

• The study identified key factors influencing the reliability of predictive models for modified peptides.
• Findings highlight the importance of appropriate similarity functions and molecular representations in computational peptide modeling.
• Dataset partitioning strategies significantly affect model generalization to modified peptide structures.

Conclusions:

• Predictive models developed using standard peptide data may exhibit reduced reliability when applied to chemically modified peptides.
• Careful consideration of similarity metrics and molecular representations is essential for accurate computational modeling of modified bioactive peptides.
• Robust validation strategies, such as similarity-based partitioning, are crucial for assessing model performance on diverse peptide datasets.