AutoPeptideML: a study on how to build more trustworthy peptide bioactivity predictors

Area of Science:

• Computational biology
• Machine learning
• Bioinformatics

Background:

• Automated machine learning (AutoML) enables experimental scientists to leverage advanced computational tools.
• Developing reliable peptide bioactivity predictors requires careful consideration of model development and evaluation steps.
• Existing methods may overestimate performance due to inadequate handling of sequence homology.

Purpose of the Study:

• To develop an automated method for generating negative peptide datasets.
• To investigate the impact of homology-based partitioning on model evaluation.
• To identify optimal peptide representation methods and machine learning algorithms for bioactivity prediction.

Main Methods:

• Developed an automated method for negative peptide generation.
• Implemented homology-based partitioning for robust data splitting.
• Systematically analyzed protein language models for peptide representation.
• Evaluated traditional machine learning algorithms against neural networks.

Main Results:

• The new automated method improves specificity and generalization for negative peptide generation.
• Homology-based partitioning reveals overestimation of performance in prior studies.
• Protein language models offer improved peptide descriptors, with no significant differences across model sizes or types.
• Ensembles of traditional machine learning algorithms match neural network performance efficiently.

Conclusions:

• AutoPeptideML provides an accessible AutoML solution for peptide bioactivity prediction.
• The tool enhances model trustworthiness through robust evaluation and interpretable results.
• Researchers can rapidly develop custom predictive models without extensive computational expertise.