Exploring the variable space of shallow machine learning models for reversed-phase retention time prediction.

Area of Science:

• Proteomics and Bioinformatics
• Computational Chemistry

Background:

• Peptide retention time (RT) prediction is crucial for understanding peptide-sorbent interactions and physicochemical properties.
• Traditional methods rely on manual feature engineering, while deep learning offers high accuracy but lacks interpretability.

Purpose of the Study:

• To develop accurate and interpretable peptide retention time prediction models.
• To isolate and understand the features learned by deep learning modules for RT prediction.

Main Methods:

• Utilized shallow convolutional neural networks (CNN) and gated recurrent units (GRU) as isolated deep learning modules.
• Correlated learned spatial features with physicochemical properties such as cross-collision sections (CCS) and accessible surface area (ASA).

Main Results:

• CNN-derived spatial features correlated with CCS and ASA.
• Identified learned parameters as 'micro-coefficients' contributing to hydrophobicity.
• GRU model with embedded CCS and ASA achieved R² = 0.981, representing 88% of the dataset with only 525 variables.

Conclusions:

• Shallow learners can achieve high accuracy and superior interpretability in peptide RT prediction compared to traditional and deep learning models.
• The identified physicochemical properties (CCS, ASA) are key drivers of peptide-sorbent interactions.