Quantitative structure retention relationship (QSRR) modelling for Analytes' retention prediction in LC-HRMS by...

Area of Science:

• Metabolomics
• Cheminformatics
• Computational Chemistry

Background:

• Retention prediction is crucial in metabolomics for analyte identification.
• Quantitative Structure Retention Relationships (QSRR) models link molecular properties to retention times.
• Machine learning algorithms are increasingly used to build QSRR models.

Purpose of the Study:

• To evaluate the performance of four machine learning regression algorithms for QSRR model development.
• To investigate the impact of collinearity in molecular descriptor datasets on QSRR model performance.
• To compare the retention prediction abilities of Bayesian Ridge Regression, Extreme Gradient Boosting Regression, and Support Vector Regression.

Main Methods:

• Development of QSRR models using Bayesian Ridge Regression, Extreme Gradient Boosting Regression, and Support Vector Regression (linear and non-linear kernels).
• Testing models on experimentally derived and publicly available chromatographic data.
• Analysis of datasets with varying levels of feature collinearity, defined by Pearson's correlation coefficient.

Main Results:

• The selected regression algorithms demonstrated an ability to effectively handle collinearity in the datasets.
• No statistically significant differences in QSRR model performance were observed across different collinearity levels in most cases.
• No single algorithm consistently outperformed others across all tested datasets.

Conclusions:

• Machine learning-based QSRR models are robust to collinearity in molecular descriptors.
• The choice of regression algorithm did not significantly impact predictive performance in the presence of collinearity.
• This study provides insights into the reliability of QSRR modeling in metabolomics.