Retention prediction of peptides based on uninformative variable elimination by partial least squares
R Put1, M Daszykowski, T Baczek
1FABI, Department of Analytical Chemistry and Pharmaceutical Technology, Pharmaceutical Institute, Vrije Universiteit Brussel-VUB, Laarbeeklaan 103, B-1090 Brussels, Belgium.
Journal of Proteome Research
|July 11, 2006
Summary
Quantitative structure-retention relationship analysis improves peptide identification in proteomics. Uninformative Variable Elimination Partial Least Squares (UVE-PLS) modeling achieved high predictive accuracy for peptide retention times.
Area of Science:
- Analytical Chemistry
- Proteomics
- Chemometrics
Background:
- Accurate peptide identification is crucial in proteomics research.
- Chromatographic retention data, combined with molecular descriptors, can inform peptide identification.
- High dimensionality and redundancy in molecular descriptors necessitate robust modeling approaches.
Purpose of the Study:
- To develop and evaluate quantitative structure-retention relationship (QSRR) models for peptide retention prediction.
- To assess the utility of UVE-PLS for improving peptide identification in proteomics.
- To compare the predictive performance of UVE-PLS against traditional PLS and other regression methods.
Main Methods:
- Quantitative structure-retention relationship (QSRR) analysis of 90 peptides using reversed-phase liquid chromatography data.
- Computation of a large set of molecular descriptors for each peptide.
- Principal Component Analysis (PCA) for descriptor space analysis.
- Variable selection using Uninformative Variable Elimination (UVE).
- Modeling peptide retention using Partial Least Squares (PLS) regression and UVE-PLS.
- Validation using an independent external test set (27 peptides).
Main Results:
- Principal Component Analysis revealed significant information overlap among 1726 molecular descriptors.
- The UVE-PLS model, utilizing fewer components (5) than the best PLS model (7), demonstrated superior predictive properties.
- The UVE-PLS model achieved an average error in retention time prediction of less than 30 seconds.
- UVE-PLS predictions were significantly better than those from stepwise regression and an empirical model.
Conclusions:
- QSRR analysis, particularly with UVE-PLS, is a powerful tool for predicting peptide retention times in chromatography.
- UVE-PLS offers improved predictive accuracy and model parsimony compared to standard PLS regression.
- This approach enhances the reliability of peptide identification in complex proteomics datasets.
