A selective genetic algorithm - PLS-DA approach based on untargeted LC-HRMS: Application to complex biomass samples

Abstract

Supervised multivariate analyses (MVA) are commonly used to extract valuable information, classify samples and highlight data trends from complex LC-HRMS datasets. However, traditional MVA methods, such as partial least-squares discriminant analysis (PLS-DA), may not provide sufficient class discrimination in the case of complex mixtures due to the presence of confounders, high-dimensionality and multicollinearity. In this regard, variable selection methods such as genetic algorithms can reduce datasets by retaining explanatory information while discarding non-informative features. The hyphenation of these selective heuristic methods and highly interpretable PLS-DA has shown significant potential for sample discrimination in spectroscopic techniques but has never been applied to complex LC-HRMS data.

Therefore, this work investigated the potential of combining variable selection by GA and classification by PLS-DA for identifying key chemical descriptors in these intricate LC-HRMS datasets. The determination of low reactivity discriminants (i.e., potential enzymatic inhibitors) in complex biomass samples was chosen as a case study and is presented in this work. Initially, PLS-DA was applied to an LC-ESI(±)-HRMS dataset containing 2167 features. However, the resulting model suffered from overfitting and yielded sub-optimal classification for low reactivity samples. Subsequently, GA was employed in triplicate, resulting in 768 final models. The most predictive model retained only 418 variables (i.e., an 81 % decrease in the dataset) and was subjected to PLS-DA. The new supervised model avoided overfitting and demonstrated significantly improved classification for low reactivity samples, achieving sensitivity, precision, and accuracy increases of 45 %, 25 %, and 38 %, respectively. Ultimately, an original strategy based on combining the variables’ importance in projection (VIP), GA-selection frequency (GA-SF) and cosine similarity was used to pinpoint species associated with low enzymatic reactivity. This study marks the first application of GA-PLS-DA for LC-HRMS data analysis. This approach effectively streamlined feature selection and enhanced class discrimination, successfully identifying 10 potential inhibitory features within the biomass samples. This methodology offers significant potential for studies on complex mixtures that encounter limited classification and/or overfitting issues with traditional supervised MVA methods.