Particle swarm optimization-based protocol for partial least-squares discriminant analysis: Application to 1H nuclear magnetic resonance analysis of lung cancer metabonomics

Abstract

The complexity of metabolic profiles makes multivariate chemometric techniques crucial for extracting mostly significant information and offering biological insight. Partial least-squares discriminant analysis (PLS-DA) was proven fruitful in metabonomic community, due to its promising properties. The issues of suboptimum and overfitting, however, often occur in PLS-DA modeling. In the current study, particle swarm optimization (PSO) was invoked to meliorate PLS-DA via simultaneously selecting the optimal variable subset as well as the associated weights and the best number of latent variables in PLS-DA, forming a new algorithm named PSO-PLSDA. Combined with 1H NMR-based metabonomics, PSO-PLSDA compared with PLS-DA was applied to recognize lung cancer patients from healthy controls. Relatively to the recognition rates of 86% and 65% for the training and test sets yielded by PLS-DA, 99% and 85% were obtained by PSO-PLSDA. Moreover, several most discriminative metabolites were identified by PSO-PLSDA to aid the diagnosis of lung cancer, including lactate, proline, glycoprotein, glutamate, alanine, threonine, taurine, glucose (α- and β-), trimethylamine, glutamine, glycine, and myo-inositol.