Development of biomarkers based on diet-dependent metabolic serotypes: practical issues in development of expert system-based classification models in metabolomic studies

Abstract

Dietary restriction (DR)-induced changes in the serum metabolome may be biomarkers for physiological status (e.g., relative risk of developing age-related diseases such as cancer). Megavariate analysis (unsupervised hierarchical cluster analysis [HCA]; principal components analysis [PCA]) of serum metabolites reproducibly distinguish DR from ad libitum fed rats. Component-based approaches (i.e., PCA) consistently perform as well as or better than distance-based metrics (i.e., HCA). We therefore tested the following: (A) Do identified subsets of serum metabolites contain sufficient information to construct mathematical models of class membership (i.e., expert systems)? (B) Do component-based metrics out-perform distance-based metrics? Testing was conducted using KNN (k-nearest neighbors, supervised HCA) and SIMCA (soft independent modeling of class analogy, supervised PCA). Models were built with single cohorts, combined cohorts or mixed samples from previously studied cohorts as training sets. Both algorithms over-fit models based on single cohort training sets. KNN models had >85% accuracy within training/test sets, but were unstable (i.e., values of k could not be accurately set in advance). SIMCA models had 100% accuracy within all training sets, 89 % accuracy in test sets, did not appear to over-fit mixed cohort training sets, and did not require post-hoc modeling adjustments. These data indicate that (i) previously defined metabolites are robust enough to construct classification models (expert systems) with SIMCA that can predict unknowns by dietary category; (ii) component-based analyses outperformed distance-based metrics; (iii) use of over-fitting controls is essential; and (iv) subtle inter-cohort variability may be a critical issue for high data density biomarker studies that lack state markers.