Metabolomics Data Analysis with TIGER

TIGER, a non-parametric method, was developed to address technical variations (e.g., plate and batch effects) in targeted and non-targeted metabolomics datasets. It integrates the random forest (RF) algorithm into a flexible ensemble learning framework, combining multiple base models with a meta-model. These base models are trained using diverse RF hyperparameter combinations, eliminating the need for manual hyperparameter tuning. This chapter highlights practical considerations for using TIGER effectively, including incorporating quality control (QC) samples into study design. When QCs are unavailable, randomly selected samples can be remeasured to facilitate cross-kit corrections. To optimize processing time, highly correlated metabolites from QC samples are selected to train the base models, with weights assigned via an exponential decay function. TIGER employs relative standard deviation (RSD) and mean absolute percentage error (MAPE) as key metrics to ensure models generalize well to unseen data while avoiding overfitting. Additionally, the developed dynamic website has been demonstrated with raw and TIGER-normalized data, enabling performance evaluation and visualization of longitudinal patterns of metabolites or metabolite ratios. This platform demonstrates TIGER's ability to accurately normalize data and uncover trends across three time points spanning a decade. With its proper application, TIGER stands to be a powerful tool for metabolomics studies.