Abstract
Metabolomics aims to perform a comprehensive identification and quantification of the small molecules present in a biological system. Due to metabolite diversity in concentration, structure, and chemical characteristics, the use of high-resolution methodologies, such as mass spectrometry (MS) or nuclear magnetic resonance (NMR), is required. In metabolomics data analysis, suitable data pre-processing, and pre-treatment procedures are fundamental, with subsequent steps aiming at highlighting the significant biological variation between samples over background noise. Traditional data analysis focuses primarily on the comparison of the features’ intensity values. However, intensity data are highly variable between experimental batches, instruments, and pre-processing methods or parameters. The aim of this work was to develop a new pre-treatment method for MS-based metabolomics data, in the context of sample profiling and discrimination, considering only the occurrence of spectral features, encoding feature presence as 1 and absence as 0. This “Binary Simplification” encoding (BinSim) was used to transform several benchmark datasets before the application of clustering and classification methods. The performance of these methods after the BinSim pre-treatment was consistently as good as and often better than after different combinations of traditional, intensity-based, pre-treatments. Binary Simplification is, therefore, a viable pre-treatment procedure that effectively simplifies metabolomics data-analysis pipelines.
Highlights
Untargeted metabolomics experiments are focused on obtaining a global picture of a system with the objectives of identifying and characterizing all its metabolites or identifying key features, characteristics, and trends in the data that can help define and discriminate the systems under study [1,2,3]
Several benchmark datasets were treated with seven different pre-treatment methods: Binary Simplification” encoding (BinSim) and combinations of different pre-treatments widely used in metabolomics data analysis—missing-value imputation by either half the minimum value of the data matrix (1/2 min imputation) or Random Forest imputation (RF imputation) [18]; normalization, by reference feature or Probabilistic Quotient Normalization [19], depending on the dataset; generalized logarithmic transformation; and Pareto scaling [4,8]
The results show that the BinSim pre-treatment does not compromise the performance of the classifiers developed for the eight problems included in this benchmark
Summary
Untargeted metabolomics experiments are focused on obtaining a global picture of a system with the objectives of identifying and characterizing all its metabolites or identifying key features, characteristics, and trends in the data that can help define and discriminate the systems under study [1,2,3]. Every biological system is unique and will have a unique metabolome. Even two genetically identical systems will have minor differences in their metabolome. This uninduced biological variation will lead to inherent variability in the data [3,4]. The metabolome is extremely sensitive to experimental manipulation and environmental factors (slight changes in pH or growth medium, stress, temperature, among others), leading to considerable changes in metabolite concentration [3]. The variability in the metabolome justifies the need for great care in the proper quenching of cellular metabolism and the efficient extraction of metabolites during sample preparation. The inherent metabolite concentration variability requires adequate statistical analysis
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
