Evaluating the effect of data merging and post-acquisition normalization on statistical analysis of untargeted high-resolution mass spectrometry based urinary metabolomics data

Urine is one of the most widely used biofluids in metabolomic studies, because it can be collected non-invasively and is available in large quantities. However, it shows large heterogeneity in sample concentration and consequently requires normalization to reduce unwanted variation and extract meaningful biological information. Biological samples like urine are commonly measured with electrospray ionization (ESI) coupled to a mass spectrometer, producing datasets for positive and negative mode. Combining these gives a more complete picture of the total metabolites present in a sample. However, the effect of this data merging on subsequent data analysis, especially in combination with normalization, has not yet been analysed. To address this issue, we conducted a neutral comparison study to evaluate the performance of eight post-acquisition normalization methods under different data merging procedures using 1029 urine samples from the Food Chain plus (FoCus) cohort. Samples were measured by a Fourier transform ion cyclotron resonance mass spectrometer (FT-ICR-MS). Normalization methods were evaluated by five criteria capturing the ability to remove sample concentration variation and preserve relevant biological information. Merging data after normalization was generally favourable for quality control (QC) sample similarity, sample classification and feature selection for most of the tested normalization methods. Merging data after normalization and the usage of probabilistic quotient normalization (PQN) in a similar setting are generally recommended. Relying on a single analyte to capture sample concentration differences, like with post-acquisition creatinine normalization, seems to be a less preferable approach, especially when data merging is applied.