Toward Automated Preprocessing of Untargeted LC-MS-Based Metabolomics Feature Lists from Human Biofluids

Maximizing the extraction of true, high-quality, nonredundant features from biofluids analyzed via LC-MS systems is challenging. Here, the R packages IPO and AutoTuner were used to optimize XCMS parameter settings for the retrieval of metabolite or lipid features in both ionization modes from either faecal or urine samples from two cohorts (n = 621). The feature lists obtained were compared with those where the parameter values were selected manually. Three categories were used to compare feature lists: 1) feature quality through removing false positives, 2) tentative metabolite identification using the Human Metabolome Database (HMDB) and 3) feature utility such as analyzing the proportion of features within intensity threshold bins. Furthermore, a PCA-based approach to feature filtering using QC samples and variable loadings was also explored under this category. Overall, more features were observed after automated selection of parameter values for all data sets (1.3- to 3.7-fold), which propagated through comparative exercises. For example, a greater number of features (on average 51 vs 45%) had a coefficient of variation (CV) < 30%. Additionally, there was a significant increase (7.6-10.4%) in the number of faecal metabolites that could be tentatively annotated, and more features were present in higher intensity threshold bins. Considering the overlap across all three categories, a greater number of features were also retained. Automated approaches that guide selection of optimal parameter values for preprocessing are important to decrease the time invested for this step, while taking advantage of the wealth of data that LC-MS systems provide.