3.2 Metabolomic profiling of no AKI vs. AKI patients
Following XCMS processing of chromatographic data collected from untargeted metabolomics, exclusion of features with high variability in pooled quality control samples, and removal of potential adducts/isotopes, 758 serum features and 484 urine features were included for subsequent multivariate analysis. The derivatized metabolomics dataset was similarly processed, without the removal of potential adducts or isotopes. Additionally, any duplicate features detected in both the derivatized and untargeted datasets were removed. Ultimately, a total of 975 serum features and 2355 urine features remained after processing of derivatized data. After cross referencing the derivatized features with a library of 263 derivatized small molecule standards, 35 serum and 117 urine features were matched bym/z and retention time with the library of standards.
To select the features most important in discriminating between no AKI and AKI patients at each timepoint, OPLS-DA models were sequentially generated, each time excluding features with VIP values < 1. This sequential exclusion of features was repeated until model statistics of OPLS-DA models were maximized. With the remaining features, PCA score plots were generated to visualize the metabolic differences between no AKI and AKI patients at each timepoint. Score plots of urine samples showed moderate separation at the pre (Figure 1A ) and 24-48h timepoints (Figure 1B ), and clear separation at the post timepoint (Figure 1C ). In serum, strong separation was observed in both the pre (Figure 2A ) and post (Figure 2C ) timepoints, with moderate separation observed at the 24-48h timepoint (Figure 2B ). Corresponding OPLS-DA models comparing no AKI and AKI patients at each timepoint mirrored the visual separation observed in the PCA score plots, with a high degree of fit (R2Y) and predictive ability (Q2Y) at the post timepoint for both urine and serum (Figure 1F, 2F ), as well as the pre timepoint for serum (Figure 2D ). Moderate model statistics were observed for the 24-48h timepoint for both urine and serum (Figure 1E, 2E ), as well as the pre timepoint for urine (Figure 1D ). Features with 0.4 < p(corr) < -0.4 and VIP > 1 in the OPLS-DA models were considered as important discriminators of AKI and were thus followed up for identification. Identified metabolites that were significantly different between no AKI and AKI groups by two-way ANOVA are summarized in Table 2 .