**3. Results**

#### *3.1. Metabolites in Urine Are Perturbed in Di*ff*erent Transplant Injuries in Kidney Transplantation*

We processed 326 urine samples for a targeted metabolomics assay that identified 266 metabolites. Figure 1 summarizes the study. Sixteen samples had missing data on more than one third of total metabolites identified following a tool called MissForest on non-parametric missing value imputation for mixed-type data [30]. Metabolomics data on the remaining 310 biopsy-matched urine samples was used for the analyses of both post-Tx injury detection and associated metabolic pathways and their enrichment. Baseline characteristics of the study subjects is provided in Table 1.

The data was used for supervised clustering to generate a heat map (Figure 2A) and z-score plot (Figure 2B). The heatmap shows heterogeneity in overall metabolome data across urine samples from di fferent phenotypes. In the z-score plot, stable-based z-scores were plotted for each of the 266 metabolites. The plots revealed robust metabolic alterations in AR (z-score range: −4.2 to 800.5) and IFTA (z-score range: −3.8 to 265.4) compared to fewer changes in BKVN samples (z-score range: −3.4 to 116.9).

**Figure 1.** Sample selection and study schematic of the study. Summary outlining study samples, assay platform, study phenotypes, analysis, and results.


**Table 1.** Patient demographic data for the discovery cohort.

\* Age in years: mean ± SD (median; range). AR, acute rejection; STA, stable graft function; IFTA, interstitial fibrosis and tubular atrophy; BKVN, BK virus nephropathy. # Estimated glomerular filtration rate (eGFR) data were unavailable for BKVN samples. Immunosuppression consisted of Tacrolimus and Mycophenolate Mofetil for all patients, with maintenance steroids for those on steroid-based immunosuppression. All patients received IL2R monoclonal antibody (Daclizumab) induction; steroid-based patients received this for 2 months and steroid-free patients received this for 6 months. Most patients were unsensitized and recipients for first allografts, with 4 repeat transplants. Of the 106 AR, 29 were ABMR. The clinical data variables that were significantly different between groups were assessed for any statistical association with their impact on metabolite signatures of different transplant phenotypes and were not found to be significant.

**Figure 2.** Metabolomic profiling of renal transplant outcomes. (**A**) Heat map representation of unsupervised hierarchical clustering by metabolite (rows) grouped by transplant phenotype (columns). Shades of black to red to orange to yellow represent continuous increases of a metabolite relative to the median metabolite levels (see color scale). (**B**) z-score plots for the data in a normalized to the mean of the stable phenotype urine samples (truncated at 25 s.d. for clarity).

#### *3.2. Metabolite Marker Panel for Alloimmune Injury*

Applying the VSURF method, a panel of 9 metabolites (Table 2) were selected out of 266 to accurately classify post-Tx alloimmune injury, combining the output from samples with either acute or chronic alloimmune injury (AR/IFTA) versus stable (STA) samples. The resulting model had a 95% accuracy of correctly discriminating between the two outcome groups (AUC = 0.950, sensitivity = 95.3%, specificity = 75.9%). This lower specificity is likely due to within group heterogeneity between AR and IFTA phenotypes. The 9 metabolite VSURF model was nearly identical in accuracy to the full 266-metabolite model, which had an AUC of 0.954. This difference in AUC values was not significant using DeLong's test (*p* = 0.731), meaning there is no significant change in classification accuracy between the full and abbreviated metabolite models (Figure 3A). This suggests that no diagnostic accuracy is lost in using the abbreviated metabolite model.


*J. Clin. Med.* **2020**, *9*, 2341

**Figure 3.** Identification of potential biomarker panel of metabolites for KTx alloimmune injury and acute rejection using VSURF method. (**A**) Two receiver operating characteristic (ROC) curves representing classification accuracies and a statistical comparison of the full and sparse RF models for alloimmune injury and the table displaying classification accuracy on the test set. The metabolites in the panel are listed on the right-hand side (**B**) Two ROC curves representing classification accuracies and a statistical comparison of the full and sparse Random Forests (RF) models for acute rejection (AR) injury and the bottom table displaying classification accuracy on the test set. The metabolites in the panel are listed on the right-hand side.

#### *3.3. Metabolite Marker Panel for Acute Rejection*

In order to identify a metabolite marker panel specific to acute rejection of KTx, we applied VSURF exclusively to the AR and STA urine metabolome datasets (*n* = 217). The resulting model contained 11 metabolites (Table 2) for AR detection. The ROC analysis resulted with an AUC of 0.985 with 92.9% sensitivity and 96.3% specificity (Figure 3B). Individual distributions for the three most significant metabolites, glycine, *N*-methylalanine, and inulobiose, are presented in the form of bean plots (Figure 4).

**Figure 4.** Significantly altered metabolites in AR versus STA. Bean plots demonstrating distribution of the 3 most significant metabolites in AR comparing to STA. The bold horizontal line represents mean value for each group.

#### *3.4. Metabolite Marker Panel for BK Virus Nephritis*

In order to identify BKVN-specific metabolites, we used VSURF on 22 BKVN urine and 288 non-BKVN urine that included AR, IFTA, and STA urine. The resulting VSURF panel contained 5

metabolites, Arabinose, 2-hydroxy-2-methylbutanoic acid, hypoxanthine, benzyl alcohol, and *N*-acetyld-mannosamine (Table 2) for BKVN classification with 72.7% sensitivity and 96.2% specificity (Table S1). When we confined our analysis to only BKVN vs. STA, VSURF resulted in a panel of 4 metabolites, arabinose, 2-hydroxy-2-methylbutanoic acid, octadecanol, and phosphate. For this panel, BKVN classification was 88.9% sensitive and 94.8% specific (Table S2). The 4-metabolite VSURF model had accuracy comparable to that of the full 266-metabolite model, which had a sensitivity of 87.5% and specificity of 93.2% (Table S3).

#### *3.5. Metabolic Pathways Associated with Graft Injury*

To explore metabolite significance by both statistical significance and magnitude of fold change in the injury group, a volcano plot with Random Forests (RF) importance score was generated (Figure 5A) that shows the relative importance of the metabolite in terms of RF score for AR-specific panel. Additionally, a volcano plot with fold changes (increased or decreased) and corresponding *p*-values displayed the significance of the various metabolites in AR (Figure S1). The plot reveals metabolites of increasing significance relative to the Random Forests classification model. Some metabolites from the 9-metabolite marker panel for alloimmune injury and the 11-metabolite marker panel for AR are among the very highly perturbed metabolites. The metabolites significantly perturbed in KTx injury with *p*-value < 0.001 (*n* = 42) were analyzed for metabolic pathway enrichment with MetaboAnalyst. Pathway analysis for enrichment identified nitrogen metabolism, ascorbate, and aldarate metabolism, and amino sugar and nucleotide sugar metabolism as the three most significantly enriched pathways (Figure 5B).

**Figure 5.** Metabolites and pathways significantly perturbed in KTx alloimmune injury. ( **A**) Volcano plot displaying fold change and significance of metabolites. Red dots denote metabolites significant at a Random Forests importance score greater than 0.5. The right half displays metabolites in the injury group with a higher signature relative to the stable group. Some metabolites from the 9-metabolite marker panel for alloimmune injury and the 11-metabolite marker panel for AR are among the very highly perturbed metabolites labeled in red dots. (**B**) Enrichment analysis of metabolic pathways using significantly altered metabolites showed enrichment in nitrogen metabolism (*p* = 0.0055), ascorbate and aldarate metabolism (*p* = 0.0083), and amino sugar and nucleotide sugar metabolism (*p* = 0.05) as significantly enriched pathways. The y-axis represents the *p*-values as the negative of their natural logarithm.
