The Assessment of immunologic Risk and Tolerance in Kidney Transplantation (ARTKT-1) study was a cross-sectional sample collection study, which enrolled 543 KTx patients at Kyung Hee University Hospital at Gangdong from 2012 to 2015, and five different hospitals (Kyung Hee Medical Center, Kyungpook National University Hospital, Samsung Medical Center, Seoul St. Mary’s Hospital of the Catholic University of Korea, and Inje University Busan Paik Hospital) from 2013 to 2015. We used the Modification of Diet in Renal Disease (MDRD) equation to estimate the glomerular filtration rate (GFR). Indication biopsy was taken upon any clinical indication, but protocol biopsy was performed at 2 weeks (KHG and KMC) or 3 months (CMC) based on the institution’s policy, and 457 biopsy-proven samples classified by pathologists at each hospital who were unaware of the analyzed results of the samples using their clinical status and Banff 2007 classification (10) and 86 non-biopsied patients were enrolled.

In 457 biopsy-proven samples, there were normal pathology (NP, n=119), acute T cell-mediated rejection (TCMR, n=75), borderline changes (BC, n=58), acute antibody-mediated rejection (ABMR, n=26), chronic active antibody-mediated rejection (cABMR, n=25), BK virus nephropathy (BKVN, n=15), acute tubular necrosis (ATN, n=30), calcineurin inhibitor toxicity (CNI, n=28), glomerulonephritis (GN, n=27), interstitial fibrosis/tubular atrophy (IF/TA, n=12), and others (n=42). In addition, we enrolled 79 non-biopsied patients defined as long-term graft survival (LTGS) with maintaining stable kidney function while on maintenance immunosuppression more than 10 years after KT and 7 non-biopsied patients defined as operational tolerance (OT) with maintaining stable kidney function (serum creatinine < 1.5 mg/dl) without any immunosuppressive drugs for more than 1 year. We excluded 42 biopsy-proven samples with acute pyelonephritis, acute TI nephritis, diabetic nephropathy, hydronephrosis, hypertensive glomerulopathy, microvascular injury, non-specific injury, unknown, and 7 OT samples because of small scale, less frequent findings, or unclearly defined specimens for pathological classification in this study.

At the time of transplantation, none of the transplant donors were from a vulnerable population and all donors or next of kin provided written informed consent that was freely given. All kidney transplant patients provided written informed consent to participate in the study, and the study was approved by the local institutional review board at Kyung Hee University Hospital at Gangdong (#2012-030, KHNMC) and registered with the Clinical Research Information Service (KCT0001010).

To select AR candidate genes, we searched the keyword “kidney rejection” in the GEO database. Among the initially queried results, we filtered further by species (Homo Sapiens) and data sets using the gene expression platform (Affymetrix GeneChip U133+2) to minimize the unexpected bias by platform difference. We manually examined the top 10 data sets in terms of the sample sizes and excluded sets that did not have acute rejection and stable samples. We found the significantly different genes between STA and AR using the GeneMeta R package in four data sets ( Supplemental Table 1 ), following the approach of Choi et al. (11). In total, 137 probe-sets corresponding to 109 unique genes were found to be significant across multiple data sets (FDR<0.01 and FC>2) by false discovery rates (FDRs) obtained from 1,000 permutations. The effective fold change of the meta-analysis was calculated as the average fold change of the four data sets weighted by the number of samples. Among them, we selected the top 10 genes in the order of fold changes with the clarity of gene annotation excluding multiple probes mapped to a gene or a probe mapped to multiple genes and with the easiness of PCR primer design ( Supplemental Table 2 ).

In addition, we selected representative genes (CD3ϵ, Foxp3, OX40, and Tim-3) that are well known as diagnostic markers for allograft rejection through literature review ( Supplemental Table 3 ).

Urine (about 50 ml) in each center were collected at the time of biopsy (for biopsied patients) or visiting the hospital (for patients with long-term survival graft). We immediately centrifuged and stored samples in each center’s laboratory after collection. All centers followed an identical protocol with urine collection at the clinical sites for the sedimentation of urinary cells. After the urine was centrifuged at 2,000x g for 20 min, the pellet was transferred into RNAlater and stored at -80°C until use.

Total RNA was extracted from the urinary pellets using a PureLink RNA Mini Kit (Invitrogen) according to the manufacturer’s recommendations. The quantity (absorbance at 260) and purity (the absorbance at 260/280) of RNA were measured using a NanoDrop^® ND-2000 UV spectrophotometer (Thermo Scientific). The median (25^th and 75^th percentile) quantity (µg) of total RNA was 0.74 (0.250-1.680), and the median (25^th and 75^th percentile) purity of RNA was 1.85 (0.50-2.01) in 494 total samples. Also, we assessed the quality of urinary RNA using 18S rRNA ≥ 1x10⁴ copies/ug and TGF-β1 ≥ 1x10² copies/ug as quality control (QC) parameters for the improvement of data quality before the measurement of urinary mRNA levels. In 494 total samples, 402 samples (81%) passed QC ( Supplemental Table 4 ).

Reverse transcription was performed with total RNA using M-MLV RT enzyme (200 U/µl; Mbiotech, Inc., Seoul, Korea), and absolute quantities of the mRNAs were measured by TaqMan probe qPCR assays without pre-amplification step. Each DNA oligo serially diluted from 1x10^-1 to 1x10^-8 ng/ul was used for a standard curve. The copy number of each mRNA was calculated using the molecular weight of DNA and the standard curve. The undetected value of mRNA was calculated by replacing 40 C_t value in the blank for data analysis. The mRNA values were then normalized by 18S rRNA copies (x10^-6) used as endogenous control and log₁₀-transformed to reduce deviation before being used in data analysis (9).

We performed in situ hybridization (ISH) assay for CXCL9 expression in renal biopsy tissues (3 NP, 3 acute TCMR, and 3 acute ABMR) using the RNAscope 2.5 assay kit (Advanced Cell Diagnostics, Hayward, CA, USA) according to the manufacturer’s recommendations. Briefly, formalin-fixed paraffin embedded sections were subjected to deparaffinization, proteolytic digestion with enzyme denaturation, and hybridization with a CXCL9 probe. The RNAscope target was retrieved at 95°C for 15 min and then incubated with RNAscope enzyme for 15 min. The hybridization of the probes for CXCL9, human peptidylprolyl isomerase B (positive control), and diaminopimelate B (negative control) were incubated at 40°C for 2 hours. The preamplifier, signal enhancer, amplifier, and label probe were sequentially incubated with samples at 40°C for 30, 15, 30, and 15 min, respectively. After each step, sections were washed two times, and hybridization signals were detected by 3,3-diaminobenzidine (DAB) staining, followed by counterstaining with Gill’s hematoxylin.

We compared the levels of transcripts between the STA and AR groups by the non-parametric Mann-Whitney test and assessed differences among the STA, AR, BKVN, and OGIs groups by the non-parametric Kruskal-Wallis test using SPSS for Windows, version 20.0 (IBM Corp, Armonk, NY).

We developed a model to predict AR by a binary logistic regression with all significantly altered genes (12). To diagnose AR using a model, we used a decision cutoff point maximizing Youden’s index, which is the sum of sensitivity and specificity (13). Receiver operating characteristic (ROC) curve analyses and decision curve analyses were conducted to evaluate the model (14, 15).

We supplemented LTGS samples for a small number of normal pathology group as STA and split into a training set of AR (n=58) and STA (n=45) groups randomly selected by a clearly defined pathology. We first measured 14 genes in 103 training samples for gene selection by absolute qPCR method without pre-amplification, and then developed an AR model by statistical analysis. In the next step, we measured 6 genes to verify the discriminative power of the model in 391 independent samples, which consist of STA (n=153), AR (n=68), borderline changes (BC, n=58) for TCMR, BKVN (n=15), and other graft injuries (OGIs, n=97) including acute tubular necrosis (ATN, n=30), calcineurin inhibitor (CNI, n=28) toxicity, glomerulonephritis (GN, n=27), and interstitial fibrosis/tubular atrophy (IF/TA, n=12). We assessed the AR prediction model with the fixed cutoff point and evaluated whether the decision curve analysis of the signature is better than biopsy for clinical management and diagnosis of AR. The work-flow chart of this study is shown in Figure 1 .

There was no significant difference in the mean age of patients and maintenance immunosuppression between the groups. The duration after KT was significantly longer in the STA group than others because this group contained the patients with long-term good graft function, whose follow-up periods after KT were over 10 years. STA group showed a lower proportion of deceased donor KT and low HLA mismatching numbers than other groups. Residual graft function was significantly high in the STA group compared to others in both the training and validation set. The baseline clinical characteristics of the study population are summarized in Table 1 .

We measured the expression levels of 14 genes using absolute qPCR without the pre-amplification step in the training set. If mRNA is undetected, the absolute value of the mRNA was calculated with the standard curve by replacing 40 C_t value in the blank with undetected value, and each mRNA level was log₁₀-transformed after normalization with 18S rRNA copies (x10^-6). The levels of CXCL9, C1QB, LCK, CD3ϵ, Foxp3, Tim-3 (P<0.001 for each mRNA), IP-10 (P<0.01), PSMB9, and FAM26F (P<0.05) were significantly elevated in the AR group compared to STA, but for OX40, IDO1, ISG20, vWF, and PTPRC there was no difference ( Figure 2A and Supplemental Table 5 ). The copy number of 18S rRNA per total RNA amount showed no difference between the AR and STA groups. Although LCK, Foxp3, and FAM26F mRNAs were statistically significant, these mRNAs were not detected in 14% (n=12), 26% (n=22), and 58% (n=49) of the QC-passed samples (n=84), respectively. Thus, we excluded these mRNAs (FAM26F, Foxp3, and LCK) due to low detectable frequency and (OX40, IDO1, ISG20, vWF, and PTPRC) with no statistical difference between AR and STA for further analysis. In addition, we investigated representatively CXCL9 level using in situ hybridization (IHS) analysis to confirm whether the mRNA level in kidney biopsy tissue is consistent with the result of urinary mRNA by qPCR assay. CXCL9 was distinctly expressed in the damaged tubules in the kidney allografts of acute TCMR and predominantly in the peritubular capillary area in ABMR groups, consistent with the results of qPCR analysis ( Figure 2B ).

We performed ROC curve analysis to evaluate each target for AR diagnosis, and the AUC values of the individual gene were not sufficient to distinguish AR from STA ( Supplemental Table 6 ). Therefore, to improve the diagnostic accuracy of AR, we developed an AR prediction model using a binary logistic regression in the training set.

The level of each mRNA was log₁₀-transformed after each mRNA copy number was normalized with 18S rRNA (x10^-6). The AUC value of the signature was 0.89 (95% CI, 0.82-0.96; P<0.001) ( Figure 3A ). The equation of a six-gene signature is shown below:

With the cutoff point (0.40889), the predicted probability of the signature also yielded 86% accuracy, 91% sensitivity, 80% specificity, 83% positive predictive value (PPV), and 89% negative predictive value (NPV) ( Figure 3B and Table 2 ). The signature improved the diagnosis of AR from STA compared to Tim-3 alone.

We measured six genes in 391 independent samples including STA, AR, BC, BKVN, and OGIs groups, and evaluated the discriminative power of the diagnostic signature. All of the 6 genes were significantly increased in the BKVN group as well as the AR group compared to STA by Kruskal-Wallis nonparametric test and Dunn’s post-hoc test for multiple comparisons. Interestingly, the expression of CXCL9 and IP-10 genes (P<0.01) in BKVN was significantly higher rather than those in the AR group. Compared to STA group, 4 genes (Tim-3, P < 0.05; CXCL9, P < 0.01; CD3ϵ, P < 0.001; C1QB, P < 0.05) in the BC group and 5 genes (Tim-3, P < 0.001; CXCL9, P < 0.01; CD3ϵ, P < 0.05; C1QB, P < 0.001; PSMB9, P < 0.001) in the OGIs group were significantly up-expressed. In addition, the expression levels of all six genes in the AR group were significantly higher than those in the BC and OGIs groups. Scatter plots of the mRNAs and the comparisons between groups by Mann-Whitney t test are shown in Supplemental Table 7 and Supplemental Figure 1 .

To evaluate the performance of the AR predicted signature with the fixed cut-off point, we tested its performance in AR (n=57) and STA (n=122) in the QC-passed samples. BC group was not included in the AR group because there are significant differences in the six genes between the AR and BC groups. The AUC value of the signature yielded 0.84 (95% CI, 0.78-0.90; P<0.001) ( Figure 4A ), and the predicted probability of the signature showed well discrimination of AR from STA, but it was difficult to discriminate BC (n=42) from AR or STA in the validation set ( Figure 4B ). With the fixed cutoff point, the signature displayed 70% sensitivity (95% CI, 57%-82%) and 80% specificity (95% CI, 72%-87%). The overall performance of the signature yielded 77% accuracy, 63% PPV, and 85% NPV ( Table 3 ). Furthermore, we assessed the potential of the signature for distinguishing the AR from the no-AR group (STA and OGIs, n=207) in the QC-passed samples excluding the BC and BKVN groups. The signature yielded an AUC value of 0.78 (95% CI, 0.72-0.85; P<0.001) with 70% sensitivity (95% CI, 57%-82%), 72% specificity (95% CI, 65%-78%), 77% accuracy, 63% PPV, and 85% NPV ( Figures 4C,D and Table 3 ). In distinguishing the AR from OGIs group (OGIs, n=85), the signature yielded an AUC value of 0.70 (95% CI, 0.62-0.79; P<0.001) with 70% sensitivity (95% CI, 57%-82%), 60% specificity (95% CI, 49%-71%), 64% accuracy, 54% PPV, and 75% NPV ( Table 3 ).

Using decision curve analysis, we assessed whether the analysis of the signature was better than biopsy for the clinical management and diagnosis of AR ( Figure 5 ). Based on the decision curve analysis, the range of reasonable threshold probabilities (p_t ) with the highest net benefit was from 0.2 to 0.5 for the diagnostic signature. Ultimately, the diagnostic signature within a reasonable threshold probability may complement or justify allograft biopsy for diagnosing AR after kidney transplantation.