A peripheral blood diagnostic test for acute rejection in renal transplantation

Abstract

Monitoring of renal graft status through peripheral blood (PB) rather than invasive biopsy is important as it will lessen the risk of infection and other stresses, while reducing the costs of rejection diagnosis. Blood gene biomarker panels were discovered by microarrays at a single center and subsequently validated and cross-validated by QPCR in the NIH SNSO1 randomized study from 12 US pediatric transplant programs. A total of 367 unique human PB samples, each paired with a graft biopsy for centralized, blinded phenotype classification, were analyzed (115 acute rejection (AR), 180 stable and 72 other causes of graft injury). Of the differentially expressed genes by microarray, Q-PCR analysis of a five gene-set (DUSP1, PBEF1, PSEN1, MAPK9 and NKTR) classified AR with high accuracy. A logistic regression model was built on independent training-set (n = 47) and validated on independent test-set (n = 198)samples, discriminating AR from STA with 91% sensitivity and 94% specificity and AR from all other non-AR phenotypes with 91% sensitivity and 90% specificity. The 5-gene set can diagnose AR potentially avoiding the need for invasive renal biopsy. These data support the conduct of a prospective study to validate the clinical predictive utility of this diagnostic tool.

© Copyright 2012 The American Society of Transplantation and the American Society of Transplant Surgeons.

Figures

Figure 1. Summary of Study Design

The gene-based biomarker discovery pipeline for an AR blood test follows a path of a) discovery by microarrays across 3 different platforms across a defined set (n=103) of AR and STA blood samples; followed by b) verification (n=34) and validation (n=47) on independent AR and STA blood samples; and c) a finally, prediction of AR (n=198) in other varying phenotypes of graft injury likely to be encountered in an outpatient clinical setting. Array data generated from the 3 platforms were compared by mapping the transcripts to Entrez Gene identifiers. Common genes regulated significantly in AR on each platform were identified using a common significance threshold (SAM; FDR

Figure 2. Single–center Verification and Validation of Gene Expression for the 5-Gene Set

Box plots of the QPCR gene expression values are shown for the selected 5 genes: DUSP1, PBEF1 And PSEN1 are upregulated in AR (red outline); NKTR and MAPK9 are downregulated in AR (green outline) in the single center Verification Set (n=34; Figure 2A) and in the single center independent Training Set 1 (n=47; Figure 2B), for building the logistic regression model on the 5 gene-set. We applied logistic regression with best subset selection to the Verification Set in order to find the minimum number of genes necessary for the proper classification of biopsy-confirmed AR. Chi-square score for logistic regression models built using these 10 genes showed that increase in the score was minimal when more than five genes were used in the model. Chi-square score for logistic regression models built using all 10 genes showed that the increase in Chi-square score from a model with 1 gene to 3 genes is 7.70; from a model with 3 genes to 5 genes is 1.87; and from a model with 5 genes to a model with 6 is only an increase of 0.48. Hence, the logistic regression model using a set of 5 genes was selected based on the best performing 5-genes set (Chi-square score = 29.63) as DUSP1, PBEF1, PSEN1, MAPK9, and NKTR. The p values for comparison of gene expression data for each gene are shown in each dataset and each value is significant (p

Figure 3. Multi-Center Validation of the QPCR Prediction Probability for AR by the 5-Gene Set

A dot plot is shown for individual percent probability prediction score for AR on 198 independent samples over the course of the 3 year follow-up (time post-transplant in months on the Axis) in the SNSO1 multicenter study. Each blood sample is paired with a biopsy for blinded, centralized, histological diagnosis of the phenotype. Based on the Receiver Operating Characteristic (ROC) curve for the logistic regression model across DUSP1, PBEF1, PSEN1, MAPK9, and NKTR, a cutoff of θ = 0.52 was selected to have the best sensitivity and specificity to discriminate between AR and STA. In other words, the prediction probability has been derived from the logistic regression model across the 5 genes (Y Axis) and percent probability prediction score of >52% predicts the sample to have an AR phenotype. In Figure 3A the 32 AR samples are shown by red dots, with 3 misclassifications (91% accuracy within class); the 94 STA samples are shown by green dots with 6 misclassifications (92% accuracy). The ROC curve for AR vs STA class is shown in Figure 3B. In Figure 3C the 72 nonAR/nonSTA samples are shown, divided into 4 categories: 12 AR borderline (pink dots), 37 CAN (light blue dots), 16 CNIT (cyan dots) and 7 other diagnoses such as reflux nephropathy (n=2), BK nephropathy (n=1), FSGS recurrence (n=1) (dark blue dots) and ischemia (n=3). Within the AR borderline class 4 samples have <50% prediction scores for AR and misclassify, giving the within class accuracy of 67% (8/12 samples) for borderline AR. The ROC curve for AR vs nonAR/nonSTA class is shown in Figure 3D.