GENOMICS AND BIOMARKERS

J. Reeve¹, J. Sellarés², M. Mengel³, B. Sis¹, L.G. Hidalgo¹, K.S. Famulski³, P. Halloran^2
1Laboratory Medicine And Pathology, University of Alberta, Edmonton/CANADA, ²Department Of Medicine, University of Alberta, Edmonton/AB/CANADA, ³Department Of Laboratory Medicine And Pathology, University of Alberta, Edmonton/AB/CANADA

Body: Assigning a probability of TCMR to kidney transplant biopsies.
Jeff Reeve, Joana Sellares, Banu Sis, Michael Mengel, Luis G. Hidalgo, Konrad S. Famulski, and Phil Halloran

INTRODUCTION: The current gold standard for diagnosing T cell mediated rejection (TCMR) in kidney transplants is the Banff system. This standard is imperfect, therefore an independent diagnostic method using molecular analysis is desirable. Several groups have used machine learning-based classifiers on microarray data to diagnose rejection. The “extreme phenotype” method (EP) has most often been used. This compares strongly rejecting TCMR samples with clearly non-rejecting kidneys. The rationale behind this design is based on the assumptions that 1) given an imperfect gold standard, the samples should all be correctly labelled with respect to their rejection status; 2) analyses using samples at the extremes of the disease spectrum will be better at detecting genes truly informative in terms of the underlying pathology. The alternative, “full-spectrum” (FS) method - using consecutive, unselected biopsies for cause presenting at the study site’s clinic - will necessarily include many misdiagnosed samples. In addition, the difference between the TCMR and non-TCMR samples will be less clear-cut, and the analysis may find less informative genes. While these assumptions are intuitively appealing, they have never been examined in the microarray literature to the best of our knowledge.
METHODS: 403 consecutive biopsies for cause, 63 with Banff TCMR, were used to build and assess two classifiers to predict the probability of TCMR. The first (EP) compared Banff TCMR that were strongly rejecting with biopsies that seemed to be clearly non-rejecting, both groups defined by histological and molecular criteria. The second (FS) compared all 63 Banff TCMR with all 340 other samples. Results based on Monte Carlo cross-validation are presented: all steps of classifier development were done from scratch within each training set for each of the 100 validation iterations. The trainingset:testset ratio was 2:1.
RESULTS: The FS classifier was able to find greater separation between TCMRs and other diseases than was the EP classifier (Figure 1). Ninety-seven genes were used at least once over the 100 FS iterations, whereas only 5 genes were used at least once in the EP classifier.
CONCLUSIONS: Our analysis suggests that the EP method is less able to separate TCMR from other diagnoses, notably ABMR, than is the FS method. Extreme samples are easy to separate accurately, but since no information from “difficult” intermediate samples is available, the classifier is unable to choose a good threshold at which the distinction between TCMR and non-TCMR should be made when faced with a clinically realistic test set.

Disclosure: All authors have declared no conflicts of interest.