Molecular Diagnosis of Antibody-Mediated Rejection in Human Kidney Transplants
Article first published online: 15 FEB 2013
© Copyright 2013 The American Society of Transplantation and the American Society of Transplant Surgeons
American Journal of Transplantation
Volume 13, Issue 4, pages 971–983, April 2013
How to Cite
Sellarés, J., Reeve, J., Loupy, A., Mengel, M., Sis, B., Skene, A., de Freitas, D. G., Kreepala, C., Hidalgo, L. G., Famulski, K. S. and Halloran, P. F. (2013), Molecular Diagnosis of Antibody-Mediated Rejection in Human Kidney Transplants. American Journal of Transplantation, 13: 971–983. doi: 10.1111/ajt.12150
- Issue published online: 30 MAR 2013
- Article first published online: 15 FEB 2013
- Manuscript Accepted: 11 DEC 2012
- Manuscript Revised: 6 DEC 2012
- Manuscript Received: 13 SEP 2012
- University of Illinois for providing biopsies
- Novartis Pharma AG
- Genome Canada, the University Of Alberta Hospital Foundation
- Roche Molecular Systems
- Hoffmann-La Roche Canada Ltd.
- Alberta Ministry of Advanced Education and Technology
- Roche Organ Transplant Research Foundation, and Astellas
- Austin Health, Melbourne, Australia
- Antibody-mediated rejection;
- kidney transplant;
- molecular diagnostics
Antibody-mediated rejection is the major cause of kidney transplant failure, but the histology-based diagnostic system misses most cases due to its requirement for C4d positivity. We hypothesized that gene expression data could be used to test biopsies for the presence of antibody-mediated rejection. To develop a molecular test, we prospectively assigned diagnoses, including C4d-negative antibody-mediated rejection, to 403 indication biopsies from 315 patients, based on histology (microcirculation lesions) and donor-specific HLA antibody. We then used microarray data to develop classifiers that assigned antibody-mediated rejection scores to each biopsy. The transcripts distinguishing antibody-mediated rejection from other conditions were mostly expressed in endothelial cells or NK cells, or were IFNG-inducible. The scores correlated with the presence of microcirculation lesions and donor-specific antibody. Of 45 biopsies with scores >0.5, 39 had been diagnosed as antibody-mediated rejection on the basis of histology and donor-specific antibody. High scores were also associated with unanimity among pathologists that antibody-mediated rejection was present. The molecular score also strongly predicted future graft loss in Cox regression analysis. We conclude that microarray assessment of gene expression can assign a probability of ABMR to transplant biopsies without knowledge of HLA antibody status, histology, or C4d staining, and predicts future failure.
donor-specific HLA antibody
panel reactive HLA antibody
T cell-mediated rejection
Antibody-mediated rejection (ABMR) has emerged as a key problem in kidney transplantation and a major cause of late graft loss [1-3]. The potential of antibodies against MHC to reject allografts was demonstrated by Peter Gorer, who described MHC antibodies and showed that they were cytotoxic and could reject mouse tumor allografts [4-6]. The clinical importance of antibody in organ transplantation became apparent with the description of hyperacute rejection of kidney transplants in recipients with preformed HLA antibodies [7-9]. Posttransplant donor-specific HLA antibody (DSA) was identified as a risk factor for late graft loss , but recognition of the clinical features of ABMR as distinct from T cell-mediated rejection (TCMR) came only with the discovery of microcirculation lesions (peritubular capillaritis (ptc) and glomerulitis (g)) in early kidney allografts in sensitized recipients [11-13]. With the demonstration of complement factor C4d in the capillaries [14-17], three features were accepted by the Banff consensus as diagnostic criteria for ABMR: microcirculation lesion, C4d staining and DSA [18-20]. These criteria led to the realization that ABMR often emerges many years after transplantation [21, 22] and leads to the development of transplant glomerulopathy with double contours (cg) and late graft failure.
The diagnosis of ABMR is problematic because of intrinsic limitations of the key diagnostic features, and because ABMR is heterogeneous and has a remarkable dynamic range, from fulminant to inactive. Microcirculation lesions have limited specificity, occurring in other diseases such as acute kidney injury (AKI), glomerulonephritis (GN) and TCMR , and like histopathology lesions in general have limited reproducibility between pathologists [24-26]. The diagnosis is heavily dependent on HLA antibody assessment, which varies between measurement platforms, reagent batches and laboratories, with uncertainty about cutoffs for positivity and quantification [27, 28]. C4d staining is performed using differing methods, giving disparate results . Most importantly, recent biopsy studies have shown that the majority of late ABMR cases are C4d negative and are missed by current Banff criteria [2, 3, 30], often resulting in incorrect diagnoses such as calcineurin inhibitor nephrotoxicity. The recognition of C4d negative ABMR revealed the true extent of ABMR in late kidney transplant deterioration , and highlighted the need for quantifiable and reproducible testing. Unfortunately, pathologists have not been able to agree on new criteria for ABMR that include C4d-negative ABMR .
The present study was based on the hypothesis that molecular changes in biopsies could provide a new dimension for assessing the probability of ABMR, adding to the existing diagnostic tests. In other areas of medicine, particularly oncology, molecular analysis of the biopsy is extending the information available to clinicians beyond what is available from traditional approaches [32-35]. The development and validation of new tests often faces the problem of a flawed gold standard. Without an independent method for assigning diagnoses, new methods will almost always look worse than the existing gold standard, even when they are closer to the “truth”. However, methods for dealing with this “reference set bias” are emerging [36-38]. The potential of molecular phenotyping to aid in the diagnosis of ABMR is supported by the demonstration that kidneys with ABMR by histology-DSA criteria often have increased expression of interferon-γ (IFNG)-induced transcripts , endothelial transcripts reflecting microcirculation stress  and natural killer (NK) cell transcripts [41, 43]. Thus in the present study we used microarrays to characterize gene expression patterns in a prospectively collected set of indication biopsies that had been phenotyped by histology and HLA antibody testing. Our goal was to develop a molecular classifier for estimating the probability of ABMR in each biopsy, similar to the classifier we recently described for TCMR .
Patients and sample collection
This study was approved by the institutional review board of the University of Alberta (issue # 5299), the University of Illinois (Protocol # 2006-0544), the University of Minnesota (Protocol #HSC#0606 M87646) and the Hennepin County Medical Center (Protocol #HSR#06-2670). Written informed consent was obtained from all study patients. All consenting patients undergoing a biopsy for clinical indications (deterioration in function, proteinuria, or stable impaired function) between September 2004 and November 2008 were included. HLA antibody testing was performed on blood drawn on or as close as possible to the day of biopsy. Biopsies were processed for microarray analysis as described previously . Briefly, one 18-gauge biopsy core was placed immediately in RNALater and stored at −20 °C. RNA extraction, labeling and hybridization to the HG_U133_Plus_2.0 GeneChip (Affymetrix, Santa Clara, CA, USA) were carried out according to the manufacturer's protocols. Microarrays were scanned using GeneArrayScanner (Affymetrix) and processed with GeneChip Operating Software Version 1.4.0 (Affymetrix).
Diagnosis of ABMR using histology and DSA assessment
Histology was assessed as previously described . Paraffin sections were processed according to Banff criteria . C4d staining was performed by indirect immunofluorescence on frozen sections using a monoclonal anti C4d antibody (Quidel, San Diego, CA, USA). Diffuse linear C4d staining in >50% of peritubular capillaries were interpreted as positive. Biopsies were classified using a new Reference Standard classification based on the Banff classification  but including C4d negative ABMR. Occasional ambiguous biopsies with microcirculation lesions and HLA antibody that was not demonstrably donor specific were designated possible ABMR. The criteria for the diagnosis of C4d-negative ABMR were DSA positive, nondiffuse C4d staining (negative or minimal focal or focal), and any of the following microcirculation lesions: peritubular capillaritis (ptc>1), glomerulitis (g>0), thromboses, or transplant glomerulopathy (cg>0). Possible ABMR was defined by the same criteria except that the HLA antibody was not demonstrably donor-specific (NDSA). Biopsies diagnosed as “no major abnormalities” were defined as having minimal interstitial fibrosis (ci-score <) and no features of a disease process . Acute kidney injury (AKI) was diagnosed in biopsies before 42 days posttransplant that lacked histologic criteria of rejection or disease . “Atrophy-fibrosis”—interstitial fibrosis and tubular atrophy (IFTA) not otherwise specified—was defined by ci-score >1 and no features of a disease process . The 20 biopsies classified as “others” included C4d staining with no pathology (n = 5), thrombotic microangiopathy (n = 2), suspicious viral changes (n = 4), transplant glomerulopathy (n = 4), posttransplant lymphoproliferative disorder (n = 1), oxalosis (n = 1), obstruction (n = 1), acute pyelonephritis (n = 1) and tubulo-interstitial nephritis (n = 1). Eleven biopsies did not meet the Banff criteria for assessment of all lesions, but were judged by our pathologists as suitable for diagnosis of ABMR and TCMR, and were included in the study.
As previously described , two pathologists, A and B, scored the lesions in all 403 biopsies. Pathologist B also assigned diagnoses in all 403 biopsies.
The rules outlined above were used to assign diagnoses from lesion and HLA antibody data, forming the Reference Standard histology-DSA diagnoses that were used to train and test the classifier. The Reference Standard diagnoses were derived from pathologist B. However, pathologist B did not use Banff rules to assign i- and t-lesions, so the scores of pathologist A were used for those lesions.
A third pathologist read 245 of the 403 biopsies to permit an estimate of the agreement of the molecular classifier with the agreement among three pathologists. These readings were not used for the Reference Standard.
HLA antibody screening
Of 403 biopsies, serum for HLA antibody testing was available in 363 at the time of biopsy. HLA antibody testing was recommended by protocol but the costs were borne by each center, whose investigators decided whether to perform HLA antibody testing on an individual basis. The anti-HLA antibody testing method varied depending on the center, as did the criteria for assigning positivity. All decisions regarding HLA testing and evaluation were done without knowledge of the molecular phenotype. Screening for HLA antibodies at the time of biopsy was performed by ELISA (n = 7), CDC-AHG panels (n = 15), FlowPRA® screen (One Lambda Canoga Park, CA, USA) (n = 318), or Luminex single antigen bead testing (n = 15). Of 124 DSA positive samples, antibody specificities were determined or confirmed by Luminex single antigen beads (n = 88), FlowPRA® single antigen I and II beads (n = 32), or positive CDC-AHG crossmatch (n = 4). Specificities to Cw were not assigned as DSA. Donor HLA typing for DP and DQA1 was not performed and therefore DSA were not attributed to DP or DQA1.
Robust Multiarray Averaging (RMA), as implemented in the Bioconductor package “RefPlus”, was used to normalize the data. Nonspecific interquartile range (IQR) filtering across the 403 samples resulted in 12,637 of the 54,675 probe sets being retained for analysis. “R” version 2.12.1 (64-bit) was used for all analyses. Microarray expression files are posted on the Gene Expression Omnibus website under the accession number GSE36059.
As previously described for a TCMR classifier , we used linear discriminant analysis (LDA) to create the classifiers, implemented with the “classification” function of the “CMA” library of “R”. The top 20 probe sets were used, based on a Welch's t-test comparing C4d positive and C4d negative ABMR to all other cases (including mixed rejection, TCMR and possible ABMR). The classifier output is a score between 0.0 and 1.0, reflecting the probability that a biopsy is ABMR. For the purpose of assessing agreement with the histology-DSA diagnoses using statistics such as accuracy, sensitivity, etc., a cutoff was imposed. We assigned biopsies above a score of 0.2 as molecular ABMR, and those below as molecular non-ABMR. This cutoff was chosen because it resulted in ∼90% specificity for diagnosing histology-DSA ABMR.
The 403 biopsies (plus 8 nephrectomies) were split randomly into 10 approximately equal sized groups (folds). One fold was left out, while the other nine were used to build the classifier. This was then used to predict the probability of ABMR for the biopsies in the left out fold. This was repeated for all 10 folds, resulting in a single predicted score for each of the 403 biopsies. This entire procedure was then repeated a further 99 times, using different random splits of the data into 10 folds, so that every biopsy was finally assigned 100 ABMR scores (Figure 1). This method is known as repeated k-fold cross-validation (k = 10 in this case, with 100 repeats). Although single k-fold cross-validation is more common, different random splits of the data into 10 folds often produce quite different results, and the single split chosen to report may not be representative. By generating repeated k-fold estimates, variation due to the random splits is averaged out, producing more stable and reliable results [50-53].
All aspects of classifier training, including gene selection, were done from scratch within each training set. No information from any left out test sample was used to build the corresponding training set derived equations. Eight nephrectomies were evaluated when in test sets, but were ignored when they occurred in training sets, and were not used to build any of the classifiers.
Graft survival analysis
Death-censored graft survival was analyzed using Cox regression. Since our 403 biopsies come from 315 patients, including multiple biopsies per patient in such analyses would be suspect. We therefore included only one biopsy per patient, chosen at random. Although the analyses we report are based on the first randomly generated set of 315 biopsies, examination of additional randomizations showed that the results were always qualitatively the same, producing the same conclusions, and differing little quantitatively.
Demographics and case mix of the population
We studied 403 prospectively collected indication biopsies obtained 6 days to 35 years posttransplant from 315 consenting kidney transplant recipients. Biopsies were collected between late 2004 and 2008 (Table 1). Patients with ABMR by histology–DSA were younger, and their primary disease was more often glomerulonephritis and less often diabetic nephropathy.
|Patient demographics (n = 315)||All patients (n = 315)||Histology-DSA ABMR (n = 56)||All others (n = 259)||p-Value (ABMR vs. all others)|
|Mean recipient age (years)||43 (5–81)||38 (13–69)||45 (5–81)||0.002|
|Recipient gender (% male) (n = 315)||196 (62%)||31 (55%)||165 (64%)||0.01|
|Race (n = 315)|
|Caucasian||201 (64%)||39 (70%)||162 (63%)||0.32|
|Black||33 (10%)||3 (5%)||30 (12%)||0.23|
|Other||81 (26%)||14 (25%)||67 (26%)||0.89|
|Primary disease (n = 315)|
|Diabetic nephropathy||64 (20%)||5 (9%)||59 (23%)||0.02|
|Hypertension/large vessel disease||29 (9%)||3 (5%)||26 (10%)||0.44|
|Glomerulonephritis/vasculitis||119 (38%)||29 (52%)||90 (35%)||0.02|
|Interstitial nephritis/pyelonephritis||20 (6%)||2 (4%)||18 (7%)||0.55|
|Polycystic kidney disease||46 (15%)||9 (16%)||37 (14%)||0.73|
|Others||15 (5%)||4 (7%)||11 (4%)||0.32|
|Unknown etiology||22 (7%)||4 (7%)||18 (7%)||1.00|
|Mean donor age (years)||40 (2–70)||37 (3–64)||41 (2–70)||0.06|
|Donor gender (% male) (n = 263)||121 (46%)||20 (43%)||101 (47%)||0.70|
|Donor type (% deceased donor transplants) (n = 310)||152 (49%)||25 (47%)||127 (49%)||0.77|
|Patients with HLA antibody testing at time of biopsy||275 (87%)||56 (100%)||219 (85%)||<0.001|
|All biopsies (n = 403)||Biopsies taken ABMR (n = 65)||Biopsies taken All others (n = 338)||p-Value (ABMR vs. all others)|
|Clinical characteristics at time of biopsy (n = 403)|
|Median and range time from transplant to biopsy (months)||21 (0.2–428)||69 (0.2–382)||21 (0.25–428)||<0.001|
|Indication for biopsy|
|Primary nonfunction||10 (2%)||0 (0%)||10 (3%)||0.38|
|Rapid deterioration of graft function||96 (24%)||24 (37%)||72 (21%)||0.01|
|Slow deterioration of graft function||150 (37%)||23 (35%)||127 (38%)||0.74|
|Stable impaired graft function||71 (18%)||1 (2%)||70 (21%)||<0.001|
|Investigate proteinuria||38 (9%)||9 (14%)||29 (9%)||0.18|
|Follow-up from previous biopsy||14 (3%)||4 (6%)||10 (3%)||0.26|
|Others||9 (2%)||1 (2%)||8 (2%)||1.00|
|Indication unknown||15 (4%)||3 (5%)||12 (4%)||0.72|
|Biopsies with HLA antibody testing at time of biopsy||362 (90%)||65 (100%)||297 (88%)||0.001|
|Maintenance immunosuppressive regimens at biopsy|
|MMF, tacrolimus, steroid||176 (44%)||27 (42%)||149 (44%)||0.94|
|MMF cyclosporine, steroid||101 (25%)||15 (26%)||83 (25%)||0.95|
|Others||126 (31%)||20 (31%)||106 (31%)||0.89|
We diagnosed ABMR using the previously published Reference Standard classification , as outlined in the Methods section. The diagnosis was based on microcirculation lesions and DSA, incorporating proposed changes to the Banff classification, including C4d-negative ABMR [2, 3] and possible ABMR. The diagnoses assigned (Table 2) included 65 ABMR (17 C4d-positive), 10 possible ABMR (all C4d-negative) and 22 mixed ABMR and TCMR (nine C4d positive). Almost all biopsies with ABMR and mixed rejection presented late, i.e. more than 1 year after transplantation, as previously reported for this unselected biopsy population . A parallel study of a cohort of early ABMR in presensitized patients with positive cross-matches is in progress.
|ABMR or mixed ABMR and TCMR (N = 87)a||C4d+ABMRb||17||3||14|
|Possible ABMR (all C4d negative)c||10||1||9|
|Other diagnoses (N = 306)||Borderline||42||29||13|
|No major abnormalitiese||72||33||39|
Because this study included indication biopsies taken up to 35 years posttransplant, the tissue typing data on the donor are sometimes incomplete. However, such patients are common in the transplant population, and must be included in a prospective study of unselected indication biopsies.
The classifier and transcript usage
We used linear discriminant analysis to assign ABMR scores to all biopsies. Repeated 10-fold cross-validation (see the Materials and Methods section) was used to generate 100 scores for each biopsy. We use the median of these 100 scores when reporting results in Tables and Figures, similar to the approach used to create a TCMR classifier .
The 100 iterations × 10-folds represent 1000 class comparisons between histology-DSA ABMR and all other conditions, with each comparison selecting the top 20 differentially expressed probe sets by Welch's t-test. Only 73 probe sets were ever selected in these 1000 top 20 probe set lists.
Table 3 shows the 30 probe sets most frequently selected in the 1000 class comparisons, and the cell types that express each transcript, based on previous annotations or the literature. A single, final, LDA model using all 403 biopsies was built. The 20 probe sets selected by this classifier (using t-test p-values based on all 403 samples) were the same as the 20 most frequently selected probe sets in the cross-validation loops. The linear discriminant scores in the final model for the 20 probe sets are shown in Table 3.
|Mean expression levelb|
|Affy ID||Gene name||Gene symbol||Probable cellular expressiona||Histology-DSA ABMR (n = 65)||Other (n = 338)||Normal kidneys (n = 8)||Number of classifiers using probe set||LD1c|
|211122_s_at||Chemokine (C-X-C motif) ligand 11||CXCL11||IFNG-induced||232||70||13||1000||0.087|
|208335_s_at||Duffy blood group, chemokine receptor||DARC||EC||226||111||44||1000||0.391|
|219584_at||Phospholipase A1 member A||PLA1A||EC, EP||317||188||135||1000||0.392|
|226028_at||Roundabout homolog 4||ROBO4||EC||244||186||160||1000||−0.018|
|221841_s_at||Kruppel−like factor 4 (gut)||KLF4||EC||279||184||276||999||0.086|
|228489_at||Transmembrane 4 L six family member 18||TM4SF18||EC, EP||240||183||178||999||0.524|
|204115_at||Guanine nucleotide binding protein (G protein), gamma 11||GNG11||EC, EP||940||734||430||970||−0.416|
|206210_s_at||Cholesteryl ester transfer protein, plasma||CETPc||(cell type unknown)||25||34||33||958||−0.754|
|226303_at||Phosphoglucomutase 5||PGM5||(cell type unknown)||189||141||114||953||0.265|
|203065_s_at||Caveolin 1||CAV1||EC, EP||366||241||140||879||0.052|
|205898_at||Chemokine (C−X3−C motif) receptor 1||CX3CR1||NK||318||191||127||719||0.059|
|223869_at||Sclerostin||SOSTc||(cell type unknown)||56||101||228||648||−0.112|
|228339_at||Endothelial cell−specific chemotaxis regulator||ECSCR||EC||153||118||82||645||−0.109|
|227188_at||Chromosome 21 open reading frame 63||C21orf63||EC||82||62||44||547||−0.387|
|204533_at||Chemokine (C−X−C motif) ligand 10||CXCL10||IFNG−induced||772||339||63||486||−0.178|
|204677_at||Cadherin 5, type 2||CDH5||EC||154||120||89||455||−0.048|
|205984_at||Corticotropin releasing hormone binding protein||CRHBPc||(cell type unknown)||131||200||322||424||−0.161|
|210873_x_at||Apolipoprotein B mRNA editing enzyme, catalytic polypeptide−like 3A||APOBEC3A||Mac||31||21||15||356||0.074|
|223836_at||Fibroblast growth factor binding protein 2||FGFBP2||NK, T||25||18||14||313||—|
|204103_at||Chemokine (C−C motif) ligand 4||CCL4||NK, T, Mac||195||116||124||302||—|
|228698_at||SRY (sex determining region Y)−box 7||SOX7||EC||87||66||70||283||—|
|205326_at||Receptor (G protein−coupled) activity modifying protein 3||RAMP3||(cell type unknown)||376||312||288||243||—|
|205619_s_at||Mesenchyme homeobox 1||MEOX1||EC||23||17||14||190||—|
|213830_at||Tcell receptor delta locus||TRD@||NK||52||42||32||163||—|
|202112_at||Von Willebrand factor||VWF||EC||490||313||124||127||—|
|204681_s_at||Rap guanine nucleotide exchange factor (GEF) 5||RAPGEF5||EC||315||239||215||87||—|
A manuscript assessing the cellular basis of each probe set in cell lines and experimental systems is in preparation, and the molecular features are only briefly discussed here. Most of the 10 probe sets selected more than 900 times (i.e. in > 90% of class comparisons) were expressed in endothelial cells: cadherin 13 (CDH13); Duffy blood group antigen (DARC) ; phospholipase A1 member A (PLA1A); roundabout homolog 4 (ROBO4); transmembrane 4 L six family member 18 (TM4SF18); and guanine nucleotide binding protein (G protein), gamma 11 (GNG11). Kruppel-like factor 4 (KLF4), a transcription factor expressed in endothelium, is involved in protection against atherogenesis . Other commonly selected probe sets included chemokine CXCL11, which is highly induced by IFNG, and several transcripts with expression in NK cells, e.g. CX3CR1, FGFBP2 and GNLY, confirming previous reports [39-41].
Relationship between the classifier scores and the histology-DSA diagnosis of ABMR
Figure 2 shows the median ABMR score for each biopsy. Biopsies are grouped on the x-axis according to their Reference Standard histology-DSA diagnoses, but are randomly arranged within each diagnostic category. Eight histologically normal kidney specimens are included for comparison.
To calculate accuracy statistics comparing ABMR scores with the Reference Standard histology-DSA diagnoses, scores were considered positive for ABMR above a cut-off of 0.2. The positive scores were compared with histology-DSA diagnosis of ABMR (C4d+, C4d−, and mixed). The statistics are as follows: accuracy 85%; sensitivity 67%; specificity 90%; positive predictive value (PPV) 64%; negative predictive value (NPV) 91%. The area under the receiver operating characteristic curve (AUC) was 0.89.
Relating the ABMR scores to histology-DSA diagnoses in late biopsies
The relationship between the ABMR score and histology-DSA diagnoses of ABMR in late biopsies is summarized in Table 4. When the score was >0.5, 39/45 (87%) had been assigned a diagnosis of ABMR or mixed rejection by histology-DSA features. Of the six exceptions, three were “possible ABMR” (ABMR histology lesions and NDSA without demonstrable DSA), and three (one with DSA and two with NDSA) had extensive atrophy-fibrosis, which could have prevented the detection of ABMR histologic lesions.
|Detailed diagnosis in non–ABMR|
|ABMR score||ABMR (C4d+ve and −ve)||Mixed||Non- BMR||Possible ABMR||TCMR||Border- line||BK||GN||Acute kidney injury||No major abnormalities||Atrophy- fibrosis||Other||Total|
Most late biopsies with ABMR scores <0.2 had been assessed as non-ABMR by histology-DSA (120/143 = 84%). Five of the nine ambiguous “possible ABMR” biopsies had scores <0.2 and one was between 0.2 and 0.5.
We examined the relationship of focal C4d staining to the ABMR score. All biopsies were classified as C4d positive or C4d negative, including focal as negative. Among biopsies that were classified as C4d negative ABMR, six had focal C4d staining: their ABMR scores were not different from those with no focal C4d staining. Among biopsies classified with diagnoses other than ABMR, 10 had focal C4d staining: their ABMR scores were similar to other non-ABMR biopsies with no focal C4d staining.
Figure 3 shows a calibration curve for the ABMR scores. The solid line represents the observed frequency of histology-DSA ABMR/mixed diagnoses corresponding to sliding windows of ABMR scores along the x-axis. A perfectly calibrated curve would fall on the diagonal line. For example, of the biopsies in a window centered on an ABMR score of 0.4, approximately 0.43 (slightly above the ideal 1:1 line) were diagnosed as histology-DSA ABMR/mixed. Throughout the range, the ABMR score was close to the observed frequency of the histology-DSA diagnosis of ABMR.
Association between ABMR scores and histologic lesions
The strength of association between the dichotomized ABMR score and the lesion scores is shown in Table 5. Because of the large number of tied values for lesion scores, the gamma statistic for ordinal data  was used. The ABMR score was strongly associated with the microcirculation lesions of ABMR (gamma scores of 0.77 to 0.81) and moderately associated with atrophy-scarring (gamma scores of 0.50 and 0.52), as well as time posttransplant. The defining TCMR lesion, tubulitis, was not significantly increased in biopsies with high ABMR scores.
|ABMR score > 0.2 (n = 90)||ABMR score <0.2 (n = 313)||Gamma statisticb|
|Median time of biopsy posttransplant (months)||63||10||0.47***|
|Mean histologic lesion scores|
|ABMR-related lesions||Peritubular capillaritis: ptc-score||1.41||0.23||0.81***|
|Transplant glomerulopathy: cg-score||1.15||0.18||0.80***|
|TCMR-related lesions||Interstitial inflammation i-score||0.6||0.43||0.21*|
|TCMR/ABMR-related lesion||Intimal arteritis v-score||0.16||0.10||0.29|
|Atrophy-scarring-related lesions||Interstitial fibrosis: ci-score||1.70||1.08||0.50***|
|Tubular atrophy: ct-score||1.72||1.15||0.52***|
|Arterial fibrous intimal thickening: cv-score||1.38||1.11||0.23**|
|Arteriolar hyalinosis: ah-score||1.41||0.88||0.36***|
Association between ABMR scores and donor-specific antibodies
In this prospective study, centers were asked to perform HLA antibody testing at their own expense but in some cases elected not to do this. HLA antibody measurements at the time of biopsy were performed in 359 biopsies (Table 6): 124 (34%) were DSA positive; 70 (19%) had HLA antibody that was not demonstrably donor-specific (NDSA); and 165 (45%) had no HLA antibody (PRA negative). HLA antibody testing was not requested in 44 biopsies.
|ABMR score||DSA positive (class II or I/II)||HLA antibody not demonstrably donor specific (NDSA)||HLA antibody negative||Not doneb||Total|
|> 0.5||40 (35)||5||0||0||45|
|0.2 - 0.5||23 (20)||4||14||4||45|
|< 0.2||61 (40)||61||151||40||313|
ABMR scores were strongly associated with DSA positivity (p<0.001), mostly anticlass II (88%). All 45 biopsies with scores >0.5 had circulating HLA antibody, either DSA (n = 40) or NDSA (n = 5).
In 44 cases, the clinicians elected not to request HLA antibody assessment because they judged that ABMR was unlikely. The mean ABMR scores in these biopsies were lower than for those that were tested (0.07 vs. 0.17, two-tailed Mann–Whitney p-value = 0.006), and the distribution of ABMR scores was similar to that of the HLA antibody-negative biopsies, with no high scores, supporting the clinical impression.
Agreement between ABMR scores and pathologist concordance
In a subset of 245 biopsies read by three pathologists, the ABMR score was strongly associated with the independent determination by all three pathologists that the diagnostic ABMR microcirculation lesions were present (ptc, g, cg, as outlined in the Methods section) (Figure 4). The mean ABMR score was 0.49 in biopsies where all three pathologists agreed the criteria were met and 0.09 when all three agreed they were not met. Thus ABMR scores reflect the likelihood of unanimity among pathologists.
Relating the ABMR score to graft survival
Table 7 shows p-values and hazard ratios for predicting death-censored graft survival in our population, using the ABMR score and various definitions of histologic ABMR. When combined with the most significant univariable histologic definition of ABMR (C4d+/−/mixed) in multivariable analysis, the ABMR score retained its significance. Thus the molecular score is an independent predictor of graft survival that performs at least as well as histologic definitions of ABMR.
|Molecular ABMR scoreb||1.50||1.76||2.06||5×10−12|
|Histology-DSA ABMR (C4d+/-/mixed)||2.33||3.71||5.90||3×10−8|
|Histology-DSA C4d+ ABMR(both alone and mixed)||1.2||2.4||4.7||0.01|
|Histology-DSA C4d- ABMR (both alone and mixed)||2.0||3.3||5.5||2×10−6|
|Molecular ABMR scoreb||1.21||1.53||1.91||2.6×10−4|
|Histology-DSA ABMR (C4d+/-/mixed)||0.93||1.79||3.43||0.08|
To develop a molecular test for ABMR, we first assigned histology-DSA diagnoses to the biopsies, including both C4d positive and negative ABMR. Classifiers were generated to assign ABMR scores to each biopsy using the molecular features that best distinguished ABMR from all other diseases present in our population of indication biopsies. The study design thus included TCMR and many diseases in the comparator group, allowing the algorithm for classifier generation to distinguish ABMR from TCMR and from the general features of kidney damage. The scores correlated with the microcirculation lesions of ABMR and with DSA: every biopsy with an ABMR score >0.5 had DSA or NDSA, and 87% had been assigned histology-DSA diagnoses of ABMR. High scores were also associated with unanimity among pathologists regarding the presence of ABMR, and predicted future graft loss at least as well as any histology-DSA definition of ABMR. Thus the ABMR score provides a new tool for assessing the probability of ABMR without prior knowledge of histology, C4d, or DSA.
The ABMR score primarily detects microcirculation abnormalities, not simply fibrosis, as shown by the observation that most biopsies diagnosed as “atrophy-fibrosis” in Figure 2 have low ABMR scores. The tight association of scarring with chronic ABMR is well known , and almost all of the ABMR cases in the present study are late and have scarring. However, we have some new unpublished biopsies from sensitized patients with early ABMR that show positive scores, which should help to define the separation of the ABMR score from fibrosis .
The 17 biopsies classified as ABMR by histology-DSA but with low ABMR scores pose an issue that merits continuing study. Discrepancies are not unexpected, given the unresolved disagreements in the histology-DSA diagnostic system , but nevertheless must eventually be explained. These biopsies may represent various combinations of false positive histology; false positive DSA; false negative molecular scores; and genuine heterogeneity in the disease, i.e. histologic and/or DSA features with little molecular activity. Indeed, heterogeneity in disease states is accepted in histologic lesions (e.g. some ABMR cases have ptc- but no g- or cg-scores, etc.) and in DSA measurements (e.g. class I versus class II), and should be anticipated in molecular tests as well. Heterogeneity is probably intrinsic in the ABMR states, based on antibody specificity (antigen targets), disease activity (which could be intermittent) and disease duration. The effect of treatment on late ABMR has yet to be determined, and this requires a dedicated prospective study in which the treatments are defined. There is no consensus on what constitutes a treatment, and there are many potential treatments, none standardized. Above all, the discrepancies illustrate the necessity for ongoing calibration and validation, as is expected with any new molecular test. Thus we have undertaken the INTERCOM prospective trial now in progress (clinical trials.gov NCT01299168).
Moreover, despite the large number of biopsies in this study, infrequent situations such as ABO incompatibility and thrombotic microangiopathy will have to be resolved in new studies. For example, in the 20 “other” biopsies, both of the thrombotic microangiopathies were C4d negative and DSA negative, and had very low ABMR scores, but the possibility that some renal microcirculation diseases could give false positive scores must be kept in mind.
The molecules most frequently selected by the class comparisons as distinguishing ABMR from all other conditions support the mechanistic model we previously proposed involving the interaction of NK cells with DSA-coated endothelium and IFNG release [41, 43]. The top genes—some selected in each of the 1000 class comparisons—include many endothelial transcripts (e.g. DARC, VWF, ROBO4), probably reflecting remodeling of the microcirculation endothelium in response to injury. The NK transcripts may reflect NK activation via CD16a Fc receptors engaging Fc portions of DSA bound to endothelium , triggering IFNG production and inducing transcripts such as CXCL11. The possibility that NK cells are engaged in antibody-dependent cellular cytotoxicity is supported by the presence of cytotoxicity-related transcripts, e.g. granulysin (GNLY)  and FGFBP2 . However, these changes may not be exclusively deleterious: some endothelial changes may reflect the ability of the microcirculation to “accommodate” to offset the impact of the antibody stress [61, 62].
The ABMR classifiers generated here will be useful for assigning ABMR scores to new biopsies, probably in an integrated diagnostic system in combination with histology and DSA. The detailed applications must await decisions on whether this will be provided as a central service, which has been the case to date with multigene molecular diagnostic tests in cancer, or as tests in individual laboratories. The ABMR score has particular promise for identifying and understanding problematic cases, for supporting multicenter therapeutic trials, and as an independent measurement that can guide the refinement and calibration of histologic lesions and DSA testing. The ABMR score will probably reveal that ABMR identified by histology-DSA is heterogeneous, which is expected as we progress toward personalized medicine . Finally, the fact that the ABMR score strikingly predicted risk for progression to failure raises the possibility that it will prove useful for predicting or monitoring response to therapy.
Based on experience with other new tests, it is likely that in practice molecular testing will present its own set of technical and logistic problems that will need to be addressed before such testing is fully developed. Nevertheless, such challenges can be addressed, and will be helped in particular by a continued clinical focus on the discrepancies and on the outcomes in individual patients.
Special thanks to Jessica Chang for statistical analyses; Vido Ramassar and Anna Hutton for technical support; and to Dr. Arthur Matas from the University of Minnesota and Dr. Bruce Kaplan from University of Illinois for providing biopsies.
This research has been supported by funding and/or resources from Novartis Pharma AG, and in the past by Genome Canada, the University Of Alberta Hospital Foundation, Roche Molecular Systems, Hoffmann-La Roche Canada Ltd., the Alberta Ministry of Advanced Education and Technology, the Roche Organ Transplant Research Foundation, and Astellas. Dr. Skene received sabbatical funding from Austin Health, Melbourne, Australia. Dr. Halloran held a Canada Research Chair in Transplant Immunology until 2008 and currently holds the Muttart Chair in Clinical Immunology.
The authors of this manuscript have conflicts of interest to disclose as described by the American Journal of Transplantation. P. F. H. holds shares in Transcriptome Sciences Inc., a company with an interest in molecular diagnostics. The other authors have no competing financial interests.
- 4The detection of antigenic differences in mouse erythrocytes by the employment of immune sera. Br J Exp Pathol 1936; 17: 42–50..
- 6The cytotoxic activity of isoantibodies in mice. Transplant Bull 1956; 3: 142–143., .
- 35The 70-gene signature as prognostic factor for elderly women with hormone receptor-positive, HER2-negative breast cancer. Breast Care 2012; 7: 19–24., , , , .
- 36Fool's gold: Why imperfect reference tests are undermining the evaluation of novel diagnostics: A reevaluation of 5 diagnostic tests for leptospirosis. Clin Infect Dis 2012; 55: 322–331., , , et al.
- 44Molecular diagnosis of T cell-mediated rejection in human kidney transplant biopsies. Am J Transplant In press 2012., , , et al.
- 50Clinical Prediction Models: A Practical Approach to Development, Validation, and Updating. 1st ed. New York: Springer Science+Business Media; 2009..
- 51Data Mining with R: Learning with Case Studies. 1st Ed. Voca Raton, FL: CRC Press, 2010..
- 53Estimating classification error rate: Repeated cross-validation, repeated hold-out and bootstrap. Comput Stat Data Anal 2009; 53: 3735–3745..
- 56Categorical Data Analysis. 2nd Ed. New York: John Wiley & Sons; 2002..
- 58Determinants of progression to transplant glomerulopathy in ABMR: A combined clinical-molecular approach. Am J Transplant 2012; 12: 44. Ref Type: Abstract., , , , .
- 63Preparing for precision medicine. N Engl J Med 2012; 366: 489–491., , .