Abstract

Background

   Antibody-mediated rejection, a leading cause of renal allograft graft
   failure, is diagnosed by histological assessment of invasive allograft
   biopsies. Accurate non-invasive biomarkers are not available.

Methods

   In the multicentre, prospective BIOMARGIN study, blood samples were
   prospectively collected at time of renal allograft biopsies between
   June 2011 and August 2016 and analyzed in three phases. The discovery
   and derivation phases of the study (N = 117 and N = 183 respectively)
   followed a case-control design and included whole genome
   transcriptomics and targeted mRNA expression analysis to construct and
   lock a multigene model. The primary end point was the diagnostic
   accuracy of the locked multigene assay for antibody-mediated rejection
   in a third validation cohort of serially collected blood samples
   (N = 387). This trial is registered with [55]ClinicalTrials.gov, number
   [56]NCT02832661.

Findings

   We identified and locked an 8-gene assay (CXCL10, FCGR1A, FCGR1B, GBP1,
   GBP4, IL15, KLRC1, TIMP1) in blood samples from the discovery and
   derivation phases for discrimination between cases with (N = 49) and
   without (N = 134) antibody-mediated rejection. In the validation
   cohort, this 8-gene assay discriminated between cases with (N = 41) and
   without antibody-mediated rejection (N = 346) with good diagnostic
   accuracy (ROC AUC 79·9%; 95% CI 72·6 to 87·2, p < 0·0001). The
   diagnostic accuracy of the 8-gene assay was retained both at time of
   stable graft function and of graft dysfunction, within the first year
   and also later after transplantation. The 8-gene assay is correlated
   with microvascular inflammation and transplant glomerulopathy, but not
   with the histological lesions of T-cell mediated rejection.

Interpretation

   We identified and validated a novel 8-gene expression assay that can be
   used for non-invasive diagnosis of antibody-mediated rejection.

Funding

   The Seventh Framework Programme (FP7) of the European Commission.

   Keywords: Kidney transplantation, Antibody-mediated rejection,
   Biomarker
     __________________________________________________________________

Research in context.

Evidence before this study

   Despite the use of powerful immunosuppression after kidney
   transplantation, allograft rejection, and most importantly
   antibody-mediated rejection, remains a strong predictor of graft
   failure. The diagnosis of antibody-mediated rejection is made on kidney
   allograft biopsies performed at the time of decline in glomerular
   filtration rate or appearance of proteinuria. Because antibody-mediated
   rejection can occur in the absence of immediate clinical signs or
   changes in these graft functional characteristics, some centers also
   perform kidney allograft biopsies at fixed time points (surveillance
   biopsies). Glomerular filtration rate and proteinuria are non-specific
   markers for antibody-mediated rejection, as many other immunological
   and non-immunological injuries can disturb graft function. More
   accurate non-invasive diagnostic markers are thus needed, with better
   sensitivity and specificity for antibody-mediated rejection than
   estimated glomerular filtration rate and proteinuria. We have
   systematically followed up and reviewed the scientific literature
   related to the development of non-invasive biomarkers for
   antibody-mediated rejection of kidney allografts. We focused on PubMed
   and Scopus for papers that report on non-invasive diagnostic markers
   for antibody-mediated kidney allograft rejection. We found only few
   independently validated non-invasive peripheral blood biomarkers for
   acute rejection. None of these markers were further developed or
   specific for diagnosis of antibody-mediated rejection.

Added value of this study

   In this study, we identified and independently validated a novel
   non-invasive biomarker for antibody-mediated kidney allograft rejection
   in peripheral blood. To our knowledge, this is the first study to
   report on a biomarker with sufficiently good diagnostic accuracy for
   antibody-mediated rejection to be useful for clinical care of kidney
   transplant recipients. The diagnostic accuracy of the 8-gene assay for
   antibody-mediated rejection was superior to that of classical clinical
   indicators, and the assay offered additional benefit in clinical
   decision-making to perform or not to perform a biopsy for diagnosis of
   antibody-mediated rejection. We were able to show that the diagnostic
   accuracy of the 8-gene assay was retained in different clinical
   scenarios: early and late after transplantation, at time of stable
   graft function and at time of graft dysfunction.

Implications of all the available evidence

   Our results indicate that kidney transplant recipients suffering from
   antibody-mediated rejection can be identified by our newly developed
   non-invasive 8-gene assay. The 8-gene assay can serve as a clinical
   decision aid whether or not to perform an invasive biopsy for
   confirmation of antibody-mediated rejection at times of clinical
   suspicion. In routine follow-up of stable kidney transplant patients,
   this biomarker shows potential for more timely detection of
   often-missed subclinical antibody-mediated rejection and prompt
   initiation of targeted therapies after confirmative histological
   diagnosis, before chronic damage has developed and before the disease
   process becomes irreversible.

   Alt-text: Unlabelled Box

1. Introduction

   Antibody-mediated rejection is recognized as a primary cause of graft
   failure after kidney transplantation [[57][1], [58][2], [59][3]]. It is
   hallmarked histologically by inflammation and C4d deposition in
   peritubular capillaries, glomerulitis, intimal arteritis and
   expansion/duplication of the glomerular basement membrane [[60]4].

   Currently, the diagnosis of antibody-mediated rejection after kidney
   transplantation is made based on histological assessment of invasive
   kidney biopsies according to the regularly updated Banff international
   consensus [[61]4]. Antibody-mediated rejection can be diagnosed in
   clinically indicated biopsies at time of graft functional problems
   (rise in serum creatinine or proteinuria), but can also occur
   subclinically, without changes in these graft functional parameters.
   Subclinical antibody-mediated rejection also associates with increased
   risk of graft failure [[62]5] but often remains undetected, unless
   protocol-specified (surveillance) kidney biopsies are performed
   [[63]2,[64]6]. Such protocol-specified biopsies are routinely performed
   in some centres, but not all, at varying time after transplantation.

   Based on the association between antibody-mediated rejection and kidney
   graft failure, and the impossibility to repeatedly perform invasive
   protocol-specified biopsies, non-invasive diagnostic markers are needed
   with better sensitivity and specificity than eGFR and proteinuria
   [[65]2,[66]5,[67]7]. Other groups have suggested non-invasive markers
   for antibody-mediated rejection, primarily assessed in urine samples
   [[68][8], [69][9], [70][10], [71][11], [72][12], [73][13]]. Additional
   validation of these markers is necessary to support their potential
   clinical value [[74][14], [75][15], [76][16]].

   Kidney allograft rejection is associated with molecular changes in
   renal allograft tissue, which reflect transcription changes in resident
   cells (e.g. interferon-gamma inducible changes in the donor
   endothelium) or changes in cell populations, like infiltration and
   activation of effector T cells and macrophages in T- cell mediated
   rejection or margination and activation of natural killer cells in
   antibody-mediated rejection [[77]17,[78]18]. As these graft
   infiltrating cells are activated primarily in lymphoid organs before
   travelling and infiltrating the allograft [[79]19], we hypothesized
   that the molecular changes that occur in renal allograft biopsies with
   antibody-mediated rejection could also be reflected by changes in
   circulating immune cells.

   Given the lack of non-invasive, sufficiently validated markers for
   antibody-mediated rejection, we aimed to develop and validate an
   mRNA-based gene set in peripheral blood that is able to non-invasively
   rule out or detect ongoing antibody-mediated rejection after kidney
   transplantation.

2. Methods

2.1. Study design, patient population and sample collection

   This part of the BIOMArkers of Renal Graft INjuries (BIOMARGIN) study
   ([80]www.biomargin.eu) is a multicentre, prospective, multiphase study,
   performed in four European transplant centres (Hôpital Necker Paris,
   France; University Hospitals Leuven, Belgium; Medizinische Hochschule
   Hannover, Germany; and Centre Hospitalier Universitaire Limoges,
   France). Samples were prospectively and consecutively collected at time
   of renal allograft biopsies, between June 2011 and August 2016. In the
   four clinical centres, protocol renal allograft biopsies were performed
   at 3, 12, and sometimes at 24 months after transplantation, according
   to local centre practice, in addition to clinically indicated biopsies
   (biopsies at time of graft dysfunction). All adult patients who had
   received a single kidney allograft at these institutions and who
   provided written informed consent, were eligible. Recipients of
   combined transplantations were excluded. All transplantations were
   performed with negative complement-dependent cytotoxicity
   cross-matches. Institutional review boards and national regulatory
   agencies (when required) approved the study protocol at each clinical
   centre.

   The study was divided in three phases. In the discovery phase, blood
   and biopsy samples were used for genome-wide expression analysis. We
   selected samples based on availability and histological criteria of
   concomitant renal allograft biopsies (excluding cases with diagnosis of
   glomerulonephritis or polyomavirus nephropathy, and cases with unclear
   diagnosis). Based on local biopsy readings, a first selection was made,
   which was then further refined by judgment of the clinical courses and
   final confirmation by central pathology, independent from the original
   centre (see the Supplementary Appendix). The same study design was used
   for the second derivation phase, for targeted validation of the results
   obtained in the discovery set, and derivation of the multigene marker.
   In the validation cohort, all samples with concomitant adequate renal
   allograft biopsy histology, prospectively collected according to the
   BIOMARGIN protocol between June 24, 2014 and July 2, 2015 were serially
   included. In this cohort, no selection was made on histology,
   demographics, time or any other factor than sample availability. In
   this validation cohort, the analyses were entirely based on central
   pathology scores.

2.2. Primary and secondary end points

   The primary end point was the diagnostic accuracy of a multigene marker
   for antibody-mediated rejection in the validation cohort in relation to
   the current gold standard of tissue pathology, based on the area under
   the receiver operating characteristic curve (ROC AUC). Secondary
   endpoints were the diagnostic accuracy in specific clinical situations
   (at time of graft dysfunction leading clinicians to perform an
   indication biopsy versus at time of stable graft function, early
   (<1 year) versus later (>1 year) after transplantation), and net
   benefit for clinical decision-making.

2.3. Sample collection and biopsy scoring

   Peripheral blood samples were collected at time of the renal allograft
   biopsies, directly in PAXgene Blood RNA tubes® (PreAnalytiX GmbH, a
   Qiagen/ BD Company, Switzerland). Two needle cores were taken at each
   kidney allograft biopsy. One was used for histology, at least half of
   the other one was immediately stored in Allprotect Tissue Reagent®
   (Qiagen Benelux BV, Venlo, The Netherlands) for RNA expression analysis
   (in the discovery set). All biopsies were rescored semiquantitatively
   according to the updated Banff 2017 classification [[81]4]. Index test
   results were not available to assessors of the histopathological
   scoring (reference standard), whereas reference standard results and
   clinical information were available to the readers of the index test.

   In the discovery cohort, RNA extracted from blood and biopsies was
   hybridized onto Affymetrix GeneChip Human Genome U133 Plus 2.0 Arrays
   (Affymetrix Inc., High Wycombe HP10 0HH, UK). In the derivation and
   validation cohorts, RNA expression analysis of mRNA extracted from
   blood samples was evaluated by real-time polymerase chain reaction
   (RT-PCR) using OpenArray® technology on the Quantstudio™ 12 K Flex
   Real-Time PCR System (Life Technologies Europe BV, Ghent, Belgium) with
   ACTB, GAPDH and SDHA as endogenous controls. Details on blood and
   biopsy sample collection and gene expression analysis are provided in
   the Supplementary Appendix.

2.4. Statistical analysis

   In the discovery phase, robust-multiarray average-normalized mRNA
   expression data of the 117 peripheral blood samples and 95 biopsies
   were analyzed in a statistical pipeline developed under the R framework
   in an extension of the biosigner R package as developed for this study
   [[82]20], with addition of Elastic-Net and Shrunken Centroids
   multivariate methods to the Sparse Partial Least Squares (SPLS), Random
   Forrest and Support Vector Machines-Recursive Feature Elimination
   (SVM-RFE) multivariate methods already available in the biosigner
   package. More information on the constructed statistical pipeline and
   determination of a multivariate score for antibody-mediated rejection
   (ABMR score) and T-cell mediated rejection (TCMR score) is given in the
   Supplementary Appendix. A multivariate score > 0·25 was considered as
   specific for antibody-mediated and/or T-cell mediated rejection.
   Ingenuity Pathway Analysis (IPA, Build: 478438 M Content version:
   44691306) was used for canonical pathway enrichment analysis.

   In the derivation phase, we identified the multivariate combination of
   transcripts that lead to the best model performance, based on the
   extended list of transcripts obtained in the discovery phase. This
   identification of the multigene signature was done by ranking a
   combination of genes according to the C-statistic of logistic
   regression models trained on this combination and estimated under a
   3-folds cross validation. Instead of identifying the best combination
   as the final multigene signature, we integrated the combinations
   obtained by the top K models (see Supplementary Appendix). The best
   multigene signature was then used to build a multivariable logistic
   regression model in a nested-cross validation approach on the
   derivation cohort. The ensuing logistic regression model (intercept and
   estimations) was then locked and represented the final multigene assay.

   The diagnostic accuracy of the locked multigene signature and logistic
   regression model calculated in the derivation phase was then evaluated
   on the validation cohort. We used receiver operating characteristic
   (ROC) curves to evaluate the C-statistic (area under the curve, AUC) of
   the multigene assay. The optimal marker threshold from the derivation
   phase, at the highest combination of sensitivity and specificity (which
   are independent of disease prevalence), was then evaluated in this
   validation cohort. In addition, arbitrarily defined low and high
   thresholds with respectively high negative and positive predictive
   values (which are dependent on actual disease prevalence) were
   evaluated in the validation cohort.

   Finally, post-hoc sensitivity analyses were performed to evaluate the
   accuracy of the marker in specific clinical situations. The net benefit
   of the 8-gene marker for clinical decision-making was evaluated using
   decision curve analysis [[83]21]. To allow for comparison and assess
   added value of the 8-gene assay compared with clinical parameters
   alone, we built a clinical model using the 8 clinical parameters that
   were differently prevalent in the ABMR vs. no AMBR group. This model
   was built and cross-validated in the discovery and derivation phases
   combined and then assessed in the independent validation phase. The
   diagnostic performances of the clinical model, the 8-gene assay and an
   integrated model (adding the 8-gene assay to the clinical model) were
   assessed and compared using ROC CONTRAST and random forest out-of-bag
   error rates (using 500 trees). For variance analysis of continuous
   clinical variables in different groups, non-parametric
   Wilcoxon-Mann-Whitney U, non-parametric ANOVA and parametric one-way
   ANOVA were used. Dichotomous variables were compared using the
   chi-square test. R [[84]22], SAS (version 9.4; SAS institute, Cary, NC)
   and GraphPad Prism (version 7; GraphPad Software, San Diego, CA) were
   used for data presentation. Normalized signal intensities and. CEL
   files of the transcriptomic data were deposited at the NIH Gene
   Expression Omnibus [85]http://www.ncbi.nlm.nih.gov/geo under the series
   accession number GSE 129166) and the microarray data were handled in
   accordance with the MIAME (Minimum Information About a Microarray
   Experiment) guidelines. The BIOMARGIN study is registered with
   [86]ClinicalTrials.gov, number [87]NCT02832661. The checklist according
   the STARD guidelines for diagnostic accuracy studies was completed and
   is included in the Supplementary Appendix (Table S9).

2.5. Study approval

   All patients provided informed written consent.

3. Results

3.1. Baseline characteristics

   We included 687 peripheral blood samples from 630 patients obtained at
   the time of a renal allograft biopsy, 120 with antibody-mediated
   rejection and 567 without ([88]Fig. 1). In the discovery phase, 117
   blood samples and 95 biopsy samples were used for genome-wide
   expression analysis from 117 patients, in a case-control study design.
   In the derivation phase, we included 183 samples from 183 individual
   patients, again in a case-control setting. In the independent
   validation cohort, 387 consecutively collected and unselected samples
   from 365 patients were included in the analyses; 56 samples were
   excluded, 49 because of inadequate biopsy histology and seven because
   of incompleteness of the gene expression profile. Patients'
   demographics and clinical characteristics of the three independent
   peripheral blood sample sets are provided in [89]Table 1 and Table S1
   in the Supplementary Appendix. There were significant differences
   between patients with versus without antibody-mediated rejection in
   some baseline characteristics, reflecting the background risk for
   development of antibody-mediated rejection as shown in [90]Table 2.
   Details on the clinical characteristics of the biopsy samples used for
   micro-array gene expression (N = 95) were provided separately [[91]18].
   Histological characteristics of the biopsies in the validation set are
   provided in Table S2.

Fig. 1.

   [92]Fig. 1
   [93]Open in a new tab

   Study design. Peripheral blood samples were obtained at the time of a
   renal allograft biopsy in four European transplant centres. In the
   discovery and derivation cohort, samples were selected based on
   availability and histological criteria of concomitant renal allograft
   biopsies (excluding cases with diagnosis of glomerulonephritis or
   polyomavirus nephropathy, and cases with unclear diagnosis), while
   graft function was not taken into account. In the validation cohort,
   all samples with concomitant adequate renal allograft biopsy histology,
   prospectively collected between June 24, 2014 and July 2, 2015, were
   serially included without selection on histology, demographics or time.
   The gene expression profile was not complete in seven of these samples,
   leading to a total of 387 cases in the validation phase.
   ABMR = antibody-mediated rejection.

Table 1.

   Characteristics of the patients and biopsies included in the validation
   cohort (N = 365 patients, 387 biopsies).
   Variable Mean (median) ± standard deviation (min – max) or no. (%)
   Transplant characteristics (N = 365 patients)
   Recipient age at transplantation (years) 50·2 (52·2) ± 15·3 (2·7–78·5)
   Recipient age at time of biopsy (years) 52·4 (54·1) ± 14·4 (19·0–79·6)
   Recipient sex (male/female) 224/141 (61·4%/38·6%)
   Repeat transplantation 64 (17·5%)
   Recipient ethnicity (European/ Asian/ African/ Other) 318/3/6/35[94]^⁎
   (87·8%/0·8%/1·7%/9·7%)
   Donor age (years) 50·6 (52·0) ± 15·5 [95]^⁎ (5·0–91·0)
   Donor sex (male/female) 177/180[96]^⁎ (50·4%/49·6%)
   Deceased/Living donor 278/83[97]^⁎ (77·0%/23·0%)
   Heart-beating/ Non-heart-beating donor 250/28[98]^⁎ (89·9%/10·1%)
   Cold ischemia time (hours) 12·1[99]^⁎ (12·6) ± 7·8 (0·27–35·8)
   


   Biopsy characteristics (N = 387 biopsies)
   Indication/protocol biopsy 134/253 (34·6%/65·4%)
   Time after transplantation (days) 908 (359) ± 1733 (6–12,564)
   Biopsy time after transplantation
   < 1 year 207 (53·5%)
   > 1 year 180 (46·5%)
   MDRD eGFR, (ml/min/1·73 m^2) 43·0 (41·8) ± 17·5 (5·8–96·2)
   Proteinuria (g/g creatinine) 0·4[100]^⁎ (0·1) ± 1·0 (0·0–8·0)
   


   Immunosuppression at time of biopsy
   Cyclosporine 40 (10·3%)
   Tacrolimus 331 (85·5%)
   Mycophenolate 320 (82·7%)
   Azathioprine 9 (2·3%)
   mTOR inhibitor 49 (12·7%)
   Corticosteroids 352 (91·0%)
   Histological diagnosis
   No rejection 330 (85·3%)
   T-cell mediated rejection
   Borderline changes 15 (3·9%)
   Grade 1 or 2 4 (1·0%)
   Antibody-mediated rejection 41 (10·6%)
   Mixed rejection^ 3 (0·8%)
   Interstitial fibrosis/tubular atrophy
   Grade 0 182 (47·0%)
   Grade 1 93 (24·0%)
   Grade 2 64 (16·5%)
   Grade 3 48 (12·4%)
   Polyomavirus-associated nephropathy 14 (3·6%)
   De novo/recurrent glomerulonephritis 26 (6·7%)
   [101]Open in a new tab

   ^Mixed rejection cases are defined by co-occurrence of
   antibody-mediated rejection and T-cell mediated rejection.
   ^⁎

   Missing data on donor age, donor gender, donor type, cold ischemia
   time, proteinuria, recipient ethnicity and induction therapy.

Table 2.

   Comparison between cases with antibody-mediated rejection vs. without
   antibody-mediated rejection in the validation cohort (N = 387). For
   variance analysis of continuous variables t-test was used; dichotomous
   variables were compared with the Chi-square test.
   Variable Antibody-mediated rejection (N = 41)
     __________________________________________________________________

   No antibody-mediated rejection (N = 346)
     __________________________________________________________________

   P value
   Mean ± SD
   or No. (%) Mean ± SD
   or No. (%)
   Recipient age at transplantation (years) 42·9 ± 17·3 50·8 ± 14·8 0·002
   Recipient age at time of biopsy (years) 50·5 ± 14·4 52·7 ± 14·4 0·37
   Recipient sex (male/female) 16/25 (39·0%/61·0%) 224/122 (64·7%/35·3%)
   0·001
   Recipient ethnicity[102]^⁎
   (European/Asian/African/Other) 38/0/0/3 (92·7%/0%/0%/7·3%) 300/3/6/34
   (86·7%/0·9%/1·7%/9·8%) 0·77
   Donor age (years) [103]^⁎ 41·2 ± 17·5 51·3 ± 14·9 <0·0001
   Donor sex[104]^⁎ (male/female) 24/14 (63·2%/36·8%) 168/173
   (49·3%/50·7%) 0·10
   Deceased/Living donor [105]^⁎ 34/5 (87·2%/12·8%) 261/83 (75·9%/24·1%)
   0·11
   Heart-beating/ Non-heart-beating donor[106]^⁎ 32/2 (94·1%/5·9%) 231/30
   (88·5%/11·5%) 0·32
   Cold ischemia time (hours)[107]^⁎ 13·3 ± 6·9 12·0 ± 7·9 0·36
   Repeat transplantation (yes/no) 60/286 (17·3%/82·7%) 9/32 (22·0%/78·0%)
   0·92
   Indication/protocol biopsy 30/11 (73·2%/26·8%) 104/242 (30·1%/69·9%)
   <0·0001
   Time after transplantation (days) 2806·0 ± 3233·7 683·5 ± 1293·5 0·0002
   MDRD eGFR, (ml/min/1.73 m^2) 34·7 ± 21·4 43·9 ± 16·7 0·01
   Proteinuria (g/g creatinine)[108]^⁎ 1·4 ± 1·9 0·3 ± 0·7 0·0006
   Presence of DSA 18 (43·9%) 46 (13·3%) <0·0001
   [109]Open in a new tab
   ^⁎

   missing data on recipient ethnicity, donor age, donor sex,
   deceased/living donor, heart-beating/non-heart-beating donor, cold
   ischemia time and proteinuria.

3.2. Discovery of transcripts specific for antibody-mediated rejection

   In peripheral blood and biopsy samples, respectively 970 and 783
   probesets (730 and 576 individual genes) had an ABMR score > 0·25.
   Pathway enrichment analysis of the biopsy signature was previously
   published [[110]18]. Based on ABMR scores in peripheral blood and
   biopsies, 2 genelists were determined (Fig. S1 in the Supplementary
   Appendix). In both genelists, there was significant enrichment in
   natural killer cell signaling, crosstalk between dendritic and natural
   killer cells, communication between innate and adaptive immune cells
   and antigen presentation pathways. Predicted upstream factor analysis
   identified interferons and interferon regulatory factors as the most
   likely upstream regulators (Table S3 in the Supplementary Appendix).

3.3. Gene signature identification and model development in the derivation
cohort

   From the genelists obtained in the discovery phase, we selected 44
   transcripts for RT-PCR analysis in the derivation cohort, based on
   combinations of ABMR and TCMR scores in blood and allograft biopsies,
   robustness of the results with different probesets of the same gene and
   by their involvement in relevant canonical pathways (Table S4 in the
   Supplementary Appendix). First, 26 genes with ABMR score > 0·25 and
   TCMR score < 0·20 in blood were selected (of these selected
   transcripts, nine also had a high ABMR score > 0·25 in kidney
   biopsies). We additionally selected 17 genes with an ABMR score > 0.25
   in biopsies and ABMR score > 0.20 in peripheral blood. Finally, given
   the biological homology of CXCL11 with CXCL10 and an ABMR score of
   CXCL11 of 0·49 in biopsy samples (but only 0·08 in blood), we added
   CXCL11 to the gene panel in the derivation phase. The univariate
   associations of the expression of these 44 genes with rejection
   phenotypes are shown in Fig. S2 in the Supplementary Appendix. From
   these 44 genes, a gene signature specific for antibody-mediated
   rejection was identified on the samples of the derivation cohort
   (N = 183), and included the following eight genes: CXCL10, FCGR1A,
   FCGR1B, GBP1, GBP4, IL15, KLRC1, TIMP1. More information on the role of
   these eight genes in ABMR is provided in Table S5. Subsequently, this
   8-gene signature was used to build a logistic regression model with
   nested loop internal cross-validation for discrimination between cases
   with and without antibody-mediated rejection in the derivation cohort.
   Applied to the samples of the derivation cohort, this gene signature
   and logistic regression model yielded a ROC AUC of 78·1% (95%
   confidence interval [CI], 70·7 to 85·6; p < 0·0001)([111]Fig. 2).

Fig. 2.

   [112]Fig. 2
   [113]Open in a new tab

   Diagnostic accuracy of the 8-gene assay for non-invasive diagnosis of
   antibody-mediated rejection in the validation cohort (N = 387). The
   left panel shows the 8-gene assay score in cases with versus without
   antibody-mediated rejection. The middle panel shows the distribution of
   cases with antibody-mediated rejection across all scores of the 8-gene
   assay. The right panel shows the ROC curves for samples with versus
   without antibody-mediated rejection, with presentation of the area
   under the receiver operating characteristic curve (ROC AUC) and the 95%
   confidence interval.

3.4. Diagnostic accuracy of the 8-gene assay in the validation cohort

   The 8-gene signature and logistic regression model built and locked on
   the derivation cohort were evaluated on the 387 samples collected in
   the validation cohort, which contained 41 cases with antibody-mediated
   rejection (10·6%), representing the natural prevalence of this
   phenotype in the cohort of biopsies performed at the participating
   centres. The 8-gene assay reached a ROC AUC of 79·9% (95% CI, 72·6 to
   87·2; p < 0·0001)(Fig. S3 in the Supplementary Appendix). When we
   evaluated the diagnostic accuracy for discrimination between pure
   antibody-mediated rejection (N = 38) and pure borderline changes and
   T-cell mediated rejection (N = 13 and 3 respectively), the 8-gene assay
   reached a ROC AUC of 82·2% (95% CI 70·7 to 93·8, p = 0·001), and 79·3%
   (95% CI, 71·6 to 86·9; p < 0·0001) for discrimination between pure
   antibody-mediated rejection (N = 38) and absence of rejection
   (N = 330). This 8-gene assay was not diagnostic for the group of T-cell
   mediated rejections (containing mostly borderline changes) (Fig. S4 in
   the Supplementary Appendix).

   We next validated the cut-off value of the 8-gene assay that was
   determined in the derivation phase (Fig. S5 in the Supplementary
   Appendix). The optimal cut-off from the derivation phase for the 8-gene
   biomarker (−1·08) had a sensitivity of 73·2%, specificity of 75·7%,
   Youden index of 0·49, positive predictive value of 26·3% and negative
   predictive value of 96·0% in the validation cohort ([114]Table 3 and
   Fig. S6 in the Supplementary Appendix).

Table 3.

   Diagnostic accuracy of the 8-gene marker for non-invasive diagnosis of
   antibody-mediated rejection in the independent validation cohort
   (N = 387).
   Diagnostic Accuracy (ROC AUC) Accuracy
   (Total correct/total) Sensitivity Specificity Positive Predictive Value
   Negative Predictive Value
   Population (N = ABMR/total) % (95% CI) % % % % %
   Validation cohort
   All biopsies (N = 41/387) 79·9% (72·7–87·2)
    Low threshold^§ (−2·50) p < 0·0001 35·1% 95·1% 28·0% 13·5% 98·0%
    Optimal threshold* (−1·08) 75·5% 73·2% 75·7% 26·3% 96·0%
    High threshold^§ (0·50) 89·1% 22·0% 97·1% 47·4% 91·3%
   


   Post-hoc sensitivity analyses
   Early biopsies <1 year (N = 10/207) 90·9% (85·3–96·4)
    Low threshold (−2·50) p < 0·0001 33·8% 100% 30·5% 6·8% 100%
    Optimal threshold (−1·08) 79·2% 90·0% 78·7% 17·6% 99·4%
    High threshold (0·50) 93·2% 20·0% 97·0% 25·0% 96·0%
   Late biopsies >1 year (N = 31/180) 73·5% (63·6–83·4)
    Low threshold (−2·50) p < 0·0001 36·7% 93·5% 24·8% 20·6% 94·9%
    Optimal threshold (−1·08) 71·1% 67·7% 71·8% 33·3% 91·5%
    High threshold (0·50) 84·4% 22·6% 97·3% 63·6% 85·8%
   Biopsies at time of graft dysfunction
   (N = 30/134) 75·3% (64·9–85·8)
    Low threshold (−2·50) p < 0·0001 36·6% 93·3% 20·2% 25·2% 91·3%
    Optimal threshold (−1·08) 71·6% 73·3% 71·1% 42·3% 90·2%
    High threshold (0·50) 79·1% 26·7% 94·2% 57·1% 81·7%
   Biopsies at time of stable graft function (N = 11/253) 83·4%
   (75·4–91·3)
    Low threshold (−2·50) p < 0·0001 34·4% 100% 31·4% 6·2% 100%
    Optimal threshold (−1·08) 77·5% 72·7% 77·7% 12·9% 98·4%
    High threshold (0·50) 94·5% 9·1% 98·3% 20·0% 96·0%
   [115]Open in a new tab

   *The optimal threshold for the ROC curve was chosen in the derivation
   phase at maximum sensitivity and specificity. ^§Low and high thresholds
   were arbitrarily selected in the independent validation cohort. ROC
   AUC = area under the receiver operating characteristic curve.

3.5. Sensitivity analysis

   The 8-gene assay retained its accuracy for antibody-mediated rejection
   in patients with stable graft function and at time of graft
   dysfunction, within the first year and also later after transplantation
   ([116]Table 3, [117]Fig. 3, Table S6 in the Supplementary Appendix). In
   all scenarios, the 8-gene assay allowed to rule out ongoing
   antibody-mediated rejection with high negative predictive values.
   However, the power was low in these sub-analyses, and these accuracy
   parameters should be interpreted cautiously. The 8-gene assay reached
   universal high specificities for antibody-mediated rejection, but the
   positive predictive value was lower due to the low prevalence of this
   disease in the validation cohort, especially in protocol-specified
   biopsies. From this, the positive predictive value was higher at time
   of graft dysfunction than in stable graft function. If restricted to
   cases with donor-specific antibodies (N = 64), the positive predictive
   value of the 8-gene assay was even higher: at the optimal threshold
   (−1·08), positive predictive value was 57·1%, with sensitivity of
   66·7%, specificity of 80·4%, Youden index of 0·47 and negative
   predictive value of 86·0%.

Fig. 3.

   [118]Fig. 3
   [119]Open in a new tab

   Diagnostic accuracy of the 8-gene assay for antibody-mediated rejection
   in specific subgroups in the validation set (N = 387). Post-hoc
   sensitivity analysis of the 8-gene marker according to time after
   transplantation is shown in panel A and according to stable graft
   function vs. graft dysfunction in panel B.

3.6. Correlation of the 8-gene assay with histological and clinical variables

   The 8-gene assay correlated with graft functional parameters like eGFR
   and proteinuria, and with histological lesions diagnostic for
   antibody-mediated rejection like glomerulitis, peritubular
   capillaritis, microvascular inflammation, and transplant glomerulopathy
   in the validation cohort (Table S7 in the Supplementary Appendix).
   Distribution of the 8-gene assay score per histological lesion grade is
   shown in [120]Fig. 4. There was no correlation with histological
   lesions of T-cell mediated rejection. The 8-gene biomarker did not
   associate with diagnosis of glomerulonephritis, polyomavirus associated
   nephropathy or interstitial fibrosis with tubular atrophy. A weak
   association with recipient gender (p = 0·04) was found. No other
   donor-recipient demographics associated with the 8-gene assay.

Fig. 4.

   [121]Fig. 4
   [122]Open in a new tab

   Distribution of the 8-gene assay score per histological lesion grade in
   the validation cohort (N = 387). The 8-gene assay score was
   significantly associated with lesions of antibody-mediated rejection.
   Significance was assessed with nonparametric one-way ANOVA and pairwise
   comparisons with t-test. Significance was apparent for higher severity
   grades of lesions associated with antibody-mediated rejection
   (glomerulitis, peritubular capillaritis, microvascular inflammation
   score, transplant glomerulopathy). No significant association was
   present with lesions of T-cell mediated rejection (tubulitis,
   interstitial inflammation) or non-specific chronic damage (interstitial
   fibrosis, tubular atrophy). g = glomerulitis; ptc = peritubular
   capillaritis; mvi = microvascular inflammation; cg = transplant
   glomerulopathy; i interstitial inflammation; t = tubulitis; v = intimal
   arteritis; C4d = C4d deposition in peritubular capillaries;
   ci = interstitial fibrosis; ct = tubular atrophy; cv = intimal
   fibrosis; ah = arteriolar hyalinosis. ns = not significant,
   *p < 0·05,**p < 0·01,***p < 0·001,****p < 0·0001.

3.7. Comparison with traditional biomarkers and added clinical value of the
8-gene assay

   The 8-gene assay associated with diagnosis of antibody-mediated
   rejection, independent of traditional factors associating with
   antibody-mediated rejection (female gender, recipient age, time after
   transplantation, presence of donor-specific antibodies and proteinuria)
   (Table S8 in the Supplementary Appendix). We next built and internally
   cross-validated a clinical model on the data from the discovery and
   derivation phase combined, which consisted of the 8 clinical parameters
   that differed between ABMR vs. no ABMR cases (donor-specific
   antibodies, proteinuria, eGFR, time after transplantation, recipient
   age at time of transplantation, donor age, recipient sex and protocol
   vs. indication biopsy). In the validation set, this clinical model
   reached an AUC of 77·3% (95% CI 68·5 to 86·0, p < 0·0001). Adding the
   result of the 8-gene assay to the clinical model increased the
   diagnostic accuracy for ABMR to 88·0% (95% CI, 82·7 to 93·3;
   p < 0·0001) (Fig. S7 in the Supplementary Appendix). Random forest
   out-of-bag error rates in the validation set were comparable for the
   8-gene assay and the clinical model (14·99% and 14·73%, respectively).
   When adding the 8-gene assay to the clinical model the error rate
   dropped to 11·37%. In protocol biopsies (at time of stable graft
   function), the clinical model was not contributive (AUC of 55·7%, 95%
   CI, 37·2–74·2). Both in indication and protocol biopsies the 8-gene
   assay had added diagnostic value on top of the clinical model (Fig. S7
   in the Supplementary Appendix). Decision curve analysis confirmed the
   net benefit of using the 8-gene assay for diagnosis of
   antibody-mediated rejection across the range of probability thresholds
   between 5% and 35% (Fig. S8 in the Supplementary Appendix).

4. Discussion

   In this multicentre, prospective, multiphase study we identified and
   validated an 8-gene expression assay in peripheral blood samples with
   good diagnostic accuracy for non-invasive diagnosis of
   antibody-mediated rejection. The 8-gene assay retained this accuracy
   both at time of stable graft function and at time of graft dysfunction,
   within the first year and also later after transplantation. The
   diagnostic accuracy of the 8-gene assay for antibody-mediated rejection
   was superior to that of clinical indicators, and the assay offered
   benefit in clinical decisions to perform or not perform a biopsy for
   diagnosis of antibody-mediated rejection.

   This study is a landmark in the field of biomarker discovery and
   development in renal transplantation in several aspects. First, its
   multiphase study design with independent discovery, derivation and
   validation sets allowed for robust biomarker development and validation
   [[123]15,[124]16]. After nested loop internal cross-validation in the
   derivation set, we performed independent external validation of the
   locked and fully specified model in a representative population with
   natural prevalence of antibody-mediated rejection, using the same
   technology platform. Second, stringent phenotypic selection with
   central re-assessment was applied, minimizing the interobserver
   variability in the current gold standard for diagnosis of rejection and
   reference standard for performance of the biomarker. Third, the
   comparison with an internally cross-validated and externally validated
   clinical model, and assessment of the net benefit of using this 8-gene
   assay indicate the clinical usefulness of this marker on top of
   routinely available clinical markers.

   The unbiased transcriptomic analysis on the discovery set illustrated
   that the molecular changes that occur in renal allograft biopsies with
   antibody-mediated rejection, primarily related to immune regulation,
   are also reflected in transcriptional differences in peripheral blood.
   The selected gene transcripts in our study come from biologically
   plausible molecular pathways, based on gene pathway enrichment
   analyses. Especially natural killer cells and their interplay with
   dendritic cells are known to be notorious players in humoral rejection
   [[125]23]. The significant enrichment of these pathways in our gene
   transcripts corroborates the specificity of our marker for
   antibody-mediated rejection. Furthermore, the eight individual genes
   have relevant pathophysiological functions and most have references to