FIELD OF THE INVENTION

The invention relates to methods and probes to detect antibody-mediated and T-cell mediated kidney transplant rejection.

Introduction

Kidney transplantation is the most cost-effective treatment for patients with end- stage kidney disease, although long-term functional outcome is challenged by the potential of rejection by the host immune system [Lamb et at. (2011) Am J Transplant. 11(3), 450-462; Coemans et a/. (2018) Kidney Int. 94(5), 964-773; Naesens et a/. (2014) Transplantation. 98(4), 427-435; Van Loon et a/. (2020) Transplantation 104(12), 2557-2566] Currently, histological interpretation of kidney transplant biopsies is the cornerstone in the diagnosis of rejection. The Banff consortium provides the internationally adopted classification for kidney transplant rejection, wherein diagnostic categories are based on semi-quantitative scoring of histological lesions in defined renal compartments [Loupy et al. (2020) Am J Transplant. 20, 2305-2317; Roufosse et al. (2018) Transplantation. 102(11), 1795-1814]. The distinction is made between antibody-mediated rejection (ABMR) and T-cell mediated rejection (TCMR), with an intermediate TCMR category termed "Borderline changes", reserved for biopsies with less severe tubulointerstitial inflammation than the full TCMR phenotype [Loupy et al. cited above].

Despite worldwide clinical use, histopathological classification of kidney transplant biopsies is hampered by poor reproducibility, sampling error and tissue heterogeneity [Mengel et al. (2007) Am J Transplant. 7, 2221-2226; Naesens 8i Anglicheau (2018) J Am Soc Nephrol. 29(1), 24-34] Correct assessment requires expertise, and even between experienced nephropathologists, poor to fair reproducibility was observed for decisive features such as tubulointerstitial inflammation severity, transplant glomerulopathy and C4d deposition [Mengel etal. (2013 ) Am J Transplant. 13, 1235- 45; Veronese et al. (2005) Clin Transplant. 19, 518-521; Gough et al. (2002) Nephrol Dial Transplant 17, 1081-1084; Furness et al. (2003) Am J Surg Pathol. 27(6), 805-810]. In addition, intermediate lesions that do not fulfill the full Banff criteria often remain unacknowledged in the dichotomous classification system. Recently, it was demonstrated that disregarding the intermediate category "suspicious for ABMR" in the Banff 2017 update resulted in decreased discriminative performance of the histological diagnosis for prediction of graft failure, which suggests that the binary histological definition insufficiently captures the heterogeneity of humoral rejection [Callemeyn et at. (2021) Am J Transplant. 21(7), 2413-2423].

Over the past two decades, charting the molecular landscapes of ABMR and TCMR by transcriptome arrays revealed ensembles of gene transcripts that correlated significantly with Banff lesions and diagnostic categories. These molecular analyses promised to overcome several issues with classic histopathological phenotyping of kidney transplant biopsies [Halloran et al. (2016) Nat Rev Nephrol. 12(9), 534-548; Sellares et al. (2013) Am J Transplant. 13(4), 971-983; Reeve et al. (2013) Am J Transplant. 13(3), 645-55; Venner et al. (2015) Am J Transplant. 15(5), 1336- 1348], and were therefore introduced in the Banff 2013 classification for ABMR diagnosis [Cendales et al. (2014) Am J Transplant. 14(2), 272-283]. In addition, transcriptional heterogeneity was found within canonical rejection phenotypes, which ultimately gave rise to the development of a commercial platform (Molecular Microscope Diagnostics system, MMDx; One Lambda, Inc, CA, USA) offering a probabilistic reporting system and alternative archetype-based classifications based on biopsy transcriptomics [Halloran et al. (2017) Am J Transplant. 17, 2851-2562; Madill-Thomsen et al. (2020) Am J Transplant. 20(5), 1341-1350; Reeve 8i Mackova (2019) Am J Transplant. 19, 2719-2731; Reeve et al. (2017) Jci Insight. 2, e94197]. Notwithstanding these advances, implementation of molecular diagnostics in the transplant clinic has remained limited. In a recent survey of surgeons, nephrologists and pathologists involved in kidney transplantation, 90% reported not to use molecular transcripts in the diagnosis of ABMR [Schinstock et al. (2020) Transpl Int. 34(3), 488-498]. Several challenges explain this lack of clinical use of molecular diagnostics in kidney transplantation, including the lack of a diagnostic gold standard, the evolving tools and algorithms, the paucity of multicentric prospective validation studies and the absence of consensus on the preferred molecular platform [Mengel etal. (2020) Am J Transplant. 20, 2305-2317]. In addition, the number of transcripts in currently established classifiers is high, potentially affecting interpretability and integration with routine clinicopathological information.

SUMMARY OF THE INVENTION

The present invention provide a robust prediction model with as few transcripts as needed, aiming at retaining high accuracy for molecular diagnosis of ABMR and TCMR while improving interpretability and implementability. In addition, the aim is to externally validate the parsimonious molecular classifiers for ABMR and TCMR, and to evaluate their added value on top of histological classification of kidney transplant biopsies.

Although the transcriptional landscapes of antibody-mediated rejection (ABMR) and T-cell mediated rejection (TCMR) have been largely elucidated, applying these gene expression signatures in transplant clinics is hampered by the large number of features and difficult integration with histological findings. Herein, sparse molecular classifiers were developed and validated for ABMR and TCMR. In a discovery cohort of 224 kidney transplant biopsies, lasso regression was applied on microarray gene expression data to derive a 2-gene classifier for ABMR ( PLA1A , GNLY) and a 2-gene classifier for TCMR (IL12RB1 , ARPC1B). External validation demonstrated maintained high diagnostic accuracy for the ABMR classifier (ROC-AUC 0.80, 95% Cl 0.75-0.85) and TCMR classifier (ROC-AUC 0.83, 95% Cl 0.77-0.89), which also permitted differentiation between pure and mixed rejection phenotypes. Complementary to their diagnostic potential, the molecular classifiers associated with accelerated graft loss and identified transplant kidneys at risk for failure with histological lesions of rejection that did not reach the Banff thresholds for ABMR or TCMR. Thus, rigid variable selection strategies can yield sparse molecular classifiers for allograft rejection phenotypes with preserved accuracy and prognostic value, which may facilitate their interpretation and clinical implementation. Combining molecular and histological information from allograft biopsies refines personalized risk assessment.

The invention is further summarised in the following statements:

1. An in vitro method of distinguishing between ABMR and TCMR in an individual at risk of a kidney transplant rejection, or having any type of kidney transplant rejection, the method comprising the steps of : a) determining in a biopsy of the renal allograft of said individual the expression level of at least two genes of a first set of genes, wherein the first set of genes consists of GNLY, PLA1A, PRF1, CCL4, KLRD1, MS4A7, XCL1, CCL3L1, BID, PRF1, and CXCL9, [These are the genes listed in the top panel of figure 7] and determining in said biopsy of said individual the expression level of at least two genes of a second set of genes, wherein the second set of genes consists of IL12RB1, TBXAS1, MS4A6A, ARPC1B, RAC 2, FYB1, Clorfl62, ARHGAP30, FCGR3B, CD1D, VAV1, JAML, ZDHHC18, GIMAP1, CD300A, ARHGDIB, FAM72A, HLA-DQB1, LAPTM5, D0K1, ADA, LRRC8C, ENTPD1 and LAIR,

[These are the genes listed in the top panel of figure 7] b) comparing the expression levels of said at least 2 genes of the first set and said at least 2 genes of the second set with their expression level in a reference biopsy of a healthy individual with a kidney transplant, c) wherein an increased expression level of said at least 2 genes of the first set is indicative of ABMR, and wherein an increased expression level of said at least 2 genes of the second set is indicative of TCMR.

It is noted that determining the expression level of the genes of the first set and the second set can be performed separately. When a conclusive diagnosis of ABMR or TCMR is obtained with the expression level of the set of genes that has been initially measures, the additional determination of expression level of the other set of genes can be omitted. However if no conclusive diagnosis is obtained the additional determination of expression level of the other set of genes is anyhow required. For the sake of efficacy, expression levels of both set of genes are determined.

2. The method according to statement 1, wherein the first set of genes consists of GNLY, PLA1A, PRF1, CCL4 and KLRD1. [These genes are the six genes with the highest relative feature selection frequency in the top panel of figure 7]

3. The method according to statement 1 or 2, comprising the step of determining in a biopsy of said individual the expression level of between 2 to 4 genes of said first set of genes.

Herein, the phrase "determining the expression level of between 2 to 4 genes of a set of genes", means that apart from the 2, 3 or 4 selected genes, for none of the other genes in that set the expression level is determined. It is an advantage of the present invention that ABMR and TCMR can be determined using a the expression level of a limited set of genes, and does not require complex assays such as microrrays and the like.

4. The method according to any one of statements 1 to 3, wherein determining the expression level of between 2 to 4 genes comprises the determining the expression level of GNLY and/or PLA1A.

5. The method according to any one of statements 1 to 4, comprising the step of determining in a biopsy of said individual the expression level of 2 genes of said first set of genes, wherein the two genes are GLNY and PLA1A. 6. The method according to statement 1, wherein the second set of genes consist of IL12RB1, TBXAS1, MS4A6A, ARPC1B and RAC2 . [[These genes are the 5 genes with the highest relative feature selection frequency in the top panel of figure 7]

7. The method according to statement 1 or 6, wherein determining the expression level of at least 2 genes of said second set of genes comprises determining the expression level of IL12RB1.

8. The method according to statements 1, 6 or 7 comprising the step of determining in a biopsy of said individual the expression level of between 2 to 4 genes of said second set of genes.

9. The method according to any one of statements 1 and 6 to 8, wherein determining the expression level of between 2 to 4 genes comprises the determining the expression level of IL12RB1 and ARPC1B.

10. The method according to any one of statements 1 and 6 to 8, wherein determining the expression level of between 2 to 4 genes comprises the determining the expression level of IL12RB1 and CD1D.

11. The method according to any one of statements 1 and 6 to 10, comprising the step of determining in a biopsy of said individual the expression level of 2 genes of said second set of genes.

12. The method according to any one of statements 1 and 6 to 11, wherein said 2 genes of said second set of genes are IL12RB1 and ARPC1B, or wherein said 2 genes of said second set of genes are IL12RB1 and CD1D.

13. The method according to any one of statements 1 to 12, comprising the steps of : a) determining the expression level of GNLY and PLA1A and b) determining the expression level of IL12RB1 and ARPC1B or determining the expression level of IL12RB1 and CD1D wherein an increased expression level of GNLY and PLA1A is indicative of ABMR,and wherein an increased expression level of IL12RB1 and ARPC1B or wherein an increased expression level of IL12RB1 and CD1D is indicative of TCMR.

14. The method according to any one of statements 1 to 13, wherein determining the expression level of said genes is performed by a multiplex PCR.

15. The method according to any one of statements 1 to 13, wherein determining the expression level of said genes is performed by nanostring technology.

16. A kit for in vitro diagnosis of ABMR and TCMR in a patient with a kidney transplant, the kit comprising: a set of probes for the detecting of the expression level of a set of genes, wherein the set of probes is limited to: probes for the detection of household genes probes for the detection of two or more of GNLY, PLA1A, PRF1, CCL4, KLRD1, MS4A7, XCL1, CCL3L1, BID, PRF1, and CXCL9 and probes for the detection of two or more of IL12RB1, TBXAS1, MS4A6A, ARPC1B, RAC 2, FYB1, Clorfl62, ARHGAP30, FCGR3B, CD1D, VAV1, JAML, ZDHHC18, GIMAP1, CD300A, ARHGDIB, FAM72A, HLA-DQB1, LAPTM5, DOK1, ADA, LRRC8C, ENTPD1 and LAIR.

17. The kit according to statement 16 comprising: a set of probes for the detecting of the expression level of a set of genes, wherein the set of probes is limited to probes for the detection of household genes probes for the detection of two or more of GNLY, PLA1A, PRF1, CCL4 and KLRD1 and probes for the detection of two or more of IL12RB1, TBXAS1, MS4A6A, ARPC1B, RAC 2.

18. The kit according to statement 16 comprising : a set of probes for the detecting of the expression level of a set of genes, wherein the set of probes is limited to probes for the detection of household genes probes for the detection of GNLY and PLA1A and probes for the detection of IL12RB1 and probes for the detection of ARPC1B and/or CD1D.

DETAILED DESCRIPTION Figure Legends

Figure 1 Internal and external validation of the ABMR classifier

A-B. ROC curves for diagnosis of ABMR in the discovery cohort by sensitivity analysis according to biopsy context (n = 139 protocol biopsies, n=85 indication biopsies) and HLA-DSA status of the recipient (n=84 HLA-DSA positive recipients, n = 140 HLA-DSA negative recipients). C. Comparison of ABMR score according to severity of microvascular inflammation. D. ABMR score in cases with versus without ABMR in the validation cohort GSE36059. E. Distribution of ABMR cases across all ABMR scores in the validation cohort. F. ROC curve for diagnosis of ABMR in the validation cohort. ABMR: antibody-mediated rejection. ROC: receiver operating characteristic, AUC: area under the curve. Values are given with 95% confidence interval. HLA-DSA: anti- HLA donor-specific antibodies. **** P < 0.0001 Figure 2 Internal and external validation of the TCMR classifier A-B. ROC curves for diagnosis of TCMR in the discovery cohort by sensitivity analysis according to biopsy context (n = 139 protocol biopsies, n=85 indication biopsies) and HLA-DSA status of the recipient (n=84 HLA-DSA positive recipients, n = 140 HLA-DSA negative recipients). C. Comparison of TCMR score according to severity of tubule- interstitial inflammation. D. TCMR score in cases with versus without TCMR in the validation cohort GSE36059. E. Distribution of TCMR cases across all TCMR scores in the validation cohort. F. ROC curve for diagnosis of TCMR in the validation cohort. TCMR: T-cell mediated rejection. ROC: receiver operating characteristic, AUC: area under the curve. Values are given with 95% confidence interval. HLA-DSA: anti-HLA donor-specific antibodies. **** P <0.0001

Figure 3 Discriminative performance of the ABMR and TCMR classifiers

A-B. Comparison of ABMR scores and TCMR scores between rejection phenotypes in the discovery cohort (N=224). For ABMR, both HLA-DSA positive and negative cases of ABMR histology were included. C-D. Comparison of ABMR scores and TCMR scores between rejection phenotypes in the validation cohort GSE36059 (N=403). Scores were calculated for each biopsy based on the classifiers derived from the discovery cohort. Box plots with range are plotted. ABMR: antibody-mediated rejection, TCMR: T-cell mediated rejection, Mixed: mixed TCMR and ABMR. * P <0.05, ** P <0.01, *** P <0.001, **** P <0.0001

Figure 4 Prognostic value of molecular classifiers for intermediate histological lesions in the discovery cohort (N=224 biopsies)

A. Kaplan-Meier survival curve for death-censored allograft failure, with biopsies stratified according to the presence of humoral lesions. Intermediate cases denote presence of humoral lesions, i.e. glomerulitis, peritubular capillaritis, intimal arteritis, c4d deposition and transplant glomerulopathy, that are insufficient to fulfill the first two Banff criteria for ABMR. ABMR histology denotes biopsies fulfilling the first two Banff criteria. B. Comparison of ABMR score between groups. Horizontal dotted lines represent predefined negative, optimal and positive thresholds. C. Graft survival rates of biopsies with humoral lesions, stratified according to the optimal ABMR molecular cutoff. For comparison, survival of biopsies without humoral lesions is plotted. D. Survival curve with biopsies stratified according to presence of borderline rejection or TCMR. E. Comparison of TCMR probability score between groups. F. Graft survival rates of biopsies with cellular lesions, stratified according to the optimal TCMR molecular cutoff value. For comparison, survival of biopsies without cellular lesions is plotted.

Figure 5 Prognostic performance of the ABMR and TCMR classifiers in validation cohort GSE21374 (N=282 biopsies)

A-B. Time-dependent area under the ROC curve for prediction of graft failure according to the ABMR score (A) or TCMR score (B), corrected for timing after transplantation. C-D. Kaplan-Meier survival curve for graft failure, with biopsies stratified according to the histological presence of rejection and ABMR score (A) or TCMR score (B). Low=below median, high=above median. AUC: area under the curve, ROC: receiver operating characteristic, ABMR: antibody-mediated rejection, TCMR: T-cell mediated rejection.

Figures 6 Parameter sweep of the LI regularization parameter C during 10 rounds of internal cross-validation

Figure 7 Frequency of appearance of non-zero coefficients in Lasso regularization across 200 realizations

Probesets are ranked according to the frequency of appearance across iterations. The corresponding gene is given for each probeset. Genes can be represented by multiple probesets in the Affymetrix GeneChip Human Genome U133 Plus 2.0 Array.

Figure 8 Accuracy parameters of the ABMR classifier and TCMR classifier in the discovery cohort (N=224 biopsies)

A. Receiver operating characteristic (ROC) curve for presence of ABMR, with presentation of the area under the ROC (AUC) and 95% confidence interval. B-C. Evolution of the accuracy parameters for detection of ABMR according to different cut-off values of the molecular ABMR score. D. ROC curve for presence of TCMR, with presentation of the AUC and 95% confidence interval. E-F. Evolution of the accuracy parameters for detection of ABMR according to different cut-off values of the molecular ABMR score. ABMR: antibody-mediated rejection, TCMR: T-cell mediated rejection.

Figure 9 Diagnostic performance of the TCMR classifier across the histological spectrum of cellular rejection

A. Receiver operating characteristic (ROC) curve for presence of cellular rejection in the discovery cohort, defined as a composite of borderline changes and TCMR, with presentation of the area under the ROC (AUC) and 95% confidence interval. B. Comparison of the molecular TCMR score between biopsies with the histological diagnosis of TCMR, borderline changes and absence of cellular rejection in the discovery cohort. C-D. Identical analyses as described in A and B were performed in the validation cohort. TCMR: T-cell mediated rejection. ** P <0.01, **** P <0.0001

Figure 10 Accuracy parameters of the ABMR and TCMR classifiers in the validation cohort (N=403 biopsies) Evolution of the accuracy parameters according to the ABMR classifier (A-B) and TCMR classifier (C-D) is shown. PPV= positive predictive value. NPV= negative predictive value. Accuracy = total correct/ total.

Figure 11 Validation of the molecular markers on a Nanostring platform

Abbreviations

ABMR: antibody-mediated rejection; ABMRh: histology of antibody-mediated rejection; ah: arteriolar hyalinosis; AUC: area under the curve; eg: transplant glomerulopathy; cv: vascular fibrous intimal thickening; Cl: confidence interval; eGFR: estimated glomerular filtration rate; FDR: false discovery rate; g: glomerulitis; HLA: human leukocyte antigen; HLA-DSA: anti-HLA donor-specific antibodies; HR: hazard ratio; i: interstitial inflammation; IFTA: interstitial fibrosis and tubular atrophy; IQR: interquartile range; MFI: mean fluorescence index; mTOR: mammalian target of rapamycin; NPV: negative predictive value; PPV: positive predictive value; ptc: peritubular capillaritis; PVAN: polyoma virus associated nephropathy; ROC: receiver operating characteristic; SD: standard deviation; SMOTE: synthetic minority oversampling technique; t: tubulitis; TCMR: T-cell mediated rejection; v: intimal arteritis. ADA: adenosine deaminase; ARHGAP30: RHO GTPase-activating protein 30; ARHGDIB: RHO GDP-dissociation inhibitor beta; ARPC1B: actin-related protein 2/3 complex, subunit IB; BID: BH3-interacting domain death agonist; CCL3L1: chemokine, CC motif, ligand 3-like 1; CCL4: chemokine, CC motif, ligand 4; CD1D: thymocyte antigen cdld; CD300A: CD300A antigen; CXCL9: chemokine, CXC motif, ligand 9; DOK1: docking protein 1; ENTPD1: ectonucleoside triphosphate diphosphohydrolase 1; FAM72A: family with sequence similarity 72, member A; FCGR3B: Fc fragment of IgG receptor Illb; FYB1: FYN-binding protein 1; GIMAP1: GTPase, IMAP family, member 1; GNLY: granulysin; HLA-DQB1: major histocompatibility complex, class II, DQ beta-1; IL12RB1: interleukin 12 receptor, beta-1; JAML: junctional adhesion molecule-like; KLRD1: killer cell lectin-like receptor, subfamily D, member 1; LAIR: leukocyte-associated immunoglobulin-like receptor 1; LAPTM5: lysosome-associated protein, transmembrane 5; LRRC8C: leucine-rich repeat-containing protein 8C; MS4A6A: membrane-spanning 4- domains, subfamily a, member 6A; MS4A7: membrane-spanning 4-domains, subfamily a, member 7; PLA1A: phosphatidylserine-specific phospholipase Al- alpha; PRF1: perforin 1; RAC2: RAS-related C3 botulinum toxin substrate 2; TBXAS1: thromboxane A synthase 1; VAV1: VAV guanine nucleotide exchange factor 1; XCL1: chemokine, c motif, ligand 1; ZDHHC18: zinc finger DHHC-Type palmitoyltransferase 18;

TCMR refers to T-cell mediated rejection.

The incidence of TCMR, which had a major impact on graft function and survival early after transplantation in the previous decades, has significantly decreased since application of more efficacious immunosuppressive regimens with tacrolimus, mycophenolate mofetil and induction therapy [Nankivell BJ 8i Alexander (2010) N Engl J Med. 363, 1451-1462]. Nevertheless, the existing immunosuppressive armamentarium is insufficient in preventing patients from developing humoral alloreactivity, with the occurrence of circulating donor-specific HLA antibodies (DSAs) and ABMR [Nankivell 8i Alexander, cited above; Loupy et al. (2012) Nat Rev Nephrol. 8, 348-357; Djamali et al. (2014) Am J Transplant. 14, 255-271; Amore (2015) Curr Opin Organ Transplant. 20, 536-542]. In recent years, DSAs were demonstrated to be a crucial prognostic factor for graft outcome, and ABMR is now recognized as a prime reason for graft failure after kidney transplantation [Naesens et al. (2014) Transplantation. 98, 427-435; El-Zoghby et al. (2009) Am J Transplant. 9, 527-535; Sellares et al. (2012) Am J Transplant. 12, 388-399; Naesens et al. (2016) J Am Soc Nephrol. 27, 281-292].

ABMR refers to antibody-mediated rejection.

ABMR is often mediated by antibodies directed against allogeneic major histocompatibility complex (MHC) molecules via the complement system. MHC molecules are interchangeably referred to as human leukocyte antigens (HLAs). HLAs are responsible for allorecognition, and without immunosuppression, allografts from a donor with different HLAs will be rejected. There are more than 1600 alleles for HLA class I and II molecules [Colvin 8i Smith (2005) Nat Rev Immunol 10, 807-817; Mandelbrodt 8i Mohamed Transplantation immunobiology. In: anovitch, ed. Handbook of Kidney Transplantation. Philadelphia, PA: Lippincott Williams 8i Wilkins (2010) 19-23.]. HLA class I molecules (e.g., HLA-A, HLA-B, HLA-C) are found on all nucleated cells, but HLA class II molecules (e.g., HLA-DP, HLA-DQ, HLA-DR) are expressed only on antigen-presenting cells (APCs). Among a recipient's anti-HLA antibodies, those specific to the donor's HLAs are DSAs. Less frequently, antibodies against other antigenic stimulants, such as ABO blood group antigens, minor histocompatibility antigens, endothelial cell antigens, and angiotensin II type 1 receptors, are responsible for ABMR [Colvin & Smith, cited above]. A complicated process mediates the development of antibodies upon exposure to antigens. Antigens are presented by either donor or recipient APCs to CD4 + T cells (i.e., T helper cells), which then activate B cells via cytokines and costimulatory factors. Immature B cells are differentiated into either memory B cells or antibody-forming plasma cells. Plasma cells subsequently produce antibodies for longer than a year without help from T cells [Shapiro-Shelef & Calame (2005) Nat Rev Immunol 3, 230-242]. Allograft cells are not destroyed by antibodies themselves, but rather via the activation of the complement system or cytotoxic cells. Therefore, the production of DSAs does not necessarily mean that a kidney transplant recipient will experience ABMR. Complement activation plays a major role in ABMR, resulting in tissue injury and thrombosis. Complement molecules (particularly Clq) bind to the antigen- antibody complex on the graft endothelium. This interaction activates a process known as the "complement dependent cascade", a complex process that occurs along the cellular membrane of a target cell (e.g., allograft endothelium and microvasculature). The presence of C4d on an allograft is evidence of complement activation. In fewer cases, antibodies can cause endothelial injury by a complement- independent mechanism via antibody-dependent cell-mediated cytotoxicity. This contributes to allograft injury through natural killer cells and macrophages, and it may be more related to chronic ABMR.

Previous exposure to foreign HLAs may predispose a kidney transplant recipient to an increased risk of ABMR. Patients are at risk of developing anti-HLA antibodies after solid organ transplant, blood infusion, pregnancy, and infection. Those with a significant level of anti-HLA antibodies prior to transplantation are referred to as "sensitized," and they are at a high risk of developing posttransplant ABMR. A calculated panel reactive antibody (cPRA) is used to identify sensitized patients prior to transplant. The cPRA estimates the probability of incompatible donors for a specific recipient based on the presence of anti-HLA antibodies pretransplant. The higher the cPRA, the more sensitized the patient is, and the less likely he or she will be offered an organ. Additionally, patients with a high cPRA are more likely to develop ABMR posttransplant compared with patients who have low cPRA. In fact, patients who developed acute ABMR had a high median pretransplant peak cPRA compared with those who did not experience ABMR. Of note, although some sensitized patients may undergo desensitization protocol pretransplant, they still remain more vulnerable to developing ABMR [Kim et a/. (2014) Pharmacotherapy. 34, 733-44].

Hyperacute ABMR is caused by a high presence of DSAs in a recipient at the time of transplantation. The diagnosis of hyperacute rejection typically relies on the timing of rejection, which occurs within minutes to hours after cross-clamps are released and the allograft is reperfused with blood. The allograft experiences severe cortical necrosis and thrombosis in the microvasculature, and in most cases, the allograft must be removed to avoid complications related to such a profound immunologic response. However, the incidence of hyperacute rejection in current practice is extremely low because of ABO antigen verification of donor and recipient and improved tissue typing methods conducted prior to transplant [Williams et al. (1968) N Engl J Med 12, 611-618; Racusen 8i Haas (2006) Clin J Am Soc Nephrol 3, 415- 420].

Acute ABMR is mediated by either DSAs that were present pretransplant or de novo DSAs that developed posttransplant. Early acute ABMR is usually seen days to weeks after transplantation, but acute ABMR can occur any time posttransplant. One study reported a case of late acute ABMR that occurred 17 years posttransplant [Halloran et at. (1990) Transplantation 1, 85-91]. Histologic findings in acute ABMR are similar to hyperacute rejection, but the severity of rejection is lower. Late acute ABMR seems to be frequently accompanied by cellular rejection features [Racusen & Haas (2006) M. Clin J Am Soc Nephrol 3, 415-420]. Studies have reported that ~5-7% of all kidney transplant recipients develop acute ABMR [Takemoto et al. (2004) Am J Transplant 7, 1033-1041]. However, the reported incidences of ABMR vary depending on factors such as the proportion of patients with preformed DSAs, detection methods and interpretation of pathohistologic findings, and the immunosuppression protocol utilized. Among sensitized patients, the incidence of acute ABMR is as high as 55% [Burns et a/. (2008) Am J Transplant 12, 2684-2694]. ABMR constitutes about a fifth to half of acute rejection cases, and it has a worse prognosis than cellular rejection [Colvin & Smith, cited above] .

Chronic ABMR develops slowly over months to years, and it is one of the important causes of chronic graft dysfunction [Colvin 8i Smith, cited above]. Chronic ABMR often causes irreversible allograft damage with a low graft survival rate and should not be confused with acute ABMR that occurs late posttransplant. In chronic ABMR, DSAs that do not lead to acute ABMR slowly activate the complement system and eventually cause histologic changes to the allograft that are distinguishable from acute ABMR and allograft dysfunction. The incidence of chronic ABMR is not known, but 60% of patients with late graft failure were found to have de novo DSAs months to years before their graft failure. In addition, concurrent cellular rejection is not uncommon in chronic ABMR [Colvin & Smith, cited above; Kim et at. cited above], "biological sample" refers to any sample taken from a subject, such as a serum sample, a plasma sample, a urine sample, a blood sample, in particular a peripheral blood sample, a lymph sample, or a biopsy. In typical embodiments, the sample is a peripheral blood sample.

Solid transplant typically refers to a kidney transplant. In other embodiments the transplanted organ can be heart, lung, liver, pancreas, or small bowel.

"expression profile" refers to the expression levels of a group of genes.

"reference expression profile" refers to a profile as obtained from a healthy subject with an solid organ transplant (such as kidney) who has been diagnosed as not having or not being at risk of developing a transplant rejection.

"housekeeping gene" refers to a gene that are constitutively expressed at a relatively constant level across many or all known conditions, because they code for proteins that are constantly required by the cell, hence, they are essential to a cell and always present under any conditions. It is assumed that their expression is unaffected by experimental conditions. The proteins they code are generally involved in the basic functions necessary for the sustenance or maintenance of the cell. Non-limiting examples of housekeeping genes include HPRT1, ubiquitin C, YWHAZ, B2M, GAPDH, FPGS, DECR1, PPIB, ACTB, PSMB2, GPS1, CANX, NACA, TAXI BP1 and PSMD2. "probes" or "set of probes" relates to oligonucleotides binding specifically to mRNA or cDNA of a target gene. Embodiments are a single probes on a micro-array binding to mRNA or cDNA, as illustrated in the below examples. Other embodiments are pairs of primers for PCR, or double pairs of primers for nested PCR. PCR using a pair of primers and an internal primer is used in e.g. Taqman PCR as illustrated in the examples. Primers can be in solution or suspension or coupled to a substrate. Primers are optionally labelled for example with a fluorescent label, magnetic label or radioactive label.

Expression level and increased expression compared to a reference can be described by determine the fold change values of the genes in the test sample.

The expression is increased if the fold change is larger than 1,5, typically larger than 2, preferably larger than 3 or 4. In the present invention a 2-gene molecular classifier is developed and validated for ABMR and a 2-gene molecular classifier for TCMR. The ABMR and TCMR molecular scores correlated with the severity of microvascular and tubulointerstitial inflammation and did not yield important differences in diagnostic performance between indication and protocol biopsies, illustrating their potential for granular and universal insights. In addition, the ABMR classifier had similar accuracy regardless of recipient HLA-DSA serostatus, confirming previous findings that ABMR histology represents a distinct molecular entity [Callemeyn et al. (2020) cited above]. In an external validation cohort, the classifiers retained good diagnostic accuracy, even providing a discriminatory capacity between ABMR and TCMR.

These are the sparsest molecular classifiers with preserved diagnostic accuracy that have been developed for any rejection phenotype to date. In comparison, the Edmonton group originally described cohort GSE36059 for the discovery of a 20-gene ABMR classifier and 30-gene TCMR classifier, which yielded slightly higher accuracies than the present 2-gene ABMR classifier and 2-gene TCMR classifier (85% vs. 79% for ABMR, and 89% vs 87% for TCMR) [Sellares (2013) cited above and Reeve (2013) cited above]. Recently, in a cohort of non-human primates, Adam et al. derived a 3- gene set for ABMR based on 34 gene transcripts using the NanoString platform, although this classifier dropped to an accuracy of 66% in the validation cohort GSE36059 [Adam et a/. Am J Transplant. 2017;17(ll):2841-50].

Starting from a high-dimensional microarray platform consisting of 54,675 probesets, several characteristics inherent to the transcriptional landscape of kidney transplant rejection allowed for this drastic reduction in the number of features. First, rejection biopsies associate with a transcriptional profile that is distinct from biopsies without rejection, which reduces the risk of bias when applying restrictive regression models. Second, pathogenesis-based transcripts associating with rejection display a high level of collinearity, as shown previously, and are therefore ideally suited for performant variable selection methods such as Lasso regularization, which aim to minimize the number of features [Mueller et al. (2007) Am J Transplant. 7(12), 2712-2722].

The molecular distinction between ABMR and TCMR is more challenging, owing to a large common background of inflammation-related signals and the frequent co occurrence of both phenotypes. Nevertheless, the TCMR classifier was able to distinguish TCMR from ABMR in the external validation cohort, and a similar trend was seen for ABMR. In addition, a distinct but lower molecular TCMR score was found in biopsies with borderline changes, corroborating the understanding of this histological category as an intermediate phenotype within the spectrum of cellular rejection. Taken together, the present results suggests that the ABMR and TCMR scores may be used as compound classifiers, thus enabling the molecular diagnosis of pristine, intermediate or mixed rejection subtypes using only 4 transcripts. Although variable selection methods do not necessarily isolate the most biologically relevant features, the discriminative capacity of these transcripts is a hallmark of a specific association with either rejection phenotype. Regarding the ABMR transcripts, GNLY, an activation marker of NK cells and cytotoxic T cells, and PLA1A, a phospholipid mediator upregulated in activated endothelial cells, have been associated in several studies across different organ systems [Hidalgo et al. (2010) Am J Transplant. 10(8), 1812-1822; Yazdani et al. (2019) Kidney Int. 95(1), 188- 98; Loupy et at. (2017) Circulation 135, 917-935] .For TCMR, an association was identified previously with IL12RB1 involved in downstream signaling of the pro- inflammatory cytokines IL-12 and IL-23 [Reeve et al. (2013) cited above], which is upregulated upon CD8+ T cell stimulation [Glassman etal. (2021) Cell. 184(4), 983- 999]. ARPC1B, involved in T cell proliferation and migration, has been reported in association with injury repair responses in murine kidney transplantation [Famulski et al. (2007) Am J Transplant 7, 2483-2495].

In addition to their diagnostic value, the ABMR and TCMR classifiers associated with allograft prognosis in the discovery cohort. The prognostic value of these molecular classifiers was further corroborated in a second external validation cohort reported by Einecke et al.(2010) J. Clin. Invest. 120, 1863-1872. Of note, a similar predictive performance for death-censored graft loss was reached with the present markers (time-integrated AUC 0.82 and 0.83 for the ABMR and TCMR classifier, respectively) as the specific molecular risk score derived by the authors in that same cohort (AUC 0.83) [Einecke et al. cited above]. The present analysis highlighted kidney transplants at increased risk for graft failure where the histological criteria for rejection were not fulfilled. In the binary Banff classification for ABMR, intermediate histological lesions that do not meet the thresholds for rejection are considered as No ABMR, but nevertheless may harbor ongoing and deleterious inflammation, which is associated with accelerated graft loss [Loupy et al. (2020) cited above; Callemeyn et al. (2021) cited above. The present invention demonstrates that molecular diagnostics can be utilized to further stratify these intermediate cases. In addition, the sparsity of these molecular classifiers requires less complex diagnostic assays, which could facilitate its implementation in transplant clinics.

The present invention identifies and validates an intragraft 2-gene ABMR classifier and 2-gene TCMR classifier for use as diagnostic and prognostic tools. Rigid variable selection strategies can yield sparse molecular classifiers for allograft rejection with preserved accuracy, which facilitates their interpretation and clinical implementation.

EXAMPLES

Example 1. Materials and methods

Population and data collection

224 renal allograft biopsy samples from four European transplant centers were included between June 2011 and March 2017 (University Hospitals Leuven, Belgium; Medizinische Hochschule Hannover, Germany; Centre Hospitalier Universitaire Limoges, France, and Hopital Necker Paris, France), in the context of the BIOMArkers of Renal Graft INjuries (BIOMARGIN) study (ClinicalTrials.gov number NCT02832661), and the Reclassification using OmiCs integration in KidnEy Transplantation (ROCKET) study. In these four clinical centers, protocol biopsies were performed at 3, 12, and sometimes at 24 months after transplantation, according to local center practice, in addition to the clinical indication biopsies (at the time of graft dysfunction). Institutional review boards and national regulatory agencies (when required) approved the study protocol at each clinical center. Each patient contributed one biopsy and gave written informed consent. Microarray gene expression data from two publicly available independent cohorts in the Gene Expression Omnibus of the National Institutes of Health were used, described originally by Sellares et al. cited above (GSE36059) and Einecke et al. (GSE21374), for external validation of the diagnostic and prognostic value of the molecular classifiers, respectively Einecke et al. (2010) J Clin Invest. 120(6), 1862- 1672.

Clinicopatholoqical assessment

Histological lesions were scored according to the Banff criteria by a local expert pathologist in each participating center. Pathological assessment of biopsies included in the BIOMARGIN study were confirmed by central pathology, independent from the original centre [Loupy et al. cited above; Roufosse et al. cited above]. The term "ABMRh" was used for biopsies that fulfilled the first two (histological) Banff 2019 criteria for ABMR, by the combination of glomerulitis, peritubular capillaritis, arteritis, thrombotic microangiopathy and C4d deposition, regardless of the HLA-DSA status. As per Banff 2019 guidelines, glomerulonephritis was considered as an exclusion criterion for ABMRh, as well as peritubular capillaritis without glomerulitis in the presence of borderline rejection, TCMR or polyoma-associated nephropathy. For the diagnosis of borderline (suspicious) for acute TCMR, the threshold of i > 0 in the presence of t > 0 was used. Microvascular inflammation severity was calculated as the sum of glomerulitis and peritubular capillaritis, and tubulointerstitial inflammation severity as the sum of tubulitis and interstitial inflammation.

HLA-DSA after transplantation were determined per local center practice. A possible presence of HLA-DSA was suspected at background-corrected MFI value around 500. At the time of an allograft biopsy, HLA-DSA positivity was determined by presence of pretransplant HLA-DSA and/or documentation of de novo HLA-DSA during follow-up after transplantation.

Biopsy samples and transcriptomic analysis

Two needle cores were taken at each kidney allograft biopsy. One was used for histomorphological assessment and at least half of the other was immediately stored in Allprotect Tissue Reagent® (Qiagen, Benelux BV, Venlo, The Netherlands). Sample processing was described in detail previously [Callemeyn et al. (2020) J Am Soc Nephrol. 31(9), 2168-2183] Briefly, fragmented cRNA was hybridized to the Affymetrix GeneChip Human Genome U133 Plus 2.0 Arrays (Affymetrix), which comprised of 54,675 probe sets covering the whole genome. The resulting image files (.dat files) were generated using the GeneChip® Command Console® Software (AGCC), and intensity values for each probe cell (.cel file) were calculated. The microarray data were handled in accordance with the MIAME (Minimum Information About a Microarray Experiment) guidelines.

Data analysis

The microarray data were analysed using Bioconductor tools in R (v3.5.3, www. rstudio.com) [Gentleman et a/. (2004) Genome Biol. 5(10), R80]. The robust multichip average method was performed on the raw expression data (.cel files) to obtain a log2 expression value for each probe set, and batch effect correction was performed for timing of the microarray analysis by use of the LIMMA package [Irizarry et al. (2003) Biostatistics 4(2), 249-264; Ritchie et al. (2015) Nucleic Acids Res. 43(7), e47]. The normalized expression data from this cohort were used for discovery of the molecular classifiers. Biopsies were classified into one of four categories: "no rejection", "TCMR", "ABMR", and "mixed rejection". For ABMR, both HLA-DSA negative and HLA-DSA positive cases of ABMRh were included, given their similar transcriptional profile, as previously described [Callemeyn et at. cited above]. Borderline changes were not included in the phenotype group of TCMR. These labels were translated into two sets of labels: a Boolean set of labels representing presence or absence of ABMR, where only biopsies classified as "ABMR" or "mixed rejection" were labeled with a positive (or "true") label, while "no rejection" and "TCMR" biopsies were labeled with a negative (or "false") label. The second set of labels corresponded to the TCMR counterpart. In both sets of labels, the number of samples with negative labels was much larger than that of samples with positive labels. To account for this imbalance, the data was oversampled using a borderline SMOTE algorithm, to ensure clear borders between positive and negative samples. Subsequently, the database and labels were split into training and test sets. The data was rescaled such that each feature in the training set had zero mean and unit variance. This scaling was later applied to the test set to prevent data leakage.

The 54,675 probesets, together with each of the two label sets were used to train two different classifiers, one to classify ABMR positive samples and another to classify TCMR positive samples. Classification was performed by using logistic regression. This classifier was used due to the simple, linear nature of the borders between classes to reduce overfitting, and for simultaneous classification and feature selection via Lasso regression. To assess the approximate diagnostic significance of different probesets for each rejection type, train/test data split and classification was performed over 200 realizations, and the probesets corresponding to non-zero coefficients were recorded for each realization.

The diagnostic accuracy of the ABMR and TCMR classifier was assessed using receiver operating characteristic (ROC) curves. The optimal threshold was derived from the discovery cohort, at the highest combination of sensitivity and specificity, and evaluated in the validation cohort GSE36059. Low and high thresholds were proposed based on respectively high negative and positive predictive values and independently evaluated in the validation cohort. The prognostic value of ABMR and TCMR scores was evaluated in the validation cohort GSE21374. Graft failure was defined as loss of kidney transplant function (i.e. return to dialysis or re-transplantation), and in case of death with a functioning graft, this was censored at the time of death. Kaplan- Meier survival curves were used to plot kidney allograft survival, and survival between groups was compared using log-rank testing. Univariable and multivariable Cox proportional hazard analyses were used to quantify the risk of rejection or allograft failure. To assess the discriminative performance of the molecular scores for graft failure after biopsy, time-dependent area under the ROC curves were calculated for a multivariable model incorporating biopsy timing, ABMR and TCMR scores. For clinicopathological features, nominal variables were compared utilizing Chi-Square test, or Fisher's exact test where appropriate. Comparison of continuous variables was performed by T test/ANOVA or Mann-Whitney U/Kruskal-Wallis test procedure for normal and non-normal distributed data, and post-hoc analysis was performed by the Tukey test and Dunn's multiple comparison test, respectively. Two-sided hypothesis tests with a significance level of less than 0.05 were considered significant. Python 3 was used with scikit learn and imblearn libraries for development of the molecular classifier, SAS (version 9.4, SAS Institute Inc., Cary, NC, United States) for further statistical analyses and GraphPad Prism (version 9.0, GraphPad Software, San Diego, CA, United States) for graphical presentation.

EXAMPLE 2. Demographics and histological description of biopsy cohort

The discovery cohort consisted of 224 kidney allograft biopsies pertaining to 224 recipients. ABMR was identified in 42 cases, TCMR in 9 cases and mixed rejection in 14 cases. An overview of the demographic and histological characteristics is presented in Table 1. As expected, biopsies with rejection were more often indication biopsies, and allograft function was worse compared to biopsies without rejection. Mixed rejection was identified earlier after transplantation than other phenotypical categories (median 14 days post-transplant vs. 369 days, 377 days and 372 days for no rejection, ABMR and TCMR, respectively).

Table 1 Demographic and histological features of the discovery cohort (N=224)

Values are noted as n (%), un ess otherwise specified. Percentages may not add up to 100% due to rounding. Denominator numbers less than the cohort size indicate incomplete information. P values are given for differences between No Rejection, TCMR, ABMR and Mixed Rejection. For ABMR, both HLA-DSA positive and negative cases of ABMR histology were included. IQR: interquartile range, eGFR: estimated glomerular filtration rate, mTOR: mammalian target of rapamycin. TCMR: T cell- mediated rejection, ABMR: antibody-mediated rejection. PVAN: polyomavirus- associated nephropathy, IFTA: interstitial fibrosis and tubular atrophy. EXAMPLE 3. Development of a molecular classifier for ABMR and TCMR

Both for TCMR and ABMR, the value of the C parameter, inversely proportional to the degree of U regularization imposed in the Lasso regression, was decided upon by calculating the mean accuracy of the classifiers as a function of the parameter, after 10 rounds of cross-validation (Figure 6). To avoid overfitting, the value of the regularization parameter for both ABMR and TCMR classifiers was set at C=0.015, the minimum value at which the TCMR classifier did not result in uninformed classification (indicated by an accuracy of 0.5, or random guessing). For the aforementioned Li regularization coefficient, the average number of non-zero coefficients across 200 realizations is reported in Table 2. The probesets associated to non-zero coefficie nts were recorded as well as their frequency of appearance over all realizations (Fig 7 and Tables 3-4). Table 2 Development of molecular classifiers for ABMR and TCMR deviation, AUC: area under the curve, Average number after 200 rounds of internal cross-validation

The final model was developed by selecting the random seeds which yielded the best ROC AUC and accuracy scores across the 200 cross-validation realizations. For ABMR, 2 probesets were selected corresponding to GNLY (37145_at) and PLA1A {219584_ at), which were both upregulated compared to absence of ABMR (Table 3).

Table 3 Differential gene expression for probesets withheld as classifiers for ABMR during 200-fold internal cross-validation in the discovery cohort The ABMR score was given by the logistic functions p _ABMR =[l+exp(- 0.0433-219584_at-0.1725-37145_at)] ¹ . Applied to the discovery cohort, the 2-gene ABMR classifier yielded a ROC-AUC of 0.89 (95% confidence interval, 0.84-0.94, Fig 8A) and accuracy of 80% for an optimal cutoff of 0.505 (Fig 8B-C). For TCMR, 2 probesets were selected corresponding to IL12RB1 (155284_at) and ARPC1B (201954_at), which were both upregulated compared to absence of TCMR (Table 4). Table 4 Differential gene expression for probesets withheld as classifiers for TCMR during 200-fold internal cross-validation in the discovery cohort

The TCMR score was given by the function p _T c _MR =[l+exp(-0.0864-1552584_at- 0.1358-201954_at] ¹ . In the discovery cohort, the 2-gene TCMR classifier yielded a ROC-AUC of 0.93 (0.86-1.00, Fig 8D) and accuracy of 83% for an optimal cutoff of 0.495 (Fig 8E-F).

EXAMPLE 4. Sensitivity analyses of the ABMR and TCMR classifiers in the discovery cohort

In order to further assess the internal validity of the ABMR and TCMR classifiers, several sensitivity analyses were performed in the discovery cohort (Fig 1A-C and Fig 2A-C). ABMR scores correlated significantly with the severity of microvascular inflammation (Spearman correlation coefficient r=0.66, P <0.0001), and TCMR scores correlated with the severity of tubulointerstitial inflammation (r=0.39, P <0.001). The ABMR classifier retained a similar accuracy according to biopsy context and recipient HLA-DSA status. Although the sensitivity analyses for the TCMR classifier were hampered by low numbers in some subgroups, also the TCMR classifier retained its diagnostic value according to biopsy context and HLA-DSA status. For the diagnosis of cellular rejection, a composite of borderline changes and TCMR, the TCMR classifier was less performant (ROC-AUC 0.82, 95% Cl 0.74-0.89, Fig 9A). Indeed, the molecular TCMR score was lower in biopsies with borderline changes than histologically confirmed TCMR (0.48±0.04 vs. 0.54±0.04, P<0.001, Fig 9B). However, the TCMR score was higher in biopsies with borderline changes compared to biopsies without cellular rejection (0.48±0.04 vs. 0.45 ± 0.04, P=0.006), supporting the understanding of borderline changes as an intermediate phenotype of cellular rejection.

EXAMPLE 5 External validation of diagnostic accuracy

In the external validation cohort GSE36059 (N=403 indication biopsies), the average ABMR score was higher in ABMR biopsies compared to No ABMR (0.48±0.04 vs. 0.44±0.03, P <0.0001, Fig 1D-E). The ABMR classifier reached a ROC-AUC of 0.80 (95% Cl 0.75-0.85, Fig IF), and an accuracy of 79.2% for the diagnosis of ABMR (Fig 10). The average TCMR score was higher in TCMR biopsies compared to No TCMR (0.47±0.05 vs. 0.42±0.04, P <0.0001, Fig 2D-E), with a ROC-AUC of 0.83 (95% Cl 0.77-0.89, Fig 2F) and accuracy of 87.8% for the diagnosis of TCMR (Fig 10). Reflecting the observations in the discovery cohort (Fig 3A-B), biopsies with TCMR or mixed rejection had a significantly higher TCMR score than biopsies with pure ABMR, suggesting that the TCMR classifier can also be applied to differentiate between rejection phenotypes (Fig 3D). Conversely, although the ABMR score was higher in ABMR and mixed rejection biopsies than TCMR, this trend did not reach statistical significance (Fig 3C). Similar to the discovery cohort, the TCMR classifier was less performant for the composite diagnosis of cellular rejection (ROC-AUC 0.74, 95% Cl 0.69-0.80, Fig 9C). The molecular TCMR score was lower in biopsies with borderline changes than histologically confirmed TCMR (0.43±0.04 vs. 0.47±0.05, P<0.001, Fig 9D), and tended to be higher compared to biopsies without cellular rejection (0.43±0.04 vs. 0.42 ± 0.04, P=0.06).

EXAMPLE 6 Prognostic value of the molecular classifiers

It was further investigated whether ABMR and TCMR scores contribute to the interpretation of equivocal histological cases. For this purpose, it was investigated whether molecular markers could aid in the prognostication of biopsies with lesions of humoral rejection, i.e. glomerulitis, peritubular capillaritis, intimal arteritis, c4d deposition and transplant glomerulopathy, that did not fulfill the Banff criteria for ABMR. In the discovery cohort, these "intermediate" cases had similar poor survival as biopsies with histology of ABMR according to the first two Banff criteria for ABMR (Fig 4A). Intermediate cases had a higher ABMR score compared to biopsies without humoral lesions, and a lower score compared to Banff ABMR (Fig 4B). However, the ABMR score varied considerably among intermediate cases, and several had an ABMR score above the optimal cut-off that was determined at 0.505 (n= 11/30, 36.7%). All biopsies with humoral lesions were assessed regardless of the Banff criteria, and stratified these according to the optimal ABMR molecular cut-off value (Fig 4C). In contrast to the histological cutoff points defined by the Banff ABMR criteria, the molecular score identified biopsies at higher risk for subsequent graft failure. A similar analysis was performed according to severity of cellular rejection, where biopsies with borderline rejection associated with intermediate survival rates and molecular TCMR scores (Fig 4D-E). Here, the optimal TCMR molecular cut-off value did not improve risk stratification for biopsies with cellular rejection (Fig 4F). These results suggested that the molecular classifiers could aid in refining the prognostic interpretation of a kidney allograft biopsy. To further corroborate this, the prognostic value of the ABMR and TCMR classifiers was examined in the external dataset GSE21374 (N=282). In a univariable survival analysis, both the ABMR score (HR 4.60 per 0.1 units increase, 95% Cl 2.77-7.63) and TCMR score (HR 3.25 per 0.1 units increase, 95% Cl 1.92-5.50) associated with graft failure (Table 5).

Table 5 Determinants of death-censored allograft failure in validation cohort GSE21374 (N=282 biopsies)

ABMR: antibody-mediated rejection, TCMR: T-cell mediatec rejection

In a time-dependent ROC curve analysis, the discriminative effect for graft failure remained stable during follow-up, with an integrated AUC of 0.82 for the ABMR score and 0.83 for the TCMR score (Fig 5A-B). To assess whether the molecular classifiers could add prognostic value to histological information, biopsies were stratified based on the presence of rejection and molecular scores. Biopsies without proven rejection but with an ABMR or TCMR score above median had similar outcome as biopsies with proven rejection, suggesting that the molecular classifier may be used to identify biopsies at risk in the absence of clear histological rejection (Fig 5C-D). In a multivariable analysis incorporating ABMR score, TCMR score, biopsy timing and presence of rejection, only ABMR score (HR 3.82 per 0.1 units increase, 95% Cl 1.85- 7.86) and biopsy timing (HR 1.10 per post-transplant year, 95% Cl 1.07-1.14) remained significantly associated with graft outcome.

EXAMPLE 7. Cross-platform validation of molecular markers for rejection

In recent years, new sequencing technologies have emerged that allow for a faster throughput time using paraffine-embedded tissue. Among these, the Nanostring platform has been advocated as a promising tool to increase molecular diagnostics in the transplant clinic. This has led to the development of the Nanostring B-HOT panel, incorporating a set of genes that have been reported in association with transplant immunology. To validate this biological pipeline, molecular classifiers were derived based on the microarray data while only using B-HOT genes (Table 6).

Table 6 Probesets withheld as classifiers using only genes from the B-HOT Nanostring panel

In a specific embodiment of the invention the above table defines an alternative set of first genes and an alternative set of second genes, which is particularly suited for the nanostring technology. Here, PLA1A and GNLY were again withheld as molecular markers for ABMR, while IL12RB1 and CD1D were markers for TCMR. To test the biological validity of these markers, a cohort of 67 biopsies was analysed with varying rejection phenotypes that were analysed on the Nanostring platform (Figure 11). Importantly, the diagnostic accuracy was retained for ABMRh (AUC 0.82) and TCMR (AUC 0.87).