Validation and clinical outcome in assessing donor-derived cell-free DNA monitoring insights of kidney allografts with longitudinal surveillance (ADMIRAL) study

Lihong Bu, MD, PhD, Gaurav Gupta, MD, Akshta Pai, MD, MPH, Sanjiv Anand, MD, MS, Erik Stites, MD, Irfan Moinuddin, MD, Victor Bowers, MD, Pranjal Jain, MD, David A. Axelrod, MD, MBA, Matthew R. Weir, MD, Theresa K. Wolf-Doty, MS, Jijiao Zeng, PhD, Wenlan Tian, PhD, Kunbin Qu, PhD, Robert Woodward, PhD, Sham Dholakia, MD, DPhil, Aleskandra De Golovine, MD, Jonathan S. Bromberg, MD, PhD, Haris Murad, MD, Tarek Alhamad, MD


Introduction
The deployment of nucleic acid-based non-invasive biomarkers within routine clinical care reflects a paradigm shift in traditional monitoring after kidney transplant. Current clinical management of transplant relies on detection of functional injury (elevated creatinine), therapeutic drug monitoring and, selectively, screening for harmful donor specific antibodies (DSA). In the absence of clinical signs, clinicians seeking to identify subclinical allograft injury and intervene prior to development of irreversible damage, were forced to rely on invasive allograft biopsies, which have inherent limitations from sampling error and variation in interpretation. 1 Routine monitoring with donor-derived cell-free DNA (dd-cfDNA) after solid organ transplantation has been shown to accurately identify and characterize allograft injury, [1][2][3] correlate with pathologic findings, [4][5][6] and assess response to therapy including treatment of rejection. [7,8] Importantly, evaluation in dd-cfDNA have been demonstrated to occur ahead of clinically apparent organ injury. [9,10] Consequently, allograft monitoring with plasma dd-cfDNA levels can support non-invasive identification of pathologies including cellular and humoral allograft rejection, viral injury, and drug toxicity. [3,6] dd-cfDNA can also be employed in the setting of acute allograft injury to guide further diagnostic testing and assess improvement following clinical intervention. [7] The routine use of dd-cfDNA to detect, characterize, or exclude ongoing allograft injury is a valuable addition in current post-transplant surveillance.
While the effectiveness of dd-cfDNA has been established in clinical trials, its utility in routine clinical practice has not been well described. The ADMIRAL study (Assessing AlloSure Dd-cfDNA, Monitoring Insights of Renal Allografts with Longitudinal Surveillance, ClinicalTrials.gov ID NCT04566055), is a large, multicenter, observational cohort study of kidney transplant (KT) recipients monitored with dd-cfDNA for up to three years. The purpose of this study was to validate clinical trial data documenting the J o u r n a l P r e -p r o o f effectiveness of dd-cfDNA in identifying allograft rejection and subclinical changes in a real-world setting and evaluate the relationship between dd-cfDNA measurements and non-immune allograft injury.
Additionally, ADMIRAL aimed to characterize the relationship between elevation in dd-cfDNA and important predictors of long-term graft survival, including estimated glomerular filtration rate (eGFR) and formation of de novo donor-specific antibodies (dnDSA).

Study Population
1092 adult KT recipients across seven transplant centers were monitored with AlloSure dd-cfDNA (CareDx Inc., Brisbane, CA) as part of their standard of care. Data was collected between June 2016 and January 2020. An IRB waiver of informed consent was obtained, the study was performed in accordance with international standards and was not part of a larger study. Patients were managed prospectively with dd-cfDNA as part of post-transplant care where data captured was retrospectively examined. Clinical events (e.g., rejection, infection) and routine laboratory testing (creatinine, donor specific antibodies) were determined using the center's electronic medical records. A full list of data collected is provided in Supplementary Table S1. Patients who had contraindications to dd-cfDNA monitoring were excluded, this includes pregnancy, multiple organ recipients, monozygous twin to twin transplant, and patients with prior bone marrow transplantation. No exclusions from the analysis and no withdrawal of patients were made as the use of dd-cfDNA was medically necessary as part of the standard of care.

AlloSure dd-cfDNA methodology
dd-cfDNA was measured at regular intervals based on each center's standard of care practice and was used both as part of surveillance testing and acutely as a diagnostic aid in patients with clinically evident graft dysfunction. A list of center management protocols is provided in Supplementary Table S2. Venous J o u r n a l P r e -p r o o f blood was collected in Streck Cell-Free DNA BCT tubes and shipped to the central Clinical Laboratories Improvements Act (CLIA)-certified laboratory at CareDx, Inc. (Brisbane, CA). Details of the standardized specimen processing and analytical methods to determine the percentage of dd-cfDNA (AlloSure ® ) have been published. [11] The targeted next-generation sequencing assay employs highly polymorphic single nucleotide polymorphisms (SNPs) to quantify dd-cfDNA without need for separate genotyping of the recipient or the donor. [11]

Diagnosis of graft dysfunction and biopsy-defined rejection
Results of protocol surveillance and for-cause renal transplant biopsies were captured. Indications for 'forcause' biopsy included change in creatinine, worsening proteinuria, and/or development of dnDSA. Initial clinical management was performed based on local biopsy interpretation at the discretion of the patient's transplant provider. Biopsy reports were subsequently examined centrally by a single pathologist, masked to the dd-cfDNA score, for study analysis. Centrally interpreted biopsy results were reported using the Banff 2019 classification scheme. Banff lesion scores were recorded and discrepancies between local and central reporting were identified. If no Banff scores or clinical diagnosis was provided on the biopsy report, or if other pathologies were reported, these rejections were excluded for the purposes of the rejection analysis. In cases of disagreement, central interpretation was included in the analysis. Mixed rejection was captured and classified as antibody mediated rejection and the TCMR group did not include borderline. A detailed breakdown of the biopsy findings is provided in Supplementary Table S3. Other concomitant pathologic diagnoses, such as calcineurin inhibitor (CNI) toxicity, glomerulopathy, or acute tubular injury/acute tubular necrosis were also captured and used for the injury analysis. They were not included in the rejection analysis. For patients diagnosed with allograft rejection, the decision to treat was made according to each center's clinical protocol. eGFR changes, dnDSA and future rejection events J o u r n a l P r e -p r o o f were also captured, along with all dd-cfDNA levels that were drawn per each center's standard protocol, before, during, and after acute events.
A paired biopsy was defined as a biopsy occurring ≤30 days after dd-cfDNA measurement. This inclusion period reflects the logistical complexity of getting patients scheduled for and completing allograft biopsy.
Biopsy results were included in the analysis only if there was no intervention performed between the time of the dd-cfDNA measurement and biopsy. A histogram of days between dd-cfDNA sampling and biopsy is shown in Supplementary Figure S1.

Statistical Analyses
Descriptive statistics were used for patient demographics and distribution of dd-cfDNA measurements obtained from blood samples at the time of clinical events. In the analysis, the discriminatory power was considered at previously published thresholds of 0.5% and 1%, [3,4] to calculate the performance characteristics of the assay (Sensitivity, Specificity, NPV, PPV). Subsequently patients were categorized as high dd-cfDNA (≥ 0.5%) versus low dd-cfDNA (<0.5%) for further analysis.
Comparisons between the high and low dd-cfDNA groupings were evaluated via Fisher's Exact Test for categorical variables and Student's t-test for continuous variables. Nonparametric comparisons of dd-cfDNA cumulative distributions between dichotomized groupings were evaluated via Kolmogorov-Smirnov two-sample tests. The area under the receiver operating characteristic curve (AUROC) was used to determine the discriminating accuracy of dd-cfDNA and other parameters of interest. Cumulative distributions curves were used to examine the relationship between dd-cfDNA level and the clinical indication for the allograft biopsy (for-cause vs. surveillance).

J o u r n a l P r e -p r o o f
Multivariate logistic regression was used to determine which independent covariates were predictive of high dd-cfDNA measurements (Supplementary Table S4  Patients and samples were included into non-mutually exclusive groups for the purpose of analysis based on the data available as shown in the flow diagram in Figure 1. The subsets of the ADMIRAL cohort included in the correlational analyses for each of the questions has been outlined in this figure, including the total biopsies taken and the breakdown of results used for analysis.

dd-cfDNA and eGFR Analyses
Renal function was determined by eGFR calculated using the MDRD equation. dd-cfDNA and eGFR for each month was assessed where present, and then was partitioned into clusters as part of an unsupervised machine learning assessment to ascertain the relationship between eGFR and dd-cfDNA using Spearman rank correlation as an alternative to regression. If more than one eGFR measurement was available each month the average was taken. Wong et al, provided an update to the analytical variation and intra-patient variation of AlloSure dd-cfDNA. [12] This was used to calculate the serial delta/RCV between dd-cfDNA results associated with pathology using the methods outlined by Lund et al. [13] Analytical variation (CVa) was defined as 2.7%, intra patient variation (CVi) = 61% and the index of individuality = 0.23%. [12] K-means Clustering (KM), [14] was used, distinct clusters representing timepoints allowed the formation of time horizons from 0 to 3-years post-transplant. The machine learning algorithm, partitioned data into monthly clusters, pre-determined by minimizing sum of squared distance using key features such as ethnicity, sex, age at transplant, evidence of BK infection, dd-cfDNA score, presence of J o u r n a l P r e -p r o o f DSA, allograft rejection and creatinine. Intra-cluster noise reduction strategies were applied to exclude interference of detection limits value. Spearman rank correlation was then used to measure the degree of association between eGFR and dd-cfDNA, with the correlation coefficient applied to determine the strength of the relationship. The clusters generated provided 3 different time horizons for assessment: 0-4 months, 4-12 months, and 12-36 months. More information is provided in Supplementary Methods.

dd-cfDNA and dnDSA Analyses
The relationship between dd-cfDNA and development of dnDSA was assessed in patients with paired dd-cfDNA and HLA DSA testing (both tests drawn at the same time). All patients started with a no identified DSA and normal AlloSure (<0.2%). Patients were defined as dnDSA positive if there was evidence of new DSA detected at a level defined as positive by the local transplant program as part of the post-transplant surveillance. Reports were then centrally read. Mean Fluorescence Intensity (MFI) of >500 was agreed to be positive, for both HLA class 1 and class 2, and was used for this analysis. [15] Freedom from dnDSA was assessed using Kaplan-Meier analysis and once patients developed DSA they were censored. Patients were categorized as having high dd-cfDNA (any measurement ≥ 0.5%) or a low dd-cfDNA (all dd-cfDNA in timeline measurement <0.5%). A multivariate statistical model and Cox proportional-hazard was used to evaluate the association of dd-cfDNA with the development of dnDSA (Supplementary Table S5 for model variables).

Quiescence and Allograft Injury Assessment
The value of dd-cfDNA as a marker of quiescence was retrospectively assessed using both biopsy and dnDSA measurement through longitudinal observation. Allograft quiescence was defined as the absence of "injury". Injury included out of range tacrolimus level (<4 ng/mL, >12 ng/mL), BK viremia, dnDSA J o u r n a l P r e -p r o o f positive, UTI, proteinuria, allograft rejection or recurrent focal segmental glomerulosclerosis (FSGS), as confirmed by paired biopsy ≤30 days after dd-cfDNA measurement.

Results
The demographic characteristics of the 1,092 ADMIRAL study patients are largely like the US adult transplant population reported to the United Network of Organ Sharing (UNOS) registry ( Table 1). The ADMIRAL cohort was comprised of a numerically higher percentage of African American recipients (28% vs 24%; p=0.78) and fewer re-transplant candidates (8% vs 13%; p=0.16). There was also a higher proportion of deceased donor recipients in this study compared with the UNOS registry (94% vs. 68%; p=0.04).

Association of dd-cfDNA level and acute rejection
The analytic sample included 5,873 dd-cfDNA measurements from 1,092 patients.  Table 2 while data on rejection is summarized in Supplementary Table S3. 16% of local biopsies were rescored by the central pathologist.
There was no statistically significant difference in the median creatinine in patients with a No Rejection biopsy (1.38 mg/dL; 95% CI 1.26-1.67 mg/dL) and patients with Banff defined Rejection (1.57 mg/dL; 95% CI 1.1-2.2 mg/dL), p = 0.096. The AUROC for creatinine was 0.492 (95% CI 0.38-0.59). In comparison, the J o u r n a l P r e -p r o o f median dd-cfDNA level among patients with a No Rejection biopsy was 0.23% (95% CI 0.22-0.38%) which was significantly lower than the median dd-cfDNA in patients with biopsies demonstrating defined cellular or humoral rejection (1.6%; 95% CI 1.1-3.7%), p<0.0001. The AUROC for all rejection dd-cfDNA was 0.798 (95% CI 0.72-0.87), which was significantly higher than the AUROC of creatinine; p<0.001 (Figure 2). The Youden's index for dd-cfDNA was 0.69%. dd-cfDNA levels differed significantly between patients with ABMR and TCMR, p<0.001. ABMR was J o u r n a l P r e -p r o o f dd-cfDNA discriminates between biopsies showing no rejection, any rejection, ABMR and TCMR biopsies (Table 3). Test characteristics differed by diagnostic threshold (0.5% vs. 1%) and identified pathology (any rejection, ABMR, TCMR). A 1% increase of dd-cfDNA was associated with a 3.3-fold increase in the risk of any rejection (p<0.001), with an overall rejection hazard ratio of 1.89 (95% CI 1.78-2.1).

Association of dd-cfDNA elevation and eGFR progression
The median number of eGFR and dd-cfDNA results per patient was 11 (

Relationship between dd-cfDNA level and identification of dnDSA
961 patients with paired dd-cfDNA and HLA DSA results had no pre-existing DSAs. The median calculated panel reactivity antibody (cPRA) was 37% (IQR 11-77%). The median number of paired DSA and dd-cfDNA samples per patient was 5 (IQR 3-9). Of these patients, dnDSA were found in 44 (4.6%) patients, 19 with class I and 25 with class II, 9 of whom also had histologic evidence of allograft rejection. dd-cfDNA above 0.5% was associated with a nearly 3-fold elevation in the risk of future dnDSA formation (HR 2.71, p=0.001) (Figure 4). In a multivariable analysis, every 1% increase in the dd-cfDNA level was associated with a 20% Furthermore, dd-cfDNA remained elevated in all cases with measurable dnDSA.

Association of dd-cfDNA elevation and graft injury
A composite state of "graft injury" defined as one or more of the following events: tacrolimus level (<4 ng/mL, >12 ng/mL), BK viremia, dnDSA positive, urinary tract infection, proteinuria, allograft rejection or recurrent focal segmental glomerulosclerosis (FSGS) was identified in 467 patients. dd-cfDNA was measured up to 30 days ahead of injury event. Another subset of 180 patients without any of these events or evidence of kidney allograft injury were grouped under "Quiescent" category. Shown in Figure 5 Table 4, a dd-cfDNA threshold value of 0.5% has a PPV of 77.5% and NPV of 71.6% for graft injury. In addition to the absolute value, the delta/RCV in dd-cfDNA was associated with allograft injury. A median increase of 149% (IQR 94-161) between serial results is indicative of graft injury (p=0.02).

4.Discussion
The large, multicenter, ADMIRAL cohort study independently validated the observation that dd-cfDNA detects both clinically evident and subclinical ABMR and TCMR in a real-world application of dd-cfDNA monitoring. dd-cfDNA was significantly more predictive of ongoing graft injury than the current standard of care measures of serum creatinine. In addition, elevated dd-cfDNA was associated with declining eGFR The median dd-cfDNA in patients with Borderline TCMR was 0.23%, with wide Confidence Intervals, suggesting heterogeneous injury within this diagnosis. Furthermore, many Borderline rejections are being treated by transplant providers without clear evidence of clinical benefit. Similar findings have been reported with histology diagnosing significantly more Borderline TCMR than tissue-based gene transcript assessment and median dd-cfDNA of 0.33%. [16] The delta/RCV between serial dd-cfDNA was also associated with clinically significant events including dnDSA formation and allograft injury. These results suggest the need to consider a deviation from J o u r n a l P r e -p r o o f baseline, in combination with an elevation of dd-cfDNA above a threshold of 0.5%, to identify significant graft injury. This finding has previously been shown by Stites et al. [4] While a measured level above 0.5% and/or the increase of 149% from baseline does not definitively prove injury, these changes suggest patients should have intensive surveillance, further diagnostic study, and/or potential intervention. Given the optimal threshold for allograft rejection was determined at 0.69%, the relative change of dd-cfDNA is very important to consider in combination with the absolute number.
The ADMIRAL study confirmed the correlation between dd-cfDNA level and rejection established by The

Circulating Donor-Derived Cell-Free DNA in Blood for Diagnosing Acute Rejection in Kidney Transplant
Recipients (DART) study (ClinicalTrials.gov Identifier: NCT02424227). [3] In the DART study, a 1% threshold was used to discriminate between rejection and no-rejection. ADMIRAL suggests that interpretation of serial change in dd-cfDNA level is also important in the interpretation of injury. These new data suggest that considering a median dd-cfDNA elevation of 149% from baseline signals a change from quiescence to potential injury. For most patients this seems to be an absolute elevation from baseline of at least 0.24% (IQR: 0.19%-0.39%). In other studies, Anand et al. demonstrated that an increase in dd-cfDNA of at least 141% was associated with abnormal pathology, supporting the 149% threshold reported here. [17] These data suggest that routine post-transplant surveillance with dd-cfDNA, which utilizes both serial changes and absolute thresholds (e.g., 0.5%), will increase the sensitivity to detect addressable injury in a timely fashion and in the absence of clinical symptoms. [18] Allograft injury is multifactorial with pathology other than alloimmune damage resulting in dd-cfDNA elevations. Allograft damage can result from recurrent disease, calcineurin inhibitor toxicity, or infection, each of which requires directed intervention. Therefore, the ability to discriminate allografts free from clinical and subclinical injury is very important, broadly "allograft quiescence". dd-cfDNA less than 0.5% J o u r n a l P r e -p r o o f was strongly correlated with allograft quiescence, potentially reducing the need for an invasive procedure.
Conversely, elevations in dd-cfDNA were specific and predictive of the study composite diagnosis of allograft injury. Thus, routine monitoring with dd-cfDNA may allow clinicians to risk stratify posttransplant patients, identify those with graft injury in need of potential further intervention, and those without injury who may benefit from reduction in immunosuppression to avoid long term drug induced comorbidity. [19] Development of dnDSA has been correlated with decreased allograft survival, even in the absence of clinically evident ABMR. [20] However, while many patients develop dnDSA, not all dnDSA results in significant allograft injury. In a recent prospective, multicenter study of 123 patients who were biopsied after the development of dnDSA in the absence of clinical rejection, only 41% had pathologic evidence of humoral rejection. [21,22] In a single center study from the Mayo Clinic, 967 patients were monitored with dnDSA screening and protocol biopsies. [23] At a median follow-up of 4.2 years, 7% of the patients developed dnDSA. 20% of patients had biopsy evidence of borderline or more severe acute cellular rejection and only 32.5% had evidence of either active or chronic active ABMR at time of dnDSA detection.
From the DART study, Jordan et al, identified 87 kidney transplant patients with 90 clinically indicated biopsies along with paired dd-cfDNA and DSA testing. [24] In patients with dnDSA with ABMR, the average dd-cfDNA was 2.9% compared with 0.34% in patients with dnDSA without ABMR, and 0.29% in patients without dnDSA. In this observational cohort, 60.7% of DSA positive patients did not have elevated dd-cfDNA, and therefore did not appear to have evidence of antibody mediated allograft injury. This supports previous findings where long-term allograft survival was not compromised in the setting of noncomplement binding DSAs. [25] These data suggest that dd-cfDNA may provide crucial incremental information which could complement dnDSA monitoring, by identifying clinical and subclinical ABMR in kidney transplant patients. Molecular sensitization as the causal injury that drives antibody formation J o u r n a l P r e -p r o o f remains an interesting prospect as does the concept of antibodies being absorbed by the allograft before being seen by Luminex, causing molecular injury. [26] Furthermore, the temporal observation between dd-cfDNA elevations and dnDSA warrants further investigation to assess both the etiology of dnDSA formation [26] and the potential for therapeutic intervention. Huang et al, previously demonstrated histologic features of ABMR in patients with elevations in dd-cfDNA that did not have any appreciable HLA antibodies. [27] In addition, non-HLA transplantation immunity revealed by lymphocytotoxic antibodies has been well published with Crespo et al showing the importance of AT1R in patients with ABMR DSA negative patients. [28,29] Thus, the utility of dd-cfDNA in the assessment of non-HLA DSA needs to be considered and, although not performed in this analysis, is planned from patients with stored serum. With the pathogenicity of non-HLA DSA still being determined, the use of dd-cfDNA in its assessment, may be a useful tool for future studies. [30] Both Clayton et al and Faddoul et al have reported that a decline in eGFR is superior to other surrogate measures of long-term for kidney transplant outcomes. [31,32] A 30% decline in eGFR between years 1 and 3 after kidney transplant is strongly associated with risks of subsequent death and death-censored allograft failure. [31,32] ADMIRAL extends our understanding of the correlation between changes in dd-cfDNA level and long-term graft outcomes. Higher levels of dd-cfDNA were correlated with subsequent declining eGFR (correlation coefficient -0.84) suggesting that early identification of injury before traditional functional changes occur could impact graft survival. The mechanisms of injury are clearly multifactorial but suggest that elevated dd-cfDNA may identify patients who would benefit from further investigation.
By using results from routine clinical care, our findings represent the largest prospective cohort of kidney transplant recipients undergoing surveillance with dd-cfDNA published to date. The limitations of this J o u r n a l P r e -p r o o f study primarily reflect its observational, real-world design. Comparison with UNOS data suggest that clinical determination did not bias inclusion of patients across these 7 centers and that this cohort truly represents the wider transplant population. As clinicians were unblinded with regards to dd-cfDNA measurements and other clinical data, clinical treatment may have altered the natural history of disease and affected the correlations reported. In addition, logistical constraints led to dd-cfDNA levels and biopsies not always being concurrently obtained. To account for these barriers, we allowed biopsies done up to 30 days after dd-cfDNA levels to be considered as paired results. While it is possible that subclinical rejection may have resolved prior to biopsy, this effect would most likely have biased the study toward the null finding and thus should not invalidate the findings reported here. Verification bias is a consideration as biopsies were performed locally and not all read or acted upon centrally. However, with data showing consistent patterns despite this heterogeneity, the results identify clear direction for future work. Missing values causing ascertainment bias in the absence of a control group in the prediction model analysis is also a consideration, but we feel the large sample size limits this, where longitudinal serial samples allow patients to be their own control. Another potential limitation is that testing is more frequently performed in the first year of transplant. Therefore, there is a natural ascertainment and selection bias as alloimmune injury and infection are more common during this period, however, this follows the routine clinic schedule so again reflects real life practice. Further investigation is needed to establish the optimal interval of monitoring as there is clear multifactorial value considering dd-cfDNA as part of the clinical assessment of the patient. Finally, heterogeneity of dd-cfDNA levels between patients, underlying pathology, effect of interventions impacting the degree of association between dd-cfDNA measurements, and clinical evidence need to be considered. In the future, Bayesian probability evaluation incorporating knowledge of the patient's past clinical course and current presentation, needs to be considered in modeling algorithms to reduce the impact of this heterogeneity.

J o u r n a l P r e -p r o o f
Disclosures Lihong Bu declares she has no conflict of interest; drafted or revised the manuscript, role in results analysis Gaurav Gupta serves on the scientific advisory board of CareDx; has received honoraria/grant support from Alexion, CareDx, Mallinckrodt, Natera, Veloxis, Gilead, NIH/NIDDK, Mendez Foundation; drafted or revised the manuscript, acquired data, role in results analysis Akshta Pai has received an educational research grant from CareDx; drafted or revised the manuscript, acquired data, role in results analysis Sanjiv Anand has received speaker honorarium from CareDx; drafted or revised the manuscript, acquired data, role in results analysis Erik Stites has received speaker and advisory board honorarium from CareDx; drafted or revised the manuscript, acquired data, role in results analysis Irfan Moinuddin declares that he has no conflict of interest; acquired data Victor Bowers declares that he has no conflict of interest; acquired data Pranjal Jain has received speaker and advisory board honorarium from CareDx and owns common stock in CareDx; acquired data and contributed to drafts Tarek Alhamad has received speaker and advisory board honorarium from CareDx; drafted or revised the manuscript, acquired data, role in results analysis         J o u r n a l P r e -p r o o f   J o u r n a l P r e -p r o o f