The Clinical Relevance of the Infiltrating Immune Cell Composition in Kidney Transplant Rejection

Visual Abstract

Abstract

Key Points

• The estimated composition of immune cells in kidney transplants correlates poorly with the primary rejection categories defined by Banff criteria.

• Spatial cell distribution could be coupled with a detailed cellular composition to assess causal triggers for allorecognition.

• Intragraft CD8temra cells showed strong and consistent association with graft failure, regardless of the Banff rejection phenotypes.

Background

The link between the histology of kidney transplant rejection, especially antibody-mediated rejection, T-cell–mediated rejection, and mixed rejection, and the types of infiltrating immune cells is currently not well charted. Cost and technical complexity of single-cell analysis hinder large-scale studies of the relationship between cell infiltrate profiles and histological heterogeneity.

Methods

In this cross-sectional study, we assessed the composition of nine intragraft immune cell types by using a validated kidney transplant–specific signature matrix for deconvolution of bulk transcriptomics in three different kidney transplant biopsy datasets (N=403, N=224, N=282). The association and discrimination of the immune cell types with the Banff histology and the association with graft failure were assessed individually and with multivariable models. Unsupervised clustering algorithms were applied on the overall immune cell composition and compared with the Banff phenotypes.

Results

Banff-defined rejection was related to high presence of CD8⁺ effector T cells, natural killer cells, monocytes/macrophages, and, to a lesser extent, B cells, whereas CD4⁺ memory T cells were lower in rejection compared with no rejection. Estimated intragraft effector memory–expressing CD45RA (TEMRA) CD8⁺ T cells were strongly and consistently associated with graft failure. The large heterogeneity in immune cell composition across rejection types prevented supervised and unsupervised methods to accurately recover the Banff phenotypes solely on the basis of immune cell estimates. The lack of correlation between immune cell composition and Banff-defined rejection types was validated using multiplex immunohistochemistry.

Conclusions

Although some specific cell types (FCGR3A⁺ myeloid cells, CD14⁺ monocytes/macrophages, and NK cells) partly discriminated between rejection phenotypes, the overall estimated immune cell composition of kidney transplants was ill-related to main Banff-defined rejection categories and added to the Banff lesion scoring and evaluation of rejection severity. The estimated intragraft CD8temra cells bore strong and consistent association with graft failure and were independent of Banff-grade rejection.

Introduction

Histological evaluation of kidney transplant biopsies remains a key component in the assessment of graft status. Experts have developed the Banff classification, a consensus rule-based system to classify cases into disease categories on the basis of lesion scores.^1–4 In this system, T-cell–mediated rejection (TCMR) is differentiated from antibody-mediated rejection (AMR) primarily on the basis of the localization of the immune infiltrates (tubulointerstitial versus microvascular) and on the presence/absence of circulating donor-specific antibodies (DSA) and complement C4d deposition as hallmarks of antibody involvement. In addition, biopsy-based molecular diagnostics discriminate between these histological phenotypes.⁵ Although convenient from a clinical and ontological point of view, the dichotomy between TCMR and AMR does not fully reflect the immunopathological reality.⁶

The link between the Banff-defined rejection phenotypes and the types of infiltrating immune cells is currently not well charted. Studying infiltrating immune cell types could improve our understanding of the rejection process in different rejection phenotypes, improve the diagnostic classification by including more causal information, and enable the development of targeted therapies. Multiplex immunofluorescence and immunohistochemical studies of kidney transplant biopsies provided the first suggestion of the mismatch between the rejection phenotypes and the composition of the infiltrating immune cells, with highly variable proportions of the cell subsets within the rejection phenotypes.^7,8 In addition, bulk transcriptomic⁹ and single-cell RNA sequencing studies¹⁰ demonstrate such heterogeneity in the immune cell repertoire of kidney transplant rejection, with an increasingly recognized role of innate immune cells.

Cost and technical complexity prohibit the use of single-cell studies in large enough cohorts, thus hindering the ability to study the relation between detailed cell infiltrate profiles and the rich heterogeneity in the histological presentation of kidney transplant rejection. However, as genome-wide transcriptomic analyses of kidney transplant biopsies are being implemented in clinical routines,¹¹ deconvolution of bulk transcriptomics data for estimating individual cellular populations enables the study of such larger sample sets.¹² We recently validated a kidney transplant–specific deconvolution matrix that can recover the proportion of intragraft cell types from their sequenced transcripts in bulk samples.¹⁰

We hypothesized that (1) there is a strong relation between the principal Banff rejection phenotypes (AMR, TCMR) and subtypes and intragraft immune cell burden and composition, (2) information on immune cell composition can further improve the diagnostic classification of kidney transplant biopsies because of a close relationship with the mechanisms of immune activation, and (3) immune cell composition contains prognostic information. To test these hypotheses, we assessed the composition of intragraft immune cells by using a recently validated kidney transplant–specific signature matrix for deconvolution of bulk transcriptomics,¹⁰ in three different kidney transplant biopsy datasets.

Methods

Datasets and Phenotypic Classification

Three datasets were used in this study to estimate the immune cell fractions from the bulk transcriptomic data. The first dataset was generated as part of the Biomargin study (BIOMArkers of renal Graft Injuries, ClinicalTrials.gov number NCT02832661) and is hereafter referred to as the Biomargin dataset. Its transcriptomics data are publicly available at the Gene Expression Omnibus (GSE)¹³ of the National Institutes of Health (www.ncbi.nlm.nih.gov/geo) under the number: GSE147089.¹⁴ The two other datasets are publicly available at the GSE under the numbers GSE36059 (N=403)¹⁵ and GSE21374 (N=282),¹⁶ respectively. Those transcriptomics datasets were chosen for their public availability and for their similarity in methodology (bulk microarray) in the specific field of kidney transplant rejection.

The three datasets used in this study exhibit discrepancies in variable availability, precluding their direct combination into a single unified dataset. As a result, and to mitigate the low prevalence of certain rejection categories in the individual datasets, the datasets were combined to increase the statistical power according to the following scheme: Combined Dataset 1 combines both the GSE36059 dataset (N=403) and the Biomargin dataset (GSE147089; N=224). It is used for analyses on the association between immune cells and main Banff categories. Combined Dataset 2 combines the Biomargin dataset (GSE147089; N=224) and the GSE21374 dataset (N=282) and is used for analyses on association with graft failure (see Supplemental Methods). Each biopsy from Combined Dataset 1 was classified into one of the following Banff categories: no rejection, AMR, TCMR, and mixed (T-cell–mediated and AMR together) rejection. The GSE36059 and Biomargin datasets did not follow the exact same classification about AMR: In Biomargin AMR, diagnosis was allowed even in the absence of demonstrated HLA-DSA, while GSE36059 seemed to use a more strict AMR classification only allowing HLA-DSA–positive AMR. As a result, to guarantee sound integration of both datasets, the HLA-DSA–negative AMR (n=21) and HLA-DSA–negative mixed rejection (n=9) biopsies were excluded from Combined Dataset 1 (N=596). Additional details on the histological Banff scoring and the phenotypic classification can be found in the studies by Reeve et al.¹⁵ and Callemeyn et al.¹⁴ for the GSE36059 and Biomargin datasets, respectively.

To confirm that our results were not driven by a single data cohort, the subsequent analyses have also been performed, when applicable, separately on each individual dataset.

Estimation of Cell Fractions from the Deconvolution of Bulk Transcriptomics

The estimated cell fractions for each bulk transcriptomics dataset were computed from the deconvolution of gene expression data with the CIBERSORTx algorithm¹² and the gene expression signature matrix KTB18 previously constructed from single-cell data¹⁰ (Figure 1 and Supplemental Methods). Additional information on the construction of and the genes involved in the signature matrix can be found in the study by Lamarthée et al.¹⁰ Two sets of cells were identified by this signature matrix: immune cells and kidney structural cells. The immune cells, as labeled by Lamarthée et al.,¹⁰ consist of CCR7⁺-naïve CD4⁺ T cells (CD4naive), CCR7⁻ memory CD4⁺ T cells (CD4mem), GZMK_low effector CD8⁺ T cells (CD8eff), GZMK_high effector memory CD8 T+ cells (CD8effmem), MKI67⁺ terminally differentiated effector memory CD8⁺ T cells (CD8temra), FCGR3A^+/− natural killer cells (NK cells, which combines both CD16_high and CD16_low, similar to ref. 10), CD14⁺ monocytes/macrophages (CD14⁺ mono/macro), FCGR3A⁺ myeloid cells (FCGR3A⁺ myeloid), and CD19⁺ B lymphocytes (B cells). The structural kidney cells consist of PLA2R1⁺ podocytes, LRP2⁺ proximal tubular cells, UMOD⁺ loop of Henle cells, TMEM213⁺ intercalated cells, TAGLN⁺ vascular smooth muscles (pericytes), PLVAP⁺ peritubular capillary cells, CLDN5⁺ vasa recta cells, and EMCN⁺ glomerular endothelial cells. The sum of these cellular fractions equals to 100% in each biopsy. All the cellular fractions mentioned throughout this study are expressed as percentage of all cell types, i.e., not relative fractions of immune or structural cell types. This allows us to report on quantification of the immune cell burden and composition of the immune cell subtypes, independent of the size of the biopsies.

Figure 1.

Overview of the immune cell estimation process with the deconvolution of bulk transcriptomics data and the multiplexed immunohistochemistry (MILAN). For the deconvolution, a signature matrix is first constructed on the basis of single-cell RNA data.¹⁰ This matrix consists of barcode genes (rows) pertaining to the individual cell types (columns). Note that the matrix also contains signature genes to identify structural kidney cell types, not represented here to simplify the figure. In a second stage, the same matrix is used to deconvolute the different bulk RNA datasets with CIBERSORTx software, resulting in estimation of cellular proportion for each individual biopsy. For the multiplexed immunohistochemistry (MILAN), the proportion of immune cells is directly observed from the images with the relevant markers. Detailed statistical methods are described in Supplemental Material. MILAN, Multiple Iterative Labeling by Antibody Neodeposition.

Observed Cell Fractions from Multiplexed Immunochemistry Data

To corroborate the results on the basis of the estimated cells from the deconvolution process, we analyzed the immune infiltrate using observed immune cells from spatial single-cell multiplexed immunohistochemistry on the basis of the Multiple Iterative Labeling by Antibody Neodeposition (MILAN) technique.¹⁷ This method uses a panel of immune cell markers to identify different cell populations, while providing a direct count of the observed cells. We used the MILAN dataset from the study by Lamarthée et al.,¹⁰ which consists of an external set of 18 biopsies (three no rejection, six AMR, four mixed rejection, and five TCMR) from kidney transplant recipients followed in the University Hospitals Leuven, Belgium. Using immune cell protein markers, Lamarthée et al.¹⁰ identified 17 cell types for a total count of 555,479 individual DAPI⁺ cells, including the following nine immune cell populations: macrophages, CD1c⁺ dendritic cells and S100+ dendritic cells, B cells, CD4 regulatory T cells (CD4reg), FcγRIII⁺ NK cells (NK cells), CD4 effector T cells (CD4eff) and CD8 effector T cells (CD8eff), and MPO⁺ neutrophil cells. These cell types are hereafter expressed as observed count over the total number of cells per biopsy. Additional details on the methodology can be found in Lamarthée et al.¹⁰

Association of Immune Cell Proportions with Banff Categories, Lesion Scores, and Clinical Features

For each biopsy, the extent of inflammation was defined as the sum of all estimated immune cell fractions. Differences in the mean cellular distribution between the rejection phenotypes and no rejection cases were assessed, as well as the association between the estimated immune cell proportions and the semiquantitative ordinal Banff lesion scores on the Biomargin dataset, for t, i, v, g, ptc, and C4d (acute lesions) and cg, ct, ci, cv, and ah (chronic lesions) (see Supplemental Methods). The activity index was computed as a reweighted sum of acute lesions, as defined by Vaulet et al.¹⁸

Relation between the Immune Cell Composition and Banff Phenotypes

Next, we performed principal component analysis on the nine estimated immune cell types, as well as unsupervised clustering algorithms. In addition, discrimination between no rejection and any rejection was evaluated with area under the receiver operating characteristic curve (AUC) and Polytomous Discrimination Index (PDI).¹⁹ This index is a natural extension of the AUC for multiclass outcome. It represents the probability of accurately identifying an instance from a set of k classes.¹⁹ Unlike AUC, it varies from 1/k (random guess, corresponding to the 0.5 baseline of AUC) to one (perfect discrimination). We report the performances with (baseline of 1/4=0.25) and without (baseline of 1/3=0.33) including the no rejection cases. To assess the overall ability of the set of immune cells to discriminate the rejection categories, we next developed multivariable models. All the metrics were corrected for optimism using the enhanced bootstrap procedure.²⁰ Details of these algorithms are provided in Supplemental Methods.

Association with Graft Failure

The association with graft failure was assessed in the Combined Dataset 2 (N=506) for each immune cell type with a Cox model, adjusted for time post-transplantation. The hazard ratios (HRs) are reported as overall measure of association between a cell type and graft failure.²¹ Kaplan–Meier estimators were used to visualize survival trend in discretized (e.g., with tertiles) immune cell variables.

Temporal Trajectory

The average temporal trajectory of each cell type of the Combined Dataset 2 (N=506) was modeled with unadjusted linear regression (i.e., immune cell proportion as a function of time). The need for nonlinear transformation of the time variable was assessed with partial F-tests against the simpler linear model. Nonlinear transformation was modeled with restricted cubic splines with four knots (placed at default quantiles: 0.05 0.35 0.65 0.95). Using splines allows us to model more flexible temporal trajectories when the data suggest a better fit for nonlinear temporal evolution rather than a simple straight line.

Results

Total Immune Cell Estimation versus Any Rejection and Rejection Subtypes

An overview of the three datasets and the demographic/descriptive variables is given in Table 1 and Supplemental Table 1. Biopsies in the Biomargin dataset (N=224) were on average earlier post-transplantation compared with the biopsies in the two other datasets (median time in months: 12.0 [interquartile range, 3.0–32.0] versus 17.0 [range, 0.2–428] and 21.2 [interquartile range, 5.0–68.7] for GSE36059 and GSE21374, respectively). Estimated immune cell distributions, including the total immune cells, were remarkably similar between cohorts, although the estimated NK cell fractions differed between three transcriptomics datasets (Table 2 and Supplemental Figure 1). The Biomargin dataset differed from the two other datasets in terms of B cell and CD4mem and CD8effmem cell distribution, while distributions were similar between the two other datasets (Table 2 and Supplemental Figure 1).

Table 1.

Descriptive statistics of the three bulk transcriptomics datasets

Variable	Individual Datasets			Combined Dataset 1	Combined Dataset 2
Dataset GSE number	GSE36059	GSE147089 (Biomargin)	GSE21374	GSE36059+GSE147089	GSE21374+GSE147089
N	403	224	282	596	506
Median time post-transplantation (range), mo	17 (0.2–428.0)a	12 (0.13–310.0)	21.2 (0.19–421.5)	NA (no details)	14.3 (0.13–421.5)
Indication biopsies, n (%)	403 (100)	85 (38)	NA	465 (78)	NA
HLA-DSA positive, n (%)	129 (36)	84 (37)	NA	213 (39)	NA
Banff diagnoses, n (%)
No rejection	281 (70)b	158 (70)	NA	439 (74)c	NA
HLA-DSA negative	198 (49)	111 (49)	NA	309 (52)	NA
HLA-DSA positive	41 (10)	47 (21)	NA	88 (15)	NA
AMRa	65 (16)d	42 (19)	NA	85 (14)d	NA
AMR HLA-DSA negative	3 (0.7)	18 (8)	NA	0 (0)	NA
AMR HLA-DSA positive	61 (15)	24 (11)	NA	85 (14)	NA
Mixed rejection	22 (5)	14 (6)	NA	27 (4)	NA
HLA-DSA negative	1 (0.2)	8 (3)	NA	0 (0)	NA
HLA-DSA positive	21 (5)	6 (3)	NA	27 (4)	NA
TCMR	35 (9)d	10 (4)	NA	45 (7)d	NA
HLA-DSA negative	28 (7)	3 (1)	NA	31 (5)	NA
HLA-DSA positive	6 (1)	7 (3)	NA	13 (2)	NA
Graft outcome, n (%)
No. of graft failures	NA	34 (15)	51 (18)	NA	85 (17)

The Combined dataset 1 (n=596) combined the GSE36059 and Biomargin datasets without the donor-specific antibodies–negative antibody-mediated rejection and mixed rejection cases because of different definitions of antibody-mediated rejection used in labeling the biopsies in those two cohorts (see Methods). The Combined dataset 2 (n=506) combined the GSE21374 and Biomargin datasets together. AMR, antibody-mediated rejection; DSA, donor-specific antibodies; NA, not available; TCMR, T-cell–mediated rejection.

^aMedian time retrieved from the study by Reeve et al.¹⁵

^bMissing donor-specific antibodies: 42 (10.4%).

^cMissing donor-specific antibodies: 42 (7.0%).

^dOne missing donor-specific antibodies.

Table 2.

Immune cell distributions per cohort, expressed as % of total cell counts per biopsy

Median (Range) in %	GSE36059 (n=403)	GSE147089 (n=224)	GSE21374 (n=282)	Combined Dataset 1 GSE147089+GSE147089 (n=596)	Combined Dataset 2 GSE21374+GSE147089 (n=506)
B cells	2.1 (0.1–19.6)	1.3 (0.4–13.0)	2.3 (0.0–15.3)	1.6 (0.1–19.6)	1.5 (0.0–15.3)
CD14+ mono/macro	3.0 (0.1–21.4)	2.3 (0.0–25.5)	3.2 (0.0–19.5)	2.8 (0.0–25.5)	2.8 (0.0–25.5)
CD4mem	11.1 (0.0–30.5)	7.7 (0.0–39.6)	12.7 (0.0–37.1)	9.9 (0.0–30.5)	10.1 (0.0–39.6)
CD4naive	0.0 (0.0–9.8)	0.0 (0.0–11.9)	0.0 (0.0–17.6)	0.0 (0.0–11.9)	0.0 (0.0–17.6)
CD8eff	1.3 (0.0–33.8)	5.5 (0.0–23.1)	0.3 (0.0–26.1)	2.4 (0.0–33.8)	2.4 (0.0–26.1)
CD8effmem	0.3 (0.0–25.4)	0.0 (0.0–8.4)	0.0 (0.0–21.9)	0.0 (0.0–25.4)	0.0 (0.0–21.9)
CD8temra	1.5 (0.1–9.6)	2.0 (0.2–5.5)	1.4 (0.0–7.4)	1.6 (0.1–9.6)	1.7 (0.0–7.4)
FCGR3A+ myeloid	0.9 (0.0–10.1)	0.8 (0.0–18.9)	0.8 (0.0–11.0)	0.8 (0.0–10.6)	0.8 (0.0–18.9)
NK cells	2.9 (0.5–7.1)	2.0 (0.0–9.8)	4.2 (0.5–10.6)	2.5 (0.0–8.6)	3.3 (0.0–10.6)
Total immune cells	28.1 (10.5–85.0)	25.6 (12.8–78.7)	29.9 (11.3–82.9)	27.3 (10.5–85.0)	27.8 (11.3–82.9)

Density plots per cell type per cohort are provided in Supplemental Figure 1.

In both deconvoluted and MILAN datasets, the any rejection biopsies were more inflamed than no rejection biopsies (deconvoluted data: 35.5%±12.3% versus 26.3%±8.6%, t test P value < 0.001; MILAN data: 40.1%±8.9% versus 21.0%±4.3%, Mann–Whitney P value: 0.002), although a significant overlap was present in the total immune cell estimation between any rejection and no rejection. No biopsy, in both deconvoluted and MILAN datasets, had a complete absence of inflammation (Figure 2, A and B); the predicted fraction of immune cells was at least 10.5%. Among the rejection categories, TCMR demonstrated, on average, the highest degree of inflammation (42.8±13.0% and 46.1±7.7% for the deconvolution and MILAN data, respectively) while DSA-positive AMR showed an average degree of inflammation closer to no rejection cases (mean±SD 30.6%±9.4% on the deconvoluted data) (Figure 2C). Similarly, the degree of inflammation for AMR cases was intermediate between no rejection and mixed rejection cases (33.4%±5.5%) (Figure 2D). For both deconvoluted and MILAN datasets, the total inflammation in (DSA-positive) mixed rejection cases fell in between (38.9%±12.2% and 42.7%±9.6%, respectively). Within each diagnostic category, including no rejection, the degree of inflammation showed a large interbiopsy variability. On the deconvoluted data, the extent of inflammation demonstrated some ability to discriminate between any rejection cases and no rejection cases (AUC, 0.75; 95% confidence interval [CI], 0.70 to 0.79), although the performance varied depending on the pair of Banff rejection phenotypes to discriminate (Figure 2E). For instance, the discrimination performance of the extent of inflammation was higher between no rejection and TCMR (AUC, 0.88; 95% CI, 0.83 to 0.92) than between no rejection and DSA-positive AMR (AUC, 0.65; 95% CI, 0.59 to 0.70). The low number of biopsies in the MILAN dataset prevented us from performing similar analysis on multiplexed immunohistochemistry data.

Figure 2.

Extent of inflammation in the main Banff categories. (A) Extent of inflammation (%) as total number of estimated immune cells compared with the total number of estimated cells in each individual biopsy of the main Banff categories in Combined dataset 1 (N=596). (B) Extent of inflammation (%) as total number of counted immune cells compared with the total number of counted cells in each individual biopsy of the main Banff categories in MILAN Dataset (N=18). The dotted line represents the mean proportion of immune cells per Banff phenotype. (C) Overall distribution of the extent of inflammation per Banff phenotype in Combined dataset 1 (N=596). Pairwise comparisons are performed with Student t tests. (D) Overall distribution of the extent of inflammation per Banff phenotype in MILAN Dataset (N=18). (E) Pairwise discrimination performance of the total of estimated immune cells for the main diagnostic categories, measured with AUC (95% CI). AMR, antibody-mediated rejection; AUC, area under the receiver operating characteristic curve; CI, confidence interval; DSA, donor-specific antibodies; NR, no rejection; TCMR, T-cell–mediated rejection.

Immune Cell Composition versus Rejection and Rejection Subtypes, Individual Banff Lesion Scores, and Clinical Parameters

Next, we evaluated the relative cellular profiles, expressed as normalized mean differences with the average no rejection profile, for each rejection category. DSA-positive mixed rejection cases were characterized by the largest difference compared with no rejection in the following cell types: NK cells (+1.7%, P < 0.001), FCGR3A⁺ myeloid cells (+2.6%, P < 0.001), B cells (+2.2%, P < 0.001), and CD4mem cells (−5.1%, P < 0.001), although large overlaps existed between the rejection subtypes (Figure 3A and Supplemental Table 2). TCMR demonstrated the highest extent of inflammation (total immune cells) (+16.5%, P < 0.001) and the largest increase in CD4naive (+1.8%, P < 0.001) and CD14⁺ mono/macro (+5.9%, P < 0.001) compared with no rejection (Figure 3A and Supplemental Table 2). CD4mem cells were the only cells whose fractions were higher in no rejection than in any rejection categories (on average +3.1%, P < 0.001). CD8eff cells were highest in TCMR and DSA-positive mixed rejection, less in the absence of an AMR component. Compared with the other cell types, the CD8effmem and CD8temra cells barely demonstrated any significant difference between no rejection and the main rejection types. Except for CD4mem and NK cells, the magnitude of the differences compared with no rejection was smaller for DSA-positive AMR than for TCMR and DSA-positive mixed rejection (Supplemental Table 2). Sensitivity analyses in each original dataset separately confirmed the main trends observed in the combined cohort, apart from the B cells, which differed between the Biomargin dataset and the GSE36059 dataset (Supplemental Figure 2).

Figure 3.

Association of the immune cell infiltration with the Banff histology. (A) Difference in normalized mean cellular contents between the rejection phenotypes and the no rejection category in the Combined dataset 1 (N=596). The immune cell types are quantified as fractions of all cell types per biopsy. DSA+ mixed rejection cases showed significant differences in various cell types compared with no rejection: NK cells (+1.7%, P < 0.001), FCGR3A⁺ myeloid cells (+2.6%, P < 0.001), B cells (+2.2%, P < 0.001), and CD4mem cells (−5.1%, P < 0.001). TCMR displayed the highest extent of inflammation (total immune cells) (+16.5%, P < 0.001), as well as notable increases in CD4naive (+1.8%, P < 0.001) and CD14⁺ mono/macro cells (+5.9%, P < 0.001). CD4mem cells were the only cells whose fractions were higher in no rejection than in any of the rejection categories (on average +3.1%, P < 0.001). CD4mem cells were the only cell type more abundant in no rejection compared with any of the rejection categories (+3.1% on average, P < 0.001). CD8effmem and CD8temra cells exhibited minimal differences between no rejection and the main rejection types. The 95% CI were constructed with bootstrap (m=2000 replications). The corresponding non-normalized mean differences and statistical tests are reported in Supplemental Table 2. (B) Difference in normalized mean cellular contents between the rejection phenotypes and the no rejection category in the MILAN dataset (N=18). The immune cell types are quantified as fractions of all cell types per biopsy. The effector cells (CD4eff and CD8eff), CD1c⁺ dendritic cells, neutrophils, and the total of immune cells demonstrate increased proportion in rejection phenotypes compared with no rejection. The CD8eff and the total of immune cells patterns are similar to those observed on deconvoluted data (A). Similarly, the NK cells also demonstrate a lower proportion in TCMR than in AMR or mixed, as found in deconvoluted data (A). CD4reg demonstrated a similar negative trend in rejection phenotypes compared with no rejection as CD4mem on deconvoluted data. The corresponding non-normalized mean differences and statistical tests are reported in Supplemental Table 3. (C) Association between the immune cell proportions, as a fraction of all cell types per biopsy, and the histological ordinal Banff lesion scores on the Biomargin dataset (N=224). The number represents the Kendall tau with the following significant levels: *<0.05, **<0.01, and ***<0.001. CD14⁺ mono/macro cells, FCGR3A⁺ myeloid cells, NK cells, and the total proportion of immune cells have a positive correlation with the set of acute lesion scores (i, t, v, g, ptc, and C4d). Tubulointerstitial lesions (t) and interstitial lesions (i) are mostly associated with CD4naive and CD8effmem cells, while CD4mem cells show a negative association with these lesions. By contrast, the different immune cell types demonstrate a weak association with the chronic lesions, except for B cells, which are significantly associated with all chronic lesion scores. (D) Association between the immune cell proportions, as a fraction of all cell types per biopsy analyzed using multiplexed immunohistochemistry, and the histological ordinal Banff lesion scores on the MILAN dataset (N=18). The number represents the Kendall tau with the following significant levels: *<0.05 and **<0.01. The effector cells (CD4eff and CD8 eff), neutrophils, and the total proportion of immune cells, and to a lesser extent B cells, macrophages, and S100+ dendritic cells, demonstrate a strong significant association with tubulointerstitial lesions (t) and/or interstitial lesions (i). NK cells are mostly associated with the g lesion and not with tubulointerstitial lesions, displaying a similar pattern as shown in the deconvoluted data. Similar to the CD4mem in deconvoluted data, the CD4reg cells show overall a negative correlation, although not reaching the significance threshold. As with the deconvoluted data, none of the chronic lesion scores were significantly associated with any of the observed immune cell types.

On MILAN data, the effector cells (CD4eff and CD8eff), CD1c⁺ dendritic cells, neutrophils, and the total of immune cells demonstrated increased proportion in rejection phenotypes compared with no rejection (Figure 3B and Supplemental Table 3). The NK cells also demonstrated a lower proportion in TCMR than in AMR or mixed rejection, as found on deconvoluted data (Figure 3, A and B). CD4reg demonstrated a similar negative trend in rejection phenotypes compared with no rejection as CD4mem on deconvoluted data (although not significative) (Figure 3B and Supplemental Table 3). Low numbers obviate performing robust statistical analyses on these MILAN data.

On the deconvoluted datasets, a group of immune cell types, dominated by CD14⁺ mono/macro, FCGR3A⁺ myeloid cells, and NK cells, was positively associated with the acute lesion scores (i, t, v, g, ptc, and C4d) (Figure 3C). CD4naive and CD8effmem cells were mostly associated with the tubulointerstitial lesions t and i, whereas CD4mem cells demonstrated a negative association with these two lesions. On the MILAN data, the effector cells (CD4eff and CD8eff), neutrophils, and total proportion of immune cells demonstrated strong significant association with tubulointerstitial lesions (t) and/or interstitial lesions (i) (Figure 3D). Overall, in both deconvoluted and MILAN datasets, the different immune cell types associated poorly with the chronic lesions (cg, ct, ci, cv, and ah), with the exception of B cells, which demonstrated a significant association with all chronic lesions scores on deconvoluted data (Figure 3, C and D).

Correlating the clinicopathological presentation of the biopsies with immune cell distributions, the strongest association between the estimated total immune cells was observed with the activity index (R: 0.38, P < 0.001, Supplemental Table 3). FCGR3A⁺ myeloid cells and CD14⁺ mono/macro demonstrated the highest association and correlation with the activity index (R: 0.57, P < 0.001 and R: 0.59, P < 0.001, respectively), followed by the NK cells (R: 0.44, P < 0.001). CD4mem cells demonstrated an inverse relationship with the activity index (R: −0.23, P < 0.001). Five cell types correlated with lower eGFR at the time of biopsy: CD14⁺ mono/macro, FCGR3A⁺ myeloid cells, NK cells, B cells, and CD8temra cells (Supplemental Table 4). These five cell types and total immune cells also significantly differed according to the indication/protocol status of the biopsy (Supplemental Figure 3A). Other continuous variables were not or only poorly correlated and/or associated with the estimated cell fractions (Supplemental Table 4).

The NK cells, FCGR3A⁺ myeloid cells, and CD14⁺ mono/macro were also significantly higher in the presence of DSA, whereas CD4mem were proportionally more present without DSA (Supplemental Figure 3B). The NK cells, FCGR3A⁺ myeloid cells, and CD14⁺ mono/macro populations, as well as the total immune cell counts, demonstrated significant differences on the basis of the C4d positivity status on a subset (n=115) of the Biomargin dataset (Supplemental Figure 3C). On the other hand, stratification of the immune cell types on the basis of the missing self indicator (Supplemental Methods) did not demonstrate significantly different distributions of the immune cells (Supplemental Figure 3D).

Univariable and Multivariable Discrimination Performance of Immune Cell Composition for Any Rejection

First, we performed an unsupervised analysis with principal component analysis built solely on the estimated immune cell fractions (Supplemental Figure 4), in relation to the Banff classification of the cases. These exploratory visualizations indicated a potential no rejection cluster, while the main Banff rejection phenotypes could not be separated using the immune cell estimates alone. Similarly, various unsupervised clustering algorithms applied on the same estimated immune cell fractions were not able to classify the biopsies according to the Banff phenotypes (adjusted R and index varying from 0.045 to 0.119, Supplemental Table 5).

In supervised analyses for the discrimination between no rejection and any rejection, every cell type demonstrated above-random ability to discriminate between no rejection and (any) rejection, with individual AUC ranging from 0.55 to 0.82 (Figure 4A). Best performance was for FCGR3A⁺ myeloid cells and CD14⁺ mono/macro with an AUC (95% CI) of 0.82 (95% CI, 0.78 to 0.86) and 0.79 (95% CI, 0.75 to 0.83), respectively, followed by the NK cells with an AUC of 0.69 (95% CI, 0.64 to 0.74) (Supplemental Table 6). A multivariable logistic regression model combining all the immune cell types had a corrected AUC of 0.84 for any rejection versus no rejection (95% CI, 0.81 to 0.87) (Table 3). Further analyses on the model's coefficients revealed that the predictions were mostly driven by four variables: CD8temra (odds ratio [OR], 0.63; 95% CI, 0.49 to 0.80), NK (OR, 1.27; 95% CI, 1.06 to 1.54), FCGR3A⁺ myeloid (OR, 1.88; 95% CI, 1.48 to 2.43), and CD14⁺ mono/macro (OR, 1.22; 95% CI, 1.09 to 1.38) cells (Supplemental Table 7). Additional variable transformation and feature selection (see Supplemental Methods) did not improve the discrimination performance (Table 3).

Figure 4.

Unadjusted discrimination performance of the individual cell type proportions in Combined dataset 1 (N=596). (A) Discrimination performance between no rejection and any rejection (TCMR, DSA+ AMR, or DSA+ mixed rejection), reported with AUC (95% CI); all the immune cells demonstrated a better discrimination performance than a random estimator (represented by the dotted line at 0.5 AUC). Both types of monocytes/macrophages had the largest discrimination ability, followed by the total immune cells. (B) Overall discrimination performance between the four Banff phenotypes, including no rejection, reported with the PDI. Because there are four different classes to discriminate, the PDI baseline, corresponding to a random guess, is equal to 1/4=0.25. The ranking of the cell types is similar to (A) with AUC, with both monocyte/macrophage types having the largest discrimination power, followed by the total immune cells. (C) Overall discrimination performance (PDI) between the three main rejection phenotypes only (TCMR, DSA+ AMR, DSA+ mixed rejection). PDI baseline is equal to 1/3=0.333. Once the no rejection cases are excluded, the total of immune cells is the best discriminator (although a large overlap is present with other individual immune cell types). The FCGR3A⁺ myeloid cells ranked lower, suggesting similarity of its distribution within the rejection categories. PDI, Polytomous Discrimination Index.

Table 3.

Discrimination performance of multivariable models to discriminate rejection and rejection phenotypes from the estimated cell composition

Metric Adjustment	Rejection versus No Rejection Discrimination		Rejection Phenotypes Discrimination
	Multivariable Logistic Regression Metric: AUC	Multivariable Logistic Regression+MFP Metric: AUC	Polytomous LR Including Non-Rejection Cases Metric: PDIa	Polytomous LR Excluding Non-Rejection Cases Metric: PDIb
Apparent performance (95% CI)	0.85 (0.82 to 0.89)	0.86 (0.82 to 0.89)	0.58 (0.53 to 0.63)	0.72 (0.64 to 0.80)
Corrected performance (95% CI)	0.84 (0.81 to 0.88)	0.83 (0.79 to 0.86)	0.52 (0.48 to 0.57)	0.63 (0.55 to 0.71)

Corrected pairwise AUCs (95% CI)
No rejection versus DSA-positive AMR			0.79 (0.74 to 0.84)	—
No rejection versus TCMR			0.90 (0.85 to 0.95)	—
No rejection versus DSA-positive mixed rejection			0.91 (0.86 to 0.96)	—
DSA-positive AMR versus TCMR			0.85 (0.78 to 0.92)	0.88 (0.78 to 0.91)
DSA-positive AMR versus DSA-positive mixed rejection			0.72 (0.63 to 0.81)	0.83 (0.66 to 0.84)
TCMR versus DSA-positive mixed			0.78 (0.68 to 0.88)	0.84 (0.67 to 0.88)

AMR, antibody-mediated rejection; AUC, area under the receiver operating characteristic curve; CI, confidence interval; DSA, donor-specific antibodies; LR, logistic regression; MFP, multivariable fractional polynomials; PDI, Polytomous Discrimination Index; TCMR, T-cell–mediated rejection.

^aBaseline Polytomous Discrimination Index: 0.25.

^bBaseline Polytomous Discrimination Index: 0.33.

Discrimination Performance of Immune Cell Infiltrates of Rejection Phenotypes

Individually, FCGR3A⁺ myeloid cells and CD14⁺ mono/macro demonstrated the best performance to discriminate between rejection phenotypes, including no rejection cases, both with a corrected PDI of 0.41 (95% CI, 0.39 to 0.43) (Figure 4B). When excluding the no rejection biopsies, the CD14⁺ mono/macro and NK cells had the highest discrimination performance, with a corrected PDI of 0.49 (95% CI, 0.42 to 0.55) and 0.47 (95% CI, 0.37 to 0.57), respectively (Figure 4C). The multivariable polytomous logistic regression model had a corrected PDI of 0.53 (95% CI, 0.478 to 0.57) when including the no rejection cases (compared with a baseline PDI of 0.25) and a corrected PDI of 0.63 (95% CI, 0.55 to 0.71) when excluding the no rejection cases (compared with a baseline PDI of 0.33) (Table 3). The pairwise corrected AUCs for the different combinations of outcomes are reported in Table 3. No rejection versus DSA+ mixed rejection cases was the easiest combination for the multivariable model to tell apart (AUC, 0.91; 95% CI, 0.86 to 0.96), whereas DSA-positive AMR versus DSA-positive mixed rejection appeared the most difficult to discriminate (AUC, 0.72; 95% CI, 0.63 to 0.81).

Temporal Trajectories and Association with Graft Failure in Combined Dataset 2

CD8temra cells and the total of immune cells did not show any significant time dependency (P values: 0.44 and 0.80, respectively) for the linear term and no significant nonlinear effect (P values partial F-statistic: 0.32 and 0.32, respectively) (Supplemental Figure 5). NK cells were modeled with a linear relationship (P value linear term: 0.002), while all the other cell types demonstrate significant nonlinear effects. We observed a temporary increase of CD4naive, CD8eff, and CD8effmem within the first 3 years post-transplantation, whereas conversely, CD4mem cells showed an initial temporary decrease (Supplemental Figure 5). B cells demonstrated an abrupt rise within the first 2 years post-transplantation.

Most of the cell types were positively associated with graft failure, even after correction for time post-transplant (Figure 5A). The estimated proportion of CD8temra cells was the most strongly associated with graft failure (HR, 1.75 [95% CI, 1.48 to 2.07], P value < 0.0001), producing a time-dependent AUC between 0.70 and 0.75 during the first 2 years after biopsy. This association remained strong in each separate cohort independently: HR: 2.00 (95% CI, 1.45 to 2.77, P < 0.001) in the Biomargin data and 1.74 (95% CI, 1.45 to 2.11, P < 0.001) in GSE21374 (Figure 5B and Supplemental Figure 6). This relation between CD8temra cell fractions and graft failure rates was independent of the Banff rejection categories because there was an even distribution of these cells across the rejection phenotypes (Figure 3 and Supplemental Table 2). FCGR3A⁺ myeloid cells and CD14⁺ mono/macro also significantly related to graft failure in both cohorts independently. CD4mem cells were the only cell type to demonstrate an inverse association with graft failure (HR, 0.95 [95% CI, 0.92 to 0.98, P = 0.001]), although this finding was not reproduced in the Biomargin data alone. Other cell types showed less robust associations with graft failure rates across the cohorts.

Figure 5.

Association with graft failure per immune cell type. The immune cell types are quantified as fractions of all cell types per biopsy. (A) HR from individual Cox models, adjusted for time post-transplantation for each immune cell type in Combined dataset 2 (N=506). CD8temra cells clearly dominate the ranking with a much stronger association (HR, 1.75 [95% CI, 1.48 to 2.073]). Note that although HR is unitless, it remains inherently linked to the original variable unit (% in this case), which provides direct interpretation of the HR. However, the comparison between cell types of different magnitude range is less apparent. (B) HR from individual Cox models, adjusted for time post-transplantation in the Biomargin (N=224) and GSE21374 (N=282) datasets. The association of CD8temra cells with graft failure remains largely superior to the other cell types in each individual dataset. Beside CD8temra, FCGR3A⁺ myeloid cells are the only cell type to remain significantly associated with graft failure in both subsets. (C) Kaplan–Meier curves for each tertile of the six individual cell types significantly associated with graft failure in Combined dataset 2 (N=506). Note the reversed ordering of the CD4mem Kaplan–Meier curves compared with the other cell types (reflecting its below-1 HR). HR, hazard ratio.

Discussion

Our study demonstrates that the estimated extent of intragraft inflammation (total of immune cells) from bulk transcriptomic datasets, despite a large heterogeneity across the biopsies, was significantly associated with kidney transplant rejection. No biopsy was without inflammation. However, some immune cell types alone had a better discrimination for rejection than the overall degree of inflammation. The degree of inflammation correlated strongly with the overall activity index,¹⁸ a validated estimator of total inflammation that is strongly associated with graft failure rates. Banff rejection relates to the high presence of CD8eff cells, NK cells, monocytes/macrophages, and to a lesser extent B cells. CD4mem cells are lower in rejection compared with no rejection. The degree of inflammation demonstrated moderate discrimination performance between the rejection phenotypes, with TCMR cases being more inflamed than AMR cases. The overall infiltrating immune cell composition could not accurately reproduce the main rejection phenotypes, which are based on the spatial distribution of the mononuclear cells in kidney transplant biopsies. Inferring causality in terms of cellular mediators (e.g., TCMR) from the histological picture is not in concordance with our finding of important heterogeneity in the immune cell composition of kidney transplant rejection phenotypes and strong overlap between these phenotypes. NK cells, FCGR3A⁺ myeloid cells, and CD14⁺ mono/macro associate significantly with DSA status and with graft dysfunction, whereas missing self did not relate to specific immune cell profiles. The immune cell composition could potentially yield predictive information for a more targeted therapy choice. Compared with the other cell types, the CD8effmem and CD8temra cells barely demonstrated any significant difference between no rejection and the main rejection phenotypes. However, we demonstrated that the estimated CD8temra cells alone bear prognostic value for graft failure, independently of the Banff phenotypes. These main results were consistent between the different cohorts studied.

Owing to the specificities of the markers used in the multiplexed immunohistochemistry data and the resolution of the signature matrix, not all the same cell types are identified on both the deconvoluted and MILAN datasets, which restricts more detailed comparison between them (for instance, no CD8temra cells are identified in the MILAN dataset). However, the main findings from the deconvoluted data are supported by the multiplexed immunochemistry data, although the strength of the conclusions is limited by the low number of biopsies in the MILAN dataset. In particular, the total amount of inflammation per main rejection phenotype was similar between both approaches (with TCMR being the more inflamed), while no immune cellular profile was strongly and uniquely associated with a given rejection phenotype.

Despite decent discrimination performance of the multivariable models when combining all the immune cell types together, the estimated immune cells alone are far from accurately depicting all the Banff categories, even under a supervised learning framework. This lack of concordance between the Banff classification and the immune cell composition, also demonstrated in earlier studies,^7,22–24 can be explained by several factors. First, in the Banff classification, no distinction is made between the mononuclear cell subtypes, while we demonstrate clear heterogeneity in the cell composition within Banff phenotypes. Second, the Banff classification uses additional spatial information to describe the location of mononuclear cells in the Banff lesion scores, e.g., whether the immune cells are in unscarred versus scarred areas. Such information is lost in the bulk transcriptomics data. The Banff lesion scores thus encompass very different information than the immune cell composition, explaining why the algorithms purely based on immune cell composition cannot reproduce the rule-based Banff system, which is based on cell localization. Future studies should further characterize the precise intragraft location of the infiltrating immune cell subtypes because it could provide an additional layer of information for better stratification. Third, we should keep in mind that the Banff classification has been developed from expert consensus using if-then-else clauses to navigate to the final diagnosis, a process that is completely different from most of the clustering algorithms.

Finally, it is possible that the heterogeneity in cell composition within Banff rejection phenotypes is explained by heterogeneity in the underlying mechanistic injury processes, not captured in the Banff system. There is increasing indication that the interplay of various allorecognition mechanisms is more complex than anticipated earlier and more complex than currently considered in the Banff classification.⁶ The lack of robust techniques to assess these additional mechanisms of allorecognition currently hampers the testing of this hypothesis. Our data suggest that the concept of missing self, recently demonstrated to be an independent risk factor for microvascular inflammation,^25,26 is not a primary explanation for the heterogeneity in NK cell composition within the phenotypes. Such lack of relation between cause (HLA-DSA status) and gene expression was also recently demonstrated in cases of microvascular inflammation.^14,27 Many other mechanisms of NK cell activation²⁸ or other types of innate allorecognition^29,30 could play a role but were not assessed. Finally, non-HLA antibody profiling is notoriously complex,³¹ and it is possible that autoreactivity plays a role in the phenotypes observed.³²

In contrast to the other immune cells, we confirm that CD8temra cells are strongly associated with graft outcome,^33,34 despite not being associated with Banff-defined rejection. Although FCGR3A⁺ myeloid cells and CD14⁺ mono/macro cells were also clearly associated with graft failure, the estimated proportion of CD8temra cells demonstrated the strongest association, even after correction for time post-transplantation and equally in both studied datasets. Estimated CD8temra cells are also correlated with eGFR and proteinuria. These findings suggest that CD8temra cells potentially hold important additional signal on the graft condition, currently not included in the Banff classification or in prognostication algorithms. A recent study has demonstrated the accumulation of CD8temra cells both in recipient blood and kidney allografts. The CD8temra cells displayed enhanced cytotoxic and migratory responses compared with CD8effmem cells and were less effectively suppressed by currently used immunosuppressive agents,³⁵ which could explain the strong association we observed with graft failure. Moreover, CD8temra cells are very versatile, both expressing donor-specific T-cell receptors and CD16 receptors able to bind antigen-antibody complexes.³⁴ CD8temra cells have also demonstrated pathogenic roles in other immune diseases, such as ANCA-associated vasculitis,³⁶ Sjögren syndrome,³⁷ and severe coronavirus disease 2019.³⁸ The exact role of these cells in human kidney transplant rejection is to be further explored, as recent data suggest the therapeutic potential of targeting these cells.³⁵

No rejection cases demonstrated a higher proportion of intragraft CD4mem cells compared with rejection phenotypes, which contrasts with all other immune cell types. Similar trend was observed for the CD4reg cells in the MILAN data. While a detailed characterization of the intragraft CD4⁺ T-cell population is necessary to better understand this apparent reversed association with rejection, several hypotheses can be formulated to describe this observation: (1) The overrepresentation of other immune cell types in rejection compared with a constant proportion of CD4mem (or CD4reg) cells would create a relative decrease of this cell type in rejection cases, in favor of CD8⁺ effector for instance; (2) lower CD4mem cells can result from the (sustained) induced immunodepression, which decreases the risk of rejection; and (3) the estimated CD4mem fraction might also contain CD4 T+ regulatory cells that are known to induce and maintain tolerance.³⁹ Finally, we observed that B cells displayed a clear temporal pattern with a drastic increase during the first years after transplantation and were clearly correlated with all histological chronic lesions. This is in line with previous studies that demonstrated an accumulation of B cells and plasma cells along with the development of chronic lesions such as atrophy-fibrosis.^40,41

Our study has several limitations. The rules of the Banff classification, based on observer-dependent semiquantitative histological lesion scores, are a topic of active discussion, which makes our reference standard of histology imperfect. Moreover, not all datasets included had sufficiently granular data to study all currently defined Banff phenotypes.³ Owing to incomplete Banff classification of certain biopsies between the publicly available datasets, our results cannot currently be extrapolated to DSA-negative, C4d-negative microvascular inflammation, or probable antibody-mediated rejection. The precise cellular composition of these groups warrants additional research. Second, the present results are based on the estimation of immune cell proportions from the deconvolution of bulk transcriptomics and do not reflect the exact cell numbers. However, the main findings from the deconvoluted data were also supported by the multiplexed immunohistochemistry data, although the strength of the conclusions is limited by the low number of biopsies. In addition, empirical measures on pseudo-bulk data demonstrated good correlation between the estimated cell count and the real cell proportion assessed by both single-cell RNA sequencing and multiplex immunohistochemistry.¹⁰ Nevertheless, the conclusions drawn should be considered within that specific framework of uncertainty. Currently, we consider the KTB18 signature matrix proposed by Lamarthée et al. as the best validated and useful specific matrix for deconvolution of bulk RNA expression data obtained from kidney transplant biopsies (including deconvolution of structural cells). Whether this matrix is also valid in other contexts (e.g., native kidney diseases) remains to be studied. Third, the number of identified immune cell populations is bound to the resolution of the kidney transplant–specific signature matrix previously developed.¹⁰ Whether more granular subtyping of the immune infiltrate could reveal further relevant associations remains an open question. Fourth, the lack of information on the spatial distribution of the immune cell infiltrates hampers the correlation with the Banff lesion scores, which are entirely related to cell localization. To this end, spatial transcriptomic studies could help map the relationship between the histological image and immune cell types, their activation states, their specific spatial distribution, their proximity to and interactions with other cells, and the underlying mechanisms and biology of allorecognition. However, the current costs of these technologies do not allow researchers to perform large enough cohort studies to cover the full spectrum of disease and to draw relevant conclusion at the population level. We should also mention that concomitant pathologies (e.g., infection or recurrent disease) might also affect the observed immune profiles, although such rare and heterogeneous events are likely to be negligible given our sample size.

Finally, whether the immunosuppressive regimen affects the immune cell composition was not assessed. Most of the patients from GSE147089 were affected to the same immunosuppressive protocol.¹⁴ Therefore, although relevant, the real effect of immunosuppressive drugs cannot be answered using the current data and methodology, but is of interest for future well-designed randomized controlled trials.

In conclusion, we showed that the overall immune cell composition of kidney transplants is ill-related to main Banff-defined rejection categories. Information on the cell composition adds to the Banff lesion scoring and evaluation of rejection severity derived from the histologic lesion scores. Some specific cell types, most notably FCGR3A⁺ myeloid cells, CD14⁺ monocytes/macrophages, and NK cells, discriminate best between rejection phenotypes. Current tools for assessing causal triggers for allorecognition are insufficient. Future studies should couple the spatial cell distribution, as currently reflected in the Banff classification, to a detailed cellular composition to better understand the relation between cause and histological/molecular phenotype. Intragraft CD8temra cells bear strong and consistent association with graft failure and independent of Banff-grade rejection, suggesting a potential therapeutic target.

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/JSN/E633.

Funding

FP7 Health (305499), Fonds Wetenschappelijk Onderzoek (1016 ROCKET, JTC2-29), and Fonds Wetenschappelijk Onderzoek (G087620N). J. Callemeyn: Fonds Wetenschappelijk Onderzoek (1196119N). E. Van Loon: Fonds Wetenschappelijk Onderzoek (1143919N). T. Vaulet: Fonds Wetenschappelijk Onderzoek (1S93918N). M. Naesens: Fonds Wetenschappelijk Onderzoek (1844019N and 1842919N). B. Lamarthée: Bpifrance (DOS0060162/00), European Regional Development Fund of the Region Bourgogne Franche-Comté (FC0013440), and Agence Nationale de la Recherche (ANR-22-CE18-0011-01).

Author Contributions

Conceptualization: Maarten Naesens, Thibaut Vaulet.

Data curation: Dany Anglicheau, Asier Antoranz, Francesca Bosisio, Jasper Callemeyn, Jill Colpaert, Tim Debyser, Wilfried Gwinner, Priyanka Koshy, Dirk Kuypers, Baptiste Lamarthée, Pierre Marquet, Maarten Naesens, Claire Tinel, Amaryllis Van Craenenbroeck, Elisabet Van Loon, Thibaut Vaulet.

Formal analysis: Asier Antoranz, Jasper Callemeyn, Baptiste Lamarthée, Thibaut Vaulet.

Funding acquisition: Maarten Naesens.

Methodology: Jasper Callemeyn, Baptiste Lamarthée, Maarten Naesens, Thibaut Vaulet.

Project administration: Maarten Naesens.

Resources: Maarten Naesens.

Supervision: Dany Anglicheau, Francesca Bosisio, Jasper Callemeyn, Wilfried Gwinner, Baptiste Lamarthée, Maarten Naesens, Amaryllis Van Craenenbroeck.

Validation: Philip F. Halloran, Pierre Marquet, Thibaut Vaulet.

Visualization: Thibaut Vaulet.

Writing – original draft: Maarten Naesens, Thibaut Vaulet.

Writing – review & editing: Dany Anglicheau, Asier Antoranz, Francesca Bosisio, Jasper Callemeyn, Jill Colpaert, Tim Debyser, Wilfried Gwinner, Philip F. Halloran, Priyanka Koshy, Dirk Kuypers, Baptiste Lamarthée, Pierre Marquet, Maarten Naesens, Claire Tinel, Amaryllis Van Craenenbroeck, Elisabet Van Loon, Thibaut Vaulet.

Data Sharing Statement

The transcriptomics datasets supporting the findings of this study are openly available at the Gene Expression Omnibus of the National Institutes of Health (www.ncbi.nlm.nih.gov/geo) under the following numbers: GSE147089, GSE36059, and GSE21374. Additional clinical data of the Biomargin dataset (with corresponding transcriptomics dataset GSE147089) are accessible from the corresponding author upon motivated request.

Supplemental Material

This article contains the following supplemental material online at http://links.lww.com/JSN/E632.

Supplemental Methods.

Supplemental Table 1. Demographics and clinical variables of the Biomargin dataset (n=224).

Supplemental Table 2. Mean differences per estimated immune cell type in % (with 95% CI) between each pair of Banff-defined phenotypes in Combined dataset 1 (N=596), with corresponding Student t test P value.

Supplemental Table 3. Mean differences per observed immune cell type in % (with 95% CI) between each pair of Banff-defined phenotypes in Combined dataset 1 (N=596), with corresponding Mann–Whitney t test P value.

Supplemental Table 4. Association and correlation of estimated immune cell proportion and external variables on the Biomargin dataset (N=224).

Supplemental Table 5. Unsupervised clustering of immune cells and their overlap with the Banff categories measured with the adjusted Rand Index (ARI) in Combined dataset 1 (N=596).

Supplemental Table 6. Pairwise AUC (with 95% CI) for each estimated immune cell type from Combined dataset 1 (N=596).

Supplemental Table 7. Odds ratios from the multivariable logistic regression model of the immune cell types for no rejection versus any rejection.

Supplemental Figure 1. Density plot per cell type per cohort.

Supplemental Figure 2. Difference in normalized mean cellular contents between subtypes of rejection and the no rejection category stratified per cohort.

Supplemental Figure 3. Association of immune cell with four binary indicators in the Biomargin dataset (N=224):

Supplemental Figure 4. Unsupervised PCA plots (with the first three principal components [PCs]) of the immune cells colored by rejection subtypes.

Supplemental Figure 5. Temporal trajectories of immune cells in the Combined dataset 2 (N=506).

Supplemental Figure 6. Association of CD8temra tertiles with graft outcome on the basis of Kaplan–Meier estimators.