Complex Linkage Disequilibrium Effects in HLA-DPB1 Expression and Molecular Mismatch Analyses of Transplantation Outcomes

Abstract

Background.

HLA molecular mismatch (MM) is a risk factor for de novo donor-specific antibody (dnDSA) development in solid organ transplantation. HLA expression differences have also been associated with adverse outcomes in hematopoietic cell transplantation. We sought to study both MM and expression in assessing dnDSA risk.

Methods.

One hundred three HLA-DP-mismatched solid organ transplantation pairs were retrospectively analyzed. MM was computed using amino acids (aa), eplets, and, supplementarily, Grantham/Epstein scores. DPB1 alleles were classified as rs9277534-A (low-expression) or rs9277534-G (high-expression) linked. To determine the associations between risk factors and dnDSA, logistic regression, linkage disequilibrium (LD), and population-based analyses were performed.

Results.

A high-risk AA:GX (recipient:donor) expression combination (X = A or G) demonstrated strong association with HLA-DP dnDSA (P = 0.001). MM was also associated with HLA-DP dnDSA when evaluated by itself (eplet P = 0.007, aa P = 0.003, Grantham P = 0.005, Epstein P = 0.004). When attempting to determine the relative individual effects of the risk factors in multivariable analysis, only AA:GX expression status retained a strong association (relative risk = 18.6, P = 0.007 with eplet; relative risk = 15.8, P = 0.02 with aa), while MM was no longer significant (eplet P = 0.56, aa P = 0.51). Importantly, these risk factors are correlated, due to LD between the expression-tagging single-nucleotide polymorphism and polymorphisms along HLA-DPB1.

Conclusions.

The MM and expression risk factors each appear to be strong predictors of HLA-DP dnDSA and to possess clinical utility; however, these two risk factors are closely correlated. These metrics may represent distinct ways of characterizing a common overlapping dnDSA risk profile, but they are not independent. Further, we demonstrate the importance and detailed implications of LD effects in dnDSA risk assessment and possibly transplantation overall.

INTRODUCTION

The development of de novo donor-specific antibodies (dnDSAs) after solid organ transplantation (SOT) is strongly associated with alloimmune processes leading to significant organ loss within a 10-year period.^1–3 Among multiple factors, the risk of developing dnDSA is related to HLA mismatches between donors and recipients.^1,4,5 In the context of organ shortage and the polymorphic nature of HLAs, absolute matching between recipients and donors is exceedingly difficult.

The definition of HLA mismatch in our field has evolved over time, based initially on serological (antigen) differences, then on incrementally more accurate assessments of HLA molecules by molecular techniques (evaluating antigen recognition domains), and finally, more recently, on epitope/amino acid (aa) differences (based on complete protein sequences). The relative degree of molecular mismatch (MM), whether expressed in aa or eplet units, has been shown to be an effective biomarker in assessing compatibility and predicting the development of dnDSA.^4,5

Another possible factor that may influence the development of dnDSA is the expression levels of mismatched epitopes in recipient and donor cells. Increased expression of HLA-mismatched alleles has been found to be associated with unfavorable transplantation outcomes.^6,7 HLA-DP expression differences appear to influence the risk of graft versus host disease (GVHD) in HLA-DPB1 mismatched hematopoietic stem cell transplantation (HSCT).⁷ We hypothesized that an analogous but inverse phenomenon may be present in SOT, in which HLA-DP-mismatched grafts with high-expression may elicit greater host immunogenic responses compared to low-expression grafts, particularly if the recipient has low expression DPB1 alleles that may render the immune system of the recipient less familiar with DP sequences and therefore more likely to perceive any DP sequence as nonself.

To investigate both the potential implications of differing grades of MM as well as expression differences on the development of dnDSA, the HLA-DPB1 locus was specifically selected for this study. HLA-DPB1 contains a molecular signature that allows assessment of expression level (low/high) for each allele. The 3′ untranslated region (UTR) of the HLA-DPB1 gene contains a known single-nucleotide polymorphism (SNP), rs9277534 G/A, which is associated with either high (G) or low (A) expression of the gene in different cells and tissues.^7–10

Utilizing well-established metrics for characterizing MM, including aa, eplet, and physicochemical approaches,¹¹ we sought to ascertain the possible effects of both HLA-DPB1 molecular-structural mismatch and DPB1 expression on HLA-DP dnDSA development. While investigating these different associations, we discovered that the inclusion of expression information (or, more specifically, a SNP that tags expression) results in unexpected complexity. It is known in genome-wide association studies that allelic heterogeneity arising from multiple causal variants at a genomic locus is frequently confounded by linkage disequilibrium (LD),¹² giving rise to spurious correlations between alleles. We uncovered a related type of correlation when investigating the association of dnDSA development with DPB1 MM and expression genotypes. We show that the relationship between these two risk factors involves highly predictable patterns, constrained and determined by specific effects of HLA-DPB1 LD structure and HLA-DPB1 allele frequencies in the general population. We additionally demonstrate that the relationships and potentially spurious correlations that have been uncovered may have important implications in not only solid organ but also hematopoietic stem cell transplantation.

MATERIALS AND METHODS

Sample Selection

Institutional Review Board approval was granted by the Institutional Review Board of the Children’s Hospital of Philadelphia (CHOP) for this retrospective study.

All SOTs that took place between February 2013 and December 2019 and were recorded in the HistoTrac database of CHOP were considered for this study (n = 259 transplant pairs). Exclusion criteria included transplant pairs with ambiguous donor identification in the HistoTrac database (n = 8; 251 transplant pairs remaining), those involving patients with multiple transplants recorded (n = 27; 224 transplant pairs remaining), those lacking 2-field HLA-DPB1 typing (n = 33; 191 pairs remaining), those involving null (n = 0) or novel (n = 1) DPB1 alleles (190 pairs remaining), matched at the DPB1 locus (n = 25; 165 pairs remaining), or those involving incomplete DPB1 IMGT protein sequences (n = 2; 163 pairs remaining). Further excluded were patients who had pretransplant HLA-DP DSA (n = 0) or <1 year between transplant and most recent DSA test (n = 60), which left 103 final transplant pairs for further analysis. See Table S1 (SDC, http://links.lww.com/TP/B928) for demographic information on the samples before and after exclusion criteria were applied.

HLA-DP DSA was considered positive if a donor HLA-DP Luminex single-antigen bead (SAB) (or closest substitute) had a mean fluorescence intensity (MFI) ≥1500 at any time during posttransplant monitoring. Because there is no standard MFI cut-off value that defines a positive DSA test result, a threshold of 1500 was chosen on the basis of our lab’s early validation studies involving concordance of results with flow cytometric crossmatch data; an MFI threshold of 1500 was determined to give the most favorable sensitivity and specificity profile for our validation dataset when utilizing the flow crossmatch results as a surrogate for a DSA truth set. It should be noted, however, that when a range of thresholds, from MFI 500 to 1400, were additionally tested in the current study (data not shown), they each produced results with statistical trends that were similar to what is reported in this study using the selected 1500 MFI threshold.

Patients were monitored for DSA at regular intervals posttransplant (wk to mo initially, less frequently during the second yr, and at least annually thereafter). Linkage information between HLA-DPB1 alleles and rs9277534 genotypes^13–15 was utilized for classification into high-expression (rs9277534-G) and low-expression (rs9277534-A) DPB1 alleles. If a typing was ambiguous (a G-group designation), an approximation was used, analogous to that used for P-groups, as described in the section “Dependence Between Molecular Mismatch and Expression Tagging Allele Combinations.” The proportion of cases by transplanted organ was 45/103 (44%) kidney, 43/103 (42%) heart, 14/103 (14%) lung, and 1/103 (1%) kidney-liver.

Assessment of Mismatches

An eplet is defined as a set of polymorphic HLA residues within a 3.0–3.5 Å radius on the molecular surface¹⁶ that theoretically constitutes the antibody binding site of the third complementarity-determining region of the immunoglobulin variable heavy chain.¹⁷ The eplets used in this analysis were those contained in the HLA Epitope Registry v.2.0.¹⁸ Full-length aa sequences and alignments for HLA-DPB1 alleles were obtained from the IPD-IMGT/HLA Database v.3.38.0.¹⁹ A mismatch was counted when a donor aa or eplet had no matching counterpart on either of the patient alleles at the corresponding position. If both donor alleles possessed the same aa/eplet mismatch at the same position, the mismatch was only counted once.

Since there exist different ways to assess MM, each with attendant advantages and disadvantages, this study analyzed MM on the basis of two common approaches: simple enumeration of (1) eplet mismatches or (2) aa mismatches. We have further added a supplemental analysis that utilizes various physicochemical parameters to augment the basic MM metric, as defined by the Grantham’s distance²⁰ or Epstein coefficient of difference²¹ (Supplemental Methods and Results: Section 1, Table S2, Figure S1, SDC, http://links.lww.com/TP/B928). These measures are calculations on the basis of aa properties that seek to provide a quantitative measure of the physicochemical change associated with specific aa substitutions in proteins. For both Grantham and Epstein metrics, the minimum score was used whenever a donor aa was mismatched to both patient amino acids at the same position, as a reflection of the degree of mismatch of that donor aa. Each individual score at the aa level was then added toward the final score at the transplant pair level. The summation of individual scores to give a total score is not meant to imply that substitutions or mismatches at different positions or in different 3D contextual space have the same immunogenicity potential. Rather, it is a simplified model, though possibly more informative than simple aa mismatch enumeration, and was meant to parallel a few related and established physicochemical metrics that have been applied in the field of histocompatibility.^5,11,22 There remains much potential and promising work in methods that seek to incorporate 3D contextual information into MM scoring systems,^23,24 though specific incorporation or contribution toward such 3D models and scores was beyond the scope of this study.

Computational and Statistical Analysis

Our dataset was analyzed as a retrospective cohort study, with HLA-DP dnDSA development as the outcome of interest and a recipient:donor HLA-DPB1 expression-linked SNP genotype of AA:GX (homozygous AA low-expression recipient paired with a donor having at least 1 high expression G allele) or MM as risk factors of interest. This AA:GX combination generally corresponds to the high-risk DPB1 expression combination described by Petersdorf et al⁷ in HSCT, in which they concluded that pairs with a low-expression A donor allele mismatched with a high-expression G recipient allele appear to be at elevated risk for GVHD.

A logistic regression model was fit to determine the marginal association of AA:GX (versus non-AA:GX) status with the development of HLA-DP dnDSA; then separately, the marginal association of MM load (in terms of either aa or eplets) with DP-dnDSA development was tested with additional logistic regression models:

$$\text{logit}[P(Y)]~{\beta }_0+{\beta }_{E}E$$

$$\text{logit}[P(Y)]~{\beta }_0+{\beta }_{m}MM$$

where the outcome Y_i is whether or not the recipient of transplant pair i developed HLA-DP dnDSA, E_i is whether or not the transplant pair is of expression combination AA:GX, and MM_i is the MM load (expressed in terms of aa or eplets).

Two additional multiple logistic regression models were fit, including AA:GX status and MM (expressed in the first model as aa mismatch and in the second as eplet mismatch) as covariates to jointly assess their association with HLA-DP dnDSA development:

$$\text{logit}[P(Y)]~{\beta }_0+{\beta }_{m}MM+{\beta }_{E}E$$

Since the 11% incidence of HLA-DP dnDSA in our study population exceeded the 10% threshold²⁵ that typically allows for the odds ratio (OR) to serve as a reasonable estimate of the relative risk (RR) in logistic regression analyses for cohort studies, we applied a correction—an approximation of the Mantel–Haenszel method^25,26—which provides a better estimate of the RR when the outcome of interest is considered common (>10%).

DPB1 expression and MM were chosen as the variables of focus in this study, without explicitly controlling for various other potentially relevant clinical covariates/risk factors, as a way to avoid over-stratification of a modest dataset. Both rs9277534 and MM, however, are biomarkers that are randomized from conception. Therefore, they allow for a closer approximation to a randomized trial, in which bias related to other covariates is reduced.^27,28

All transplant pairs that passed our basic inclusion and exclusion criteria were included, which provides a relatively stable institutional treatment background from which the AA:GX and non-AA:GX pairs were selected. The AA:GX cohort consisted of 31 pairs, while the non-AA:GX cohort consisted of 72 pairs.

Transplant pairs were examined according to the type of organ transplanted, with no clear aberrations in the general pattern observed (Figure S2, SDC, http://links.lww.com/TP/B928). It was decided to evaluate all SOT data together, since stratifying by organ would result in relatively few cases for each organ type.

Computational procedures and data analysis were performed using custom programs written in Python v.3.7.4²⁹ and R 3.6.1.³⁰

Dependence Between Molecular Mismatch and Expression Tagging Allele Combinations

Because it is difficult to generalize about patterns that may appear between patient-donor expression SNP genotypes and MM distributions in a limited-size dataset, we sought to simulate a very large population of transplant pairs to shed light on this issue. Such an exercise was performed by leveraging population-level HLA-DPB1 allele frequencies and computational means of simulating transplants between all possible patient and donor types in a given population. Specifically, the following procedure was performed: (1) identification of the most frequent DPB1 alleles³¹ comprising the top 99% of the cumulative allele frequency, which we refer to as the common alleles, (2) linking of these common alleles to their respective rs9277534 genotypes by published linkage^13–15 or based on exon 3 information¹³ (Tables S3 and S4, SDC, http://links.lww.com/TP/B928), (3) permutation of every possible patient-donor DPB1 allele combination, (4) calculation of corresponding MM loads, and (5) weighting the results by each simulated transplant pair’s population-level allele frequencies. Since the reported allele frequencies were based on pre–April 2010 HLA nomenclature,³² the alleles were translated into the corresponding current P-groups and a representative allele was used for each P-group, which in most cases comprised >99% of the unambiguously typed alleles of the P-group, according to our institutional database of 11 132 high-resolution unambiguous DPB1 typings from all clinical and research databases (see Table S4, SDC, http://links.lww.com/TP/B928 for details). For those P-groups with a less extreme split of constituent alleles, the alleles differed by a minimal amount of aa, and the rs9277534 linkage was usually identical (Tables S3 and S4, SDC, http://links.lww.com/TP/B928). This procedure produced an analytical projection of the MM distribution for every patient-donor rs9277534–SNP genotype. We termed this the frequency-weighted permutation (FWP) dataset, in contrast to our CHOP dataset. Mann-Whitney U tests (MW-U) were performed to assess the likelihood that the CHOP and FWP AA:GX versus non-AA:GX samples were selected from populations with the same MM distribution.

The common alleles were further analyzed in a LD analysis by extracting their sequences from the IPD-IMGT/HLA Database v.3.38.0.¹⁹ LD between each biallelic aa residue and the rs9277534 SNP was then computed³³ (Table S5, SDC, http://links.lww.com/TP/B928). Likewise, LD between the rs9277534 SNP and each HLA-DPB1 eplet defined in the HLA Epitope Registry v.2.0¹⁸ was computed by treating each eplet as a biallelic allele (either present or absent in each DPB1 sequence) (Table S6, SDC, http://links.lww.com/TP/B928). The frequency of each polymorphic aa variant and eplet within each of the rs9277534-A versus -G clades was also characterized, as shown in Figures S3 and S4 (SDC, http://links.lww.com/TP/B928).

HLA Typing

HLA-DPB1 typing was performed as part of a general HLA typing procedure based on NGS technology (Holotype HLA, Omixon Biocomputing Ltd., Budapest, Hungary), as described previously.³⁴

Despite the relevance of both DPB1 and DPA1 mismatches to HLA-DP dnDSA development, DPA1 is known to be much less variable and polymorphic than DPB1 and not typically typed or used in transplant matching. As a result, it was not considered in the main study. We have, however, provided a supplemental analysis in which DPA1 is included, which decreases the number of available cases for analysis, but the general results appear to be similar to the results of the main analysis (Supplemental Methods and Results: Section 2, Table S7, Figures S5 and S6, SDC, http://links.lww.com/TP/B928).

Accounting for SABs Known to Have a High False-positive Rate

Certain HLA-DP SABs have been described as having relatively high levels of false positives.³⁵ The prevalence of DP alleles that correspond to the problematic beads was examined among the donors of DP-dnDSA positive and negative pairs. The prevalence of these problematic alleles was found to be no higher among the DP dnDSA+ pairs than among the DP dnDSA− pairs (Table S8, SDC, http://links.lww.com/TP/B928).

RESULTS

Within the set of 103 transplant pairs, 11 (10.7%) patients developed HLA-DP dnDSA, with a median follow-up of 34 months (range, 12–72). Figure 1 depicts MM distributions for HLA-DP dnDSA+ and HLA-DP dnDSA−transplant pairs, with color-coding according to patient-donor rs9277534 A/G expression combinations. The AA-patient:GA-donor and AA-patient:GG-donor classes make up all but 1 of the DSA+ cases. These 2 groups collectively constitute a high-risk {low-expression recipient (AA) | high-expression donor (GX)} combination class, with AA representing homozygous low-expression and GX representing at least 1 rs9277534 high-expression (G)-linked allele. This high-risk patient-donor expression combination will hereafter be referred to as “AA:GX.” When fitting a simple logistic regression with HLA-DP dnDSA development as the outcome of interest and AA:GX (versus non-AA:GX) as the risk factor of interest, a RR of 23.2 was observed (P = 0.001) (Table 1).

FIGURE 1.

Distribution of molecular mismatch counts for HLA-DP dnDSA± groups, color-coded by patient:donor rs9277534 expression genotypes. (A) amino acid differences, (B) eplet differences. Results are color-coded by patient:donor rs9277534 A/G genotypes (A = low expression, G = high expression). (C) Count of HLA-DP dnDSA+ and HLA-DP dnDSA−outcomes among AA:GX and non-AA:GX pairs. dnDSA, de novo donor-specific antibody.

TABLE 1.

Testing association of MM and AA:GX with dnDSA

Coefficient	Estimate	SE	z	P	OR (95% CI)	RR (95% CI)
logit [P(Y)] ~ β0 +βEE
AA.GX	3.521	1.078	3.267	0.00109**	33.8 (5.98–63.9)	23.2 (5.59–64.8)
logit [P(Y)] ~ β0+βm MMaa
Amino acid MM	0.24608	0.08237	2.987	0.00281**	11.7 (2.75–74.7)a	–
logit [P(Y)] ~ β0 +βm MMepl
Eplet MM	0.24300	0.09041	2.688	0.00719**	11.4 (2.16–79.4)a	–
logit [P(Y)] ~ β0+βm MMaa +βEE
Amino acid MM	0.07741	0.11781	0.657	0.511134	2.17 (0.21–23.5)a	–
AA.GX	2.99006	1.30002	2.300	0.021447*	19.9 (2.35–515)	15.8 (2.31–63.3)
logit [P(Y)] ~ β0+βmMMepl +βEE
Eplet MM	0.07036	0.11998	0.586	0.557597	2.02 (0.19–22.9)a	–
AA.GX	3.20888	1.18251	2.714	0.006655**	24.8 (3.51–531)	18.6 (3.39–63.5)

^*P < 0.05

^**P < 0.01.

The RR estimation cannot be applied to the MM covariates in a straightforward and intuitive way because it is meant to be applied to risk factors that involve a small number of discrete categories. Therefore, RR has been omitted for the MM calculations (indicated by the ‘–’ symbol in the RR column), since the MM measures are represented in a ratio/count scale, each encompassing many values.

^aThe OR for MM is calculated in terms of 10 mismatch increases.

dnDSA, de novo donor-specific antibody; MM, molecular mismatch; OR, odds ration; RR, relative risk.

The mean HLA-DPB1 aa mismatch for transplant pairs in DP-dnDSA+ and DP-dnDSA− categories was 12.7 and 7.1, respectively, while the mean eplet mismatch was 11.4 and 7.6, respectively. Both aa and eplet mismatch were associated with DP dnDSA development when evaluated using separate logistic regression models (P = 0.003 and 0.007, respectively; Table 1). It has been demonstrated that aa mismatch and eplet mismatch are correlated,⁴ which is also observed in our sample, with r² = 0.81 (Figure S7, SDC, http://links.lww.com/TP/B928). The differential distribution of MM load for DP-dnDSA+ and DP-dnDSA−outcome groups is depicted in Figure 2.

FIGURE 2.

Molecular mismatch distributions according to DP-dnDSA+ and DP-dnDSA− outcome. (A) Amino acid and (B) eplet mismatch distributions according to DP-dnDSA+ and DP-dnDSA− outcome, expressed as box plots with individual points representing transplant pairs. The lower and upper box limits represent the first and third quartiles of the data, respectively, with the internal horizontal line indicating the median. The whiskers indicate the range of the data, out to a maximum of 1.5 times the interquartile range beyond the upper and lower quartiles. Points more extreme than the whiskers are depicted as diamond-shaped outliers. dnDSA, de novo donor-specific antibody.

When combining patient-donor AA:GX status as a covariate together with either of the MM covariates in multiple logistic regression, we find that MM is no longer associated with DP-dnDSA development (P = 0.51 and 0.56 for the aa and eplet mismatch covariates, respectively), whereas AA:GX status is still significantly associated with DP-dnDSA development (P = 0.02 or 0.007, when regressed with either the aa or eplet mismatch covariate, respectively). The RR of developing DP dnDSA+ for AA:GX pairs was estimated to be 15.8 (P = 0.02) when analyzed together with the aa mismatch covariate and 18.6 (P = 0.007) when analyzed together with the eplet mismatch covariate (Table 1). An interaction term (β_interact (MM×E)) was added to the multivariable regression models but was not found to be statistically significant (P > 0.6).

When separately evaluating each of the 2 risk factors involving MM and expression, both demonstrate significant association with DP-dnDSA. However, when we attempt to determine the individual and distinct effects of MM versus expression by computing their relative contributions at the same time (in multivariable regression), the effect of the AA:GX expression risk factor remains strong, but the MM is no longer significantly associated with DP-dnDSA. The instinct might be to interpret these results as indicating that expression is the more important factor. An important caveat, however, in the above analysis is that the expression and MM covariates are correlated, which is understandable since the rs9277534 expression-tagging SNP and polymorphisms along the length of DPB1 are in LD^13,36 (Tables S5 and S6, SDC, http://links.lww.com/TP/B928).

As a result of such a correlation, it is challenging to disentangle the true effects of expression and MM. An interesting side effect of the LD, however, is that certain aa residues and eplets in high LD with the expression SNP (Tables S5 and S6, Figures S3 and S4, SDC, http://links.lww.com/TP/B928) represent nearly fixed aa/eplet differences between rs9277534 A and G allele types, resulting in a subset of specific and frequently occurring mismatched aa/eplets within the AA:GX pairs. These specific mismatches may potentially contribute significantly to the overall elevated DP-dnDSA risk among AA:GX pairs, and thus, their individual risk contributions will need to be disentangled also from the bare MM enumeration and AA:GX classification risk factors themselves.

To assess the dependence between MM and rs9277534-linked allele combinations, population-level allele frequencies were leveraged (see “Dependence Between Molecular Mismatch and Expression Tagging Allele Combinations” in Materials and Methods). An analytical projection of the MM distributions for all the types of patient:donor rs9277534 genotype combinations was performed by permuting all possible patient:donor combinations of the 99% most frequent DPB1 alleles and weighting the results by the corresponding population-level allele frequencies. The FWP MM distributions were found to closely mirror the distributions of our CHOP-based dataset for every patient-donor rs9277534 expression combination (Figure 3, Table 2, Figure S8, SDC, http://links.lww.com/TP/B928), which indicates that similarly constrained and characteristic MM patterns would be expected to be observed for each patient-donor rs9277534–SNP genotype in any large study population. The distributions of MM across AA:GX category (AA:GX versus non-AA:GX) within both CHOP and FWP datasets are significantly different (CHOP, AA:GX versus non-AA:GX MW-U P < 0.001; FWP, AA:GX versus non-AA:GX MW-U P < 0.001), whereas when looking within AA:GX category but across datasets—CHOP versus FWP—the distributions of MM do not appear statistically significantly different (AA:GX, CHOP versus FWP MW-U P = 0.29; non-AA:GX, CHOP versus FWP MW-U P = 0.31) (Figure 4).

FIGURE 3.

FWP of most frequent patient-donor DPB1 alleles to define characteristic MM distributions for each patient:donor rs9277534 genotype. An analytical projection of the MM distributions for all the types of patient:donor rs9277534 genotype combinations was performed by permuting all possible patient:donor combinations of the 99% most frequent DPB1 alleles and weighting the results by the corresponding population-level allele frequencies. The FWP MM distributions are correlated with patient-donor rs9277534 expression genotypes in a way that closely mirrors the correlation seen in our CHOP population. Colors correspond to those of Figure 1 for ease of comparison. The width of each violin plot is proportional to the size of the data within it; these violin plots are, therefore, quantitatively comparable across the plots within the same CHOP or FWP dataset. The lower and upper limits of the miniature boxes within the violin plots represent the first and third quartiles of the data, respectively, with the internal white dot indicating the median. The whiskers indicate the max/min of the data or 1.5 times the interquartile range beyond the upper and lower quartiles, whichever is less extreme. The full range of the data is indicated by the upper and lower limits of the violin figures that enclose the box plots. FWP, frequency-weighted permutation; MM, molecular mismatch; CHOP, Children’s Hospital of Philadelphia.

TABLE 2.

Mean ± SD of CHOP vs FWP amino acid mismatch distributions for each patient:donor rs9277534 genotype

Patient:donor rs9277534 genotype	CHOP aa MM (mean ± SD)	FWP aa MM (mean ± SD)
AA:AA	4.0 ± 3.2	3.4 ± 2.7
AA:GA	12.8 ± 3.0	11.8 ± 2.8
AA:GG	13.4 ± 2.4	14.3 ± 3.1
GA:AA	2.6 ± 1.8	2.7 ± 2.0
GA:GA	4.1 ± 4.1	5.2 ± 2.7
GA:GG	7.5 ± 4.8	6.3 ± 2.8
GG:AA	11.1 ± 2.4	10.4 ± 2.8
GG:GA	10.3 ± 3.1	11.6 ± 3.5
GG:GG	4.7 ± 3.4	6.9 ± 3.7
non-AA:GX	5.4 ± 4.3	5.8 ± 4.0
AA:GX	12.9 ± 2.8	12.4 ± 3.1

CHOP, Children’s Hospital of Philadelphia; FWP, frequency-weighted permutation; MM, molecular mismatch.

FIGURE 4.

FWP to define characteristic MM distribution of AA:GX and non-AA:GX patient:donor genotype groups. The distributions of MM across AA:GX category (AA:GX vs non-AA:GX) within both CHOP and FWP datasets are significantly different (CHOP, AA:GX vs non-AA:GX MW-U P < 0.001; FWP, AA:GX vs non-AA:GX MW-U P < 0.001), whereas, when looking within AA:GX category but across datasets—CHOP vs FWP—the distributions of MM do not appear statistically significantly different (AA:GX, CHOP vs FWP MW-U P = 0.29; non-AA:GX, CHOP vs FWP MW-U P = 0.31). The width of each violin plot is proportional to the size of the data within it; these violin plots are, therefore, quantitatively comparable across the plots within the same CHOP or FWP dataset. The lower and upper limits of the miniature boxes within the violin plots represent the first and third quartiles of the data, respectively, with the internal white dot indicating the median. The whiskers indicate the max/min of the data or 1.5 times the interquartile range beyond the upper and lower quartiles, whichever is less extreme. The full range of the data is indicated by the upper and lower limits of the violin figures that enclose the box plots. CHOP, Children’s Hospital of Philadelphia; FWP, frequency-weighted permutation; MM, molecular mismatch; MW-U, Mann-Whitney U test.

To determine whether specific aa/eplet mismatches might clearly associate with the split of AA:GX cases into DP-dnDSA+ and DP-dnDSA− outcome groups, all of the individual aa/eplet mismatches were identified and compared between the AA:GX DP-dnDSA+ and DP-dnDSA−groups (Figure S9, SDC, http://links.lww.com/TP/B928). There did not appear to be especially notable or divergent mismatch candidates identified between the two groups.

To expand our MM analysis to include methods that incorporate physicochemical parameters to attempt to quantify the degree of difference in each aa mismatch/substitution, we made use of the Grantham²⁰ and Epstein²¹ physicochemical metrics. These methods appear to be potentially useful augmentations of the basic MM metrics (Figure S1, Supplemental Methods and Results: Section 1, Table S7, SDC, http://links.lww.com/TP/B928).

To provide some internal validation of our results, we trained a logistic regression classifier to determine its dnDSA discrimination/prediction ability when using the investigated risk factors as predictors (DPB1 MM and AA:GX expression status, separately and in combination; Figure S10, SDC, http://links.lww.com/TP/B928). Stratified-shuffled-split^43,44 cross-validation (ie, Monte Carlo cross-validation⁴⁵) using a 75%–25% train–test split with 500 runs for each classifier produced average area under the ROC curves (AUCs)⁴⁶ ranging from 0.78 to 0.86 (AA:GX status as the predictor, AUC = 0.83 ± 0.09 [mean ± SD]; aa MM as the predictor, AUC = 0.83 ± 0.09; AA:GX status and aa MM as predictors, AUC = 0.86 ± 0.08; eplet MM as the predictor, AUC = 0.78 ± 0.10; AA:GX status and eplet MM as predictors, AUC = 0.86 ± 0.07), reflecting the high performance of the logistic regression classifiers that use the MM and expression risk factors as predictors, supporting the strong associations we have reported; the performance and variance appear to improve when combining predictors, though larger studies with external validation sets are needed to confirm such results.

DISCUSSION

This work demonstrates that HLA expression analysis may have an important role to play in the assessment of immunological responses in SOT, beyond the traditional role of structural differences alone. However, disentangling the contribution of HLA expression from that of molecular–structural mismatch in dnDSA risk is not straightforward. Due to LD, certain polymorphisms are constrained to each of the rs9277534-A and -G expression-associated clades, which in turn constrain the possible MM load of A/G patient-donor combinations. Even with larger sample sizes, in an attempt to better control for separate covariates, the problem of spurious correlation among MM and expression is still expected to propagate, as is demonstrated through our analytical permutating of all common allele combinations (Figure 3). In such settings, machine learning techniques^37–39 along with fine mapping^12,40 and colocalization approaches^41,42 specifically designed to address LD patterns and complex correlations could be adopted to hone in on particular causative mismatches and eQTLs. This would not only provide critical insight into potential true biologic influences that could be used to improve clinical testing but also improve the classification of benign mismatches, allowing for more flexibility in donor–patient matching.

The development of dnDSA remains a major risk factor for chronic graft rejection/failure and a serious management challenge in SOT. Assessing the risk of dnDSA development can be instructive for pretransplant assessment of donor–recipient compatibility as well as for posttransplant monitoring of DSA. The relevance and impact of the degree of HLA matching on dnDSA development, T-cell responses, and overall graft survival have been well described.^4,5 Not much is known, however, as to whether the differential expression of different HLA loci and alleles within a particular locus play a role in graft survival and, more specifically, in either dnDSA development or T-cell immune responses. The realization that DP expression may affect the development of GVHD in HSCT suggests that a similar phenomenon may occur in SOT as well. In the setting of HSCT, presence of a HLA-DPB1 mismatch between a G-linked patient allele and an A-linked donor allele was shown to be associated with GVHD.⁷ In SOT, MM load has been established as a strong predictor of DSA development and subsequent graft failure.^4,5 In this study, we do not attempt to dispute or affirm either of these concepts in the context of HLA-DPB1, but instead, we pose the interesting observation that the true causal factors of HLA-DP dnDSA and their relative effect sizes from among DPB1 MM load and AA:GX expression status are difficult to disentangle. When considered by themselves, both risk factors appear to be tagging dnDSA risk in our SOT cohort, but LD structure makes it difficult or impossible at the current sample sizes to differentiate true effects. It has been shown that the region from exon 3 to the 3´-UTR of HLA-DPB1, encompassing rs9277534, represents an evolutionarily conserved region and that stratified analysis of high- and low-risk expression groups and T-cell epitope (TCE) permissive versus nonpermissive mismatches is effective in risk assessment in the context of GVHD,¹⁵ but the possible interactions and underlying mechanisms have not been explicitly characterized. Further, it is possible that both the polymorphisms that underlie MM and relative levels of expression are under coevolutionary selective pressures. These results should serve as motivation for future large-scale studies and development of enhanced techniques to better understand the individual effects and to better isolate true biologic factors. Although we did not find a significant association when adding an interaction term (which tests for a situation in which the effect of one risk factor is influenced by the level of another one) to the multivariable regression models, we believe that there is potential to uncover such an interaction in future larger studies should one exist and that such larger studies would also help to improve the risk estimates for each of the individual risk factors.

We reflect on an example of the problem posed by this tangled association of DPB1 MM and expression with transplantation-related outcomes. In the analysis of various DPB1-mismatched rs9277534-linked patient-donor allele combinations and their association with GVHD outcomes in HSCT, Petersdorf et al⁷ conclude that “among recipients of transplants from donors with rs9277534A-linked HLA-DPB1, the risk of acute GVHD was higher for recipients with rs9277534G-linked HLA-DPB1 mismatches than for recipients with rs9277534A-linked HLA-DPB1 mismatches.” The comparison they are making is similar to comparing the Figure 3 orange AA:GA transplant pairs to the blue AA:AA pairs. If the AA:GA orange transplant pairs are reported to correspond to higher rates of adverse outcomes, an observer who lacks the MM information may simply presume that such an outcome is due to expression-related effects alone. However, it is clear that the high-MM AA:GA orange distributions and the relatively low-MM AA:AA blue distributions represent two quite opposite ends of the MM spectrum. Therefore, there is good reason to believe that the Petersdorf et al DPB1 GVHD study may also have been affected by such a tangled association between MM and expression-tagging, given our FWP generated results. The FWP results should, by definition, generalize to any population with similar underlying DPB1 allele frequencies.

Quantity of aa mismatches in a sense undergirds both the direct TCE⁴³ and indirect PIRCHE⁴⁴ methods of alloimmune risk stratification, since increasing the number of aa mismatches theoretically increases the chances of both an unfavorable TCE combination as well as PIRCHE score. Basic MM may therefore correlate with both of these T-cell measures to some degree (we demonstrate such correlations using our FWP approach: Supplemental Methods and Results: Section 3, Figures S10, S11, S12, S13, S14, SDC, http://links.lww.com/TP/B928). Both of these measures are associated with GVHD risk, each contributing in an independent capacity.^10,45,46 Consequently, there is reason to believe that simple aa mismatch count may also have implications in HSCT and may correlate with adverse outcomes, as in SOT.

Our previous work involving computational assessment of miRNA targeting of the 3ÚTR of HLA-DPB1 indicates that rs9277534 is likely only a marker of expression, in LD with causative factor(s) and not necessarily a causative SNP itself.⁴⁷ There is a possibility that several SNPs in LD with rs9277534 may be driving the differential expression between the 2 major HLA-DPB1 clades³⁶ through such mediators as miRNAs. Therefore, there may be additional layers of complexity to unravel when considering HLA expression effects, especially when dealing with other loci involving more complex expression-associated allele clades. There remains also a question of homozygous versus heterozygous haplotype effects on expression and whether HLA haplotypes found in homozygous versus heterozygous individuals can display more complex interactions, such as those based on differentially encoded miRNAs per haplotype or various types of enhancer–promoter chromatin interactions. Besides the consideration of allele and haplotype-specific expression patterns of HLA molecules, tissue- and sex-specific expression differences can result in variable or even opposing expression patterns for one and the same SNP genotype.^48,49 Therefore, a significant amount of work remains as we attempt to develop a comprehensive understanding of HLA expression patterns and their role in HLA gene functions and histocompatibility.⁵⁰

This study involved certain limitations. First, it focused only on pediatric transplants. Given that pediatric transplants are more limited in number and have population-specific factors that may influence dnDSA development,⁵¹ additional analysis of data from adult transplants would be very appropriate. Due to the size limitations of this study, all SOTs were combined and analyzed together, although it is recognized that organ-specific factors may affect dnDSA development and that the significance of dnDSA may differ according to the transplanted organ.⁵² Additionally, DSA positivity in this study was based solely on an MFI threshold of 1500. It is possible that, had more subtle factors or criteria been considered, some of the DSA assignments would have changed. Further, this study did not control for certain clinical and patient variables that could contribute to bias such as sex, age, medication nonadherence, or physician-directed changes in immunosuppression (see Figure S15 though to see that sex is relatively balanced among DSA± groups but that controlling for all relevant donor-recipient sex/unknown combinations would likely have overstratified the dataset). It should be noted, however, that traditional sources of bias would not be thought to preferentially affect specific patient-donor DPB1 expression classes or MM combinations in our study, as such molecular markers should be invisible to the clinician. Mendel’s law of segregation and independent assortment, often invoked in settings of Mendelian randomization, likely also apply in this case where the observed risk allele may be independent of other clinical covariates, offsetting the explicit need to control for such factors.^27,28 One caveat, however, is that there are differences in frequencies of rs9277534 A and G alleles in different ethnic groups, which may contribute to bias. We would seek to control for such potential biases in future large-scale studies and see this as further motivation for more diverse study samples. A recent study by Philogene et al, however, suggests that such mixed-ethnicity factors may have less impact on dnDSA development than what might otherwise have been expected.⁵³

As we work to disentangle and acquire a more nuanced understanding of additional and specific risk factors that may influence the development of dnDSAs or GVHD, the ultimate goal would be to integrate all relevant factors into a comprehensive risk analysis scheme. Specifically, this integrative approach would likely include both quantitative and qualitative components, reflecting numbers of epitope mismatches as well as immunogenicity of each specific epitope (whether based on physicochemical^{11,20,23,24,54–56} or other characteristics and capacity of epitopes to be presented by the responder’s HLA molecules^57,58), levels of expression of these epitope-containing HLAs, as well as aspects of their regulation in various contexts. It is possible that additional metrics may be included in the future. This integrative approach should generate an improved system for assessing HLA mismatches and allow for a clearer determination of permissible mismatches and therefore influence both the longevity of transplants as well as the ability to perform a greater number of longer-lasting transplants.