The study cohort consisted of 274 pediatric heart transplant recipients enrolled in the Clinical Trials in Organ Transplantation in Children (CTOTC) with available recipient and donor DNA. All patients in the CTOTC trial were chronologically recruited by physicians at participating centers over a period of two years and were treated using the same immunosuppressant protocol (19). This cohort provided HLA genotyping data to generate descriptive statistics for HLA-EMMA and PIRCHE-II. One-hundred thirty-one cases had clinical data for DSA and ABMR. After excluding 31 cases with pre-formed DSA, the remaining 100 cases were used to determine cut-off values (molecular mismatch thresholds) for the risk of DSA and ABMR. Characteristics of this 100-patient cohort are described in Table 1 and are similar to previously reported characteristics of a large portion of the CTOTC study’s patients (20). The 100 patients have a mean follow up time of 3 years post-transplant. Descriptive statistics for HLA-EMMA and PIRCHE-II were generated for the entire cohort of 274 patients with genotype. The 143 cases for whom we do not currently have access to the clinical data will be used to validate the cut-off values generated in this pilot study. A flow diagram of this study cohort was previously published (see Figure 1 of Mangiola et al., 2022) (17).

Approval from the IRB for the CTOTC studies was obtained for all clinical sites and informed consent was obtained based on each participating center’s IRB-approved consent form. When appropriate based on age and developmental maturity, assent was obtained. The study is registered at ClinicalTrials.gov (NCT02752789). No animal research was performed.

Class I and Class II HLA typing (HLA-A, B, C, DRB1/3/4/5, DQA1, DQB1, DPA1, DPB1) was performed by next generation sequencing (NGS) using Immucor MIA FORA NGS Flex 11 Kit (Immucor^®, Peachtree Corners, GA) according to the manufacturer’s instructions. Sequencing was conducted on an Illumina MiSeq^® using a standard v2 flow cell and a 300-cycle MiSeq^® V2 reagent kit (Illumina, San Diego, CA). Sequencing data in FASTQ format were demultiplexed and analyzed using MIA FORA HLA Flex software v4.5 with IMGT database 3.36.0 (Immucor). Ambiguities were resolved by sequence-specific oligonucleotide and/or sequence-specific primer genotyping.

HLA allele and antigen matching was conducted by comparing donor and recipient typing results after null alleles had been excluded from the dataset. AAMM and SAAAMM scores were determined for each HLA locus using HLA-EMMA version 1.04 (https://hla-emma.com), which predicts B-cell response. HLA-EMMA AAMM score is determined by matching HLA proteins using their individual amino acid residues with mismatches in the donor relative to the recipient being counted to produce a score. HLA-EMMA SAAAMM score is the number of amino acids included in the AAMM score that are solvent accessible (5). Locus scores were summed to calculate Class I (HLA-A+HLA-B+HLA-C), DR (HLA-DRB1+HLA-DRB3+HLA-DRB4+HLA-DRB5) and DQ (HLA-DQA1+HLA-DQB1) AAMM and SAAAMM scores.

PIRCHE-II scores were determined for each HLA locus using the PIRCHE-II algorithm (https://www.pirche.com, version 3.3.57 with database version 3.40) and summed (as stated for HLA-EMMA above) to achieve Class I, DR and DQ scores. PIRCHE-II is a prediction algorithm that estimates the amount of non-self, donor-derived HLA peptides that can be presented by recipient HLA-DRB1/3/4/5, HLA-DQA1/HLA-DQB1, and HLA-DPA1/HLA-DPB1 antigens to the recipient’ CD4 T cells. The number of peptides from the donor containing amino acid mismatches to the recipient that can be presented by recipient determine the PIRCHE-II score (6).

Quantile analysis was used to determine cut-offs for the various algorithms. This process uses the distribution of scores from HLA-EMMA or PIRCHE-II for DSA or ABMR cases only, and determines a position in the distribution (cut-off) at which 95% of patients with the outcome of interest (DSA or ABMR) have a score above this position and 5% score below. This is done by identifying the 0.05 position in the probability distribution. The numerical value of this position is used as the cut-off for risk of DSA or ABMR for a given HLA antigen.

The CTOTC core antibody laboratory (University of Pittsburgh) performed all patient serum testing using the Luminex LABScreen^® Single Antigen (One Lambda, Thermo Fisher, West Hills, CA) system, as previously described (19, 21). Testing was performed at ten timepoints: at enrollment, while on waitlist, immediately pre-transplant, 7 days, 1, 3, 6, 12, 24, and 36 months post-transplant. Additional testing was performed at the time of predefined clinical events such as acute rejection. A positive result for post-transplant DSA was defined as ≥1000 mean fluorescence intensity, and for the analysis presented we only considered persistent DSA, defined as the same specificity observed in two or more post-transplant sera. Allele-level HLA antibodies such as HLA DQA1*/DQB1* were identified by re-analyzing single antigen bead results compared to the allele-level donor genotype.

All subjects were followed at frequent intervals for assessment of graft function using clinical findings, diagnostic test results (e.g. electrocardiogram and echocardiogram), as well as by serial endomyocardial biopsy performed at the same timepoints indicated for DSA surveillance above and at the time of clinically suspected rejection episodes. For the present study episodes of ABMR were only included when proven by endomyocardial biopsy using local site interpretation, as previously described (19, 21, 22). All patients in the ABMR group that were used for determining HLA molecular mismatch cut-offs had at least one biopsy-proven episode of ABMR.

All graphing and statistical analysis were conducted in R version 4.1.2 (https://www.r-project.org/). For correlation statistics the Pearson’s correlation was calculated and the rho (r) value is reported. Quantile analysis was implemented via the R stats package. Kaplan-Meier analysis and log-rank tests were performed using the survival and survminer packages. For log-rank tests a p-value of ≤0.05 is considered statistically significant. For Kaplan-Meier analysis all patients were censored at the 1000 days mark because no data were available for any patient past this timepoint. Freedom from DSA was calculated as the time to the first appearance of a persistent, post-transplant DSA. Freedom from ABMR was calculated as the time to the first ABMR diagnosis for a given patient. Data manipulation was accomplished via the reshape2 package and base R functions. All confusion matrix statistics were calculated in R using the confusion Matrix function in the caret package.

Two different molecular characterizations of donor/recipient HLA compatibility were performed with HLA-EMMA: all mismatches at the amino acid level (AAMM), and solvent-accessible amino acid mismatches (SAAAMM), which are considered to be more “accessible” to anti-HLA antibodies (5). The total AAMM load ranged from 26-193 and SAAAMM load ranged from 18-140 ( Table 2 ). We observed a large degree of heterogeneity in this population with respect to AAMM and SAAAMM values, which can also be observed within discrete allele mismatch categories for Total HLA (Class I + Class II; Figures 1A, D ), Class I (HLA-A+HLA-B+HLA-C; 1B and 1E) and Class II (HLA-DRB1+HLA-DRB3+HLA-DRB4+HLA-DRB5+HLA-DQA1+HLA-DQB1+HLA-DPA1+HLA-DPB1; 1C and 1F).

The 100 transplant cases with clinical data and no pre-formed DSA were used in this study to determine HLA-EMMA SAAAMM cut-offs. Thirty-eight patients (38%) tested positive for DSA post-transplant (29% Class I, 22% Class II, 13% Class I/II), with 26% of DSAs being transient, 26% being persistent, and 14% categorized as both. We only used persistent post-transplant DSA in the following analyses. Among the 26% persistent post-transplant DSA cases the distribution was 9% Class I, 10% Class II, 7% Class I/II, and 7% were diagnosed with biopsy proven ABMR ( Supplementary Tables 1 , 2 ). The risk cut-off for developing DSA for each class and specificity was ≥19 for HLA Class I (n=16 patients used to establish cut-off, Figure 2A ), ≥14 for DRB1/3/4/5 (n=9, Figure 2B ), and ≥10 SAAAMM for DQA1/DQB1 (n=11, Figure 2C ). However, the ABMR (n=7) risk cut-off for Class I, DRB1/3/4/5, and DQA1/DQB1 were ≥18 ( Figure 2D ), ≥7 ( Figure 2E ), and ≥27 SAAAMM ( Figure 2F ), respectively. PIRCHE-II risk cut-offs were previously reported on the same dataset and are represented by the horizontal lines shown in Figure 2 (17).

When examining the distribution of PIRCHE-II scores based on the HLA-EMMA cut-offs, we observed that cases below and above the SAAAMM cut-offs have a wide range of PIRCHE-II scores. Similarly, cases with high or low PIRCHE-II scores have a wide range of SAAAMM values ( Supplementary Table 3 ).

To further assess the ability of HLA-EMMA to risk stratify cases, we conducted Kaplan Meier analysis to determine freedom from DSA and ABMR ( Figure 3 ). Cases with SAAAMM scores above or below the cut-off value were designated as high- or low-risk, respectively. Patients with high-risk for DSA to DRB1/3/4/5 demonstrated significantly lower freedom from DSA (p=0.027, Figure 3B ). Although not statistically significant, we observed a trend toward lower freedom from DSA in the high-risk groups for Class I ( Figure 3A ) and DQA1/DQB1 ( Figure 3C ). With respect to ABMR, we also observed significantly lower freedom of ABMR in the high-risk group based on DQA1/DQB1 SAAAMM score (p=0.0049; Figure 3F ). However, the differences in freedom of ABMR between the high- and low-risk groups were diminished when categorized based on Class I ( Figure 3D ) and DRB1/3/4/5 ( Figure 3E ).

The two algorithms show correlation values of 0.73 for HLA-DRB1/3/4/5, 0.75 for DQA1/DQB1 and 0.51 for HLA Class I ( Figure 4 ). To further examine this, the cohort was divided into discrete incremental categories based on SAAAMM values and PIRCHE-II scores plotted within these categories ( Figure 4 ). For each SAAAMM group, large variations in PIRCHE-II score were observed for HLA Class I ( Figure 4A ), DRB1/3/4/5 ( Figure 4B ) and DQA1/DQB1 ( Figure 4C ). For example, cases with a Class I SAAAMM of 21-25 had PIRCHE-II scores ranging from 67 to 568 ( Figure 4A ). For both DSA and ABMR risk, cases with SAAAMM scores above our risk cut-offs for SAAAMM load ( Figure 2 ) had PIRCHE-II scores ranging above and below the PIRCHE-II risk cut-off value ( Supplementary Table 3 ). Thus, for transplant recipients with high-risk SAAAMM scores, the risk for DSA and ABMR may be further divided based on the PIRCHE-II score cut-off.

To determine if SAAAMM and PIRCHE-II high-risk cases could be re-stratified using a second method, for each predictive algorithm, cases with above cut-off loads were split into intermediate- or high-risk based on the score of the second method ( Table 3 ; Supplementary Figure 1 ). As expected, the low-risk cases by individual algorithm or the combined method had the lowest incidence of DSA and ABMR ( Table 3 ) and the high-risk cases had the highest incidence ( Table 3 ). Interestingly, the intermediate-risk cases, defined as high-risk by one algorithm but low-risk by the other, demonstrated a DSA and ABMR incidence like the low-risk cases ( Table 3 ). Although the results of these analyses were not statistically significant, our data appear to indicate that high-risk cases could be further divided into higher and lower risk groups based on the second score. DSA and ABMR risk scores generated from molecular mismatch data for DRB1/3/4/5 ( Figure 3B ) and DQA1/DQB1 ( Figure 3F ), respectively, suggest that patients considered to be high-risk by one algorithm, but low-risk by the other, show a lower tendency to develop DSA and ABMR and could be reclassified as intermediate-risk.

To determine the impact of using both molecular mismatch scoring methods together, we performed Kaplan Meier analysis after dividing the 100 cases into three discrete categories: high-risk patients with an above cut-off score for both algorithms (top right quadrant of graphs in Figure 2 ), intermediate-risk patients with an above cut-off score by one but not both algorithms (top left and bottom right quadrants of graphs in Figure 2 ), and low-risk patients with below cut-off scores from both algorithms (bottom left quadrant of graphs in Figure 2 ).

When assessing freedom from Class I DSA ( Figure 5A ), low- and intermediate-risk categories behaved similarly while high-risk clearly showed a decreased freedom from DSA. However, DRB1/3/4/5 ( Figure 5B ; p=0.035) and DQA1/DQB1 ( Figure 5C ) showed clear separation between the low, intermediate, and high-risk categories, with the low-risk category having the longest time free of DSA, the high-risk having the least, and the intermediate falling between the two. Using this method, 16 cases for Class I, 24 for DRB1/3/4/5, and 14 for DQA1/DQB1 could be reclassified as intermediate risk. We also performed log-rank test to compare high-risk to a combined low- and intermediate-risk category. This resulted in an increase in statistical significance at DRB1/3/4/5 (p=0.018) and nearly reached statistical significance for Class I (p=0.051), but still failed to reach statistical significance at DQA1/DQB1 ( Figure 5 ).

Like the ABMR risk stratification by HLA-EMMA ( Figure 3 ) or PIRCHE-II (17) alone, the combined algorithms split the cohort to a lesser extent for Class I ( Figure 5D ) and DRB1/3/4/5 ( Figure 5E ) but did so significantly (p=0.0068) for DQA1/DQB1 molecular mismatch ( Figure 5F ). For ABMR risk, a higher number of cases may be needed to assess the role of combining HLA-EMMA and PIRCHE-II for Class I and DRB1/3/4/5 molecular mismatch; this analysis method will be tested again when the clinical data for the full cohort is available. However, with respect to ABMR risk at DQA1/DQB1, the two-algorithm method allowed us to reclassify 25 cases from high-risk to intermediate-risk. Combining the low- and intermediate-risk groups and comparing them to the high-risk group again led to similar results with a slight decrease in p-value for all HLA loci tested.

To assess the differences between the performance of our method for defining high risk patients and the low or intermediate risk patients we calculated various statistics using a confusion matrix ( Supplementary Table 4 ). To do this we combined the low- and intermediate-risk groups into a single category and compared those patients to our high-risk patients. We focused on the negative predictive value (NPV), which tells us the likelihood that a patient below our cutoff does not have DSA or ABMR. We found high NPVs for the combined and individual methodologies when assessing DSA (Class I: SAAAMM = 0.93, PIRCHE-II = 0.96, Combined = 0.94; DRB1/3/4/5: 0.98, 0.97, 0.97; DQA1/DQB1: 0.94, 0.96, 0.96) and ABMR (Class I: 0.96, 0.97, 0.97; DRB1/3/4/5: 0.96, 0.97, 0.97; DQA1/DQB1: 0.98, 0.98, 0.97), meaning that the likelihood that a patient in our low- and intermediate-risk categories does not have DSA or ABMR is extremely high.