This retrospective cohort study was conducted in pediatric patients undergoing their first haploidentical transplantation following the Beijing protocol at our hospitals from 2013 to 2023. The study included children diagnosed with acute myeloid leukemia (AML), acute lymphoblastic leukemia (ALL), or other malignant hematological diseases. All patients received modified myeloablative conditioning regimens that included total body irradiation, busulfan, or fludarabine, depending on the patient’s comorbidities.

Graft-versus-host disease (GVHD) prophylaxis consisted of anti-thymoglobulin (ATG), cyclosporin A (CsA), methotrexate (MTX), and mycophenolate mofetil (MMF). Patients were excluded if they had previously undergone solid organ transplantation, had incomplete KIR genotyping data, or lacked samples for retesting. The study was approved by the ethics committee (No. S2025-014-01), and informed consent was obtained in accordance with the Declaration of Helsinki.

All donors and patients were tested for KIR genes. The presence or absence of various KIR genes (KIR2DL1, KIR2DL2/3, KIR2DL4, KIR2DL5A, KIR2DL5B, KIR2DS1, KIR2DS2, KIR2DS3, KIR2DS4, KIR2DS5, KIR3DL1/3DS1, KIR3DL2, KIR3DL3) and two KIR pseudogenes (KIR2DP1 and KIR3DP1) was determined using a commercially available KIR-SSO typing kit (Immucor Transplant Diagnostics, USA). Amplicons were quantified with the Luminex LABScanTM 100 flow analyzer. Allelic subtypes for KIR2DS4 were resolved using probe-based PCR (probes 145, 175, and 234).

Three inhibitory KIR genes—KIR2DL1, KIR3DL1, and KIR2DL3—were prioritized for high-resolution genotyping using next-generation sequencing (NGS) to identify allelic variants (GenDx, Netherlands) in both donors and patients. Class I and II HLA loci were typed for all recipients and donors using sequencing-based typing (SBT) with GenDx excellerator kits (GenDX, Netherlands). Epitope ligand groups (HLA-C1/C2, -Bw4/Bw6, -A3/A11) were assigned, and HLA-B alleles were grouped into Bw4-80I/Bw-80T/Bw6 epitope-bearing ligands based on information retrieved from the database (https://www.ebi.ac.uk/ipd/kir/ligand.html) and relevant publications (12, 13). HLA typing results for patients and donors were also used to calculate subtype frequencies with the R package midasHLA version 1.14.0 (14).

Various models have been developed to explain how KIRs and their cognate ligands influence clinical outcomes (2, 3). The ligand–ligand model compares the KIR ligands of the donor and recipient without considering KIR genotyping; donor NK cells will be alloreactive toward host cells if the recipient lacks HLA class I molecules that are present in the donor. The receptor–ligand (or missing-ligand) model takes into account both donor KIR and its HLA ligand in the recipient: if at least one KIR does not recognize its cognate HLA molecules, NK cell inhibition in the donor is reduced, leading to increased cytotoxic activity. This model does not consider the education of donor NK cells.

The educational model encompasses both donor and recipient HLA class I molecules, as well as donor KIR typing. If the donor carries both a KIR and its ligand, the donor NK cells will be licensed and exhibit full alloreactivity. Conversely, if the donor has the KIR but lacks its ligand, the NK cells will remain unlicensed.

The haplotype-based model evaluates the composition of activating and inhibitory genes within KIR haplotypes (2, 19), which are split into two parts: centromeric (Cen, A/B motif) and telomeric (Tel, A/B motif). KIR A/B haplotypes can be distinguished by their characteristic gene content. The A haplotype comprises seven genes (KIR3DL3, KIR2DL3, KIR2DL1, KIR2DL4, KIR3DL1, KIR2DS4, and KIR3DL2) and two pseudogenes (KIR2DP1 and KIR3DP1), while the B haplotype is characterized by the presence of at least one of eight genes (KIR2DS2, KIR2DL2, KIR2DL5B, KIR2DS3, KIR3DS1, KIR2DL5A, KIR2DS5, and KIR2DS1) (2).

To elucidate the combined effects of KIRs, scoring systems have been developed to integrate data on KIRs and their respective ligands. The B content score, which ranges from 0 to 4, is derived by counting the number of B motifs. The ct-KIR score is assigned based on the presence of CenB and TelB motifs: a score of 0 for donors with CenB– and TelB+, a score of 1 for those with either CenB+ and TelB+ or CenB– and TelB–, and a score of 2 for those with CenB+ and TelB–. The KIR matching model quantifies the presence of aKIR and/or iKIR genes in the donor that are absent in the recipient, and vice versa.

These models have primarily focused on the presence or absence of KIR genes, neglecting the implications of allelic polymorphism, which can result in KIR molecules with significant biological differences.

The presence of NK cell alloreactivity was initially assessed using the educational model, which is predominantly applied to interactions between the four main iKIR genes and their HLA ligands in the recipient: 2DL1–C2, 2DL2/3–C1, 3DL1–Bw4, and 3DL2–A3/A11. An additional educational pathway involves the activating KIR2DS1 and its cognate C2 ligand (15). Donors with the 2DS1+/C1+ genotype produce NK cells that are fully educated and capable of recognizing their ligands on C2+ recipient cells. In contrast, donors with the 2DS1+/C1– genotype generate hyporesponsive NK cells, regardless of the ligands expressed by leukemic cells (16).

In high-resolution typing, KIR3DL1 allotypes were classified as high expression (KIR3DL1*001, KIR3DL1*015015), low expression (KIR3DL1*005, KIR3DL1*007), or null/lack-of-expression (KIR3DL1*004) (17). The dimorphism between isoleucine and threonine at position 80 of HLA-Bw4 was categorized as high (Bw4-80I) or low (Bw4-80T) expression. Further comparisons were conducted to assess the inhibitory effects of various allelic combinations (80I/T).KIR2DL1 allotypes were categorized as KIR2DL1*002/001g, KIR2DL1*003, KIR2DL1*004g, and null allele group (KIR2DL1*0320102N). Notably, alleles with arginine at amino acid position 245 (R²⁴⁵) show lower expression and weaker inhibitory signaling than those with cysteine at the same position (C²⁴⁵) (18). KIR3DL1, KIR2DL1, and KIR2DL3 were present in nearly all donors and recipients, making receptor–receptor mismatch analysis based on gene presence or absence at the allelic level unfeasible.

Secondly, models integrating information on varying degrees of iKIR-mediated inhibition and KIR2DS1 activation were evaluated for their ability to predict relapse risk and overall survival, following the methodology of Dhuyser (3). The qualitative value of each individual KIRxDLx education model (Education–2DL1, –2DL2/3, –3DL1, –3DL2) was first assessed. This was followed by a quantitative analysis in which the number of iKIR models predicting alloreactivity was counted, yielding an integer value from 0 to 4. Because of the limited sample size, statistical power was insufficient for subgroup analysis based on these values. Therefore, patients without predicted alloreactivity were compared with those showing predictions from 1 to 4 models (indicating at least one model predicted alloreactivity).

The study also validated the association between haplotype motif–based scores and clinical outcomes (19, 20). Given the interaction between donor KIRs and recipient ligands, KIR functionality was contingent on the presence of cognate ligands expressed by the recipient’s HLA molecules. The count functional inhibitory KIR score (CF-iKIR score) was calculated as follows: CF-iKIR score = (1 if functional KIR2DL1) + (1 if functional KIR2DL2 and/or functional KIR2DL3) + (1 if functional KIR3DL1) (21). The weighted inhibitory score was calculated as: inhibitory score = (1 if functional KIR2DL1) + (1 if strong functional KIR2DL2 or 0.5 if weak functional KIR2DL2) + (0.75 if functional KIR2DL3) + (1 if functional KIR3DL1) (8, 21).

Additionally, the ct-KIR score was tested (1), as it has been identified as a relapse risk factor in childhood acute lymphoblastic leukemia. The KIR B content score was calculated using an online calculator (https://www.ebi.ac.uk/ipd/kir/matching/b_content/). All individuals were categorized as homozygous A haplotype (A/A) or carrying at least one group B haplotype (B/x). Bx haplotypes with a prevalence exceeding 20 donors in the cohort were assessed and labeled according to the Allele Frequency Net Database (http://www.allelefrequencies.net/kir6001a.asp) for further study (22, 23).

The primary endpoints were relapse incidence and overall survival (OS). OS was defined as the time from transplantation to death from any cause, with censoring at the last follow-up for patients who were alive. Kaplan–Meier curves were used to compare survival rates across KIR-defined subgroups. The cumulative incidence of relapse (CIR) was estimated using Fine and Gray’s method, with NRM treated as a competing risk. Unadjusted outcomes were compared between groups with or without predicted alloreactivity using an indicator variable for transplants. Multivariate analysis was performed using Cox proportional hazards regression models, adjusting for clinical covariates such as donor age, donor sex (female), patient disease type, and HLA matching within donor/recipient pairs. Each KIR-based variable was individually examined by incorporating each into the multivariate model. Cases with incomplete data were excluded from the analysis.

Descriptive statistics were used to summarize population characteristics. Continuous variables are referred to as medians, and categorical variables as counts (%). Direct counting based on the presence or absence of each KIR gene was used to determine observed gene frequencies in the studied population. All statistical tests were two-sided, and p values less than 0.05 were considered statistically significant. Data were collected up to December 31, 2023. Statistical analysis was performed using the graphical user interface for R, EZR version 1.32 (24).

This analysis included 189 patients with a median age of 6.0 years (range, 1–16 years). Demographic and transplant-related characteristics are summarized in Table 1 . The initial diagnoses were acute myeloid leukemia/myelodysplastic syndrome in 96 patients (50.8%), acute lymphoblastic leukemia in 74 patients (39.2%), juvenile myelomonocytic leukemia (JMML) in 10 patients, mixed phenotype acute leukemia in 5 patients, and other diseases in 4 patients (including 1 case of acute non-lymphocytic leukemia, 1 case of T-lymphoblastic lymphoma, and 2 cases of myeloid sarcoma).

There were 47 relapses, with a median time to relapse of 84 days (range, 38–1,446 days), and 26 deaths, including 4 cases of non-relapse mortality (NRM). The median follow-up duration was 812 days (range, 85–3,193 days). The inhibitory KIR score and CF-iKIR score, derived from the presence of donor inhibitory KIRs and cognate HLA ligands in the recipient, had median values of 1.5 (range, 1–3) for the CF-iKIR score and 1.75 (range, 0.75–4.75) for the inhibitory KIR score.

The number of patients with positive predicted alloreactivity based on the four educational iKIR models ranged from 12 to 37. Collectively, 86 patients demonstrated predicted cytotoxicity with at least one positive iKIR alloreactivity, referred to as Synthesis-iKIR. In the activating KIR2DS1–education model, 36 patients exhibited predicted alloreactivity. In total, 107 patients (56.6% of the cohort) had positive predicted alloreactivity. These results are summarized in Table 2 .

Four inhibitory genes (KIR2DL1, KIR2DL3, KIR3DL1, and KIR3DL2) were present in nearly all donors, while the activating KIR genes KIR2DS4, KIR3DS1, and KIR2DS1 were found in 94.7%, 38.1%, and 40.7% of donors, respectively ( Figure 1C ). HLA genotyping within our cohort was counted and compared across different populations, as illustrated in Figures 1A and B . Consistent with previous reports, the allotype distributions of HLA-A, -B, and -C loci in the Chinese population differed from those in European Caucasians, despite our limited sample size.

Among the Chinese cohort, the most frequently observed subtypes were A*11:01, B*40:01, and C*01:02, whereas A*01:01, B*07:02, and C*07:02 were predominant among European Caucasians. The frequency of A*01:01, A*03:01, B*07:02, B*44:02, C*05:01, and C*04:01 among the Caucasian population was significantly higher than in the Chinese cohort. In contrast, A*33:03, A*02:07, B*46:01, B*13:01, C*08:01, and C*01:02 were more common in the Chinese cohort. According to epitope statistics, 77 patients had HLA-A3/A11, 118 had Bw4, and 174 were C1+.

In the context of allelic distribution, KIR2DL1*003 was the predominant allele among donors, exhibiting a positivity rate of 81.0%, followed by KIR2DL1*002 (13.7%), KIR2DL1*004 (2.5%), and KIR2DL1*069 (2.2%). For the KIR3DL1 locus, KIR3DL1*015 was the most prevalent allele, with a positivity rate of 63.7%, followed by KIR3DL1*005 and KIR3DL1*007, which had positivity rates of 14.8% and 9.9%, respectively. The allele distribution among recipients was generally similar to that of donors, except for three novel low-frequency alleles—KIR3DL1*038, KIR3DL1*077, and KIR2DL3*006—present in recipients at frequencies of 0.6%, 0.3%, and 0.6%, respectively. Allele frequencies for the three common KIR genes in donors and recipients are summarized in Table 3 .

We tested potential alloreactivity through the interaction of individual KIRs with HLA ligands within an educational model framework. NK cell alloreactivity was classified as ALLO (indicating the presence of predicted alloreactivity) or Non-ALLO (indicating the absence of predicted alloreactivity). Unlicensed NK cells, which typically exhibit hyporesponsiveness to stimuli, were classified as Non-ALLO and compared with NK cells showing predicted alloreactivity.

With the exception of the KIR3DL2–A3/A11 combination, patients whose donors exhibited non-inhibiting KIR–ligand interactions or those with activating KIR2DS1–ligand interactions did not show a lower risk of relapse, challenging their utility as prognostic indicators in pediatric haploidentical transplantation. The 5-year cumulative incidence of relapse (CIR) for patients classified as Non-ALLO based on KIR3DL2–A3/A11 interactions was significantly higher (0.318; 95% CI, 0.236–0.403) compared with those classified as ALLO (0.091; 95% CI, 0.004–0.347) (p = 0.012) ( Table 2 ). The influence of the KIR3DL2–A3/A11 combination on relapse risk remained significant after adjusting for clinical factors, with an 86.4% reduction in relapse risk for patients with negative A3/A11 receiving both KIR3DL2+ and A3/A11+ grafts (adjusted HR, 0.136; p = 0.0489) ( Table 4 , Figure 2 ).

In terms of KIR2DS1–C2 educational interactions, patients exhibiting activating KIR–HLA interactions had a higher 5-year CIR (0.352; 95% CI, 0.179–0.530; n = 36) compared with those without predicted alloreactivity (0.274; 95% CI, 0.190–0.365; n = 151), although this difference did not reach statistical significance (p = 0.495) ( Table 2 ).

Given that nearly all donors possessed the KIR2DL1-R²⁴⁵ type, a grouping analysis was performed for patients based on the presence or absence of allele mismatches. Among patients with the C2+ antigen, the presence of KIR2DL1 allele mismatches within donor–recipient pairs was strongly associated with an elevated risk of relapse (CIR, 0.482; 95% CI, 0.223–0.702; p = 0.023). Conversely, the absence of such mismatches correlated with a lower relapse risk (CIR, 0.296; 95% CI, 0.206–0.393), although no statistically significant difference in CIR was observed across the entire cohort ( Table 5 ). Furthermore, KIR2DL1 allele mismatches did not appear to influence OS ( Table 5 ). Patients were subsequently grouped according to donor KIR2DS1 status (activating versus nonactivating) and combinations of KIR3DL1–Bw4 (strong inhibiting versus weak/noninhibiting). The CIR curves for the four groups were almost superimposable ( Figure 3 ), and further analysis indicated no statistically significant differences in OS among these groups.

We then assessed the influence of KIR haplotypes on relapse incidence and survival. Previous studies have indicated that patients receiving Bx grafts have a lower risk of relapse compared with those with AA donors. However, our analysis did not reveal a significant trend toward a lower CIR for patients with Tel- or Cen-Bx grafts compared with Cen- or Tel-AA donors (p > 0.05) ( Table 2 ). Due to the limited number of individuals classified as the best cases, we combined the better and best individuals into a preferred group for comparison with a neutral group. This comparison yielded no significant differences in OS or CIR between the two groups. Similarly, Kaplan–Meier survival curves showed no significant differences in OS among patient groups categorized by donor Cen/Tel haplotypes (p > 0.05; data not shown).

An additional widely used algorithm for donor selection in HSCT is the KIR B content score, which classifies donors as neutral, better, or best, with increasing weighting allocated to donors exhibiting high Cen-B content. The donor cohort in this investigation exhibited a comparable proportion of neutral individuals (70.4%, n = 133) to that reported by Cooley et al. (69.0%), but a higher percentage of better individuals (27.5% vs. 20.2%) and a lower proportion of best individuals (1.6% vs. 10.8%) (19). Given the limited number of best individuals, we designated the better and best individuals as the preferred group and compared them with the neutral group. This comparison also revealed no statistically significant differences between the two groups with respect to OS and CIR (p > 0.05) ( Table 2 ). Furthermore, we completed donor Bx genotype analysis online (AFND, http://www.allelefrequencies.net/kir6001a.asp) to determine potential KIR gene combinations that may affect relapse and survival. Two Bx genotypes with a higher prevalence among donors in the cohort were screened: Genotype 2, characterized by the presence of all KIRs except for KIR2DL2, KIR2DS2, and KIR2DS3, and Genotype 8, which includes all KIRs except for KIR2DL2, KIR2DS2, and KIR2DS5. When compared to other Bx or AA genotypes, regardless of HLA antigen, neither Genotype 8 nor Genotype 2 was associated with CIR or OS across the entire cohort (p > 0.05) ( Table 2 ). Utilizing data on KIR haplotypes, we further explored additive models; however, neither the ct-KIR nor the weighted inhibitory score exhibited any significant correlation with CIR or OS ( Table 2 ).

Finally, multivariate Cox proportional hazards regression analysis demonstrated that the KIR3DL2–A3/A11 combination significantly reduced the incidence of relapse across the whole cohort (adjusted HR, 0.136; p < 0.005) ( Table 4 , Figure 2 ). Synthesis-iKIR, alone or in conjunction with KIR2DS1, had a significant independent effect on OS, with a clinical factor–adjusted HR of 0.305 (95% CI, 0.122–0.763; p < 0.005) and 0.316 (95% CI, 0.137–0.728; p < 0.005), respectively ( Table 4 , Figure 2 ). As anticipated, the disease state at the time of transplantation emerged as an independent prognostic factor ( Table 4 ). Other alloreactivity models did not successfully predict relapse or survival among pediatric recipients.