Systematic identification of minor histocompatibility antigens predicts outcomes of allogeneic hematopoietic cell transplantation

Abstract

T cell alloreactivity against minor histocompatibility antigens (mHAgs) — polymorphic peptides resulting from donor-recipient (D-R) disparity at sites of genetic polymorphisms, including SNPs, indels, and frameshifts — is at the core of the therapeutic effect of allogeneic hematopoietic cell transplantation (allo-HCT). Despite the crucial role of mHAgs in graft-versus-leukemia (GvL) and graft-versus-host (GvHD) reactions, it remains challenging to consistently link patient-specific mHAg repertoires to clinical outcomes. Here, we devise an analytic framework to systematically identify autosomal and Y-encoded mHAgs, including their detection on HLA class I ligandomes and functional verification of immunogenicity. The method relies on the integration of polymorphism detection by whole exome sequencing of germline DNA from D-R pairs with organ-specific transcriptional- and proteome-level expression. Application of this pipeline to a cohort of 220 HLA-matched allo-HCT D-R pairs uncovered associations with acute GvHD and chronic pulmonary GvHD occurrence, and defined promising GvL targets, confirmed in a validation cohort of 58 D-R pairs, for prevention or treatment of post-transplant disease recurrence.

Genomic analyses to quantify neoantigens arising from somatic tumor mutations have been tremendously impactful towards advancing cancer immunology and immunotherapy, and implementing personalized immune-based treatments in clinical practice. A widely used individualized form of immunotherapy that is potentially curative for many blood disorders is allogeneic hematopoietic cell transplantation (allo-HCT)¹, wherein a suitable stem cell donor is selected for each patient based on HLA matching. In this scenario, the driving principle underlying response is alloreactivity, primarily originating from immune responses against minor histocompatibility antigens (mHAgs), HLA-binding peptides derived from polymorphic protein sequences differing between donor and recipient (D-R) pairs². Like tumor neoantigens, mHAgs are sensed as foreign by donor T cells and are expected to be highly immunogenic due to the lack of central tolerance against them³. However, mHAgs are inherited as germline traits encoded by polymorphic genes rather than presenting as somatic events, hence they are not tumor-specific antigens per se. The pathogenesis of graft-versus-host disease (GvHD), the most detrimental immune-related complication after allo-HCT^{4, 5}, can be thus attributed to a donor-derived immune response directed against mHAgs either broadly expressed across tissues or specifically in GvHD-affected tissues. Conversely, the curative graft-versus-leukemia (GvL) effect can be conceptualized as the result of productive donor immune responses against mHAgs expressed on hematopoietic cells, including, but not limited to epitopes with hematopoietic tissue restriction^{2, 3}. While more than 50,000 allogeneic transplants are performed annually worldwide⁶ – with numbers still rising –, the beneficial effect of allo-HCT is too often hampered by disease relapse or complicated by GvHD, which together account for >50% of post-transplant mortality^{7, 8}. Hence, identification of molecular determinants to aid in predicting transplant outcomes is urgently needed.

Currently, only D-R HLA matching and the activity of GvHD prophylaxis strategies are available to help clinicians in this challenge⁹. We hypothesized that post-transplant outcomes could be impacted by genome-wide mHAg load, delineated in an organ- and malignancy-specific fashion. While >100 individual mHAgs have been identified thus far worldwide^{2, 10–13}, only few have been linked to increased risk of GvHD^14–17, and such associations have been only inconsistently validated^{18, 19}. More recently, the use of high-throughput sequencing technologies to catalogue the mHAg repertoire has been reported^20–23, but only one study linked mHAg burden to 1-year GvHD mortality in a large yet non-contemporary transplant cohort²⁴. Certainly, large-scale sequencing capabilities^{25, 26}, coupled with the availability of robust HLA class I epitope prediction models^{27, 28}, now offer an unprecedented opportunity to delineate the mHAg repertoire of allo-transplanted patients in a personalized fashion which can then be linked to clinical outcomes.

In the present study, we have developed a novel analytic framework to predict, starting from WES of germline DNA from D-R pairs, candidate mHAgs either expressed in organs that are frequently targeted by GvHD or in hematopoietic cells. By applying this pipeline to a cohort of 220 HLA-matched D-R pairs, we uncovered associations with GvHD outcomes, as well as defined potential GvL targets for prevention or treatment of post-transplant disease recurrence.

Results

Building a pipeline for systematic mHAg discovery

To identify single nucleotide germline variants (both autosomal and Y chromosome-encoded) present exclusively in the recipient and resulting in nonsynonymous alterations in protein-coding regions, we analyzed whole-exome sequencing (WES) data from paired D-R DNA. We devised an analytic pipeline that further incorporated: (i) filtering for variants in genes of interest, either expressed in GvHD-targeted tissues [skin, liver, gastro-intestinal (GI), lung, oral mucosa and lacrimal gland]; or in GvL genes, i.e. genes preferentially expressed in malignant (and non-malignant) hematopoietic cells, but not in non-hematopoietic tissues, and (ii) predicting variant-containing peptide 8–11mers (k-mers) binding to patient-specific HLA class I alleles using the tool HLAthena²⁸ (Fig 1a, Extended Data Fig 1).

Fig 1. Building a pipeline for systematic mHAg discovery.

a, Overview of the analysis workflow for prediction of minor histocompatibility antigens (mHAgs): starting from whole exome sequencing (WES) obtained from germline DNA of donor and recipient pairs, single-nucleotide polymorphisms (SNPs) altering the protein sequence and present only in the patient are identified. Through the application of tissue-specificity filters (see b), the polymorphisms present in genes expressed in the tissues of interest are selected and all possible k-mers encompassing the SNPs are computed and then subjected to HLA class I binding prediction. Resulting candidate mHAgs can be used for downstream applications. b, Tissue-specificity filters of the pipeline: top left, GvHD filter to comprehensively catalog the genes expressed in GvHD target organs. Single-cell RNA-Seq datasets were used to define an expression atlas of tissue-resident cell types from skin, liver, colon (GI), lung, oral mucosa and lacrimal gland (in addition to hematopoietic cell lineages); colored vertical bars on the upset plot – enumeration of genes with exclusive organ-specific expression; bottom left, GvL filter to identify genes with preferential expression in acute myeloid leukemia (AML): a single-cell based molecular classifier³⁹ was applied to the Beat AML dataset⁴⁰ to define distinct AML expression clusters (ECs). AML genes with evidence of expression at the RNA or protein level in the GTEx repository of human adult healthy tissues were excluded, to define a list of 259 candidate genes with preferential expression in AML (heatmap). The second component of the GvL filter (‘Hematopoietic filter’) is detailed in Extended Fig 4; right, Y mHAg filter, designed to identify mHAgs arising from genes in the male-specific region of the Y chromosome (MSY) which serve as additional targets of allorecognition in female-to-male transplants. The flowchart summarizes the principal steps of Y-chromosome pipeline.

A critical component of the pipeline was building a reliable expression atlas for acute or chronic GvHD-targeted tissues (‘GvHD filter’). To capture less frequent albeit biologically relevant cell types consistently underestimated in bulk RNA expression profiles, we evaluated multiple external single-cell datasets of healthy human skin²⁹, liver^{30, 31}, lung^{32, 33}, GI^{34, 35}, oral mucosa³⁶, and lacrimal gland³⁷ (Fig 1b- top left). For each tissue, we merged the corresponding single-cell data sets and then clustered and annotated tissue-resident cell types (Extended Data Fig 2), for which specific expressed genes were then identified. Using a cut-off of 5 counts per million for positive expression (based on the expression levels of lineage-defining markers, such as MLANA, SFTB and ALB, Extended Data Fig 3), we identified 13,512 genes expressed in ≥1 GvHD target tissue (Supplementary Table 1). Consistent with the notion that the identified genes were representative of the post-HCT setting, the vast majority (97.5%) of transcripts expressed in a newly generated dataset of post-HCT GI-resident cells were present in the GvHD filter (Extended Data Fig 4, Supplementary Fig 1).

Conversely, to predict candidate mHAgs with an acceptable safety profile if targeted therapeutically, we defined a set of genes with exclusive or near-exclusive expression within the hematopoietic compartment (‘GvL filter’). We focused on acute myeloid leukemia [AML] and myelodysplastic syndromes [MDS], as they represent the most common indications for allo-HCT in adults³⁸ (Fig 1b-bottom left). To capture the transcriptional heterogeneity of malignant myeloid cells, we applied a single-cell based classifier³⁹ to define genes expressed by leukemic cells in the Beat AML cohort⁴⁰, and then removed those expressed in non-hemopoietic tissues, per the GTEx database (both RNA⁴¹ and/or protein⁴²), applying a gender-specific tolerance for reproductive organs based on the inputted patient gender (Extended Data Fig 5a-b). Similarly, a broader ‘Hematopoietic filter’ was built by incorporating gene expression profiles from 18 mature hematopoietic lineages⁴³ and hematopoietic stem and precursor cells^{44, 45} (Extended Data Fig 5a-b). Through this stringent process, we identified 259 genes with preferential expression in AML, and 615 broadly expressed in the hematopoietic compartment (Supplementary Table 2). Notably, these genes were distributed: (i) across all chromosomes, thereby providing targetable options even in the presence of chromosomal aberrations, that frequently occur in myeloid malignancies; and (ii) across distinct biological pathways and cellular compartments (Extended Data Fig 5c-e). We further tuned the pipeline to incorporate RNA-Seq data from leukemic blasts, if available, to confirm patient-specific expression (Methods).

In the ~25% of allo-HCT cases consisting of male recipients paired with female donors (F→M)⁴⁶, chromosome Y-encoded genes represent an additional source of potentially immunogenic epitopes^{47, 48}, since the female immune system lacks central tolerance to these. To systematically evaluate Y-encoded mHAgs, we selected those genes from the 78 harbored in the male-specific region (MSY) of the Y chromosome with expression >1 TPM in ≥1 adult GvHD target tissue per GTEx (Fig 1b-right; Extended Data Fig 6a), generated corresponding in silico proteomes, and filtered out all homologous peptides (100% BLAST identity) arising from other chromosomal locations (primarily gene paralogues). HLA class I binding prediction of the remaining set of unique k-mers resulted in a median of 62 (range: 24 – 107) predicted epitopes per allele (Extended Data Fig 6b). The number of predicted epitopes per MSY gene varied greatly across the different HLA alleles, when grouped based on their peptide binding motifs (Extended Data Fig 6c).

We confirmed the ability of our pipeline to predict a set of known mHAgs (12 autosomal² and 9 Y-encoded⁴⁹, Supplementary Figure 2) in a training dataset of 19 D-R pairs with available WES data^{50, 51} (Supplementary Table 3). We verified that: (i) genes harboring the causative SNPs were included in the GvHD and Y filters, respectively; (ii) the SNPs were detected in various combinations among the D-R pairs in accordance with their reported allelic frequencies⁵²; (iii) in D-R pairs with correct SNP configuration (i.e. D^neg, R^pos), peptides corresponding to mHAg epitopes were part of the in silico generated k-mers; and iv) mHAg-corresponding peptides were predicted as HLA binders. Overall, our pipeline predicted 18 of 21 known mHAgs. Detection failures were due to the epitope originating from an alternative open reading frame⁵³ (not captured by our pipeline, which focuses on mHAgs deriving from SNPs, indels and frameshifts) or to HLA prediction rank above the threshold of positivity.

Antigenicity and immunogenicity of predicted mHAgs

To ascertain whether our predictions were supported by evidence of HLA presentation, we evaluated a previously generated HLA class I ligandome dataset of 60 single-HLA class I expressing B721.221 cell lines that encompassed all peptide-binding motifs identified by Sarkizova et al.²⁸ (Fig 2a, Extended Data Fig 7a). From whole-genome sequencing of parental B721.221 cells⁵⁴, we identified exonic non-synonymous SNPs present in the B721.221 cells (surrogate ‘HCT recipient’) that were absent in the reference genome (surrogate ‘donor’). Epitopes predicted from these sites were searched against the immunopeptidomes from these monoallelic cell lines. With the caveat of the limited sensitivity of MS-based immunopeptidome analysis, 517 were confirmed to be presented across all 60 HLA alleles (median of 8 [range: 1–20] MS-supported peptides per allele) (Fig 2b). To evaluate Y mHAg antigenicity, we generated and interrogated male immunopeptidomes from 11 male cell lines^55–57 and from the IEDB database⁵⁸ for epitopes predicted from the 9 MSY genes (since B721.221 cells are of female origin, Fig 2c). Indeed, 81 predicted Y mHAgs were presented across 37 HLA alleles (Fig 2d).

Figure 2. Antigenicity and immunogenicity of predicted mHAgs.

a, Schematic showing the workflow for identifying predicted epitopes from B721.221-specific SNPs and estimating their antigenicity. b, Antigenicity evaluation of predicted autosomal mHAgs in the B721.221 model: top stacked histograms show the total number of epitopes that are predicted per HLA allele (white), and the fraction of those with MS evidence of presentation in the corresponding monoallelic B721.221 cell line (colored based on the HLA allele: -A in blue, -B in teal, -C in orange); bottom histograms summarize the percentage of predicted epitopes that are detected by MS at the allele-specific level. c, Schematic showing the workflow for estimating the antigenicity of the Y-encoded mHAgs. d, Top stacked histograms show the total number of Y epitopes predicted per HLA allele (white), and the fraction of those having MS support (color-coded based on the HLA alleles) in male immunopeptidomes. Bottom histograms summarize the percentage of predicted Y epitopes with evidence of presentation. Unbiased immunogenicity validation of all predicted Y epitopes across 6 common HLA class I alleles using T cells from 3 female healthy donors: e, Representative flow cytometry plots depicting dextramer staining for Y mHAg-specific CD8⁺ T cells upon in vitro expansion. f, Histograms (showing mean +/− SEM) summarizing the number and frequency of expanded mHAg-specific T cells for each HLA tested. Grey - the threshold below which the result was considered negative (0.2% of CD8⁺ T cells). For each HLA, epitopes are ordered based on their prediction rank (range: 0 – 0.5%). g, Summary of the cumulative percentage of immunogenic predicted epitopes per allele. h, Representative flow cytometry plots showing CD137 upregulation (top graphs) and IFNγ production (bottom graphs) in response to stimulation with autologous antigen-presenting cells (LCLs) pulsed with the Y mHAg peptide (right) or irrelevant ovalbumin (OVA) peptide (left). Cells are gated on CD8⁺ cells. i, Scatter plots (mean +/− SEM) summarizing the results of the functional validation of 9 selected Y mHAg specificities assessed by CD137 upregulation (top) and IFNγ production (bottom). Significance of antigen-specific target recognition compared to OVA control (paired t test, with 2-tailed p values).

To estimate the fraction of predicted binders that were immunogenic, we focused on Y-encoded mHAgs in the setting of F→ M transplants, since SNP genotyping was not required for functional validation, and this facilitated ready and unbiased testing of all predicted epitopes for any given HLA allele. We isolated peripheral blood mononuclear cells (PBMCs) from 3 healthy female donors carrying 6 common class I HLA alleles (HLA-A*02:01, -A*01:01, -B*07:02, -B*18:01, -C*05:01, and -C*07:02), and challenged purified T cells with pools of synthetic peptides encompassing all 410 predicted binders (HLAthena rank <0.5, n=53 for HLA-A*02:01; n=97 for HLA-A*01:01; n=47 for HLA-B*07:02; n=68 for HLA-B*18:01; n=94 for HLA-C*05:01; and n=51 for HLA-C*07:02). Peptides were pulsed on CD3-depleted PBMCs, and cocultured with autologous female-derived T cells. Cells were similarly restimulated after 7 days; on day 14, they were screened for antigen specificity by dextramer staining.

Of 410 peptides tested, 110 (14, 21 and 75 HLA-A, -B and -C restricted, respectively) elicited antigen-specific T cell responses, with a preponderance of HLA-C-restricted epitopes (Fig 2e-g). For 9 such Y mHAgs, we confirmed antigen-specific recognition by CD137 upregulation and IFNγ production in response to the cognate epitope presented on autologous immortalized B cell lines (Fig 2h-i). To investigate the unexpectedly high number of immunogenic HLA-C-restricted epitopes, we evaluated epitope hydrophobicity, a property associated with immunogenicity⁵⁹. Observed inter-allele differences were mainly driven by the richness of hydrophobic residues in the allele-specific binding motifs, with HLA-A0201 having the highest frequency of hydrophobic predicted binders (Kyte-Doolittle hydrophobicity score, Extended Data Fig 7b-c). Comparison between peptides with or without experimental evidence of immunogenicity across HLA alleles revealed immunogenicity correlated with higher hydrophobicity scores only for HLA-A*01:01 and HLA-C*05:01 (Extended Data Fig 7d). Thus, hydrophobicity potentially contributed but was not the sole driver of HLA-C epitope immunogenicity. We considered additional contributing factors, including: (i) the in vitro stimulation conditions potentially favoring HLA-C restricted responses, as antigen-presenting cells express higher cell surface levels of HLA-C than other cell types⁶⁰; and (ii) a less stringent central tolerance for HLA-C epitopes, since cortical and medullary thymic epithelial cells tend to express less HLA-C than HLA-A and -B^{61, 62} (Extended Data Fig 7e). Nevertheless, by longitudinally tracking Y-encoded mHAg-specific T cells ex vivo in a patient undergoing a F→M transplant and experiencing severe chronic GvHD, we detected a sizeable population (accounting for up to 3% of the circulating CD8⁺ cell pool) of T cells specific for HLA-C*05:01-restricted epitopes with in vitro evidence of immunogenicity (Extended Data Fig 8a-b), thereby supporting the in vivo impact of HLA-C-restricted mHAgs. Notably, T cells specific against the dominant ZFY_186–194 C*05:01-restricted epitope were already detectable at high frequency in the donor pre-transplant, a result which we confirmed in the analysis of 7 additional HLA-C*05:01⁺ female donors of male patients (Extended Data Fig 8c).

mHAgs shape GvHD outcomes

To determine how mHAg load relates to GvHD risk, we used our mHAg pipeline to evaluate paired whole-exomes (≥85% target base coverage at ≥20x depth) generated from 220 D-R pairs treated with a matched-related donor (MRD) allo-HCT for AML/MDS (Table 1; Supplementary Fig 3-4, Supplementary Table S4). We considered the impact of diagnosis, prognostic risk score, status at transplant, conditioning regimen, and graft source against median mHAg load on risk for acute and chronic GvHD. Across these patients, total autosomal mHAg load > median, together with donor age, was associated with increased risk of developing grade II-IV acute GvHD (HR = 2.59, 95% confidence interval: 1.04 – 6.46, p = 0.04, Fig 3a-b). While the total autosomal mHAg load was not associated with development of chronic GvHD (Extended Data Fig 9a-b), subgroup analysis of 55 F→M transplants revealed that a combined score of autosomal and Y-encoded mHAgs > median was associated with development of NIH moderate/severe chronic GvHD (p = 0.042, Fig 3c).

Table 1. Characteristics of the clinical cohorts.

While no statistically significant differences were observed in the distribution of patient, donor or transplant characteristics across the DFCI-MRD and the HP-MRD cohorts, the latter was characterized by a more favorable disease status at transplant and preferential use of myeloablative conditioning. CR: complete remission; PBSC: peripheral blood stem cells; BM: bone marrow. PTCy: post-transplant cyclophosphamide; n.a.: not available.

	DFCI-MRD (n=220)	HP-MRD (n=58)
Patient Demographic
Age at transplant (years), median (range)	58 (20 – 75)	56 (26 – 72)
Gender, n (%)
Male	133 (60)	33 (57)
Female	87 (40)	25 (43)
Disease, n (%)
AML	153 (70)	37 (64)
MDS	67 (30)	21 (36)
Disease status at transplant, n (%)
CR1	125 (57)	24 (44)
CR>1	19 (8)	6 (10)
active disease	63 (29)	8 (14)
upfront	13 (6)	6 (10)
missing info	-	14 (24)
HCT Comorbidity Index, median (range)	3 (0 – 13)	n.a.
Donor Demographics
Age at transplant (years), median (range)	56 (18 – 69)	53 (17 – 72)
Gender, n (%)
Male	115 (52)	32 (55)
Female	105 (48)	26 (45)
Female to Male transplants, n (%)	59 (27)	13 (22)
Transplant characteristics
Conditioning regimen, n (%)
Myeloablative	127 (58)	41 (71)
Reduced intensity	93 (42)	17 (29)
Graft source, n (%)
PBSC	206 (94)	57 (98)
BM	14 (6)	1 (2)
GvHD prophylaxis, n (%)
CNI-based	212 (96)	58 (100)
PTCy-based	7 (3)	0 (0)
Other	1 (<1)	-
Outcomes
Follow-up of surviving patients (months), median (range)	60 (15 – 122)	76 (17 – 126)
Non-relapse mortality, n (%)
day +100	3 (1)	3 (5)
overall	19 (9)	14 (24)
Acute GvHD, severity, n (%)
grade II - IV	34 (15)	13 (22)
grade III - IV	22 (10)	8 (14)
Chronic GvHD, n (%)
patients evaluable for analysis	210 (95)	50 (85)
mild	22 (10)	8 (16)
moderate	64 (31)	13 (26)
severe	35 (17)	9 (18)
Relapse, n (%)	107 (49)	15 (26)
Graft failure, n (%)	3 (1)	0 (0)

Figure 3. mHAgs shape GvHD outcomes.

a, Cumulative incidence (CI) of grade II-IV acute GvHD stratifying patients based on the overall autosomal mHAg load below (light red) or above (dark red) the median value (median autosomal mHAg load = 632): 1-year CI of acute GvHD was 6.3% (95% confidence interval: 2.8% - 12%) and 16% (95% confidence interval: 9.5% - 23%), respectively; 2-sided p = 0.029 (Gray’s test). b, Hazard Ratios (HR with 95% confidence interval) from multivariate Cox proportional hazards regression modeling of variables influencing post-transplant outcomes. Statistically significant p values are in bold. c, Cumulative incidence of NIH moderate/severe chronic GvHD in the subgroup of female-to-male transplants, stratifying patients based on the load of autosomal and Y mHAgs. Patients with both autosomal and Y mHAg load above median (green) have a higher incidence of chronic GvHD compared to patients with all other combinations (grey): 3-year CI of chronic GvHD were 71% (95% confidence interval: 37% - 89%) versus 40% (95% confidence interval: 25% - 55%), 2-sided p = 0.042 (Gray’s test). Median autosomal mHAg load: 629; median Y mHAg load: 87. d, Distribution of 14 patients experiencing lung chronic GvHD (NIH grade 2–3) across deciles of lung mHAgs (i.e., load of predicted mHAgs in genes expressed in the lung). e, Cumulative incidence (CI) of lung chronic GvHD stratifying patients based on the lung mHAg load below (grey) or above (light blue) the median value (median lung-specific mHAg load: 424): 5-year CI of lung chronic GvHD are 0.93% (95% confidence interval: 0.08% - 4.7%) versus 12% (95% confidence interval: 6.7% - 20%), 2-sided p < 0.001 (Gray’s test). f, HR (and 95% confidence interval) from multivariate Cox proportional hazards regression modeling of variables influencing post-transplant outcomes. Statistically significant p values are in bold. HCT-CI: HCT Comorbidity Index.

We asked whether organ-specific mHAg load could predict the risk for organ-restricted acute or chronic GvHD. By evaluating organ-specific GvHD occurrence across deciles of mHAg load (Extended Data Fig 9c-d), we identified 2 scenarios with distinct patterns of association. In the case of chronic lung GvHD, all 14 positive cases occurred in patients with a lung-specific autosomal mHAg load > median (Fig 3d-e). By multivariable modelling with other key clinical variables, lung mHAg load > median was the sole factor associated with increased risk of lung chronic GvHD (OR = 10.45, 95% confidence interval: 1.31 – 83.31, p = 0.03, Fig 3f).

In contrast, for acute GvHD of the liver, we observed a different pattern, where 12 of 13 affected patients had a liver mHAg load < median (Fig 4a). Here, we hypothesized that a limited set of immunodominant liver mHAgs might drive the clinical manifestations in these patients. By analyzing the mHAg landscape of patients affected by liver acute GvHD, we observed an enriched representation of 7 polymorphisms in genes with preferential expression in liver (Fig 4b). Notably, these SNPs were less frequent in patients developing acute GvHD without liver involvement (p = 0.002, Fig 4c), and co-occurred at a lower frequency in patients experiencing chronic liver GvHD (p = 0.0009, Extended Data Fig 9e-f). This last observation could be partly explained by the presence of interferon-responsive elements in the promoter region of 4 of 7 of these genes (Extended Data Fig 9g), potentially leading to their increased expression (and presentation) in the inflammatory milieu characterizing the early peri-transplant period. For 3 patients experiencing liver acute GvHD, we could trace ex vivo T cells specific for putative driver liver mHAgs, which, in contrast to polyclonal or viral-specific T cells, showed an activated phenotype (measured by CD69 expression) that was temporally associated with GvHD onset (Fig 4d-g, Supplementary Fig 5-7). Restricting our analysis to patients receiving CNI-based GvHD prophylaxis (n = 212) did not alter the overall pattern of associations (Supplementary Fig 8). Overall, our data show the potential of mHAg prediction for refining the risk prediction of GvHD.

Figure 4. mHAgs in acute liver GvHD.

a, Distribution of 13 patients experiencing liver acute GvHD (grade II-IV) across deciles of liver mHAgs (i.e., load of predicted mHAgs in genes expressed in the liver). b, Heatmap showing the expression profile of the 7 genes with recurrent SNPs in patients experiencing acute liver GvHD in single-cell RNA-Seq clusters from liver, colon (GI) and skin. c, The number of co-occurring driver liver mHAgs in patients with acute GvHD with (green, n = 13) or without (grey, n = 21) liver involvement. Boxplots show min to max, with all individual values represented; 2-tailed p = 0.002 with Mann Whitney test. d, Direct ex vivo longitudinal tracking of T cells specific for the ATXN2 p.L107P driver liver mHAg (HLA-B0702-restricted) in MRD044, one of the patients in the DFCI-MRD cohort experiencing liver acute GvHD. Flow cytometry plots show BV785-CD8 staining on the X axis and dextramer on the Y axis, at the indicated post-transplant time-points. T cell reactivity against cytomegalovirus (CMV) pp65 (HLA-A0201-restricted) is shown as control (full flow cytometry panel is reported in Supplementary Fig 4). e, Timeline showing MRD044 clinical course: after early post-transplant relapse, the patient received salvage chemotherapy (CT) achieving complete remission which was consolidated with a single unmanipulated donor lymphocyte infusion (DLI). Soon after the DLI, patient developed late onset acute GvHD primarily involving liver as shown by serum bilirubin levels (in green). GvHD was treated with steroids (light blue) with gradual normalization of liver function. Under immune suppression, the patient experienced CMV reactivation (in red). Diamonds indicate peripheral blood (PB) samplings used for flow cytometric antigen-specific T cell quantification. f, Flow plot showing CD69 expression on mHAg-specific (orange), CMV-specific (blue) and polyclonal (yellow) CD8⁺ T cells at day +236 post allo-HCT (GvHD onset). g, Quantification of CD69 expression on liver mHAg-specific CD8⁺ T cells circulating at GvHD onset in 3 patients from the DFCI-MRD cohort experiencing liver acute GvHD (paired t test, with 2-tailed p values).

Predicted GvL mHAgs as targets for immunotherapy

To discover mHAgs with optimal GvL potential for therapeutic targeting, we focused on the subgroup of 45 patients in our cohort with long-term survival in the absence of both relapse and GvHD requiring systemic treatment (i.e. GvHD-free and relapse-free survival [GRFS⁶³], Supplementary Fig 3f). For these patients experiencing purely GvL, we analyzed the individual repertoire of predicted GvL mHAgs, and found 87 SNPs recurring in ≥5 GRFS patients (Fig 5a), that gave rise to predicted epitopes across multiple HLA restrictions (Supplementary Fig 9). Notably, these GvL SNPs were enriched in the subgroup of GRFS patients versus those without GRFS (p<0.0001, Fig 5b), suggesting that their overrepresentation was not merely dictated by high allelic frequency in the overall cohort. Using the median number of recurring GvL mHAgs in the GRFS subgroup (‘GRFS mHAgs’) as a cutoff to stratify the entire study cohort, we found that a higher GRFS mHAg load was associated with a protective effect against relapse in both univariate and multivariable analysis (Fig 5c-d), and conferred a benefit in 2-year overall survival (p = 0.03, Fig 5e). Analysis of an external cohort of 58 MRD allo-HCT D-R pairs (‘HP-MRD cohort’, Table 1, Supplementary Table 5) confirmed that GRFS mHAgs co-occurred at a higher frequency in patients not experiencing relapse nor GvHD (p = 0.009, Fig 5f) and showed a trend towards protection against relapse (p = 0.1, Fig 5g). Analysis of the 2 cohorts combined further validated the protective effect of GRFS mHAg load on relapse incidence and OS (Supplementary Fig 10). Finally, HLA class I immunopeptidome analysis of 5 AML cell lines revealed evidence of presentation for 7 epitopes predicted from genes included in the overall GvL filter (Fig 5h), three of which were also part of the GRFS mHAg set, i.e., APOBEC3F p.A108S (Fig 5i), MCPH1 p.R256I (Supplementary Fig 9) and CCDC171 p.K1069R.

Fig 5. Predicted GvL mHAgs as targets for leukemia immunotherapy.

a, Heatmap with 87 recurrent polymorphisms in GRFS patients. Right histograms - patients with same polymorphism (color-coded according to filter of origin); Bottom histograms - recurrent GvL SNPs/patient. b, Recurrent GvL SNPs in GRFS patients (GRFS^yes, pink) versus all other patients (GRFS^no, grey); boxplots show min to max and median values (2-tailed p<0.0001, Mann Whitney test). c, Cumulative incidence (CI) of relapse stratifying the DFCI-MRD cohort based on median number of recurrent GvL mHAgs in the GRFS subgroup (‘GRFS mHAgs’); 2-sided p=0.009, Gray’s test. d, Hazard Ratios (HR, with 95% confidence interval) of variables influencing post-transplant outcomes (significant p values in bold). e, Kaplan-Meier curve stratifying the DFCI-MRD cohort on the GRFS mHAg load below (grey) or above (pink) the median (p = 0.03 at 2 years, log-rank test). f, ‘GRFS mHAgs’ in the HP-MRD cohort stratified on GRFS outcome; boxplots show min to max and median values (2-tailed p = 0.006, Mann Whitney test). g, CI of relapse stratifying the HP-MRD cohort on median number of GRFS mHAgs; 2-sided p = 0.1, Gray’s test. h, HLA class I immunopeptidome analysis for 5 AML lines. For each, from left to right, dots show: (i) total predicted epitopes across GvL filter genes; (ii) predicted epitopes with RNA expression >10 TPM; (iii) predicted epitopes with evidence of HLA presentation. i, Mass spectrum of a detected HLA-B4403-presented GRFS mHAg (SET2 cell line). Red, blue, and green peaks represent y-, b-, and internal ions, respectively. j, Schematic of allo-HCT simulation: from the 1000 Genomes repository individuals identical at HLA class I alleles were identified across ethnicities (EUR: pink; EAS: yellow; SAS: green; AFR: teal; AMR: purple) to model population coverage for GRFS mHAg-targeting immunotherapeutic strategies. DRP: donor-recipient pair. k, For each GRFS SNP: top - allelic frequency (AF) per population (black bar: AF in the overall simulation cohort; blue shade: AF range with the highest probability of D-R mismatch); bottom - informative DRPs, where only individuals serving as ‘recipient’ have the SNP and an HLA restriction presenting the SNP-encompassing epitope. Stacked histograms are colored based on ‘recipient’ ethnicity. l, Population coverage for a GRFS-based vaccine using a 5-, 10- or 15-SNP design.

Given the limited size and ethnic diversity of our discovery and validation cohorts (composed mainly of patients with European ancestry), we used genomic data from the 1000 Genomes project⁶⁴ to estimate the feasibility of targeting our GRFS mHAg set in a broader population via in silico modeling of allo-HCT. From 2,504 individuals, we identified those with identical HLA class I alleles (with tolerance for 1 mismatch), mirroring a typical unrelated donor search in the National Marrow Donor Program (NMDP) registry. We found 844 individuals who could be matched to generate 2270 D-R pairs across 5 populations of disparate ancestry (Africa [AFR], East Asia [EAS], Europe [EUR], South Asia [SAS], and Americas [AMR]; Fig 5j, Supplementary Table 6). All 87 GvL SNPs were detected in the 844 individuals, albeit with differences in their relative representation across the distinct ancestry backgrounds. The only exceptions were GTPBP10 p.L85F absent in AFR and TLR7 p.Q11L absent in EAS (Fig 5k-top). By evaluating the number of informative D-R pairs per SNP, defined as those where only the individual serving as ‘recipient’ was positive for the SNP (Fig 5k-bottom), we found that the vast majority of GRFS SNPs were represented across the different ancestry backgrounds.

To gauge the feasibility of generating personalized mHAg-specific immunotherapy (i.e. cancer vaccines, adoptive cellular therapy^{65, 66}) based on the GRFS mHAg set, we estimated that a 10-SNP design (i.e., the number of GRFS SNPs used for the outcome correlative analyses in our discovery cohort, Fig 5c-d) would be possible for ~90% of the 1000 Genomes simulated cohort (EUR: 99%, EAS: 72%, SAS: 91%, AFR: 69%, and AMR: 74%, Fig 5l). Through a similar process, we calculated population coverage simulating a T cell-based approach targeting ≥1 epitope from the GRFS mHAg set. To this end, we identified 18 HLA alleles that could ensure ~99% coverage across diverse ethnic backgrounds (Supplementary Fig 11-14). For each HLA allele, the 3 most frequently predicted (and with highest agretopicity) epitopes from the GRFS SNP set were selected (Extended Data Fig 10a), defining a pool of 54 epitopes. We thus calculated that ≥1 epitope could be potentially targeted for 81% of the simulated D-R pairs (EUR: 91%, EAS: 78%, SAS: 71%, AFR: 62%, and AMR: 85%, Extended Data Fig 10b).

Discussion

The analytic pipeline described herein overcomes the limitations of previous efforts to delineate the patient-specific mHAg repertoires due to the incorporation of multiple novel features. First, our pipeline provides genome-wide assessment of the entire mHAg landscape, as it uses WES as data input for mHAg prediction (rather than SNP genotyping, utilized in most prior studies). Second, we more stringently filter mHAgs to be incorporated within the GvHD or GvL set than any prior study^{67, 68} on the basis of expression on GvHD targeted tissues or on malignant myeloid cells. This was achieved through: (i) intensive incorporation of information from high resolution single-cell RNA-seq data (analyzing 354,606 individual cells across 9 datasets covering 6 GvHD-targeted organs) to curate an expression atlas of all resident cell types present in organs frequently targeted by acute and chronic GvHD; and (ii) usage of RNA and protein GTEx data from >15,000 samples across 49 non-hematopoietic organs to selected genes with preferential expression in malignant and non-malignant hematopoietic cells and limited – if not absent – representation in healthy tissues. Third, the pipeline was designed to predict not only autosomal but also Y-encoded mHAgs. Fourth, it substantially reduces false positive predictions by eliminating all possible redundant k-mers. These are peptide sequences that, in addition to arising from the discordant SNP, are also generated from other sites in donor or patient exomes. This was accomplished through extensive pruning via sequential BLAST searches against the patient- and donor-specific proteomes reconstructed in silico from WES; indeed, we found that ~40% of SNP-encompassing k-mers were redundant. As a result of this pruning step, our unbiased screening of Y-encoded mHAgs predicted a higher percentage of epitopes with detectable immunogenicity than similar screenings performed for cancer neoantigens⁶⁹.

Our robust pipeline enabled us to gain numerous novel insights with several clear translational implications through its application to a cohort of 220 patients transplanted from HLA-matched related donors. For one, we were already able to identify genomic attributes that could aid in the prognostication of GvHD risk. We found that: (i) autosomal mHAg burden was predictive of grade II-IV acute GvHD incidence, in univariate and multivariable analysis; (ii) the presence of both autosomal and Y-encoded predicted mHAg load > median was associated with increased risk of chronic GvHD in F→M transplants, with the caveat of the limited number of sex-mismatched patients available for this subgroup analysis; (iii) two mechanisms of organ-specific mHAg expression were associated with risk of distinct types of GvHD, namely predicted organ-specific mHAg burden (for chronic lung GvHD) and presence of immunodominant organ-specific mHAgs (for acute liver GvHD). It has been long debated whether GvHD is driven by the cumulative load of mHAgs or rather, by a limited set of immunodominant ones. Our results suggest that both mechanisms underlying GvHD can be operative. Overall, our findings indicate that molecular characterization of D-R pairs to define GvHD risk could facilitate the design of personalized post-HCT treatments to minimize this highly morbid condition, including incorporation of additional immunosuppressive therapies, introduced early for high-risk patients, or reduced-intensity approaches in low-risk settings. This is particularly valuable for lung GvHD, which is associated with the highest morbidity and mortality⁷⁰.

A corollary to such molecular-based risk prognostication is the notion that quantification and qualitative characterization of mHAgs per D-R pair could inform donor selection, when more than one candidate is available.

Finally, we have devised an analytical framework to potentially identify candidate targets for post-transplant immunotherapy by defining GvL mHAgs that were significantly recurrent and had a protective effect against relapse in a subgroup of patients who had clinical evidence of a pure GvL effect. Our in silico modelling via analysis of the 1000 Genomes project⁶⁴ suggested that broad coverage across diverse ancestries with a limited set of 87 SNPs is feasible, with the caveat that such modelling does not fully recapitulate the complexity of allo-HCT. Such analyses pave the way for generating ‘off the shelf’ GvL mHAg-targeting vaccine or adoptive T cell therapeutics. While we focused our analysis on AML/MDS, the scalable design of the GvL filter enables incorporating expression profiles from other hematological malignancies with indication to allo-HCT but only a paucity of immunotherapy targets (such as T-ALL or PTCLs). We anticipate that systematic identification of GvL mHAgs may lead to their effective use to prevent or treat post-transplant relapse in an analogous fashion as has been tested for tumor neoantigens⁷¹.

The robustness and modular design of our pipeline facilitates the high throughput analysis of additional large-scale patient datasets, from which we can anticipate gaining greater sensitivity to further refine the algorithm for prognostication and prediction of response to HCT in future studies. Such datasets could include the analysis of ethnically diverse allo-HCT study populations to expand the list of actionable GvL mHAgs to improve population coverage – especially for individuals of Asian or African ancestry. Extension studies could also evaluate other transplant modalities, such as different donor types (i.e., unrelated donors, versus our current study of related donors) and GvHD prophylaxis regimens, such as post-transplant cyclophosphamide⁷² (versus our current study of transplants using predominantly CNI-based GvHD prophylaxis), that is expected to substantially impact the mHAg-specific T cell repertoire by in vivo purging of alloreactive T cells. Additional studies on larger cohorts would furthermore aid in providing increased power to potentially pinpoint other organ-specific driver mHAgs in settings beyond lung and liver GvHD.

In summary, our study demonstrates the potential applications of personalized mHAg prediction in allo-HCT. We note the potential technical limitations intrinsic to epitope prediction tools, such as high false positive rates and non-linear relationship between epitope antigenicity and immunogenicity. We also note that future refinements to the pipeline could expand prediction capabilities to address other mHAg species (i.e. resulting from gene deletions⁷³, intron retention⁷⁴, 5’ or 3’ UTR SNPs⁷⁵, or alternative ORFs⁵³) that are beyond the currently included SNPs, indels and frameshifts. Another refinement could relate to extending the evaluation of mHAgs to include HLA class II prediction, given the recognized role of CD4-restricted responses in alloreactivity^76–78, although this remains less accurate than HLA class I prediction⁷⁹. HCT represents a model setting of precision medicine: one individual donor is selected for one individual recipient based on genetic findings, and optimal donor matching to modulate the GvHD/GvL effects has been an inherent point of debate since its inception. We assert that immunogenetic models, such as the one proposed herein, will have important implications for precision immuno-oncology to improve patient outcomes.

Methods

Patient samples

Peripheral blood mononuclear cells (PBMCs) were collected from allo-HCT patients and donors following written informed consent through sample collection protocols approved by the Dana-Farber and Hospital de la Princesa Institutional Review Board in accordance with the principles of the Declaration of Helsinki. All samples were processed by Ficoll-Paque PLUS (Fisher Scientific) density gradient centrifugation and then cryopreserved with FBS/10% DMSO and stored in liquid nitrogen until time of analysis.

For the DFCI-AML-MRD cohort, we analyzed matched donor and recipient DNA, as well as PBMCs in some cases, collected from all adult patients who underwent first T-replete allo-HCT from a matched related donor (MRD) between January 1^st, 2013 and December 31^st, 2020 at Dana-Farber Cancer Institute, Boston. For the HP-MRD cohort, we analyzed matched donor and recipient DNA collected from 58 adult patients who underwent first T-replete MRD allo-HCT for myeloid disease (AML/MDS) between January 1^st, 2011 and April 30^th, 2021 at Hospital de la Princesa, Madrid. All patients were considered in complete remission if disease activity could not be documented by BM evaluation. Patients not falling within this definition were categorized as having active disease. Minimal residual disease was not considered for the definition of disease status at transplant, as the information was not available for all study patients. Patients were stratified by disease-specific prognostic risk scores: ELN 2017⁸⁰ and R-IPSS⁸¹ for AML and MDS, respectively. For multivariate analysis the ELN and R-IPSS scores were consolidated in a single variable, termed ‘overall risk score’ and comprising 3 categories: i) favorable, including ELN favorable and R-IPSS very low/low; ii) intermediate, encompassing ELN intermediate and R-IPSS intermediate; and iii) adverse, including ELN adverse and R-IPSS high/very high. Clinical diagnosis and grading of acute GvHD were annotated according to consensus criteria^{82, 83}. Chronic GvHD diagnosis and grading were based on the National Institutes of Health consensus criteria⁸⁴.

Exome sequencing, processing and analysis.

Library preparation and sequencing. A total of 556 DNA samples originating from 278 D-R pairs were processed and sequenced by whole-exome sequencing (≥85% target base coverage at ≥20x depth; Genomics Platform, Broad Institute). For DFCI-MRD D-R pairs, genomic DNA (250 ng) was provided by the HLA typing Lab at the Brigham and Women Hospital (Boston, MA). Libraries were constructed as previously described⁸⁵ and sequenced on NovaSeq S4 platform using the NovaSeq 6000 Xp workflow with a paired-end reads of 2×151bp. Quality control identification check was performed using fingerprint genotyping of 95 common SNPs by Fluidigm Genotyping. Alignment and quality control. All DNA sequence data were processed through Broad Institute pipelines. Outputs from Illumina software were processed by the Picard data-processing pipeline to yield BAM/CRAM files. Raw sequence data were aligned to the human genome hg19 genome assembly (v.b37, using BWA-MEM [v.0.7.15-r1140]) provided by the Picard and Genome Analysis Toolkit (GATK) developed at the Broad Institute⁵⁵, a process that involves marking duplicate reads, recalibrating base qualities and realigning around indels.

Single-cell analysis

For all publicly available datasets of skin²⁹, liver^{30, 31}, GI^{34, 35}, lung^{32, 33}, lacrimal gland³⁷ and oral mucosa³⁶, files were downloaded from the appropriate repository (GEO or EGAS). For each dataset, only data generated from healthy subjects were extracted and imported into Seurat-compatible objects. All quality control, normalization, and downstream analyses were performed using the R package Seurat⁸⁶ (v.4.3.0, https://github.com/satijalab/seurat). Low quality cells were excluded from downstream analyses based on percentage of mitochondrial reads (<20), features per cell (>200 and <4,000), and number of reads per cell (<20,000). For GI, liver and lung, 2 independent datasets were merged, and data integration and batch correction were performed using Harmony. Louvain clustering was then performed on all cells with the ‘FindClusters’ function using the first 50 PCs and a resolution of 1. Through manual annotation, clusters of hematopoietic origin were identified using standard lineage markers (PTPRC, CD3E, MS4A1, CD79B, KLRB1, KLRG1, LYZ, CD68, CD14, KIT) and excluded from further analysis (with the only exceptions of Langerhans cells in skin and Kupffer cells in liver). Clustering was repeated after the removal of immune cells, and non-immune cell types were manually annotated using the same set of lineage-defining genes used in the original publications. For each cluster, gene expression profiles were compiled, upon exclusion of non-coding and HLA genes. Genes with a sum count >5 counts per million (CPM) were retained to create the final gene list genes to be used for the GvHD filter.

For the thymic single-cell dataset from Park et al⁶², files were downloaded from https://developmentcellatlas.ncl.ac.uk and loaded into a Seurat object as described above. EPCAM⁺ thymic epithelial cells were clustered using a resolution of 0.03 to define 3 macro-clusters (corresponding to cortical, medullary, and myo/neuroendocrine thymic epithelial cells). Pseudo-bulk expression of AIRE, HLA-A, HLA-B and HLA-C within the 3 clusters was calculated using the ‘AggregateExpression’ function in Seurat and resulting expression levels were normalized on the number of cells present in each cluster.

For colon biopsies from allo-HCT patients (n=3), samples were collected at time of diagnostic coloscopy for suspected GI acute GvHD, upon written informed consent to a DFCI/Boston Children’s Hospital IRB-approved protocol. Samples were obtained at a median time of 89 days post allo-HCT (range: 21 – 102). and enzymatically digested to generate single-cell suspensions. Sample processing for scRNA-seq was performed (Chromium Single Cell 5′ Library and Gel Bead Kit, 10x Genomics), following manufacturer’s recommendations, and sequenced on Illumina NovaSeq S4 platform. scRNA-seq data were processed with Cell Ranger software (version 3.1.0) and resulting count matrices were read into Seurat (version 4.3.0). Cells were filtered to retain those with 20% or less mitochondrial RNA content and with a unique molecular identifier (UMI) content comprised between 250 and 10,000. Overall, scRNA-seq data comprised 20,405 cells that passed quality filters, of which 10,159 (allo-HCT pt#1: 5,235; allo-HCT pt#2: 640; and allo-HCT pt#3: 4,284; Supplementary Fig 2) were annotated as non-immune and utilized for subsequent analyses performed as described above for the external datasets. To evaluate interferon-related signature enrichment in allo-HCT vs. healthy subjects, we used AUCell⁸⁷, a tool specifically designed to evaluate enrichment of user-defined gene signatures in sc-RNASeq data. Three interferon (IFN)-related signatures were used: one derived from patients with IFN-mediated autoinflammatory diseases⁸⁸ and the IFNα and IFNγ signatures from MSigDB⁸⁹. Despite observing an enrichment of IFN-related genes in GI cells from allo-HCT patients, 6,894 (97.5%) of the 7,070 genes expressed in the allo-HCT GI dataset were already present in the GvHD filter.

Minor Histocompatibility Antigen pipeline

Input files for the pipeline were donor and recipient exomes in the form of BAM/CRAM files aligned to the hg19 reference genome, that were processed using DeepVariant⁹⁰ version 1.1.0 and Funcotator (part of the GATK package, v4.2.6.1) to define germline variants (including SNPs, indels and frameshifts) and proceed to their annotation, respectively. By comparing resulting VCF files from each D-R pair, only germline variants present exclusively in the recipient and producing nonsynonymous alteration in protein-coding regions were retained. Subsequent steps included filtering for variants present in genes of interest through the use of ‘GvHD’, ‘GvL’ and/or ‘Y mHAg’ filters. The ‘GvHD filter’ has been already described in the previous section. The GvL filter features 2 main components: the ‘AML filter’ (to define genes expressed by malignant myeloid cells) and the ‘Hematopoietic filter’ (to define genes expressed across the different hematopoietic lineages). With respect to the ‘AML filter’ design, to address the challenge posed by AML heterogeneity, a molecular classifier based on AML single-cell expression profiles³⁹ is used to define 7 expression clusters in the Beat AML cohort⁴⁰. Genes expressed with TPM ≥2 in each cluster are retained. Next, to select genes with preferential expression in AML, all genes with median expression >5 TPM and/or max expression >8 TPM in any of the normal tissues present in the Genotype-Tissue Expression (GTEx) Project RNA repository were filtered out. Based on the patient gender (a required input information when running the pipeline), the list of GTEx tissues varies to include either female reproductive organs (for male patients) or male reproductive organs (for female patients); for both female and male patients, whole blood, spleen and lung are excluded given the inherent large proportion of hematopoietic cells present. The resulting list of genes is then subjected to a second filtering round using the GTEx proteomics dataset⁴², with all genes with a tissue specificity score (TS) >2 in any GTEx tissue being removed; in instances when protein data was missing, we applied a second RNA-Seq filtering step, using a cutoff of log2(TPM)>5 in any GTEx tissue, a threshold shown to reliably correspond to protein expression.⁴² The same filtering steps are applied for the generation of the ‘Hematopoietic filter’, using as starting point publicly available RNA-Seq expression profiles of 18 purified mature hemopoietic cell types⁴³ and of hematopoietic stem and precursor cells^{44, 45}. Overall, the resulting gene sets are composed by 259 genes for the ‘AML filter’, and 615 for the ‘Hematopoietic filter’ (with 224 overlapping genes). The list of variant-encompassing k-mers are then blasted against custom in silico proteomes inferred from the exomes from both donor and recipient, sequentially. All k-mers found elsewhere in the proteomes (100% homology) are discarded, and remaining unique k-mers are subjected to binding prediction to the patient-specific class I HLA alleles through HLAthena²⁸ using a threshold of 0.5% prediction rank to define binders. Whenever available, pre-transplant AML RNA-Seq expression can be provided as an optional input file, in order to filter prediction results based on actual expression in the specific patient analyzed. The pipeline is entirely docked on Terra, one of the NCI Cloud Pilots systems, to ensure fully reproducible analyses.

Immunopeptidome analysis

B721.221 monoallelic cell line immunopeptidome analysis.

WGS from parental B721.221 was processed using BWA-MEM [v.0.7.15-r1140], subsetted to coding regions, and annotated with DeepVariant. A modified version of the mHAg pipeline was used in order to predict all k-mers deriving from non-synonymous polymorphisms present in the B721.221 (serving as surrogate ‘recipient’) and not in the reference hg19 genome (serving as surrogate ‘donor’). Genes were filtered based on RNA-Seq expression in B721.221 (TPM>10). Raw mass spectra from 60 HLA-monoallelic B721.221 cell lines were interpreted with the Spectrum Mill (SM) software package, v.8.01 (Broad Institute; proteomics.broadinstitute.org) as previously described²⁸. MS/MS spectra were searched against the B721.221-specific SNPs appended to a base reference proteome composed by 98,298 entries, including all University of California Santa Cruz Genome Browser genes with hg19 annotation of the genome and its protein-coding transcripts (63,691 entries), 602 common laboratory contaminants, 2043 curated smORFs (lncRNA and uORFs), 237,427 nuORF DB v1.037, and the JPT iRT peptides (JPT Peptide Technologies, Berlin, Germany, RTK-1–10 pmol) for a total of 303,803 entries⁵⁴. Target-decoy FDR estimation was enabled by Spectrum Mill with on-the-fly generation of decoy sequences during searches. For each candidate sequence passing the precursor mass tolerance filter, the internal sequence was reversed, while holding fixed the second position and the peptide C terminus, to maintain not only equal size target and decoy search spaces, but also comparable HLA class I binding motifs among the sequence candidate population. Peptide spectrum matches (PSMs) within <1% false discovery rate (% FDR) were confidently assigned for individual spectra via the target decoy estimation of the SM Autovalidation module. PSMs were filtered for precursor charges of +1 to +5, sequence lengths ranging between 9 to 40 amino acids, and a minimum backbone cleavage score of 5. PSMs were consolidated to the peptide level to generate lists of confidently observed peptides for each allele using the Spectrum Mill protein/peptide summary module’s peptide-distinct mode with filtering distinct peptides set to case sensitive.

IEDB search for Y-encoded mHAgs.

A curated set of previously identified HLA class I ligands was downloaded from IEDB at http://www.iedb.org/downloader.php?file_name=doc/mhc_ligand_full.zip (accessed on September 19, 2022). Records were filtered to Organism = homo sapiens, Epitope Object Type = Linear peptide, Parent Protein and Antigen Name = MSY genes of interest and Allele Name consistent with human HLA class I nomenclature.

AML cell lines.

Mono-MAC6, MUTZ-3, OCI-AML3 and SET2 were purchased from DSMZ; MOLM-13 were kindly gifted by the Genovese Lab (BCH/DFCI). Cells were cultured as detailed in Supplementary Table 7. All cell lines are part of the LL-100 panel⁹¹. For each, paired WES and RNA-Seq data were downloaded from ENA (accession numbers: PRJEB30297 and PRJEB30312, respectively). WES was run through a modified version of the mHAg pipeline, using hg19 as the surrogate ‘donor’ to define the GvL mHAg landscape of each AML cell line. RNA-Seq data was used to evaluate the expression levels of the genes harboring the GvL SNPs and to infer HLA class I typing with OptiType⁹². Up to 50 million or 0.2 g of each AML cell line were immunoprecipitated, as previously described²⁸. Peptides of three immunoprecipitations were combined, acid eluted, and analyzed using LC/MS-MS on Orbitrap Exploris 480 equipped with a FAIMSpro interface (Thermo Fisher Scientific)⁹³. MS spectra were interpreted with Spectrum Mill as detailed in the ‘B721.221 monoallelic cell line immunopeptidome analysis’ section, with the only differences being the MH+ in the range of 700–3000 and a precursor charge of +1 to +5, and Ensembl protein reference databases with hg19 annotation of the genome proteome and the JPT iRT peptides were customized with AML and cell line specific the GvL SNP entries.

Immunogenicity assays

Peptides.

All synthetic 8–11mers peptides used throughout the study were purchased as microscale libraries with purity >70% (median purity: 93.8%, range: 70.1% – 99.8%) from GenScript, and dissolved in ultrapure DMSO (Sigma Aldrich) to a stock concentration of 10 mM.

Antigen-specific stimulation.

PBMCs for immunogenicity testing were isolated by Ficoll-Paque PLUS (Fisher Scientific) density gradient centrifugation from peripheral blood samples of healthy donors. DNA was extracted with the DNA mini kit (Qiagen) following manufacturer’s instructions and used for HLA typing (through the Brigham and Hospital HLA typing lab), and for SRY PCR to select those from female donors with appropriate HLA restrictions. For the SRY PCR, the following primers were used: 5’-CATGAACGCATTCATCGTGTGGTC-3’ and 5’-CTGCGGGAAGCAAACTGCAATTCTT-3’, following the protocol described by Cui et al⁹⁴. T cells were enriched from PBMCs using PanT cell selection kit (Miltenyi Biotech) and then stimulated at 1:1 ratio with autologous CD3-depleted PBMCs pulsed with pools of 5 µM synthetic Y-encoded mHAg peptides, in RPMI-1640 supplemented with 5% AB-positive heat-inactivated human serum (Gemini Bioproduct), in presence of 30 ng/ml of IL-21 (Peprotech) till day 3, then replaced by 5 ng/ml of IL-7 and IL-15 (Peprotech) from day 4 on. On day 7 cultures were restimulated with CD3-depleted PBMCs pulsed with 2 µM peptide pools, in the presence of 5 ng/ml of IL-7 and IL-15. Half-medium change and supplementation of cytokines were performed every 3 days.

Quantification of mHAg-specific T cells.

The presence of antigen-specific T cells was determined on day 14 post stimulation, by staining with HLA-A0201, -A0101, -B0702, -B1801, -C0501 and -C0702 easYmers (Immunaware) first loaded with the relevant peptides, and then docked on U-Load dextramers (Immudex), following manufacturer’s instructions. mHAg specificity was assessed by analyzing up to 3 specificities at a time using triplets of FITC-, PE- and APC-conjugated U-Loads. Prior to staining, the pools of dextramer reagents were prepared in the presence of 100 µM d-Biotin per Immudex protocol to reduce the risk of staining artifacts. Cells were then stained with anti-CD3 antibody (conjugated with BV510, clone UCHT1, Biolegend; dilution 1:100), CD8 (BV785, clone RPA-T8, Biolegend, dilution 1:50), CD4 (PerCP-Cy5.5, clone RPA-T4, Biolegend, dilution 1:50), Zombie Violet (vitality dye, Biolegend, dilution 1:500), and CD69 (APC-Cy7, clone FN50, Biolegend, dilution 1:50) for quantification of circulating liver mHAg-specific T cells. All stainings were performed in 100 µL using 0.5 – 2 × 10⁶ cells. Samples were acquired on a high throughput sampler (HTS)-equipped Fortessa cytometer (BD Biosciences) using the BD FACS Diva software v8.0 and analyzed using Flowjo v10.8 software (BD Biosciences). The gating strategy is reported in Supplementary Fig 15. For each sample, non-specific binding to more than one dextramer was manually evaluated, and if present double-positive cells were excluded from further analysis.

Generation of EBV-immortalized B cell lines (EBV-LCLs).

CD19⁺ cells were isolated from PBMCs of the same female healthy subjects used for the immunogenicity studies through magnetic selection (Miltenyi Biotech). 0.2 – 0.5 × 10⁶ purified B cells were then incubated with a 1:1 mix of RPM1–1640 supplemented with 20% FBS and 1% penicillin/streptomycin, and EBV supernatant (ATCC) in a total volume of 200 µL. Every 3–4 days thereafter, cultures were examined for signs of transformation and fresh medium was added. EBV-immortalized cell lines were established in 3–4 weeks and were then maintained in culture with RPM1–1640 supplemented with 10% FBS at a maximum density of 1–2 × 10⁶ cells/mL.

CD137 and IFNγ catch assays.

For 9 mHAg specificities reported in Fig 2, selected based on the expansion level of mHAg-specific T cells, functional validation of antigen-specificity was performed used 2 assays: CD137 upregulation and IFNγ production upon coculture with autologous EBV-LCLs pulsed with the relevant (or control) peptides. Peptide pulsing of target cells was performed by incubating EBV-LCLs in serum-free medium at a density of 2 × 10⁶ cells/ml for 2 h in the presence of 5 µM peptides. Ovalbumin (OVA) peptide was used as control. For CD137 assay, upon overnight co-incubation of effector and target cells, mHAg specificity was assessed by flow cytometric detection of CD137 upregulation on CD8⁺ T cells, using the following antibodies: anti-human CD8a (BV785, clone RPA-T8, Biolegend), CD4 (PerCP-Cy5.5, clone RPA-T4, Biolegend), CD137 (PE, clone 4B4–1, Biolegend, dilution 1:50), Zombie Violet (vitality dye, Biolegend), CD19 (APC-Fire750, clone HIB19, Biolegend, dilution 1:50) and CD14 (APC-Fire750, clone M5E2, Biolegend, dilution 1:50). All stainings were performed in 100 µL using 0.5 – 2 × 10⁶ cells. Data were acquired on a high throughput sampler (HTS)-equipped Fortessa cytometer (BD Biosciences) using the BD FACS Diva software v8.0 and analyzed using Flowjo v10.8 software (BD Biosciences). To detect the antigen-specific production of IFNγ, T cells were co-incubated with autologous EBV-LCL pulsed with the appropriate peptides for 6 hours, and then labeled with the IFNγ secretion Assay detection kit in PE (Miltenyi Biotech) following manufacturer’s instructions. Cells were then counterstained with the following antibodies: anti-human CD8a (BV785, clone RPA-T8, Biolegend), CD4 (PerCP-Cy5.5, clone RPA-T4, Biolegend), Zombie Violet (vitality dye, Biolegend), CD19 (APC-Fire750, clone HIB19, Biolegend) and CD14 (APC-Fire750, clone M5E2, Biolegend). Data were acquired on a high throughput sampler (HTS)-equipped Fortessa cytometer (BD Biosciences) using the BD FACS Diva software v8.0 and analyzed using Flowjo v10.8 software (BD Biosciences). Gating strategies are reported in Supplementary Fig 16 and 17.

Peptide hydrophobicity calculation.

The GRAVY hydrophobicity index with the Kyte-Doolittle scale was calculated for the 410 tested Y mHAgs using the ‘Peptides’ R package (v 2.4.4).

1000 Genomes in silico allo-HCT simulation

The WGS bam files for the 1000 Genomes Project (1000G), mapped to the GRCH37 reference genome, were obtained from the Google Brain Genomics repository. Suitable D-R pairs were selected based on the HLA typing information available for each individual in the 1000G dataset. The criterion for selection was that ≥5 alleles in the HLA class I genes (HLA-A, HLA-B, HLA-C) had to be matched between donor and recipient. A total of 844 individuals combined into 2270 unique D-R pairs satisfying this criterion were identified. The study cohort comprised 239 individuals from EUR, 216 from EAS, 160 from SAS, 140 from AFR and 89 from AMR ancestry. WGS bam files from the selected individuals were reduced to include only the coding region of the genes included in the AML and Hematopoietic filters. Variant calling was performed on the reduced bam files using DeepVariant (version 1.1.0)⁹⁰, generating germline variant call files. Allelic frequency (AF) of each GvL SNP was calculated for the 1000G cohort and compared to the gnomAD allele frequency which was incorporated into the Funcotator task. Population coverage analysis was performed as previously described by Bui et al.⁹⁵

Statistical analyses

All statistical analyses were conducted using Prism v.9.5 (GraphPad Software) or R v.4.2.2 (https://www.r-project.org/). The following statistical tests were used in this study, unless otherwise indicated: paired t test, unpaired t test or Mann Whitney test, based on prior assessment of data normality (using Shapiro–Wilk test or Kolmogorov–Smirnov test, depending on sample size). The minimum threshold for significance was defined as p < 0.05, and all statistical tests were two-sided.

Clinical outcomes are reported as of June 2022 (lock date: 25 June 2022). Overall survival (OS) was defined as the time from stem cell infusion to death from any cause. Patients who were alive were censored at the time last seen alive. GvHD-free/relapse-free survival (GRFS) was defined as time to first occurrence of grade III-IV acute GvHD, chronic GvHD requiring systemic treatment, relapse, or death, whichever occurred first. Probabilities of OS and GRFS were estimated with the Kaplan-Meier method using the ‘survival’ (version 3.4–0) R package, while cumulative incidences were estimated using the ‘tidycmprsk’ (version 0.2.0) R package. Cumulative incidence of non-relapse mortality (NRM), relapse, acute GvHD, and chronic GvHD were computed to take into account the presence of competing risks. Specifically, in calculating the cumulative incidence rates of NRM, the competing risk was relapse; when calculating relapse, the competing risk was NRM. For acute and chronic GvHD, the competing risks were relapse and death from other causes. All patients were considered evaluable for acute GvHD analysis, and those who had a documented engraftment and a follow-up >100 days were evaluated for chronic GvHD (n = 210). To link mHAg load with GvHD outcomes (i.e., overall acute and chronic, as well as organ-specific GvHD), initial evaluation of deciles of mHAg load was used to guide subsequent analyses and define stratification criteria (mHAg load median vs. analysis of individual SNPs). The log-rank test and the Gray test were used for group comparison of OS and cumulative incidence of relapse and GvHD, respectively. The risk factor analysis on acute GvHD-, lung chronic GvHD- and relapse-specific hazard was firstly assessed by the Cox proportional-hazards model, and hazard ratio (HR) with the associate 95% confidence interval were calculated for each variable. Forest plots were performed by the forest_model function in the ‘forestmodel’ (version 0.6.2) R package. Analysis of promoter region for interferon-responsive elements in the genes associated with liver acute GvHD was performed using the open access interferome database (www.interferome.org)⁹⁶, using the following search conditions: all interferon types, homo sapiens species (all systems and sample types). All datasets used in the study are summarized in Supplementary Fig 18.

Extended Data

Extended Data Fig 1. Pipeline details.

a, Detailed workflow for the prediction of autosomal (left) and Y-encoded (right) mHAgs. b, Pipeline outputs for the training AML cohort composed of 11 D-R pairs (see Supplementary Table 3). Shown are the median number (and interquartile range) of hits for each step of the pipeline for matched-related donor (MRD, blue; n=2) and unrelated donor (URD, orange; n=9) transplants. The number of discordant variants between donors and recipients was, as expected, higher in URD than in MRD transplants. Pie charts (on the right) - distribution of the types of discordant variants in MRD (top) and URD (bottom) D-R pairs, i.e., SNPs: single nucleotide polymorphisms, DEL: deletions, INS: insertions. GvHD denotes graft-versus-host disease; GvL: graft-versus-leukemia.

Extended Data Fig 2. Single-cell data analysis to define the GvHD filter gene set.

a, Summary of the single-cell datasets used, related to the following organs target of GvHD: oral mucosa, lacrimal gland (eye), skin, liver, colon (GI), and lung. For each organ analyzed, the number of datasets and their accession numbers are shown, together with the total number of cells after standard single-cell data QC analysis, as well as after removal of immune cells. The ID of the second lung dataset has been abbreviated for ease of visualization, but the full identifier is reported in panel c legend. b, UMAP plots showing clustering of the resident cell types for each organ. Note that for ‘eye’, the dataset includes both primary cells and organoids (derived from ductal cells) which were both analyzed. c, Violin plots of lineage-defining markers for each cell type across the different organ datasets. d, UMAP plots showing the relative contribution of individual datasets, for those organs with 2 available. Vasc.: vascular; Lymph.: lymphatic; KC: keratinocyte; LSEC: liver sinusoidal endothelial cells; CT: crypt top.

Extended Data Fig 3. Threshold definition for single-cell based expression atlas.

a, To minimize the drop-out effect common in single-cell RNA-Seq data, gene expression was analyzed in a pseudo-bulk fashion for each cluster. To define the threshold for positive expression having the best signal-to-noise ratio, the expression levels of lineage-specific markers such as MLANA (melanin, expressed only in melanocytes), SFTB (surfactant B, expressed only in the lung), and ALB (albumin, expressed only in the liver) were analyzed, using as control a pan-expressed gene, B2M. The dot plot shows the expression levels for all single-cell clusters (with the tiles next to their names indicating the organ of origin: green for ‘liver’, light blue for ‘lung’, yellow for ‘skin’, orange for ‘GI’, red for ‘oral mucosa’ and navy blue for ‘eye’). b, Comparison of the expression profile of fibroblasts from 2 independent single-cell dataset (oral mucosa and skin). Results of the linear regression analysis (R squared and p value) are reported and show a substantial transcriptional identity between fibroblasts across different anatomical sites and datasets. c, Comparison of the expression profile of fibroblasts derived from single-cell sequencing data versus bulk sequencing available through the GTEx repository. Fibroblasts were chosen as they were the only purified cell type available in both GTEx and single-cell datasets. Results of the linear regression analysis are reported, showing a significant transcriptional similarity across single-cell versus bulk RNA sequencing. Vasc.: vascular; Lymph.: lymphatic; KC: keratinocyte; LSEC: liver sinusoidal endothelial cells; CT: crypt top; CPM: counts per million; TPM: transcript per million.

Extended Data Fig 4. GI single-cell RNA-Sequencing of allo-HCT patients.

a, Schema depicting the patients included in the allo-HCT dataset and the analytic pipeline for the single-cell RNA-Sequencing analysis. Briefly, single-cell RNA-Sequencing data was generated from the biopsies of 3 allo-HCT patients undergoing diagnostic colonoscopy for suspected GI GvHD at a median time from transplant of 90 days (range: 22 – 103). Upon standard processing and QC, viable cells were clustered and manually annotated (see Supplementary Figure 2). Immune cell clusters were excluded, and remaining resident non-immune cells were merged and harmonized with the healthy subject GI dataset used for the generation of the GvHD filter. CPM: counts per million. b, UMAP showing cluster annotations from the merged Seurat object containing both allo-HCT and healthy subject-derived GI cells (top) and Violin plots depicting the lineage-defining markers used for cluster annotation (bottom). c, UMAP depicting the clusters colored based on the dataset of origin. d, Venn diagram showing the number of genes that were present in the allo-HCT GI dataset vs. the GI healthy subject dataset. e, Venn diagram showing the overlap of the genes expressed in the allo-HCT GI dataset vs. the overall GvHD filter. f, Enrichment analysis of interferon-related signatures (Kim et al⁸⁷, MSigDB IFNa and IFNg⁸⁸) in the allo-HCT vs. GI healthy subject datasets (p < 0.0001, 2-tailed Mann Whitney test).

Extended Data Fig 5. GvL filter details.

a, Schematic depicting the generation of the 2 components of the GvL filter, i.e., AML and Hematopoietic filters. For the ‘AML filter’, a single-cell based classifier³⁹ was applied to bulk RNA-Seq data from the Beat AML cohort⁴⁰, to fully capture the AML transcriptional heterogeneity. For the ‘Heme filter’, bulk RNA-Seq from 18 purified mature hemopoietic cell types⁴³ as well as from hematopoietic stem and progenitor cells (HSPCs)^{44, 45} were the starting source. From the list of expressed genes (TPM >2), all those with expression in adult non-hematopoietic tissues per the GTEx RNA and protein repositories were excluded to define a set of 650 genes with preferential expression in AML and/or hematopoietic cells. b, Gender-specificity of the GvL filters: the GTEx filtering step was performed in a gender-specific fashion, as genes expressed in the male reproductive organs are not filtered out if the patient is female, and vice versa genes expressed in the female reproductive organs are maintained if the patient is male. c, Histograms depicting the chromosomal location of the 650 genes comprising the ‘AML’ and ‘heme’ filters; below each bar, relative chromosome size is depicted. For the X chromosome, only the pseudo-autosomal regions have been included in the analysis. d, Subcellular localization of the genes in the ‘AML’ and ‘heme’ filters. e, Biological functions of the genes included in the filters. Biological functions (from GO and Superpaths) have been manually clustered in macro-groups as specified in Supplementary Table 2.

Extended Data Fig 6. Y-encoded mHAg filter.

a, Schematic depicting the structure of the Y chromosome with a special focus on the genes in the male-specific region (MSY). Heatmap showing the expression pattern of the genes in the MSY across different healthy adult tissues: only the first 9 genes (RPS4Y1, DDX3Y, KDM5D, EIF1AY, ZFY, USP9Y, TMSB4Y, UTY and NLGN4Y) have evidence of expression (≥1 TPM) in ≥1 adult tissue site of GvHD. PAR: pseudo-autosomal region. b, Stacked histograms showing the number of predicted Y epitopes across individual HLA-A, -B, and -C alleles, and divided based on the MSY gene of origin. c, Bubble plot showing the median number of predicted epitopes for each MSY gene grouped based on the HLA peptide binding motif from Sarkizova et al.²⁸

Extended Data Fig 7. Antigenicity and immunogenicity of Y mHAgs.

a, Correlation matrix showing the peptide binding motifs of the HLA-A, -B, -C alleles from Sarkizova et al.²⁸; lateral panels display the individual HLA alleles belonging to each peptide binding motif, whose corresponding monoallelic B721.221 immunopeptidomes have been analyzed in Fig 2b. b, Hydrophobicity scores of the 410 Y mHAg peptides tested for immunogenicity, grouped by individual HLA restrictions: HLA-A0201 had the highest number of predicted binders with a score > 0 (Boxplots show min to max and median values; Kruskal-Wallis test with Dunn’s multiple comparisons test). c, Hydrophobicity scores of the 410 Y mHAgs grouped by HLA groups. Whiskers indicate min and max values, with all individual values shown (Kruskall-Wallis test with Dunn’s multiple comparisons test). d, Hydrophobicity scores of the predicted binders for each HLA allele, grouped based on the experimental evidence of T cell immunogenicity (per Fig 2f): only for HLA-A0101 and -C0501 was hydrophobicity significantly associated with immunogenicity (assessed with 2-tailed unpaired t test). Whiskers indicate min and max values, with all individual values shown. e, UMAP showing cluster annotations of single-cell thymic epithelial cells (TECs) from Park et al.⁶² (left), with feature plots of cluster-defining markers (middle); normalized expression of HLA-A, -B and -C genes per cluster (right).

Extended Data Fig 8. Tracking of Y mHAg-specific T cells ex vivo.

a, Flow cytometry plots showing the percentage of circulating CD8⁺ T cells specific for the indicated Y mHAgs at the listed time-points, including donor before allo-HCT in a patient transplanted from his HLA-identical sister and experiencing severe chronic GvHD. An irrelevant epitope from the EBV EBNA3A protein (HLA-B0702-restricted) was used as control, as both patient and donor were EBV seropositive. b, Timeline depicting the patient clinical course, highlighting the onset and course of the severe chronic GvHD, involving primarily skin and liver as shown by the liver function tests (ALT in red, total bilirubin in green). Triangles - peripheral blood samples used for Y mHAg-specific T cell tracking; diamonds - EBV reactivation. MMF: mycophenolate mofetil; tx: transplant. c, Quantification of ZFY-C0501-specific T cells in the leukapheresis (LK) products of additional 7 female donors to male patients (F to M), compared with T cells stained with a control C0501-dextramer. Boxplots show min to max and median values; significance assessed with 2-tailed Wilcoxon paired t test.

Extended Data Fig 9. Autosomal mHAgs and GvHD.

a, Normal distribution of the autosomal mHAg load in the DFCI-MRD cohort. b, Cumulative incidence of NIH moderate/severe chronic GvHD stratifying patients based on the overall autosomal mHAg load below (orange) or above (yellow) the median: no differences in 5-year cumulative incidences are observed: CI are 42% (95% confidence interval: 33% - 52%) and 39% (95% confidence interval: 30% - 49%) for < median and > median, respectively, 2-sided p = 0.8 (Gray’s test). c, Distribution of patients experiencing grade II-IV skin (left) and GI (right) acute GvHD across deciles of skin and GI mHAgs, respectively. d, Distribution of patients experiencing NIH moderate/severe organ-specific chronic GvHD across deciles of mHAgs expressed in the indicated GvHD target organs: from left to right – skin, GI, liver, eye, and oral. e, Heatmap depicting the co-occurrence of the 7 SNPs associated with liver acute GvHD in: from left to right, patients with liver acute GvHD, patients experiencing acute GvHD without liver involvement, and patients with chronic liver GvHD. f, Number of co-occurring driver liver mHAgs in the 3 patient groups outlined in e and defined with the same color code. Boxplots show min to max and median values (Kruskall-Wallis test with Dunn’s multiple comparison test). g, Promoter analysis of the genes harboring the SNPs associated with liver acute GvHD: 4 of 7 genes have interferon-responsive elements in their promoter region. Transcription factor binding site locations within 1500 base pairs (bp) upstream of the transcription start site (position 0) and the 5’ UTR are indicated.

Extended Data Fig 10. Population coverage simulating a T cell-based immunotherapy approach targeting GRFS mHAgs.

a, Heatmap showing the donor-recipient pairs (DRPs) that are informative for the pool of 54 GRFS epitopes indicated in the columns. DRPs (rows) are grouped by population of origin of the simulated recipient, as shown in the inner right bar. The outer right bar shows the number of predicted epitopes per DRP. Grey histograms on the bottom indicate the number of informative DRPs for each epitope. b, Population coverage analysis for: from top to bottom, overall simulation cohort, EUR, EAS, SAS, AFR, AMR. The histogram bars denote the percentage of DRPs that are informative for the indicated number of epitope hits, while the open circles indicate the cumulative percentage of population coverage for each number of epitope hits. The percentage indicated in the top-right corner of each graph shows the % of population for which ≥1 GRFS epitope could be potentially targeted. The red line denotes the 90% threshold of population coverage, which is considered optimal.

Data Availability

WES and RNA-Seq data from the training HCT, DFCI-MRD and HP-MRD cohorts is available through dbGaP portal (accession number phs003394.v1.p1). The original mass spectra, peptide spectrum matches, and the protein sequence databases used for searches have been deposited in the public proteomics repository MassIVE (https://massive.ucsd.edu) and are accessible at ftp://massive.ucsd.edu/v08/MSV000095025/. Original mass spectrometry data for the previously published B721.221 monoallelic immunopeptidomes are accessible at ftp://massive.ucsd.edu/MSV000080527. The GI GvHD single-cell RNA-Seq dataset is available from the corresponding author upon reasonable request. All external datasets used in this study (identified by reference number) are summarized in Supplementary Fig 18 and their accession numbers are the following: [29] GSE164403; [30] GSE124395; [31] GSE115469; [32] EGAS00001002649; [33] GSE123904; [34] GSE116222; [35] GSE125970; [36] GSE164241; [37] GSE164403; [39] GSE116256; [40] dbGAP study ID 30641, accession ID phs001657.v1.p1; [62] E-MTAB-8581 accessed online through https://developmentcellatlas.ncl.ac.uk; [43] www.proteinatlas.org/about/download; [44] GSE109093; [45] GSE113046. GTEx data was accessed from https://gtexportal.org/home/, IEDB from https://www.iedb.org/database_export_v3.php, and 1000 Genomes project from https://www.internationalgenome.org/data. Additional databases/datasets used include: interferome (www.interferome.org), MutSigDB (https://www.gsea-msigdb.org/gsea/msigdb/).