Functional Class I, Class II, and nonclassical HLA variation drives lymphoma and myeloma risk in the all of us research program

Background: HLA-mediated immune surveillance involves classical, non-classical, and class I-like pathways. Because lymphoma and myeloma rely on different immune mechanisms, we evaluated HLA class I/II, non-classical, and class I-like loci across ancestries using allele-level amino acids(AA), heterozygosity (HET), evolutionary divergence (HED), IEDB-derived peptide-binding breadth (PBW; 9-mer binding entropy), and positional amino-acid heterozygosity (residue mismatch). Methods: Hodgkin lymphoma (HL), NHL and subtypes, and multiple myeloma (MM) in All of Us Cohort were identified using ICD/SNOMED; controls lacked hematologic malignancy. HIBAG-imputed allele with n>20 cases were analyzed in EUR, AFR, and AMR. Models adjusted for age, sex, PCs, and admixture proportions tested ORs, 95% CIs, and FDR<0.1. We used joint PRS-HLA models to test whether HLA metrics contributed risk information beyond the lymphoma PRS. Results: We replicated EUR HET and/or HED class I/II associations with NHL (N=1668), DLBCL (N=490), HL (N=393), and FL (N=435), and extended these patterns using PBW and AA-level metrics. In NHL, we confirmed that HET-C reduced risk (p=0.027) and found greater PBW was protective (OR=0.49; p=0.09). Similarly, in CLL, we replicated the HET-A association (p=0.027), identified PBW-A effect (OR=0.25; p=.03), and a positional signal at A-163 (OR=.83, p=0.015) that also appeared in NHL(p=.015). Neither locus showed HED effects.Novel MM associations with HET-B and -C (OR≈0.74; p<0.03), were supported by a positional hit at B-42 (OR=0.77; P=0.033) and a PBW-C effect (OR≈0.24; p=0.01), while HED remained null. In contrast, DLBCL showed protection across all class I metrics as well as at MICB (OR=0.75; p=0.006) and MHC-like HET (OR=0.73; p=0.084). In EUR DLBCL, A*01:01 (OR=1.23; p=0.076) and A26:01 (OR=1.56; p=0.051) showed risk effects consistent with some prior reports reinforced by multiple A-locus positional signals (P<0.02).Novel AFR-specific findings included A*68:02 in: HL (OR=2.88; p=1.8×10⁻⁵), mirrored by a positional hit at A-30 (OR≈3.0, p=.0004), NHL at B*07:02 (OR=1.69; p=0.018) and for C1-motif alleles C07:01 (OR=1.51, p=.04) and C*07:02 (OR= 2.02, p=.00017) consistent with a KIR-C1 mechanism.The DLBCL PRS replicated (which includes an HLA-B variant) in EUR (OR=1.25, 95% CI 1.15-1.36; p=7.3×10⁻⁷) and joint models showed HED (p=0.0028) and HET (OR=0.86, 95% CI 0.80-0.93, p=7.0×10⁻ 5 ) remained significant when combined with PRS. AIC identified HET + PRS with covariates as the best model. The PRS did not replicate in AFR or AMR. Conclusion: We replicated classical HLA associations and identified novel non-classical and functional HLA features that shape B-cell malignancy risk across ancestries. Furthermore, adding orthogonal HLA metrics provided substantial, independent information beyond the PRS.