Machine learning with immunogenetics for the prediction of hematopoietic cell transplant outcomes

ABSTRACT Allogeneic hematopoietic cell transplantation (HCT) is the only curative treatment for most forms of acute myelogenous leukemia (AML), but its 50% failure rate remains unacceptably high, with the principal causes of death due to disease relapse and graft-versus-host disease.

When successful, HCT prevents leukemic relapse due to a graft versus leukemia effect, co-mediated by T cell and natural killer (NK) cell immune functions.

Selection of donors whose allografts will provide higher NK anti-leukemic response potential but low GVHD risk remains a major unmet need in HCT.

The polygenic, polymorphic KIR receptors, in combination with their HLA ligands, control NK function, dictating NK repertoire content and establishing thresholds for NK cell response in a process called ?NK education?.

Large retrospective studies in HCT have demonstrated that specific KIR-HLA allele combinations associated with NK education are predictive for relapse control, but they represent only a fraction of known KIR-HLA interactions.

Furthermore, out of the thousands of phenotypes present in the NK repertoire, the NK population(s) responsible for leukemia control in HCT is unknown and they likely differ between transplant pairs.

Aim 1 proposes a machine learning approach to integrate NK genotype, phenotype, and function to identify how genotype determines overall repertoire response and which subpopulations contribute most to global response.

Parallel statistical modeling of NK genotypes and HCT outcome in a cohort of 2800 AML patient may confirm the same genotypes that are potent for global response also play a role in HCT outcomes but may also identify unexpected ones. HLA is the most important determinant of GVHD risk.

Precise HLA matching lowers the risk for GVHD, but for patients who lack HLA-compatible donors, predicting permissible HLA mismatches is a paramount and unmet need.

Two lineages of HLA-B allotypes exist based on the M and T leader peptide dimorphism, and GVHD risk in HLA-mismatched HCT differs depending on the match status of the leader.

The division of the HLA-B locus into two lineages provides a novel approach for mapping functional motifs in transplantation that removes reduces the sheer numbers of polymorphic positions that previously precluded examination of more than 1 residue at a time.

Machine learning approaches using HLA data from more than 11,000 transplant patients will permit assessment of the full spectrum of lineage variation and the relationship between T-cell and NK alloresponses.