Modeling coverage gaps in haplotype frequencies via Bayesian inference to improve stem cell donor selection

Abstract

Regardless of sampling depth, accurate genotype imputation is limited in regions of high polymorphism which often have a heavy-tailed haplotype frequency distribution. Many rare haplotypes are thus unobserved. Statistical methods to improve imputation by extending reference haplotype distributions using linkage disequilibrium patterns that relate allele and haplotype frequencies have not yet been explored. In the field of unrelated stem cell transplantation, imputation of highly polymorphic human leukocyte antigen (HLA) genes has an important application in identifying the best-matched stem cell donor when searching large registries totaling over 28,000,000 donors worldwide. Despite these large registry sizes, a significant proportion of searched patients present novel HLA haplotypes. Supporting this observation, HLA population genetic models have indicated that many extant HLA haplotypes remain unobserved. The absent haplotypes are a significant cause of error in haplotype matching. We have applied a Bayesian inference methodology for extending haplotype frequency distributions, using a model where new haplotypes are created by recombination of observed alleles. Applications of this joint probability model offer significant improvement in frequency distribution estimates over the best existing alternative methods, as we illustrate using five-locus HLA frequency data from the National Marrow Donor Program registry. Transplant matching algorithms and disease association studies involving phasing and imputation of rare variants may benefit from this statistical inference framework.

High-resolution HLA phased haplotype frequencies to predict the success of unrelated donor searches and clinical outcome following hematopoietic stem cell transplantation

“…Interestingly, the comparative analysis undertaken in this study showed that only a mere 39% assignation concordance was achieved between the real phased haplotypes in our cohort and a "blind" estimation with the EM algorithm. The discrepancies were mostly due to rare haplotypes and this problem has recently been discussed [20], but this also concerned the frequent haplotype A*02:01~B*08:01~DRB1*03:01. This illustrates the usefulness of family data for characterizing high-resolution multi-locus haplotypes [26] when sample sizes are not huge (meaning hundred thousand, or even millions of individuals).…”

“…Some pairs like B~C and DRB1~DQB1 are found in tight association on chromosome 6 [17], whereas other loci are defined by weaker (HLA-A) or non-significant linkage (HLA-DPB1), due to more distant location or recombination hotspots [18,19]. These characteristics represent a significant hindrance to determine HLA multi-locus haplotypes and their corresponding frequencies [20]. In consequence, powerful methodologies have been developed to assess haplotype frequencies in various populations [21,22] and in large cohorts of unrelated donors [23], but such approaches need to rely on representative sample sizes and on assumptions that are not always met in practice [22].…”