HLA*IMP--an integrated framework for imputing classical HLA alleles from SNP genotypes

Abstract

Motivation: Genetic variation at classical HLA alleles influences many phenotypes, including susceptibility to autoimmune disease, resistance to pathogens and the risk of adverse drug reactions. However, classical HLA typing methods are often prohibitively expensive for large-scale studies. We previously described a method for imputing classical alleles from linked SNP genotype data. Here, we present a modification of the original algorithm implemented in a freely available software suite that combines local data preparation and QC with probabilistic imputation through a remote server.

Results: We introduce two modifications to the original algorithm. First, we present a novel SNP selection function that leads to pronounced increases (up by 40% in some scenarios) in call rate. Second, we develop a parallelized model building algorithm that allows us to process a reference set of over 2500 individuals. In a validation experiment, we show that our framework produces highly accurate HLA type imputations at class I and class II loci for independent datasets: at call rates of 95-99%, imputation accuracy is between 92% and 98% at the four-digit level and over 97% at the two-digit level. We demonstrate utility of the method through analysis of a genome-wide association study for psoriasis where there is a known classical HLA risk allele (HLA-C*06:02). We show that the imputed allele shows stronger association with disease than any single SNP within the region. The imputation framework, HLA*IMP, provides a powerful tool for dissecting the architecture of genetic risk within the HLA.

Availability: HLA*IMP, implemented in C++ and Perl, is available from http://oxfordhla.well.ox.ac.uk and is free for academic use.

Fig. 1.

Visualization of the L&S Hidden Markov Model (HMM) states for a group of reference chromosomes carrying the A allele. Usually, the computation of an emission probability for a given chromosome c would involve filling the corresponding forward table (Rabiner, 1989) from s₁ to s_n and summing over the entries in s_n. However, the emission probability can also be calculated at any point s in the HMM, by combining the forward- and backward-tables up to s. In our parallelization approach, we compute both tables for each chromosome in advance (gray cells in the figure, polymorphisms highlighted in dark gray) and add the specific transition probabilities for any given SNP s (middle column), which can be performed in parallel without changing the precomputed table values.

Fig. 2.

The front end of HLA*IMP controls for missing data, aligns complementary SNPs and phases haplotypes in a largely automated manner. In this screen shot: graphical output from the alignment procedure, comparing SNP allele frequencies in the user dataset to HapMap allele frequencies, before (left) and after (middle) alignment. Complementary SNPs are aligned using an expectation-maximization (EM)-based procedure. A straight line of data points (right) indicates that there are no gross deviations between EM estimated and HapMap frequencies.