HLA-VBSeq v2: improved HLA calling accuracy with full-length Japanese class-I panel

HLA-VBSeq is an HLA calling tool developed to infer the most likely HLA types from high-throughput sequencing data. However, there is still room for improvement in specific genetic groups because of the diversity of HLA alleles in human populations. Here, we present HLA-VBSeq v2, a software application that makes use of a new Japanese HLA reference panel to enhance calling accuracy for Japanese HLA class-I genes. Our analysis showed significant improvements in calling accuracy in all HLA regions, with prediction accuracies achieving over 99.0, 97.8, and 99.8% in HLA-A, B and C, respectively.


Introduction
The human leukocyte antigen (HLA) system is located within the 6p21.3 region on chromosome 6 and encodes the major histocompatibility complex (MHC) proteins that are essential to the immune system. HLA genes are associated with a wide range of disorders, including cancer, organ transplants, and autoimmune and infectious diseases [1][2][3] , and the current version of the IPD-IMGT/ HLA Database, which indexes HLA sequences, contains more than 20,000 alleles of HLA subtypes, illustrating the extreme polymorphism of HLA alleles 4 . The accurate inference of HLA genotypes from whole-genome sequencing (WGS) data is therefore expected to significantly contribute not only to disease studies but also to fields such as pharmacogenomics. However, inferring HLA genotypes (HLA calling) remains challenging due to the substantial sequence similarity within the cluster and the exceptionally high variability of the loci.
In the past, we proposed a statistical method that infers the most probable HLA types from high-throughput sequencing data for each individual based on the optimal alignment of reads to the reference HLA sequences 5 and developed a computational tool called HLA-VBSeq 6 . While HLA-VBSeq uses the reference sequences of the IPD-IMGT/HLA Database for the alignment of read sequences, it is still possible to improve the prediction accuracy in specific populations since the allele distribution and haplotype frequencies of HLA genes largely depend on geographic location and genetic group 7,8 . In this study, we present an updated version, HLA-VBSeq v2, which adds a Japanese reference panel of class-I HLA genes (the ToMMo HLA panel 9 ) to the IPD-IMGT/HLA Database to form a new reference panel.
To evaluate the performance of HLA-VBSeq v2 with the new reference panel, we compared the accuracy of HLA-VBSeq and HLA-VBSeq v2 by using WGS data in the Japanese population. Additionally, we compared the performance of HLA-VBSeq v2 with that of another HLA calling tool, HLA*PRG:LA, which calls for the alleles at G group resolution (a group of HLA alleles that have identical nucleotide sequences across the exons encoding the peptide-binding domains) 10 .
We also provide an exhaustive list of the ambiguous samples observed in our validation datasets. Ambiguous HLA alleles occurred because of the typing strategies. The PCR-SSOP-Luminex method only uses exons 2 and 3 for HLA class-I analysis (the HLA class-I region contains 8 exons), and this may result in more than one possible typing result within a given region. By using the WGS data, for which the HLA region is fully phased, HLA-VBSeq v2 displayed high performance in ambiguous samples; we therefore suggest that HLA-VBSeq v2 can be used to validate the typing results obtained with the PCR-SSOP-Luminex method.
The ToMMo HLA panel is a full-length Japanese reference panel of class-I HLA genes constructed from 208 samples at the Tohoku Medical Megabank Organization 9 . The ToMMo HLA panel consists of 139 alleles that were extended from knowns in the IPD-IMGT/HLA Database and includes 40 novel alleles compared with the closest subtypes in the IPD-IMGT/HLA Database. We added the ToMMo HLA panel to the IPD-IMGT3.31.0 reference sequences to create a new set of sequences to be used as a custom reference panel.
We used two independent datasets to evaluate and validate the prediction accuracy of HLA-VBSeq v2: the Tokyo Healthy Controls (THC) dataset and the Stevens-Johnson syndrome (SJS) dataset, using WGS data in both cases, i.e. with the HLA region fully phased. The THC dataset consists of 418 healthy Japanese individuals from the Tokyo area, and the SJS dataset consists of 117 Japanese individuals with cold medicine-related Stevens-Johnson syndrome with severe ocular complications.
To evaluate the performance of each calling tool, we calculated the prediction accuracy, defined as the percentage of inferences that are correct relative to the 'true' HLA types (the expected HLA types). In this study, we defined the 'true' HLA types by using the PCR-SSOP Luminex method and next-generation sequencing (NGS)based HLA typing, which we implemented in both datasets. NGS-based HLA typing was performed with a NXType™ Class-I NGS HLA typing kit and the AllType™ NGS 11-loci Amplification Kit (Thermo Fisher Scientific, Waltham, MA, USA), which phased the full genes. The PCR-SSOP-Luminex method is a high-resolution genotyping method that combines polymerase chain reaction (PCR) and sequence-specific oligonucleotide probe (SSOP) protocols with the Luminex 100 xMAP technology at the HLA-A, B, and C loci, amplified by PCR on only exon 2 and exon 3 11 .
In this study, if PCR-SSOP-Luminex and NGS-based HLA typing gave the same result, it was considered to be the 'true' HLA types. By contrast, when samples displayed inconsistent typing results ('ambiguous' typing results), they were reexamined by Sanger sequencing-based typing (SBT). Since in all cases SBT gave the same result as NGS-based HLA typing, NGS-based HLA typing was considered to be the most reliable method to determine the 'true' HLA types.
In our analysis, we assessed the prediction accuracy of Japanese HLA calling from WGS data at the 2-field resolution, which specifies the amino acid sequence of the encoded protein. We evaluated two HLA alleles (either heterozygous or homozygous) in three HLA regions (HLA-A, B and C) or six HLA alleles in total for all individuals. Furthermore, we compared the performance of HLA-VBSeq v2 with that of HLA*PRG:LA, another HLA calling software application that infers alleles at G group resolution (the group of HLA alleles that have identical nucleotide sequences across the exon encoding the peptide-binding domains and are denoted by G at the 3-field resolution., e.g., A*01:01:01 G) 10 . HLA*PRG:LA reports a group of alleles for each gene at the 2-field resolution, but here we only used the first allele as the calling result.
The results obtained using HLA-VBSeq v2 show significant improvements in prediction accuracy in all HLA regions after including the ToMMo HLA panel in both the SJS and THC datasets (Table 1). Regarding the performance of HLA-VBSeq v2 compared to HLA*PRG:LA, though both programs achieved prediction accuracies of over 97.8%, HLA-VBSeq v2 displayed slightly better prediction accuracies for the HLA-A and HLA-C regions (99.0 and 99.8% vs. 98.1 and 98.4%, respectively) but lower performance for HLA-B genes (97.8 vs. 99.6%). HLA-VBSeq v2 displayed slightly better prediction accuracies for the HLA-A and -C regions in the THC and SJS datasets but lower performance for HLA-B genes. Moreover, in the SJS dataset, HLA-VBSeq v2 exhibited 100% prediction accuracy in the HLA-C region (Table 1).
This study shows that including the ToMMo HLA panel substantially improves the calling accuracy of HLA class-I genes from WGS data in the Japanese population. In Table 2, we summarized the inference results obtained with HLA-VBSeq and HLA-VBSeq v2 that displayed inconsistencies with the 'true' genotype. We observed that the performance improved significantly with HLA-VBSeq v2 in some alleles, such as B*40:06, B*54:01, B*55:02 and C*14:03. For instance, HLA-VBSeq failed to correctly call C*14:03 (it was identified as C*14:02 in a high number of cases) in both datasets. The reference panel of HLA-VBSeq consists of reference sequences of C*14:02 and C*14:03 but not full genomic sequences (including exons, full introns and regulatory regions). By contrast, HLA-VBSeq v2 uses the ToMMo HLA panel, which contains the full-length reference sequences of C*14:02 and C*14:03 for the Japanese population, and therefore identified C*14:03 correctly. This indicates that full genomic sequences are highly informative in HLA calling and that using HLA-VBSeq v2 leads to more accurate calling results in the Japanese population.  Ambiguous allele combinations occurred in all HLA loci; a possible reason for this may lie in the presence of heterozygous sequences, i.e., the presence of more than one possible pair of alleles within the region analyzed or the existence of a polymorphism outside of the region analyzed 12 . Because the PCR-SSOP-Luminex method only covers exons 2 and 3 for HLA class-I analysis (there are 8 exons in HLA-A, B and C regions) 11 , we also observed some ambiguous samples in our datasets (Table 3). Table 3 shows that HLA-VBSeq v2 provided outstanding performance in ambiguities in the Japanese population. For instance, since C*04:01 and C*04:82 have identical nucleotide sequences in exons 2 and 3, it is impossible to distinguish them only by analyzing exon-2 and -3 sequences, leading to an ambiguous result. Because the ToMMo HLA panel includes the full-length reference sequences of C*04:82, HLA-VBSeq v2 distinguished C*04:01 from C*04:82 and made correct calls. Table 3 also shows that in most instances, HLA*PRG: LA and Luminex inferred the same allele type when dealing with ambiguous cases. Similarly, NGS-based HLA calling and HLA-VBSeq v2 reached the same result in a high number of cases. The reason why HLA*PRG:LA and Luminex tend to provide the same calling results likely lies in the design of HLA*PRG:LA to call for the alleles at G group resolution 10 , which means that the calling methods focus on exons 2 and 3. By contrast, HLA-VBSeq v2 is designed for 8-digit resolution and includes the fulllength HLA sequences for the analysis, making it more likely that the results will coincide with those from NGSbased HLA typing ("true" HLA types). Because of these differences, the performance of HLA calling in ambiguous cases varies greatly depending on the software application used. Given the high performance of HLA-VBSeq v2 in ambiguities, we suggest reexamining PCR-SSOP-Luminex typing results with HLA-VBSeq v2 to obtain more reliable results regarding the correct HLA type in the Japanese population.
In conclusion, the addition of a reference panel for the Japanese population was highly effective for improving the calling accuracy of HLA class-I genes from WGS data in the Japanese population. This solution is promising for These typing results were used as the "true" types the inference of HLA class-II genes and other highly diverse regions, e.g., killer-cell immunoglobulin-like receptors (KIRs). Additionally, this study suggests that the ToMMo HLA panel is a valuable resource whose use is not limited to HLA-VBSeq v2 but could be expanded to other HLA calling tools.