Adverse drug reactions (ADRs) pose a significant burden to the healthcare system, comprising a significant portion of hospital admissions, mortality, and overall healthcare costs (Baker et al., 2004; Pirmohamed et al., 2004). Globally, the cost of drug-related morbidity and mortality in 2000 was estimated at $177.4 billion, a figure that nearly doubled the previous 1995 estimate (Ernst and Grizzle, 2001). In addition to the strictly economical cost, the distribution of ADRs across the population highlights disparities within the healthcare system. In a United States study, it was found that older individuals, men, Black persons, and individuals residing in extremely rural areas (counties with less than 10,000 residents) experienced higher mortality associated with ADRs (Shepherd et al., 2012). This indicates the need for action to prevent ADRs to ensure that subsets of our population are not facing disproportionate challenges in their access to and quality of care.

Pharmacogenomics provides the opportunity to prevent ADRs. HLA-A and HLA-B are genes encoding human leukocyte antigens (HLAs), part of the class I major histocompatibility complex (MHC) (Wieczorek et al., 2017). This complex is a group of cell surface proteins involved in the presentation of antigenic peptides to T cells to induce an immune response (Wieczorek et al., 2017). The region that encodes the HLA complex consists of ∼260 genes in a ∼4-Mb span on chromosomal region 6p21.3, base pair positions 29,640,000–33,120,000 from the Genome Reference Consortium Human Build 37, hg1 (Trowsdale and Knight, 2013). The most polymorphic HLA class I and class II proteins are each expressed from three gene regions (MHC class I: HLA-A, -B, -C; MHC class II: (HLA-DR, -DP, -DQ), which are all highly polymorphic (Trowsdale and Knight, 2013). The HLA complex is the most polymorphic human genetic system, with over 8000 functional class I variants (Choo, 2007; Illing et al., 2017). Variations in the HLA genes play an important role in determining susceptibility to infection and to autoimmune disease, and can contribute to the recognition of drugs and their metabolites as foreign antigens by T-cell receptors with a consequent immunological reaction (Fan et al., 2017).

Specifically, the HLA haplotypes HLA-B*15:02 and HLA-A*31:01 are predictive of severe cutaneous reactions to the anticonvulsant drug carbamazepine, also prescribed as a mood stabilizer in psychiatry (Fan et al., 2017; Gui et al., 2018; Carvalho Henriques et al., 2020c). Cutaneous ADRs are one of the most common types of ADRs, typically occur in the first two to three months of drug use, and can range from a mild skin rash to a life-threatening reaction such as Stevens-Johnson syndrome and the more severe form of this, known as toxic epidermal necrolysis (SJS/TEN). SJS/TEN is characterized by rapid blistering, cutaneous and muscle detachment, and has a mortality rate of 10–30% (Fan et al., 2017; Mullan et al., 2019). Performing testing for these haplotypes can therefore be potentially life saving, with a Hong Kong study finding that testing for HLA-B*15:02 reduced the incidence of SJS/TEN induced by carbamazepine from 0.24 to 0% (Chen, Liew, and Kwan, 2014). The HLA-A*31:01 haplotype has a frequency of 2–5% in Northern European populations, with its presence increasing the risk of carbamazepine-induced SJS/TEN from 5 to 26% (Mccormack et al., 2011). As a result, before prescribing carbamazepine, it is advised that a genetic test be performed to ensure the patient does not carry either of these risk haplotypes (Amstutz et al., 2014).

Although genetic testing can greatly improve patient care and reduce the potential for ADRs, the uptake of screening has been limited because of the technical and economic barriers associated with conventional HLA haplotyping methods. Due to the highly polymorphic nature of the complex, the “gold-standard” for HLA haplotyping is sequencing-based methods and other specialized techniques, such as polymerase chain reaction-based sequence-specific priming and microbead hybridization (Dunckley, 2012; Lazaro et al., 2013; Edgerly and Weimer, 2018). However, these strategies are time consuming and expensive; the mean cost for testing in North America is $363.65 USD (Verbelen, Weale, and Lewis, 2017). This may limit the availability of such testing in patient groups in which it should be offered. Current United States Food and Drug Administration (FDA) recommendations for carbamazepine state that “prior to initiating carbamazepine therapy, testing for HLA-B*1502 should be performed in patients with ancestry in populations in which HLA-B*1502 may be present” and that “HLA-B*1502 is found almost exclusively in patients with ancestry across broad areas of Asia” (US National Library of Medicine, 2021). As a result, in many jurisdictions, HLA-B*15:02 testing for hypersensitivity to carbamazepine is offered only to Asians (Fang et al., 2019; Alberta Precision Laboratories, 2021; US National Library of Medicine, 2021).

SNV based methods of genotyping permit high throughput economical typing (Erlichster, Goudey, Skafidas, and Kwan, 2019) and therefore using such to “tag” the HLA haplotypes offers a potential route of making the identification of the at risk haplotypes more widely available (Petersdorf and O’hUigin, 2019). While previous studies had identified potential haplotype tagging SNVs for HLA-B*15:02 and HLA-A*31:01, there is no consensus regarding which SNVs can be used across ancestral groups. In fact, some of the literature is contradictory. For example rs2844682 and rs3909184 have been previously suggested as tag SNVs for HLA-B*15:02 (de Bakker et al., 2006; He et al., 2015), while another study concluded that these SNVs had no clinical value for identifying HLA-B*15:02 carriers (Zhu et al., 2015). In addition, many studies have restricted their populations to individuals from a particular ethnic group (de Bakker et al., 2006; Liu et al., 2015; Maekawa et al., 2015; Zhou et al., 2016; Gui et al., 2018). Further, methodologies vary greatly between studies, with some estimating validity based on databases (Erlichster et al., 2019) and others using genotyping (He et al., 2015; Fang et al., 2019).

A haplotype-tagging SNV method of screening US patients for HLA-B*15:02 without ethnicity-based preselection was able identify more than twice the number of carriers at risk of a carbamazepine-related ADRs than screening patients of Asian ancestry alone (Fang et al., 2019). This supports a prior suggestion that the FDA recommendations be revised to capture the broader range of patients who may be at risk for ADRs in response to carbamazepine (Payne, 2014). The aim of this study was therefore to generate validated SNV assays for the identification of HLA-B*15:02 and HLA-A*31:01 across a range of ancestral groups.

Our data are consistent with rs10484555 and rs17179220 being the best tagging SNVs for HLA-B*15:02 and HLA-A*31:01 respectively, across diverse ancestral backgrounds.

Assay ANPRV67 for rs10484555 had 100% sensitivity and specificity for HLA-B*15:02 (excluding one sample, 755, that failed to amplify on multiple attempts). We suggest further investigation, such as sequencing, of sample 755 and one sample that required manual calling (NA17019) is warranted. Our data are consistent with the database prediction of Erlichster et al. (2019) of 100% sensitivity and 98.7% specificity of rs10484555 for tagging HLA-B*15:02 in a diverse population, and with the 100% sensitivity and 99.3% specificity for tagging HLA-B*15:02 in the Han Chinese population reported by Liu et al. (2015).

Assay AN33GCK for rs144012689 with the nucleotide base G at rs2596496, in combination with manual calling, resulted in 100% sensitivity and 100% specificity for the identification of HLA-B*15:02. However, if another nucleotide other than C or G is present at rs2596496, which is theoretically possible (National Center for Biotechnology Information, 2020), the assay may not work. Even though there are no frequency data available for the other possible nucleotide variants, we suggest further custom assays for the other nucleotide possibilities at rs2596496 be developed to rule out this possibility. The presence of the C or G nucleotide within our sample set is not population-specific, as the aforementioned group of samples we hypothesize to have the C nucleotide at rs2596496 are from European, Han Chinese, and Hispanic populations. As well, there is the potential for false positives in individuals with the HLA-B*15:13 haplotype. This haplotype has a global frequency of 0.08%, in comparison to a frequency of 0.91% for HLA-B*15:02 (Fang et al., 2019), which is small but not negligible. Overall, for an HLA-B*15:02 haplotyping strategy, we suggest using assay ANPRV67 for rs10484555 first, and if any output such as that seen with samples 755 and NA17019 results, assay AN33GCK for rs144012689 may be additionally used.

Assay C__33415939_10 for rs17179220 had 100% sensitivity and specificity for HLA-A*31:01. This is comparable to the database prediction of Erlichster et al. (2019) of 100% sensitivity and 99.6% specificity of rs17179220 for tagging HLA-A*31:01. The assay performs best when the number of PCR amplification cycles is increased to 50 (increasing total run time from 90 to 110 min). Assay ANNK49J for rs1061235 also identified HLA-A*31:01 with 100% sensitivity, with its specificity impacted by this SNV also being present in the HLA-A*33:01 and HLA-A*33:03 haplotypes (Thorstensen et al., 2014). A study in an ancestrally-diverse group of Canadian children (Amstutz et al., 2013) found that out of 20 children carrying the minor allele of rs1061235, 12 had a haplotype of HLA-A*31:01, five had a haplotype of HLA-A*33:03, and three had a haplotype of HLA-A*33:01. As a result, they concluded that rs1061235 is not an optimal tag for HLA-A*31:01 in ancestrally diverse populations. However, they did find that the rs1061235 minor allele was overrepresented among carbamazepine hypersensitivity cases in patients not carrying the HLA-A*31:01 haplotype, suggesting a possible association of the HLA-A*33 haplotypes with such hypersensitivity. Further research regarding this is required.

The selection of positive controls, especially for minor alleles that are less represented within a sample set, has the potential to improve clustering and the ability of the QuantStudio software to make accurate automatic calls. For this purpose, we recommend NA23093 as a homozygous HLA-B*15:02 control and HG00463 or NA23090 as a heterozygous HLA-B*15:02 control. Numerous Coriell samples would be suitable for heterozygous HLA-A*31:01 controls (NA17039, NA17229, NA17235, NA17277, NA17281). A limitation of our study is that owing to the limited availability of reference samples with relevant haplotypes, our analysis was limited to only 38 samples, primarily of European ancestry, with Hispanic, Han Chinese, Filipino, and Black American representation. It has been previously noted that population sets that include larger numbers of those of different ancestry may see higher rates of false positives (Liu et al., 2015). Nonetheless, as the ADRs associated with these haplotypes on prescribing carbamazepine or oxcarbazepine may be fatal, individuals may not know the full details of their ancestry, and certainly testing HLA-B*15:02 is indicated in individuals had physician-reported ethnicity other than Asian (Fang et al., 2019) we recommend using these SNVs to identify these haplotypes in individuals of any ancestral group. Another potential limitation in the translation of this work to the clinical laboratory setting is the need for occasional manual calls. This is in fact not unusual with this and other techniques in this setting: data would need to be reviewed before exporting files or reporting to patients. This process is not particularly lengthy or challenging but does require some knowledge of the TaqMan assays and use of the QuantStudio software. Different population background and sample quality may be relevant to the amount of data that requires manual manipulation. As mentioned, the proximity of nearby SNVs to the SNVs of interest can interfere with assay performance, and certain SNVs may be more prevalent in certain ancestral groups. Poor sample quality or low concentration may also increase the prevalence of data that requires manual calls.

In summary, we have validated a method of detection of HLA-B*15:02 and HLA-A*31:01 that at <$1USD per sample plus labour of less than 2 h is orders of magnitude lower than the cost of haplotyping by the current “gold-standard” methods. A recent study that aimed to assess the cost-effectiveness of pre-emptive pharmacogenomic testing for HLA haplotypes, based on their relative frequencies worldwide, concluded that for carbamazepine, pre-emptive genotyping of HLA-B*15:02 would be cost-effective across most of East and South Asia, whereas HLA-A*31:01 testing would be likely to be cost-effective globally (Zhou et al., 2021). However, their calculations assumed the cost per sample to be $40, which is much greater than our estimated cost of TaqMan SNP genotyping.

Our findings are relevant for other technologies that rely on SNV detection for the identification of HLA haplotypes. For example, the Ion AmpliSeq Pharmacogenomics Panel uses rs1061235 to tag HLA-A*31:01. Although we and other investigators have found that the specificity of this SNV as a haplotype tagging marker is not 100%, as carbamazepine hypersensitivity has been found in patients with the rs1061235 minor allele but not carrying the HLA-A*31:01 haplotype (Amstutz et al., 2013), the use of this SNV to screen for carbamazepine hypersensitivity should be adequate. Further, the sensitivity is 100% in our data and that of others, and given the severity of the potential ADRs, the absence of false negatives is more important than the absence of false positives. Agena Biosciences use markers including rs10484555 for HLA-B*15:02 and rs1061235 for HLA-A*31:01, which should be fine. Pharmacoscan includes rs1061235 as a screen for HLA-A*31:01 as well. Cross-validation of our method could be undertaken versus the multiple technologies above described in a manner that we have previously described (Carvalho Henriques, et al., 2020a; 2020c).