The phylogeny of 48 alleles, experimentally verified at 21 kb, and its application to clinical allele detection

Exact matching for alleles improved survival following bone marrow transplantation [1] and reduced alloimmunization in chronically transfused patients [2‐4]. Using computational algorithms, the large genotype datasets from next generation sequencing (NGS) can be phased into alleles or haplotypes [5, 6]. Using family relationships or applying experimentally confirmed allele information improves the inference accuracy, as routinely demonstrated in clinical HLA typing [7]. Blood group genes are less polymorphic than the highly variable, often shorter, HLA genes. Out of the 36 blood group systems and the genes encoding them, experimentally confirmed alleles are known for short genes only, such as ICAM4 [8] and ACKR1 [9]. For longer genes, such as ABO and ERMAP of more than 20 kb, and linked genes, such as RHD and RHCE, most haplotypes had only been computationally predicted [10‐12].

The ERMAP gene, located on chromosome 1, encodes the glycoprotein carrying the antigens of the Scianna blood group system (SC; ISBT 013) in humans [13‐15]. The single-pass transmembrane glycoprotein is likely involved in cell adhesion and recognized by immune cells [13, 16, 17]. The gene belongs to the butyrophilin (BTN) family which is a type 1 membrane protein of the immunoglobulin (Ig) superfamily [18]. The butyrophilin and butyrophilin-like proteins have recently been studied as potentially important immune regulators [19, 20].

We have previously assessed the nucleotide variations in the ERMAP gene and unambiguously identified 48 alleles at 21,406 nucleotides each in 50 unrelated individuals from 5 different populations [21]. We propose using the phylogeny of this set of 48 alleles and identifying evolutionary steps to derive the observed alleles [22]. We predicted unobserved alleles at every internal node and their posterior probabilities. These inferred alleles, represented by sequences identified in the nodes, are possible candidates for alleles segregating in the population. Our new approach proposes a method of utilizing not-yet-observed alleles, predicted by phylogeny, for phasing patient genotypes in clinical diagnosis and therapy.

The sequence information for 48 ERMAP alleles was retrieved from GenBank (KX265189–KX265236) [21]. The phylogenetic tree was rooted using the chimpanzee ERMAP sequence as outgroup (GenBank number NC_006468.4; range 42,268,258 to 42,295,767). Full-length sequences were aligned using the MAFFT version 7 program [23]. All of the 72 variable sites were extracted manually from the 48 ERMAP alleles [21]. The Bayesian coalescent algorithm implemented in BEAST v1.8.3 [24] was used to estimate a coalescent phylogeny for these variants and the allelic ancestral states at the internal nodes of the phylogeny. All analysis was done using default parameters. Internal node is a theoretical representation of a common ancestor between sampled alleles and are often extant in population level studies [25]. If more than one mutational or recombinational step is required to join some nodes, predicted alleles are incorporated to complete the tree [26].

We executed 4 independent runs of the program, each using the Tamura-Nei substitution model [27], a lognormal relaxed clock model [28], and a constant-size coalescent model [29]. After 40 million generations the parameter estimates were examined and determined to have converged for each run. The allelic ancestral states at each node and their posterior probabilities were extracted manually from the maximum clade compatibility tree estimated from 9001 Markov chain Monte Carlo samples generated by the BEAST software. For the ancestral allele reconstructions, we generated a set of all possible ancestors for each node and selected the predicted allele with the highest posterior probability.

A Bayesian phylogeny of 48 previously published ERMAP alleles was calculated (Fig. 1). Based on this phylogenetic tree, we predicted alleles, many of which may be extant in the population, particularly those of greater posterior probability. Our approach applied standard methods of phylogenetic inference, ancestral character reconstruction and aimed to enrich the repertoire for a focal genomic region, of specific clinical interest.

A phylogenetic analysis was applied to a set of 48 physically confirmed ERMAP alleles covering 5 populations worldwide [21]. We predicted 4243 unobserved alleles and their distinct posterior probabilities. The relatively small number of predicted alleles contrasted to the vastly larger number of theoretically possible alleles. The predicted alleles have a stronger support for being correct and extant in the population because they are more likely ancestral to the observed alleles. We propose the concept of detecting unobserved, likely novel, alleles based on the phylogeny of verified alleles.

Previous computationally driven algorithms to phase NGS data such as read-backed phasing [30] and haplotype improver [31] remain very useful for phasing haplotypes and alleles in a population sample but may fail when applied to a single observation in an individual patient. Our approach utilizes predicted alleles and their posterior probabilities along with the verified alleles as templates for phasing the genotypes detected in high-throughput sequencing (Fig. 2), complementing the computationally driven algorithms. This approach increases the effective number of templates available for phasing and thus the accuracy of phased haplotypes and alleles. When a previously unobserved allele matches one of the predicted alleles, its posterior probability allows to quantify the reliability of the estimate for clinical decisions, such as in transfusion and transplantation settings, in a patient who bears a new allele. While the validation of the predicted alleles by applying our protocol was not performed in this study, the novel approach illustrates the potential use of phylogenetic data in a clinical diagnostic setting.

Our approach relies heavily on the accuracy of the alleles used for the phylogenetic analysis. Hence, reference sequences from online databases such as GenBank should be avoided as long as the information is not sufficiently replicated or independently verified [32]. The prevalence of the 48 alleles derived from 5 populations worldwide, but may still bias the imputation of novel alleles. Hence, addition of other alleles that are considered accurate, although computationally rather than physically derived, will strengthen the phylogenetic analysis and contribute to phasing of haplotypes and alleles, such as computed from the 1000 Genomes project [33] and similar online databases.

In our previously published set of long range ERMAP alleles with 72 single nucleotide polymorphisms (SNPs), the number of theoretically possible alleles was 2⁷² [21]. However, it is known that the majority of the haplotype diversity is constituted by only few common haplotypes, which is constant in a given population [34]. Our algorithm restricts the possible ERMAP alleles from 2⁷² to 4243 only, some associated with greater probability of being correct, but all as potential precursors of the experimentally verified extant alleles. With only 72 variable nucleotide positions in our set of 48 ERMAP alleles [21], the vast majority of positions remained uninformative (21,334 of 21,406 nucleotides: 99.66%).

Our observation contrasts with the 2353 SNPs, including 66 out of our 72 SNPs, reported for this DNA stretch covering the ERMAP gene [35], most of them being rare and often not validated to the extent needed for clinical decision making. Increasing the sample size will result in the confirmation of many or most of the previously reported 2353 SNPs and also the identification of novel SNPs in this DNA stretch. However, many of these SNPs will be specific for a small number of individuals resulting in a small global allele frequency.

While initially disregarding recombination as a major contributor, the subsequent analysis of the ERMAP sequences using the ClonalFrameML software [36] was also unable to detect any recombination event among the 48 confirmed alleles. This observation could be explained by the small sample size, which will resolve with the accumulation of more data. Our observation may, however, be an actual feature of ERMAP alleles in the population, because it is similar to the ABO gene, for which the detected recombinant alleles are also of low frequency [37]. As ERMAP alleles caused by recombination will eventually be found, they can be incorporated in the set of alleles used to compute the phylogenetic analysis.