Star alleles explained: how CYP2D6*4 and CYP2C9*3 become a metabolizer group

One copy of a gene, one name

Everyone carries two copies of most genes, one from each parent. A star allele names one of those copies by the particular changes it carries, read together. CYP2D6*4 is a copy of CYP2D6 whose defining change, 1847G>A (rs3892097), breaks a splice site so that the protein comes out truncated^[4]. A pair of names, written *1/*4, is the diplotype: the two copies side by side^[2].

The numbers are handed out by PharmVar's expert panels, and they follow a rule. A haplotype gets a new star number only if it carries a change that alters the protein, disrupts splicing or has been shown to alter function. A newly found haplotype that differs only by silent changes becomes a suballele under an existing number, such as *2.001, and every suballele under one number is assumed to work the same^[4]. That is why *4 can have 28 suballeles and still be one allele: they all share 1847G>A, the single change in its core definition^[4].

*1 is what a test reports when it finds nothing

In most genes *1 is the reference haplotype, the version the gene was first described from. That does not make it the most common one everywhere, and in some genes the reference is another number: for NAT2 it is *4^[2]. The more important point is how a test arrives at *1. An international working group on test reporting put it plainly: laboratories report an allele when they find one of its defining changes, and default to the reference when they find none^[2].

Defaulting works in the other direction too. The change 100C>T (rs1065852) is part of the core definition of several CYP2D6 alleles, among them *10, *36 and *49. A test that checks it without the positions that tell those alleles apart may call *10 by default, and PharmVar's CYP2D6 review notes that defaulting to *1 when nothing is found inflates that allele's apparent frequency^[4].

From two names to a metabolizer group

The step from diplotype to phenotype is a lookup, and CPIC publishes the tables. For CYP2D6 it uses an activity score, proposed in 2008 and tested on 672 people whose CYP2D6 activity was measured with the probe drug dextromethorphan^[5]. Each allele gets an activity value; the two are added; the sum falls into a band^[3]. A duplicated allele counts once per copy, so *1x2 is worth 2^[6].

The bands themselves were renegotiated. Before a consensus published in 2020, CPIC counted a score of 1, such as *1/*4, as a normal metabolizer, while the Dutch guideline group DPWG did not, so the same genotype could carry two labels. A Delphi panel of CYP2D6 experts settled on one system, with 82 percent agreeing to the final method: a score of 1 became intermediate, and *10 was downgraded from 0.5 to 0.25^[3]. A report printed before that and one printed after can disagree about the same DNA.

The frequency column shows how uneven the gene is. In CPIC's European group the no-function *4 is about as common as the normal *2, each near 18.5 percent of alleles, so nearly one copy of CYP2D6 in five is a *4^[7]. The figures are estimates, built from published studies that often tested few positions^[4].

CYP2C9: the same arithmetic, fewer surprises

CYP2C9 uses the same method on a simpler gene. CPIC gives each allele 0 for no function, 0.5 for decreased and 1 for normal; a score of 2 is a normal metabolizer, 1 or 1.5 intermediate, and 0 or 0.5 poor^[8]. In CPIC's European group the two commonest reduced-function alleles are *2, at 12.7 percent of alleles, and *3, at 7.6 percent^[9]. With *1 they make six diplotypes.

The table is short because the gene behaves. Each of *2 and *3 is defined by a single change, rs1799853 and rs1057910^[11], and CPIC's CYP2C9 function table lists no deletion or duplication alleles. It does rate 75 alleles in all, and for 34 of them the function is uncertain; a pair that includes one is classed as indeterminate^{[8, 12]}. A test that reads only *2 and *3 reports anyone carrying a rarer allele as *1.

When the variant is the gene itself

CYP2D6 sits next to two close relatives, CYP2D7 and CYP2D8, which do not make working enzyme. The similarity lets whole stretches of the region be lost, copied or swapped. PharmVar lists the results as structural alleles: *5, a deletion of the entire gene; duplications and multiplications such as *1xN, *2xN and *4xN; and hybrids built partly from CYP2D6 and partly from CYP2D7, such as *13, which can occur alone or in tandem with another copy^[4].

CPIC's tables carry these arrangements by name. Tandems are written with a plus sign: *68+*4 is a non-working hybrid in tandem with a *4 on the same chromosome, and the table also lists combinations as long as *13+*68x2+*4, all scored 0^[6]. The non-working *4 reaches its highest frequency in CPIC's European group and duplications of *1 theirs in the Oceanian group^[4], at 18.5 and 11.9 percent of alleles respectively^[7].

Why a chip cannot count copies

A genotyping chip measures fixed positions, and a sequencing VCF lists the positions that differ from the reference. Neither records how many copies of a gene were present. Take three people whose CYP2D6 copies carry none of the defining changes. One has the usual two copies, one has lost a copy to *5, and one has three copies from *1x2. At every defining position, all three read as reference.

Where it has been measured, the gap is large. In about 50,000 UK Biobank participants with both chip and exome data, CYP2D6 phenotypes inferred from the imputed chip data agreed with the study's combined call set for 64.9 percent of people and the diplotypes for 34.2 percent, against 99.4 percent agreement on phenotype for CYP2C19. Structural variants were not called for CYP2D6 in either data set^[13].

Aligned reads keep what a chip throws away: depth. By simple arithmetic, three copies of a region collect about one and a half times the reads that two copies do, and a single copy about half. Stargazer, a caller built on this idea, estimates copy number from read depth and was run on 28 pharmacogenes in whole genomes^[15]. PyPGx, the caller our Deep Read uses, was written by Stargazer's first author and, in its documentation's words, heavily inspired by it. It adds a machine-learning classifier trained on copy-number profiles of real and simulated genomes, because reads from CYP2D6 and CYP2D7 can be placed on the wrong gene^[16].

Reads are not a guarantee. On 75 reference samples from the GeT-RM programme, the best of three read-based callers matched 67 CYP2D6 diplotypes, and the weakest 54^[14]. PharmVar's CYP2D6 review gives a sharper example: a sample known to carry the *5 deletion on both chromosomes may be called *2/*2 from sequencing files processed without a structural-variant caller^[4].

Same change, different names

Even with a shared dictionary, two tests can name one sample differently, because a name depends on which positions were looked at. When the GeT-RM programme compared seven pharmacogenetic test panels across 28 genes, no two panels tested the same set of variants for any gene, and different haplotypes were often reported for the same allele in the same sample^[2]. The laboratories did not always use the same nomenclature either^[2].

Our own report meets this inside one person. CYP4F2, one of the genes in CPIC's warfarin guideline, has a common change, V433M (rs2108622), named *3^[17]. CPIC's allele definitions show that the same change also sits in *4, together with W12G (rs3093105), and that W12G alone is *2^[18].

The chip and VCF card reads only rs2108622 and names what it finds *3. From aligned reads, PyPGx reads both positions, so a copy that also carries W12G comes back as *4. The two results describe the same V433M under different names. Our report now matches them by the defining change: when a reads-called allele carries exactly one of the changes the variant-file card uses, it takes that card's name for the comparison. The card then says the reads found the same change and why they name it otherwise, for example that the reads name the genotype *1/*4 because the same copy also carries rs3093105, a variant the card does not use.

What our report shows for each file

From a chip export or a plain VCF the report calls nine genes from fixed defining positions and looks the diplotype up in a CPIC-derived table; it does not print an activity score. On a chip a gene is called only when every defining position was typed and can be read on the right strand. A plain VCF has no record for most positions, so a missing record is read as the reference allele and the card labels the result an unconfirmed inference. CYP2D6 is left out of both, and the pharmacogenomics section says so. The VCF, gVCF, BAM and CRAM guide explains why a plain VCF cannot do better.

A gVCF raises the count to 25. PyPGx calls seventeen genes from the variant file, ABCG2 among them, which then replaces the fixed-position ABCG2 card, and each PyPGx card says how many of the gene's defining positions the file examined. A gene is reported only when at least 95 percent were. There is still no read depth, so no copy number, and CYP2D6 stays out.

A BAM or CRAM runs Deep Read: PyPGx on the aligned reads, with CYP2D6 depth compared against a control gene to call copy number. The CYP2D6 card shows the diplotype, the structural call and the activity score, and the CYP2C9 card shows its activity score too. When PyPGx calls an allele the report holds no activity value for, such as the hybrids *13 and *68 or a tandem like *68+*4, the card says so and assigns no phenotype rather than guess one. The CYP2D6 page has the drugs it affects; statins and antidepressants have their own guides on SLCO1B1 and ABCG2 and on antidepressant testing.