Warfarin genetics: CYP2C9, VKORC1, CYP4F2 and what the dosing trials found

What each gene does

Warfarin is a mixture of two mirror-image forms, and S-warfarin is the more potent^[1]. CYP2C9, a liver enzyme, is the main route by which the body clears it, so a copy that works poorly leaves more drug behind^[1]. VKORC1 makes the protein that warfarin blocks. That protein converts vitamin K epoxide back to vitamin K, the rate-limiting step of vitamin K recycling^[1]. CYP4F2 pulls the other way: it oxidises vitamin K to hydroxy-vitamin K1 and so takes it out of the cycle, which is why CPIC calls it a counterpart to VKORC1^[1].

The shares are for people of European ancestry; in other groups the same variants explain less, because the common alleles differ^[1]. In a genome-wide scan of 1,053 Swedish patients, the strongest signals sat next to VKORC1 (p < 10^-78) and in CYP2C9 (p < 10^-31), and nothing else reached genome-wide significance until the authors adjusted for those two genes, age and sex, which brought out CYP4F2 at p = 8.3 x 10^-10^[2].

CYP2C9: *2, *3 and the alleles most tests leave out

The two common reduced-function alleles in people of European ancestry are *2 (p.Arg144Cys, rs1799853) and *3 (p.Ile359Leu, rs1057910)^[1]. CPIC puts their impairment of S-warfarin metabolism at about 30 to 40 percent for *2 and 80 to 90 percent for *3, and says carriers are at greater risk of bleeding on warfarin, need lower doses and take longer to reach a stable INR^[1]. In gnomAD's Finnish group *2 is on 11.4 percent of chromosomes and *3 on 6.2^{[3, 4]}. By our Hardy-Weinberg estimate about 32 percent of Finns carry at least one, and about 1.8 percent carry *3/*3 or *2/*3, the two pairs the guideline gives as examples of poor metabolism^[1]. The star alleles guide explains the scoring and the CYP2C9 page the gene's other drugs.

CPIC lists four further decreased-function alleles, *5, *6, *8 and *11, found mostly in people of African ancestry and collectively more common there than *2 and *3^[1]. In gnomAD's African and African American group *8 is on 5.7 percent of chromosomes, *11 on 1.8, *6 on 1.2 and *5 on 1.1, about 9.8 percent added together (our estimate, assuming four separate haplotypes) against 3.4 percent for *2 and *3^{[3, 4, 5, 6]}. They are not confined to that group: *11 is on 0.4 percent of Finnish chromosomes, and CPIC notes carriers who do not identify as, or know of, African ancestry^{[5, 1]}. A fifth marker, rs12777823 in the CYP2C gene cluster, came from a genome-wide study in African Americans and acted independently of *2 and *3; CPIC reports the association, and a moderate-grade recommendation, only in that group; it did not replicate in a cohort of Egyptians^[1].

VKORC1 -1639G>A, the largest single piece

Laboratories typically report VKORC1 by one variant upstream of the gene, c.-1639G>A (rs9923231), rather than by star alleles; it sits on a haplotype that changes how much VKORC1 protein is made^[1]. The 2005 study that defined the haplotypes studied European-American patients on long-term warfarin and sorted them into a low-dose group, A, and a high-dose group, B. Patients with two A haplotypes had a mean maintenance dose under half that of patients with two B haplotypes (our arithmetic on the paper's means), and the groups explained about 25 percent of the variance in dose^[7]. CPIC's summary is that one or two -1639A alleles call for progressively lower doses than G/G, and that the variant largely explains average dose differences between white, black and Asian patients^[1].

The name causes confusion. VKORC1 is read off the other strand of the chromosome, so the same change appears as C>T in gnomAD, at 16-31096368 on build GRCh38^[8]; a file listing C and T and a report listing G and A can describe one genotype. CPIC adds that laboratories report either -1639G>A or the linked 1173C>T (rs9934438), a different position^[1]. In gnomAD's genomes 16 percent of Finns carry two copies, 46 percent one and 38 percent none, from a small sample of 5,279 people^[8].

CYP4F2 adds a modest correction

CYP4F2 came out of a genome-wide scan. In the Swedish study its V433M variant, rs2108622, emerged after adjusting for VKORC1 and CYP2C9, and was confirmed in 588 more patients^[2]. CPIC takes its best estimate from two large meta-analyses, one in Han Chinese: A-allele carriers took 8 to 11 percent higher doses, a significant but modest effect seen in people of European and Asian ancestry and not African^[1]. CPIC's database gives all three gene-warfarin pairs level A, yet the guideline makes the CYP4F2 adjustment optional and has no CYP4F2 recommendation for people of African ancestry^{[1, 9]}.

The allele is less common in Finland than elsewhere in Europe: 19.6 percent of Finnish chromosomes in gnomAD, against 29.8 percent in the non-Finnish European group^[10]. By gnomAD's counts 31 percent of Finns carry one copy and 4.0 percent two. The variant is called *3 when only rs2108622 is read and *4 when a second change, rs3093105, sits on the same copy; the star alleles guide explains why one test can give the same person either name.

How common the variants are

Each group orders the four variants differently. The table below counts Finnish people rather than chromosomes.

The CPIC guideline and the dosing algorithms

CPIC's 2017 guideline updates its 2011 one with CYP4F2, rs12777823 and ancestry-specific advice^[1]. It recommends that dosing go through one of two published algorithms, the Gage algorithm and the one from the International Warfarin Pharmacogenetics Consortium (IWPC), which CPIC says draw on more than 5,000 subjects between them and give very similar doses^[1]. For people of non-African ancestry whose CYP2C9 genotypes predict poor metabolism, or who combine VKORC1 -1639 A/A with poor CYP2C9 metabolism, it says an alternative oral anticoagulant might be considered^[1].

In the IWPC study 21 research groups in nine countries on four continents contributed 5,700 patients, and the algorithm was derived in 4,043 and tested in 1,009 others^[11]. The two algorithms between them draw on age, sex, race, weight, height, smoking status, the reason for treatment, the target INR and interacting drugs, plus the VKORC1 and CYP2C9 genotypes^[1]; the IWPC version considered only *2 and *3 for CYP2C9^[11]. In the validation cohort the pharmacogenetic version put its estimate within 20 percent of the eventual stable dose more often than a clinical-only version for patients who needed little warfarin (49.4 against 33.3 percent) and for those who needed a lot (24.8 against 7.2 percent)^[11]. The derivation cohort was 55 percent white, 30 percent Asian and 9 percent Black^[11].

Three trials, three answers

EU-PACT randomised 455 people starting warfarin for atrial fibrillation or venous thromboembolism and dosed the genotype arm by algorithm for five days. The median time to a therapeutic INR was 21 days against 29^[12]. CPIC describes the participants as a homogeneous European population^[1]. COAG kept patients and clinicians blind to the dose for four weeks, and 27 percent of its patients were Black. Overall there was no difference, but the result differed by race (interaction p = 0.003). In 255 Black patients time in range was 35.2 percent with the genotype algorithm and 43.5 with the clinical one; in 700 non-Black patients it was 48.8 against 46.1^[13]. CPIC attributes this to overdosing in carriers of *5, *6, *8, *11 and rs12777823, which COAG did not genotype; that reading was not tested in the trial^[1].

GIFT enrolled 1,650 people aged 65 or over having a hip or knee replacement, 91 percent of them white, and dosed the genotype arm for 11 days^[14]. The composite fell from 14.7 to 10.8 percent (relative rate 0.73, 95% CI 0.56 to 0.95), mostly through INR of 4 or more, 56 events against 77. Major bleeding was 2 against 8, with a ratio whose interval ran from 0.05 to 1.15, and venous thromboembolism 33 against 38^[14]. CPIC calls it the first warfarin pharmacogenetics trial powered for clinical outcomes^[1].

The trials differ in four ways that matter. The comparison was a fixed loading regimen in EU-PACT and a clinical algorithm in the other two; CPIC says it is unknown whether loading-dose differences contributed to the different results^[1]. Outcomes ran from four to twelve weeks and from time in range to clinical events. The patients ranged from homogeneous European to a quarter Black to older orthopaedic cases. And GIFT alone tested CYP4F2, while none tested the African-ancestry alleles^{[12, 13, 14]}. A pooled analysis of 27 trials and 9,906 patients found that genotype-guided dosing beat fixed dosing on time in range, time to a stable dose and other efficacy outcomes, while against clinically adjusted dosing only the time to a first therapeutic INR was shorter, with fewer major bleeds^[15]. CPIC's own summary is that trial data are equivocal on whether better dose prediction improves clinical outcomes^[1].

Why newer anticoagulants changed the picture

Warfarin's gene guidance has no counterpart for the newer drugs. Dabigatran, rivaroxaban, apixaban and edoxaban have no CPIC guideline; a search of its database on 10 October 2026 found none^[9]. A meta-analysis of the four pivotal atrial fibrillation trials, 71,683 participants, found that the four drugs cut stroke or systemic embolism by 19 percent against warfarin (relative risk 0.81, 95% CI 0.73 to 0.91), and raised gastrointestinal bleeding (1.25, 1.01 to 1.55)^[16]. The major-bleeding advantage was larger where warfarin control was poorer: 0.69 at centres below 66 percent time in range, 0.93 at or above it^[16].

In Finland the direct oral anticoagulants passed warfarin in number of users in 2019, after an uptake the authors call gradual and slower than in many countries, tied to cautious guidelines and restricted reimbursement^[17]. A trial of dabigatran in people with mechanical heart valves stopped after 252 patients for excess thromboembolic and bleeding events: ischaemic or unspecified stroke in 5 percent against none, major bleeding in 4 percent against 2^[18].

What our report shows for each file

On a chip export, a plain VCF or a gVCF, CYP2C9, VKORC1 and CYP4F2 are three of the nine genes the report reads from fixed positions and looks up in a CPIC-derived table. CYP2C9 gets a metabolizer group (normal, intermediate or poor), VKORC1 a sensitivity label (normal, increased or highly increased) and CYP4F2 a function label; each card carries CPIC's wording (the warfarin page lists the CYP2C9 lines), and the VKORC1 and CYP4F2 wording points to CPIC's algorithm rather than an amount. A chip gene is normally called only when every defining position was typed and readable. A plain VCF has no record where nothing varies, so a missing record reads as reference and the card calls it an unconfirmed inference; a gVCF reads the same positions and leaves the gene unresolved when its reference blocks call nothing there.

From a BAM or CRAM, Deep Read runs PyPGx on the reads; the CYP2D6 guide covers the one gene where it also counts copies. The CYP2C9 card shows a diplotype and an activity score; an allele outside CPIC's activity table, which includes *5, *6, *8 and *11, leaves it without a phenotype. VKORC1 and CYP4F2 appear as genotypes, and beside a variant file that reads the same genotype the variant file's card, with its label and warfarin line, replaces the reads' row. The CYP2C rs12777823 genotype is reported without guidance. Nothing combines the three genes into an estimate, because CPIC's algorithms need clinical inputs the report does not hold.