ABCB1 genotype and clinically relevant bleeding risk during extended anticoagulation therapy for pulmonary embolism: a two-phase study Authors: Zhu Zhang, Haobo Li, Bingzhang Zou, Xiaofan Ji, Yu Zhang, Yuzhi Tao, Linfeng Xi, Kuangxun Li, Huiwen Li, Mengjie Duo, Li Zhao, Xianbo Zuo, Zhenguo Zhai, Chen Wang, CURES investigators Published: 2026 ABSTRACT AbstractBackgroundExtended direct oral anticoagulant (DOAC) therapy reduces recurrent venous thromboembolism, but clinically relevant bleeding (CRB) remains a safety concern. DOACs exhibit substantial pharmacokinetic variability among individuals, and genetic determinants of bleeding risk during extended anticoagulation remain unclear.ObjectivesTo explore the association between pharmacogenomic variants and CRB during extended rivaroxaban therapy and explore their potential contribution to bleeding risk prediction in patients with pulmonary embolism (PE).MethodsIn phase 1, 633 rivaroxaban-treated PE patients from the multicenter registry with genome-wide genotyping data were included. Thirteen candidate single-nucleotide polymorphisms across 9 genes previously reported to influence DOAC pharmacokinetics were assessed for associations with CRB during extended anticoagulation. In phase 2, 101 PE patients from an independent cohort underwent rivaroxaban plasma concentration measurement and genotyping for pharmacokinetic validation.ResultsIn phase 1, 70 patients (11.1%) experienced CRB during the 1-year follow-up. Three ABCB1 variants showed nominal associations with CRB; none remained significant after study-wide false discovery rate correction (adjusted P = .074), although they met the gene-level false discovery rate threshold (adjusted P = .030). Among these, rs1045642 was also associated with dose-normalized peak plasma concentration in phase 2 (P = .001), without affecting recurrent venous thromboembolism or mortality. In phase 2, ABCB1 rs1045642 TT carriers had significantly higher dose-normalized peak concentrations (P = .001). Incorporating rs1045642 into 5 established bleeding scores (Kuijer, RIETE, VTE-BLEED, PE-SARD and VTE-PREDICT) consistently improved reclassification (category-free net reclassification improvement, 0.214-0.312; integrated discrimination improvement, 0.022-0.035), although the corresponding improvements in area under the receiver operating characteristic curve did not reach statistical significance.ConclusionABCB1 rs1045642 showed a nominal association with CRB and higher rivaroxaban peak plasma concentration during extended therapy. Adding this genotype to existing bleeding scores improved risk reclassification but not overall discrimination, and its clinical utility requires independent prospective validation.Keywords: DOACs, extended anticoagulation, pharmacogenomics Introduction Pulmonary embolism (PE) affects approximately 1 in 1000 adults annually and remains a leading cause of cardiovascular mortality [1]. In China, the incidence of PE among hospitalized patients has increased 5-fold over the past decade [2,3]. For patients with persistent or unprovoked risk factors, extending anticoagulation beyond the initial 3 to 6 months reduces recurrent venous thromboembolism (VTE) without increasing major bleeding (MB) [[4], [5], [6], [7]]. However, clinically relevant non-major bleeding (CRNMB) remains significantly elevated with prolonged therapy [8], and this risk appears particularly pronounced with rivaroxaban, which was associated with more than twice the rate of clinically relevant bleeding (CRB) compared with apixaban in the COBRRA trial [9]. This is of particular concern in China, where rivaroxaban is the most widely prescribed direct oral anticoagulant (DOAC) for PE [10,11], and where East Asian patients exhibit a distinct bleeding-efficacy balance compared with Western populations [12,13]. Accurately identifying which rivaroxaban-treated patients are most susceptible to bleeding during extended therapy is therefore a clinical priority. Five bleeding-risk scores currently inform anticoagulation decision-making in VTE: Kuijer [14], RIETE [15], VTE-BLEED [16], PE-SARD [17], and VTE-PREDICT [18]. However, none was specifically calibrated for rivaroxaban during the extended phase, and East Asian populations were under-represented in all derivation cohorts. The transportability of these scores to Chinese or East-Asian PE patients receiving extended rivaroxaban therefore remains uncertain. Pharmacogenomic stratification has been successfully implemented for warfarin and clopidogrel [19,20], and several studies have explored genetic determinants of DOAC pharmacokinetics, although results remain inconsistent [21,22]. Prior rivaroxaban pharmacogenomic studies have focused on atrial fibrillation or acute VTE cohorts, with no study specifically evaluating PE patients or CRB during extended anticoagulation. Whether incorporating genotype into existing bleeding scores can improve extended-phase risk prediction in PE patients remains unknown. In this study, we used a 2-phase design to explore pharmacogenomic associations with CRB during extended anticoagulation therapy in Chinese PE patients from the China pUlmonary Thromboembolism REgiStry (CURES) multicenter PE registry. In phase 1, we screened candidate variants for association with bleeding outcomes in the CURES. In phase 2, we examined the association between the lead variant and rivaroxaban peak plasma concentration in an independent pharmacokinetic (PK) cohort and evaluated changes in model performance after incorporating genotype into established bleeding scores. Methods Study design and setting This study comprised 2 complementary phases. Phase 1 was a secondary analysis of genotyping and clinical outcome data from the CURES [23]. CURES is a nationwide, prospective, multicenter registry of patients hospitalized for PE across tertiary centers in China [11]. In phase 1, we restricted the analysis to the rivaroxaban-treated subset of the CURES cohort to examine pharmacogenomic determinants of bleeding risk. Phase 2 was a prospective, single-center PK substudy conducted at the China-Japan Friendship Hospital (Beijing, China), designed to examine the association between the identified genetic variant and rivaroxaban peak plasma concentration. Participants For phase 1, a total of 1268 CURES PE patients with available genotyping data who were screened for eligibility and patients who received rivaroxaban for extended anticoagulation were enrolled after excluding: (a) anticoagulant treatment duration ≤3 months (n = 43); (b) non-rivaroxaban anticoagulant regimen (n = 508); and (c) follow-up duration <3 months (n = 84). Phase 2 was a prospective, single-center PK substudy conducted at the China-Japan Friendship Hospital between June 2023 and October 2024. Eligible participants were consecutive adults (≥18 years) with confirmed PE who had completed at least 3 months of initial anticoagulation and were continuing rivaroxaban for extended-phase therapy at the discretion of the treating physician. Rivaroxaban dose during the extended phase was recorded from medical records at each study visit. Patients received either the full treatment dose (20 mg once daily) or a reduced dose (10-15 mg once daily) at the treating physician’s discretion based on individual bleeding risk. All participants provided written informed consent. To minimize confounding of rivaroxaban plasma concentration measurement, patients were excluded if they were receiving strong inhibitors or inducers of P-glycoprotein or CYP3A4 (itraconazole, ketoconazole, voriconazole, posaconazole, ritonavir, rifampicin, phenytoin, phenobarbital, or carbamazepine) within 14 days before enrollment, or if they were pregnant or lactating. Patients who subsequently discontinued anticoagulation or declined blood sampling for plasma concentration measurement and genotyping were not included in the final analysis data. Rivaroxaban plasma concentrations were measured at steady state, defined as ≥5 days of uninterrupted rivaroxaban treatment (Figure 1). Flowchart of patient selection. Outcomes The primary outcome was a composite of MB and CRNMB, defined according to the International Society on Thrombosis and Haemostasis (ISTH) criteria [24]. A central adjudication committee, blinded to genotype and baseline characteristics, reviewed all bleeding events and required full consensus for final classification. Secondary outcomes were recurrent VTE and all-cause mortality. Outcomes were assessed by telephone or outpatient follow-up at 3, 6, and 12 months, following the CURES study protocol. Genotyping and quality control Genotyping of participants has been described in detail previously [23]. Briefly, genomic DNA was extracted from peripheral whole blood. All samples were genotyped at WeGene Lab using a customized Illumina WeGene V2 Array (Illumina iScan System) comprising approximately 700,000 markers, including ∼560,000 genome-wide backbone markers optimized for Chinese population genome-wide association study and additional markers for clinical research. Genotype calling was performed with Illumina GenomeStudio software. Genotype imputation was conducted using the ChinaMAP imputation server. Quality control procedures excluded individuals with gender mismatches, call rate >5% missingness, evidence of relatedness (second-degree or closer by KING), inbreeding coefficient outside (−0.2, 0.2), or non-Han Chinese ancestry. Single-nucleotide polymorphisms (SNPs) were excluded for call rate <95%, minor allele frequency (MAF) <1%, or deviation from Hardy-Weinberg equilibrium (P < 1 × 10−6). Population stratification was assessed using the top 10 principal components (PCs) calculated by Genome-wide Complex Trait Analysis. Candidate gene association analysis Candidate SNPs were based on a literature review [25] and pharmacogenomic databases (PharmGKB, CPIC) for variants previously reported to influence the pharmacokinetics or pharmacodynamics of DOACs. Selected genes encompassed drug transporters involved in DOAC absorption and efflux and metabolizing enzymes involved in DOAC biotransformation, among others, all previously implicated in rivaroxaban or DOAC-class PK variability. A total of 13 SNPs across 9 candidate genes were analyzed. Each variant was assessed under both dominant and recessive genetic models using Cox proportional hazards regression adjusted for age, sex, renal function, active gastrointestinal tumors, and prior bleeding. Multiple testing was controlled using the false discovery rate (FDR) method (Benjamini-Hochberg procedure), with an FDR threshold of q < 0.05. PK analysis In phase 2, blood samples for determining peak plasma concentration (Cmax) were collected 3 hours after observed rivaroxaban administration, whereas samples for trough concentration (Cmin) were collected immediately before the next scheduled dose. Blood samples were collected into EDTA-coated tubes, and plasma was separated by centrifugation at 3000 rpm for 10 minutes at 4°C within 2 hours of collection. Plasma samples were stored at −80°C until analysis. Rivaroxaban plasma concentrations were quantified using a validated ultraperformance liquid chromatography–tandem mass spectrometry method (UPLC-MS/MS), as previously described [26]. Bleeding risk scores Baseline data from the phase 1 cohort were used to calculate the Kuijer [14], RIETE [15], VTE-BLEED [16], PE-SARD [17], and VTE-PREDICT (Supplementary Table 1). Variable definitions from each original derivation study were used whenever possible. Where original definitions differed from the available database variables, the closest standardized definitions were substituted. Each score was entered as a continuous predictor in subsequent analyses. Statistical analysis Continuous variables are presented as mean ± SD or median (interquartile range), and categorical variables as frequencies and percentages. Hardy-Weinberg equilibrium was tested for each candidate SNP by Pearson χ2; all variants satisfied equilibrium. Missing baseline data were handled using multiple imputation by chained equations (MICE), with 5 imputed datasets combined using Rubin’s rules. Genetic associations with the composite bleeding endpoint were estimated by cause-specific Cox proportional hazards regression under dominant and recessive inheritance, adjusted for age, sex, renal function, active gastrointestinal tumors, and prior bleeding, with all-cause death treated as a competing event. Multiple-testing correction was applied at 2 levels: a primary Benjamini-Hochberg FDR adjustment across all 13 candidate SNPs, and a gene-level correction in which FDR was recalculated separately among SNPs within each gene region to examine locus-specific signals among biologically related and correlated variants. Cumulative incidence functions were estimated using Fine-Gray subdistribution hazard models and compared across genotypes by Gray test. In the phase 2 PK substudy, dose-normalized peak concentrations (Cmax/Dose) were compared across ABCB1 genotypes using the Kruskal–Wallis test with Dunn pairwise post-hoc comparisons. Predicted probabilities for the composite endpoint at the cohort follow-up time were estimated by entering each score as a continuous predictor in a Fine-Gray model with death treated as a competing event. The lead variant was then added to each score, and discrimination of the base and enhanced models was compared using the time-dependent area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (DeLong method). Calibration was assessed using calibration plots. Statistical significance was defined as a two-tailed P < .05 unless otherwise specified. Analyses were performed using R version 4.0.2. Results Study population The patient selection process is shown in Figure 1. In phase 1, 1268 patients with PE were screened from the CURES multicenter registry. After excluding patients with < 3 months of anticoagulant treatment (n = 43), those not receiving rivaroxaban (n = 508), and those with follow-up < 3 months (n = 84), 633 patients receiving rivaroxaban were enrolled after completing 3 months of anticoagulation for PE. In phase 2, 210 patients who completed 3 months of rivaroxaban treatment were consecutively recruited from June 2023 to October 2024 at China-Japan Friendship Hospital. After excluding patients who discontinued anticoagulation as judged by the treating physician (n = 44), those who declined plasma concentration testing and genotyping (n = 61), and those using strong P-gp/CYP3A4 inhibitors or inducers (n = 4), 101 patients underwent rivaroxaban plasma concentration and genotyping testing. Baseline characteristics Baseline demographic and clinical characteristics are presented in Table 1, anticoagulation regimens and duration in Supplementary Table 2, and additional characteristics of the index PE event and initial anticoagulant treatment in Supplementary Table 3. In phase 1 (n = 633), the mean age was 63.8 ± 4.8 years and 49.0% were female. During the extended treatment phase (from 3 to 12 months after PE diagnosis, censored at month 12, 70 patients (11.1%) experienced a CRB, including 10 MB events and 60 CRNMB. Thus, CRNMB accounted for 85.7% of the composite bleeding events. And 90 patients (14.2%) died from any cause, and 19 (3.0%) had symptomatic recurrent VTE. Table 1Demographic and clinical characteristics of the phase 1 and phase 2 groups.Phase 1Phase 2Standardized differencesAllN = 633CRBN =70Non-CRBN =563N = 101Demographic Age, y, mean ± SD63.8 ± 4.864.5 ± 4.863.4 ± 4.963.4 ± 5.20.08 Female, n (%)310 (49.0)32 (45.0)278 (49.4)48 (47.5)0.03 BMI, kg/m224.1 ± 1.223.6 ± 1.324.3 ± 1.224.9 ± 4.60.24Comorbidities, n (%) Diabetes76 (12.0)9 (13.4)67 (11.9)5 (5.0)0.26 Hypertension83 (13.1)13 (18.1)70 (12.5)10 (9.9)0.10 Coronary artery disease102 (16.1)11 (15.7)91 (16.2)8 (7.9)0.25 Stroke64 (9.7)10 (14.8)54 (9.6)5 (5.0)0.20 Heart failure155 (24.5)19 (27.5)136 (24.1)9 (8.9)0.43 Major surgery or trauma (within 3 months)77 (12.2)10 (14.8)67 (11.9)1 (1.0)0.46 Active cancer53 (8.4)7 (10.0)46 (8.2)6 (5.9)0.09 Gastrointestinal tumors20 (3.0)4 (5.4)16 (2.8)1 (1.0)0.15 Urologic neoplasms3 (0.5)03 (0.5)00.10 Others30 (4.7)3 (4.3)27 (4.8)5 (5.0)0.01Persistent immobility55 (8.7)6 (8.6)49 (8.7)4 (4.0)0.20Chronic kidney disease49 (7.7)9 (12.9)40 (7.1)1 (1.0)0.34 Chronic liver disease12 (1.8)3 (4.0)9 (1.6)00.20Bleeding events occurred within 3 months of anticoagulation, n (%)16 (2.5)3 (4.0)13 (2.3)1 (1.0)0.12Continuous variables are presented as mean ± SD; categorical variables as n (%). CRB = clinically relevant bleeding (a composite endpoint consisting of major bleeding and clinically relevant non-major bleeding during follow-up). Standardized differences (SMD) compare phase 1 (overall, n = 633) with phase 2 (n = 101). Demographic and clinical characteristics of the phase 1 and phase 2 groups. Continuous variables are presented as mean ± SD; categorical variables as n (%). CRB = clinically relevant bleeding (a composite endpoint consisting of major bleeding and clinically relevant non-major bleeding during follow-up). Standardized differences (SMD) compare phase 1 (overall, n = 633) with phase 2 (n = 101). Patients who experienced CRB (n = 70) were comparable to those without CRB across all baseline variables. In phase 2 (n = 101), baseline characteristics were broadly similar to those in phase 1, with SMDs below 0.20 for most variables (Table 1). Notable exceptions included heart failure (24.5% vs 8.9%; SMD 0.43) and recent major surgery or trauma (11.6% vs 1.0%; SMD 0.46), both lower in phase 2, possibly reflecting a survivor bias inherent in the cross-sectional enrollment design. A higher proportion of phase 2 patients received reduced-dose rivaroxaban (10-15 mg/day: 25.7% vs 13.9%; Supplementary Table 2). Pharmacogenomic associations with CRB (phase 1) The associations of 13 candidate SNPs with CRB are presented in Table 2. Under the dominant model, 3 ABCB1 variants rs1045642, rs4148738, and rs2032582 showed nominally significant associations with CRB. After FDR correction across all 13 SNPs, none reached the 0.05 threshold (adjusted P = .074 for all 3), however, all 3 met the gene-level FDR threshold in a secondary exploratory analysis within the ABCB1 locus (adjusted P = .030). Given the known high linkage disequilibrium among these variants, these signals may reflect a shared underlying effect. Absolute risks of CRB by genotype are provided in Supplementary Table 4. MB occurred in 2 (0.84%) CC carriers, 6 (2.06%) CT carriers, and 2 (1.90%) TT carriers. No other SNP reached nominal significance under either genetic model. Kaplan-Meier curves stratified by ABCB1 genotype are shown in Figure 2. Table 2Association of candidate single-nucleotide polymorphisms with the composite bleeding endpoint under dominant and recessive genetic models in the phase 1 cohort.SNPChromosome: positionCohort MAFEAS reference MAFAmerican reference MAFEuro reference MAFDominant modelRecessive modelHR (95% CI)PP1P2HR (95% CI)PP1P2ABCB1 rs10456427:871386450.3880.3980.4280.4822.01 (1.13-3.56).017.074.0301.30 (0.69-2.44).413.596.413ABCB1 rs11285037:871796010.3460.3730.4160.4030.62 (0.38-1.03).066.215.0660.54 (0.21-1.39).203.596.413ABCB1 rs20325827:871606180.4420.3980.4100.3690.42 (0.21-0.84).015.074.030----ABCB1 rs41487387:871630490.4150.4060.3820.4352.17 (1.18-4.00).013.074.0301.29 (0.71-2.34).410.596.413ABCG2 rs22311424:890523230.3030.2910.1410.0940.95 (0.58-1.57).845.845.8450.73 (0.28-1.90).523.596.523ABCA6 rs721250617:671017180.1450.0670.2290.0671.26 (0.73-2.17).414.544.414----AKR7A3 rs17380231:196112410.1970.1490.1350.1420.81 (0.47-1.38).435.544.4350.63 (0.15-2.72).536.596.536AKR7A3 rs17380251:196092540.2020.1590.1330.1500.79 (0.46-1.34).382.544.4350.58 (0.13-2.48).460.596.536CES1 rs224461316:558446090.3800.3980.2720.1611.22 (0.72-2.09).460.544.4601.64 (0.87-3.10).127.596.127CYP3A4 rs22424807:993614660.2490.2680.3930.0811.39 (0.84-2.29).196.425.1961.85 (0.82-4.18).138.596.138CYP3A5 rs7767467:992705390.3110.2790.2010.0541.56 (0.94-2.57).083.216.0831.51 (0.71-3.22).290.596.290SUSD3 rs762925449:958417410.1100.1230.0230.0561.18 (0.64-2.17).600.650.600----SLCO1B1 rs414905612:213315490.1120.1240.1340.1611.31 (0.74-2.32).360.544.3600.98 (0.12-8.02).988.988.988Chromosomal positions refer to the GRCh38 human reference assembly. The 1000G EAS MAF column reports the minor allele frequency in the East Asian superpopulation (n = 504) of the 1000 Genomes Project phase 3.Hazard ratios were estimated using cause-specific Cox proportional hazards regression for clinically relevant bleeding, adjusted for age, sex, renal function, active gastrointestinal tumors, and prior bleeding, with all-cause death treated as a competing event.A dash (–) under the recessive model indicates that no homozygous variant carriers were observed in the cohort, precluding estimation.SNP, single-nucleotide polymorphism; HR, hazard ratio; CI, confidence interval; MAF, minor allele frequency; P, unadjusted P value; P1, Benjamini–Hochberg false discovery rate–adjusted P value; P2, P value computed within each gene region (gene-level correction). Association of candidate single-nucleotide polymorphisms with the composite bleeding endpoint under dominant and recessive genetic models in the phase 1 cohort. Chromosomal positions refer to the GRCh38 human reference assembly. The 1000G EAS MAF column reports the minor allele frequency in the East Asian superpopulation (n = 504) of the 1000 Genomes Project phase 3. Hazard ratios were estimated using cause-specific Cox proportional hazards regression for clinically relevant bleeding, adjusted for age, sex, renal function, active gastrointestinal tumors, and prior bleeding, with all-cause death treated as a competing event. A dash (–) under the recessive model indicates that no homozygous variant carriers were observed in the cohort, precluding estimation. SNP, single-nucleotide polymorphism; HR, hazard ratio; CI, confidence interval; MAF, minor allele frequency; P, unadjusted P value; P1, Benjamini–Hochberg false discovery rate–adjusted P value; P2, P value computed within each gene region (gene-level correction). Kaplan-Meier curves for clinically relevant bleeding by ABCB1 genotype. Kaplan-Meier curves are presented for visualization of event patterns. Hazard ratios were estimated using Fine-Gray subdistribution hazard regression with all-cause death as a competing event, adjusted for age, sex, renal function, active gastrointestinal tumors, and prior bleeding. (A) ABCB1 rs1045642; (B) ABCB1 rs2032582; (C) ABCB1 rs4148738. Association of ABCB1 variants with rivaroxaban plasma concentration (phase 2) To explore a PK basis for the observed genetic association, we examined the 3 nominally significant ABCB1 variants against dose-normalized peak concentration (Cmax/Dose) in phase 2 (n = 101). rs1045642 was significantly associated with Cmax/Dose (P = .001; Figure 3A). Pairwise comparisons showed that TT carriers had significantly higher Cmax/Dose than CC carriers (P < .01). Neither rs4148738 (P=0.512) nor rs2032582 (P=0.837) was significantly associated with Cmax/Dose (Figure 3B, C). No significant associations were observed between any of the 3 ABCB1 variants and dose-normalized trough concentrations (rs1045642, P = .333; rs4148738, P = .349; rs2032582, P = .925; Supplementary Figure 1). ABCB1 rs1045642 carrier status was not associated with recurrent VTE or all-cause death (Table 3). However, these secondary analyses were limited by the number of outcome events. Association of ABCB1 variants with rivaroxaban peak plasma concentrations. Cmax/Dose values (ng·mL−1per mg) by ABCB1 genotype. P values were derived from the Kruskal–Wallis test with Dunn’s post-hoc pairwise comparisons. (A) rs1045642; (B) rs4148738; (C) rs2032582. Ns, not significant; ∗∗ P < .01. Table 3Secondary clinical outcomes during follow-up in the phase 1 cohort, stratified by bleeding status and ABCB1 rs1045642 genotype.OutcomeAll (n = 633)Stratified by bleeding statusStratified by ABCB1 rs1045642 (dominant model)CRB (n = 70)Non-CRB (n = 563)HRa(95% CI)PCT + TT (n = 396)CC (n = 237)HRb (95% CI)PSymptomatic recurrent VTE19 (3.0)5 (7.1)14 (2.5)1.4 (0.7-2.8)0.3012 (3.0)7 (2.9)1.0 (0.4–2.6)0.96All-cause death90 (14.2)10 (14.3)80 (14.2)1.5 (0.6-2.9)0.2356 (14.1)34 (14.3)1.0 (0.6–1.5)0.93Categorical variables are presented as n (%).CRB, clinically relevant bleeding, defined as a composite endpoint consisting of major bleeding and clinically relevant non-major bleeding during follow-up according to International Society on Thrombosis and Haemostasis (ISTH) criteria; VTE, venous thromboembolism.aHazard ratios compare patients who experienced CRB with those who did not, with bleeding modelled as a time-varying exposure to avoid immortal time bias. Cox proportional hazards regression was adjusted for age and sex; for recurrent VTE, all-cause death was treated as a competing event using a Fine–Gray subdistribution hazards model.bHazard ratios compare carriers of the variant allele (CT + TT, dominant model) with non-carriers (CC) at ABCB1 rs1045642. Cox proportional hazards regression was adjusted for age and sex, with all-cause death treated as a competing event for non-fatal outcomes. Secondary clinical outcomes during follow-up in the phase 1 cohort, stratified by bleeding status and ABCB1 rs1045642 genotype. Categorical variables are presented as n (%). CRB, clinically relevant bleeding, defined as a composite endpoint consisting of major bleeding and clinically relevant non-major bleeding during follow-up according to International Society on Thrombosis and Haemostasis (ISTH) criteria; VTE, venous thromboembolism. Hazard ratios compare patients who experienced CRB with those who did not, with bleeding modelled as a time-varying exposure to avoid immortal time bias. Cox proportional hazards regression was adjusted for age and sex; for recurrent VTE, all-cause death was treated as a competing event using a Fine–Gray subdistribution hazards model. Hazard ratios compare carriers of the variant allele (CT + TT, dominant model) with non-carriers (CC) at ABCB1 rs1045642. Cox proportional hazards regression was adjusted for age and sex, with all-cause death treated as a competing event for non-fatal outcomes. Changes in bleeding risk model performance after adding ABCB1 rs1045642 We first evaluated the discriminative performance of 5 established bleeding risk scores (Kuijer, RIETE, VTE-BLEED, PE-SARD and VTE-PREDICT) in our cohort. The base AUCs ranged from 0.524 to 0.613 (Table 4), indicating modest discrimination. Adding rs1045642 (dominant model) consistently improved reclassification metrics: the category-free net reclassification improvement was significant for all 5 scores (range, 0.214-0.312), as was the integrated discrimination improvement (range, 0.022-0.035). AUC improvements ranged from +0.061 to +0.097 (Figure 4) but did not reach significance by the DeLong test, warranting cautious interpretation. All enhanced models showed adequate calibration (Hosmer–Lemeshow P > .05; Supplementary Figure 2), and DCA demonstrated greater net benefit across the 5% to 25% threshold range (Supplementary Figure 3). Table 4Discrimination, reclassification, and calibration of established bleeding risk scores with and without ABCB1 rs1045642 in patients with pulmonary embolism receiving rivaroxaban.Risk scoreAUC of base score (95% CI)AUC of enhanced model (95% CI)ΔAUCP valueacfNRI (95% CI)IDI (95% CI)CalibrationP valuebKuijer0.524 (0.470-0.578)0.621 (0.534-0.708)+0.0970.0510.312 (0.062-0.554)0.035 (0.008-0.071)P > .05RIETE0.560 (0.504-0.636)0.648 (0.583-0.713)+0.0880.0630.286 (0.048-0.523)0.032 (0.007-0.065)P > .05PE-SARD0.613 (0.526-0.700)0.674 (0.591-0.756)+0.0610.3190.214 (0.006-0.438)0.022 (0.001-0.052)P > .05VTE-BLEED0.546 (0.492-0.600)0.629 (0.555-0.702)+0.0830.0740.268 (0.035-0.501)0.030 (0.006-0.062)P > .05VTE-PREDICT0.602 (0.520-0.684)0.671 (0.590-0.751)+0.0690.2200.234 (0.017-0.462)0.025 (0.003-0.056)P > .05Data are presented as point estimates (95% confidence intervals) unless otherwise indicated. Enhanced model = base risk score plus ABCB1 rs1045642 genotype (dominant model: CT+TT vs CC).aP value for the comparison of AUCs between base and enhanced models, calculated by the DeLong test.bCalibration was assessed by the Hosmer–Lemeshow goodness-of-fit test; P > .05 indicates adequate calibration. AUC, area under the receiver operating characteristic curve; cfNRI, category-free net reclassification improvement; IDI, integrated discrimination improvement; NRI, net reclassification improvement; CI, confidence interval. Discrimination, reclassification, and calibration of established bleeding risk scores with and without ABCB1 rs1045642 in patients with pulmonary embolism receiving rivaroxaban. Data are presented as point estimates (95% confidence intervals) unless otherwise indicated. Enhanced model = base risk score plus ABCB1 rs1045642 genotype (dominant model: CT+TT vs CC). P value for the comparison of AUCs between base and enhanced models, calculated by the DeLong test. Calibration was assessed by the Hosmer–Lemeshow goodness-of-fit test; P > .05 indicates adequate calibration. AUC, area under the receiver operating characteristic curve; cfNRI, category-free net reclassification improvement; IDI, integrated discrimination improvement; NRI, net reclassification improvement; CI, confidence interval. Comparison of discrimination for bleeding risk scores with and without ABCB1 rs1045642. Receiver operating curves for the base risk score (blue) and the enhanced model incorporating ABCB1 rs1045642 (red): (A) Kuijer; (B) RIETE; (C) VTE-PREDICT; (D) PE-SARD; (E) VTE-BLEED. Discussion In this 2-phase study of Chinese PE patients receiving rivaroxaban, ABCB1 rs1045642 was nominally associated with CRB during extended anticoagulation and with higher dose-normalized peak plasma concentrations. Adding rs1045642 to established bleeding risk scores improved reclassification but not overall discrimination; therefore, it may provide incremental value in refining risk stratification despite the absence of a significant change in AUC. Extended anticoagulation reduces VTE recurrence by over 80% without increasing MB, as established in multiple large-scale trials [7,27], benefits that extend beyond unprovoked VTE to provoked events with persistent risk factors. However, extended anticoagulation consistently increases CRNMB, as evidenced by the EINSTEIN [8] (hazard ratio [HR], 5.19; 95% CI, 2.3-11.7), HI-PRO (4.8% vs 1.7%; P = .06) [7], and ExACT trials (adjusted HR, 2.99; P = .10) [5]. Although rarely fatal, persistent CRNMB increases health care costs, imposes a considerable patient burden, and is a recognized driver of premature anticoagulation discontinuation [18], which is why we adopted CRB as the primary safety endpoint. Because 60 of the 70 bleeding events (85.7%) were CRNMB, the composite endpoint and the observed genetic association primarily reflected CRNMB rather than MB. As the population eligible for extended anticoagulation expands, accurate bleeding risk stratification becomes essential to identify patients who can safely continue prolonged therapy. However, 5 established bleeding scores demonstrated only modest discrimination in our cohort (AUC, 0.524-0.613), likely reflecting their derivation from vitamin K antagonist-treated Western populations and the distinct thrombosis-bleeding balance in East Asian patients. Incorporating ABCB1 rs1045642 consistently improved reclassification across all 5 scores (category-free net reclassification improvement, 0.214-0.312; integrated discrimination improvement, 0.022-0.035), with the clinically meaningful gain observed for VTE-BLEED and VTE-PREDICT, the only 2 scores designed for extended anticoagulation treatment period and recommended by guidelines. These findings suggest that this genotype may contribute to risk reclassification alongside existing clinical bleeding scores, although its impact on overall discriminative performance remains to be confirmed. Pharmacogenomics offers a complementary approach. In the antiplatelet domain, the CHANCE substudy demonstrated that CYP2C19 genotype modified clopidogrel efficacy without affecting bleeding [20] directly informing the genotype-guided CHANCE-2 trial [28]. For rivaroxaban, existing pharmacogenomic data are limited to small PK studies in which genotype-concentration associations did not translate into clinical bleeding differences [21,26], and even the largest DOAC candidate-gene study to date (n = 2364) found no significant association between any PK-related variant and clinical bleeding after multiple-testing correction [29]. In our PE cohort (n = 633), which exclusively evaluated bleeding beyond 3 months of anticoagulation, ABCB1 rs1045642 was the only variant nominally associated with CRB after gene-level correction (HR, 2.01; adjusted P = .030), while all previously reported positive findings were not replicated. Although rs1045642 did not survive study-wide FDR correction (adjusted P = .074), this is consistent with the current state of DOAC pharmacogenomics, in which no candidate variant has yet achieved study-wide significance for clinical bleeding endpoints. Our finding is, however, supported by convergent evidence: a recent meta-analysis of 39 studies (n = 13,300) confirmed that rs1045642 was consistently associated with altered rivaroxaban trough concentrations and bleeding risk across multiple populations [30], a large European biobank study reported associations between ABCB1 haplotypes and clinical outcomes in DOAC users [31], and a clinical case of rivaroxaban-induced hemorrhage linked to homozygous ABCB1 variant alleles demonstrated a 2- to 3-fold prolongation of elimination half-life [32]. This selective association with peak but not trough concentrations may reflect the fact that peak levels are more sensitive to first-pass intestinal absorption efficiency, whereas trough levels represent a composite of absorption, distribution, and multipathway elimination (hepatic and renal), potentially diluting the influence of a single transporter variant. Taken together with our phase 2 PK data showing significantly higher dose-normalized Cmax in rs1045642 TT carriers, these findings support a biologically coherent genotype-Cmax-bleeding association that warrants prospective validation. Notably, rs1045642 in our study did not survive FDR correction across all 13 SNPs (adjusted P = 0.074), consistent with the broader DOAC pharmacogenomics literature in which no candidate variant has achieved genome-wide or study-wide significance for clinical bleeding endpoints to date, even in cohorts exceeding 2000 patients [29]. The exploratory gene-level association, together with the concordant PK findings in phase 2, supports further investigation in independent prospective cohorts. There are several limitations in our study. The observational design of phase 1 remains susceptible to unmeasured confounding (eg, concomitant antiplatelet use, renal function trajectory) and channeling bias in dose selection. Because no association survived study-wide FDR correction and phase 2 did not validate bleeding outcomes, the gene-level findings remain exploratory and susceptible to chance. The small number of MB events and nonsignificant AUC improvements limit conclusions regarding MB and predictive utility. Plasma concentration data were available only in phase 2, precluding formal mediation analysis in the larger cohort. The cohort was exclusively Han Chinese, and replication in other populations is required before clinical implementation. Conclusion The ABCB1 rs1045642 was nominally associated with CRB during extended rivaroxaban therapy in Chinese PE patients, with variant carriers showing a higher bleeding risk but no difference in VTE recurrence or mortality. Adding this genotype to 5 established bleeding scores consistently improved reclassification; pharmacogenomic information may complement clinical risk assessment but requires external validation and prospective evaluation before clinical application can be considered. Acknowledgments The authors thank the participants of the China Pulmonary Thromboembolism Registry Study (CURES) for their important contributions. Funding This research was funded by the National Key Research and Development Program of China (No. 2023YFC2507200, 2024YFE0101900), the Chinese Academy of Medical Sciences Innovation Fund for Medical Sciences (No.2024-I2M-TS-035), Noncommunicable Chronic Diseases National Science and Technology Major Project (No.2024ZD0528700). Ethics statement The study protocol was approved by the independent ethics committee of China-Japan Friendship Hospital (phase 1, approval No. 2016-SSW-7; phase 2, approval No. 2022-KY-238) and was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants. Author contributions Study concept and design: Z. Zhang and Z. Zhai. Data were collected and cleaned by H. Li, B. Zou, Y. Zhang, Y. Tao, L. Xi, K. Li, H. Li, M. Duo, and L. Zhao. Data analysis: H. Li, B. Zou, X. Ji. Drafting of the manuscript: Z. Zhang, H. Li, and B. Zou. All authors provided final approval of the version to be published. Relationship disclosure There are no competing interests to disclose. Data availability For original data and code, please email the corresponding author. Footnotes Zhu Zhang, Haobo Li, Bingzhang Zou, and Xiaofan Ji contributed equally to this manuscript as first authors. Handling Editor: Professor Michael Makris The online version contains supplementary material available at https://doi.org/10.1016/j.rpth.2026.106925. Contributor Information Zhu Zhang, Email: zhuperfectlife@126.com. Zhenguo Zhai, Email: zhaizhenguo2011@126.com. Supplementary material Supplementary Material References Associated Data Supplementary Materials Supplementary Material Data Availability Statement For original data and code, please email the corresponding author.