Synergistic Interaction Between PEAR1-FGB and the Dominant Role of ABCB1 in Thrombosis Risk: A SHAP and MDR-Based Study in Vietnamese Pedigrees Authors: Hai Bui Hoang, Trang Thi Nguyen, Anh Ngoc Thi Le, Dan Hoang Le, Huy Quang Do, Sang Thanh Nguyen, Ha Thu Thi To, Thuy Thu Thi Nguyen, Hanh Minh Thi Nguyen Published: 2026 ABSTRACT AbstractBackgroundThrombosis is a multifactorial disorder with a substantial genetic component, yet gene-gene interactions contributing to thrombotic risk remain insufficiently characterized in Southeast Asian populations. Identifying both individual genetic determinants and interaction effects may improve risk stratification, particularly in families with early-onset thrombotic disease.ObjectiveThis study aimed to explore individual and interactive genetic contributions to thrombotic events in Vietnamese pedigrees with early-onset venous thromboembolism (VTE) using an integrated analytical framework combining Multifactor Dimensionality Reduction (MDR) and SHapley Additive exPlanations (SHAP).MethodsA pedigree-based study was conducted among 88 individuals from families with a history of early-onset VTE (<50 years), including 19 affected cases. Genetic variants were identified using next-generation sequencing. Random Forest modeling with SHAP was used to quantify global and local feature contributions, while MDR was applied to detect potential non-linear gene–gene interactions beyond single-locus effects. The Random Forest model was not explicitly adjusted for familial clustering or relatedness; results were therefore interpreted as exploratory.ResultsSHAP analysis identified rs1045642 (ABCB1) as the strongest contributor to model-predicted baseline thrombotic risk. MDR analysis suggested that this background susceptibility may be further modified by a two-locus interaction between rs12041331 (PEAR1) and rs1800790 (FGB). Individuals classified within the high-risk MDR genotype combination showed increased odds of thrombosis (OR = 8.00, 95% CI: 1.67–74.96; p = 0.0028), compared with the corresponding single-locus model (OR = 3.73, 95% CI: 0.93–21.48; p = 0.0405). Confidence intervals were wide, reflecting limited precision.ConclusionIn this exploratory pedigree-based study, rs1045642 (ABCB1) was associated with baseline thrombotic susceptibility, while a potential interaction between PEAR1 and FGB may further modulate risk in specific genotype combinations. The integration of MDR and SHAP provides an interpretable framework for investigating complex genetic architectures. However, given the small sample size and limited number of events, these findings should be considered hypothesis-generating and require validation in larger independent cohorts.Keywords: venous thrombo-embolism, rs12041331, rs1800790, PEAR1 gene, FGB gene, SHAP analysis, MDR analysis Introduction Thrombosis, encompassing both venous and arterial thrombotic events, represents a major global health burden and remains one of the leading causes of cardiovascular morbidity and mortality worldwide.1 Recent epidemiological data indicate that thrombotic disorders are increasingly diagnosed in younger individuals, suggesting a shifting demographic pattern. Although advances in prophylaxis and treatment have improved outcomes, the overall burden of thrombosis remains substantial, particularly in cases of early-onset disease.2 Of particular concern is evidence from the United States and Europe highlighting a rising incidence among young adults (under 50 years), a population traditionally considered low-risk.3 Unlike elderly cohorts, where VTE is typically secondary to acquired triggers such as surgery, malignancy, or prolonged immobilization, VTE in younger individuals is frequently unprovoked or strongly linked to genetic predisposition. The clinical impact on this demographic is profound due to high recurrence rates and the requirement for extended anticoagulation, which significantly affects both quality of life and workforce productivity. Recent literature suggests that hereditary thrombophilia was detected in 50% of VTE patients younger than 20 and in 21.8% of those over the age of 70.4 The familial aggregation of thrombosis provides robust evidence for its underlying genetic architecture. Cohort studies have confirmed that the incidence of VTE among siblings was 2.2 cases per 1000 person-years, representing a relative risk of 3.08 as compared with the general population.5 Although VTE is highly heritable, with heritability estimates ranging from 35% to 60%, established monogenic variants-such as Factor V Leiden and Prothrombin G20210A-account for only a small fraction (approximately 5%) of the overall genetic risk.6 In addition, arterial thrombotic diseases such as coronary artery disease are also clearly influenced by genetic factors. Studies have shown that genetic variants encoding lipoproteins related to thrombosis formation are associated with the risk of developing coronary artery disease.7 Attention has increasingly shifted toward genes involved in platelet activation, coagulation pathways, and pharmacogenetic regulation. Inherited thrombophilia may arise from genetic variants affecting both procoagulant and anticoagulant pathways. Gain-of-function variants in procoagulant genes, such as F2 and F5, can enhance thrombin generation and promote clot formation. Conversely, loss-of-function variants in natural anticoagulant genes, including PROC, PROS1, SERPINC1, and THBD, may impair physiological anticoagulant mechanisms, thereby increasing susceptibility to venous thromboembolism. Additional genetic factors involved in platelet activation, fibrin formation, homocysteine metabolism, and endothelial regulation have also been implicated in thrombosis risk. The PEAR1 gene serves as a critical regulator of platelet aggregation signaling through cell-cell interactions. While certain PEAR1 single nucleotide polymorphisms (SNPs) have been shown to enhance platelet aggregability, individual studies have yielded inconsistent associations with clinical thrombotic events.8,9 Meanwhile, the FGB (Beta-Fibrinogen) gene encodes the fibrinogen beta chain, and variants within its promoter region (eg, −455G>A) directly influence plasma fibrinogen concentrations and fibrin network architecture.10 However, prior research has largely focused on the isolated effects of these SNPs. Preliminary evidence suggests that the interplay between PEAR1-mediated platelet signaling and FGB-driven fibrin stability may increase thrombotic risk. In addition to coagulation and platelet-related genes, polymorphisms in drug transporter genes such as ABCB1 (rs1045642) may influence thrombotic risk indirectly by modulating the pharmacokinetics of anticoagulant and antiplatelet agents.11 Variability in drug transport may affect therapeutic efficacy and, consequently, clinical outcomes in genetically predisposed individuals. Elucidating these gene-gene interactions is essential not only to address the “missing heritability” problem but also to explain why certain carriers of risk variants remain asymptomatic unless a complementary genetic modifier is present.12 Detecting non-linear gene-gene interactions in complex pathologies such as thrombosis remains a formidable challenge for conventional statistical frameworks, including logistic regression, which may have limited power to capture high-order interactions and can become unstable as dimensionality increases. To address these limitations, Multifactor Dimensionality Reduction (MDR) has emerged as a robust, non-parametric alternative capable of identifying high-risk genetic combinations without prior assumptions regarding the underlying inheritance model. Meanwhile SHapley Additive exPlanations (SHAP) allow quantification of the individual contribution of each SNP within a predictive model and facilitate visualization of how interactions between genes such as PEAR1 and FGB may influence predicted thrombotic risk. The integration of MDR and SHAP provides an interpretable framework for exploring complex genetic architectures. While sporadic studies have utilized this dual approach in cardiovascular research between 2024 and 2025, its application to specifically investigate the PEAR1-FGB interaction in thrombosis remains limited. In light of the escalating thrombosis incidence among younger individuals and the emergence of novel risk factors, identifying genetic markers may contribute to improved risk stratification. This is particularly vital for idiopathic or recurrent cases in which standard thrombophilia screening often yields negative results.13 The present study aims to investigate thrombosis-associated genetic variants in Vietnamese pedigrees with early-onset VTE, with a specific focus on evaluating the interaction between PEAR1 and FGB variants using a combined MDR and SHAP analytical framework. Materials and Methods Participants This study included 88 members from three generations within four families, selected based on the following criteria: Inclusion Criteria +Family members with a blood or marital relationship (spouse). Family members with a blood or marital relationship (spouse). +Families were eligible if at least two biologically related members had been diagnosed with venous thromboembolism (VTE) before the age of 50, a criterion used to enrich the cohort for potential hereditary thrombophilia. Families were eligible if at least two biologically related members had been diagnosed with venous thromboembolism (VTE) before the age of 50, a criterion used to enrich the cohort for potential hereditary thrombophilia. Exclusion Criteria +Individuals with one or more of the following conditions: -Hematological disorders: Leukemia, polycythemia, lymphoma.-Cardiovascular diseases: Valvular heart disease (mechanical or prosthetic valves), endocarditis, vasculitis, varicose veins.-Atherosclerosis or other vascular diseases such as stenting, and venous mesh.-Nephrotic syndrome. Individuals with one or more of the following conditions: -Hematological disorders: Leukemia, polycythemia, lymphoma.-Cardiovascular diseases: Valvular heart disease (mechanical or prosthetic valves), endocarditis, vasculitis, varicose veins.-Atherosclerosis or other vascular diseases such as stenting, and venous mesh.-Nephrotic syndrome. -Hematological disorders: Leukemia, polycythemia, lymphoma. Hematological disorders: Leukemia, polycythemia, lymphoma. -Cardiovascular diseases: Valvular heart disease (mechanical or prosthetic valves), endocarditis, vasculitis, varicose veins. Cardiovascular diseases: Valvular heart disease (mechanical or prosthetic valves), endocarditis, vasculitis, varicose veins. -Atherosclerosis or other vascular diseases such as stenting, and venous mesh. Atherosclerosis or other vascular diseases such as stenting, and venous mesh. -Nephrotic syndrome. Nephrotic syndrome. +Members who do not consent to participate or cannot provide a blood sample for testing. Members who do not consent to participate or cannot provide a blood sample for testing. +Currently or previously participating in other studies that may affect genetic test results. Currently or previously participating in other studies that may affect genetic test results. +Do not meet the technical requirements of the study. Individuals whose biological samples did not meet laboratory quality-control requirements for genetic analysis, including insufficient DNA quantity or quality, or inadequate sequencing quality metrics. Do not meet the technical requirements of the study. Individuals whose biological samples did not meet laboratory quality-control requirements for genetic analysis, including insufficient DNA quantity or quality, or inadequate sequencing quality metrics. Study Procedures Participant Identification -Review the list of patients under 50 years old diagnosed with VTE at Hanoi Medical University Hospital. Review the list of patients under 50 years old diagnosed with VTE at Hanoi Medical University Hospital. -Explore family history to identify high-risk pedigrees. Explore family history to identify high-risk pedigrees. Data Collection -Conduct direct interviews to collect information on personal and family medical history. Conduct direct interviews to collect information on personal and family medical history. -Perform physical examinations and record information on thrombosis treatment. Perform physical examinations and record information on thrombosis treatment. Blood Sampling and Gene Sequencing -Collect venous blood samples from all eligible study members. Collect venous blood samples from all eligible study members. -Utilize Next-Generation Sequencing (NGS) technology to analyze a panel of 34 genes associated with thrombosis risk, selected from international databases (ClinVar, Varsome, OMIM). Utilize Next-Generation Sequencing (NGS) technology to analyze a panel of 34 genes associated with thrombosis risk, selected from international databases (ClinVar, Varsome, OMIM). Evaluation and Analysis -Analyze genetic variants in two main groups: Analyze genetic variants in two main groups: +Group 1: Genes associated with thrombosis susceptibility. These included genes involved in coagulation pathways, platelet function, endothelial regulation, and anticoagulant mechanisms. Gain-of-function variants in procoagulant genes, such as F2 and F5, may enhance thrombin generation and increase thrombosis risk, whereas loss-of-function variants in natural anticoagulant genes, including PROC, PROS1, SERPINC1, and THBD, may impair physiological anticoagulant pathways and predispose individuals to venous thromboembolism. Additional genes related to fibrin formation (FGB, FGG, F13A1), platelet activation (ITGA2B, ITGB3), homocysteine metabolism (MTHFR, MTR, MTRR), and vascular biology (BMP2, ACVRL1, CAV1) were also evaluated. Group 1: Genes associated with thrombosis susceptibility. These included genes involved in coagulation pathways, platelet function, endothelial regulation, and anticoagulant mechanisms. Gain-of-function variants in procoagulant genes, such as F2 and F5, may enhance thrombin generation and increase thrombosis risk, whereas loss-of-function variants in natural anticoagulant genes, including PROC, PROS1, SERPINC1, and THBD, may impair physiological anticoagulant pathways and predispose individuals to venous thromboembolism. Additional genes related to fibrin formation (FGB, FGG, F13A1), platelet activation (ITGA2B, ITGB3), homocysteine metabolism (MTHFR, MTR, MTRR), and vascular biology (BMP2, ACVRL1, CAV1) were also evaluated. +Group 2: Genes affecting the efficacy and safety of anticoagulant and antiplatelet therapies. These included pharmacogenetic genes involved in drug metabolism, transport, and therapeutic response. Variants in VKORC1 and CYP2C9 may influence individual sensitivity to vitamin K antagonists, particularly warfarin, by affecting drug target activity and metabolic clearance. Variants in platelet receptor genes, such as P2RY12, may alter responsiveness to antiplatelet agents and contribute to interindividual variability in treatment outcomes. Group 2: Genes affecting the efficacy and safety of anticoagulant and antiplatelet therapies. These included pharmacogenetic genes involved in drug metabolism, transport, and therapeutic response. Variants in VKORC1 and CYP2C9 may influence individual sensitivity to vitamin K antagonists, particularly warfarin, by affecting drug target activity and metabolic clearance. Variants in platelet receptor genes, such as P2RY12, may alter responsiveness to antiplatelet agents and contribute to interindividual variability in treatment outcomes. -Statistical analysis: Multifactor Dimensionality Reduction (MDR) was applied to identify gene–gene interaction models associated with venous thromboembolism risk. The choice of MDR over traditional parametric approaches, such as logistic regression, was based on its capacity to detect potential high-order interactions in modest datasets without requiring specification of interaction terms. As a non-parametric and model-free framework, MDR does not require prior assumptions regarding the underlying genetic architecture or mode of inheritance, effectively mitigating the risk of model misspecification. Given the pedigree-based design, family clustering was considered in the interpretation of results; however, no explicit variance correction for intra-family correlation was applied, and findings should therefore be interpreted cautiously. The stability and reproducibility of the identified interaction models were further ensured through 10-fold cross-validation (CV) and the assessment of CV consistency. Shapley Additive exPlanations (SHAP) were subsequently employed to quantify and interpret the contribution and interaction effects of individual variants within a Random Forest framework, providing an additional layer of interpretability to the non-linear genetic associations. Because both Random Forest and MDR are flexible, high-capacity methods applied to a modest sample, we took several steps to limit overfitting and to keep the analysis exploratory. The Random Forest was grown with a restricted number of features considered at each split and a limited tree depth, and its outputs were used only to rank the relative contribution of variants (via SHAP) rather than to generate calibrated individual risk predictions. For MDR, the number of candidate loci entered into the interaction search was deliberately constrained to variants with prior biological plausibility, and every model was evaluated by 10-fold cross-validation with reporting of cross-validation consistency to penalise combinations that did not reproduce across folds. We did not perform independent external validation; consequently, the balanced-accuracy and odds-ratio estimates should be regarded as internal, optimistic bounds that require confirmation in larger cohorts. Statistical analysis: Multifactor Dimensionality Reduction (MDR) was applied to identify gene–gene interaction models associated with venous thromboembolism risk. The choice of MDR over traditional parametric approaches, such as logistic regression, was based on its capacity to detect potential high-order interactions in modest datasets without requiring specification of interaction terms. As a non-parametric and model-free framework, MDR does not require prior assumptions regarding the underlying genetic architecture or mode of inheritance, effectively mitigating the risk of model misspecification. Given the pedigree-based design, family clustering was considered in the interpretation of results; however, no explicit variance correction for intra-family correlation was applied, and findings should therefore be interpreted cautiously. The stability and reproducibility of the identified interaction models were further ensured through 10-fold cross-validation (CV) and the assessment of CV consistency. Shapley Additive exPlanations (SHAP) were subsequently employed to quantify and interpret the contribution and interaction effects of individual variants within a Random Forest framework, providing an additional layer of interpretability to the non-linear genetic associations. Because both Random Forest and MDR are flexible, high-capacity methods applied to a modest sample, we took several steps to limit overfitting and to keep the analysis exploratory. The Random Forest was grown with a restricted number of features considered at each split and a limited tree depth, and its outputs were used only to rank the relative contribution of variants (via SHAP) rather than to generate calibrated individual risk predictions. For MDR, the number of candidate loci entered into the interaction search was deliberately constrained to variants with prior biological plausibility, and every model was evaluated by 10-fold cross-validation with reporting of cross-validation consistency to penalise combinations that did not reproduce across folds. We did not perform independent external validation; consequently, the balanced-accuracy and odds-ratio estimates should be regarded as internal, optimistic bounds that require confirmation in larger cohorts. -Outcome: The primary analytical outcome was venous thromboembolism (VTE) status (present/absent). Arterial thrombotic events, when present, were descriptively reported but were not included as primary endpoints in the MDR and Random Forest modeling. VTE status served as the dependent variable in MDR and Random Forest analyses. Outcome: The primary analytical outcome was venous thromboembolism (VTE) status (present/absent). Arterial thrombotic events, when present, were descriptively reported but were not included as primary endpoints in the MDR and Random Forest modeling. VTE status served as the dependent variable in MDR and Random Forest analyses. Research Ethics This study was approved by the Ethics Committee of Hanoi Medical University (Approval No. 8014/QD-DHYHN, dated October 23, 2025). All procedures performed in this study involving human participants were conducted in accordance with the ethical standards of the institutional research committee and with the Declaration of Helsinki. Informed consent was obtained from all individual participants included in the study. Results The baseline demographic and clinical characteristics of the study population are summarized in Table 1. Data are presented as mean ± standard deviation (SD), median and range, or number (percentage). The number of thrombotic events exceeds the number of patients because some patients experienced multiple hospitalizations. Table 1Baseline Characteristics of the Study Population (N = 19)CharacteristicValueAge, yearsMean ± SD48.1 ± 16.9Median (IQR)47 (-)Range (min-max)18-87Age (Patient group), yearsMean ± SD48.11 ± 16.9Median (IQR)47Range (min-max)18 - 87Age at onset (Patient group), yearsMean ± SD32.4 ± 19.6Median (IQR)28.5 (-)Range (min-max)6-75Sex, n (%)Male11 (57.9)Female8 (42.1)Type of thrombosisVenous thrombosis, n (%)30 (68.2)Arterial thrombosis, n (%)14 (31.8) Baseline Characteristics of the Study Population (N = 19) Comment: +The study population included individuals aged between 2 and 95 years, with a mean age of 40.08 years. Participants in the 40–60-year age group accounted for the highest proportion of the study population. The study population included individuals aged between 2 and 95 years, with a mean age of 40.08 years. Participants in the 40–60-year age group accounted for the highest proportion of the study population. +The patient group ranged in age from 18 to 87 years, with a median age of 47 years. The age at disease onset, as documented in medical records, ranged from 6 to 75 years, with a median onset age of 28.5 years. The patient group ranged in age from 18 to 87 years, with a median age of 47 years. The age at disease onset, as documented in medical records, ranged from 6 to 75 years, with a median onset age of 28.5 years. +The patient group consisted of 11 males (57.9%) and 8 females (42.1%). The patient group consisted of 11 males (57.9%) and 8 females (42.1%). +Analysis of hospitalization frequency by thrombotic site, showed that venous thrombosis accounted for 68.2% (30 episodes), whereas arterial thrombosis accounted for 31.8% (14 episodes). The most common venous thrombotic sites were deep veins of the lower extremities, mesenteric veins, cerebral veins and pulmonary arteries. The most frequently affected arterial sites was cerebral arteries. Analysis of hospitalization frequency by thrombotic site, showed that venous thrombosis accounted for 68.2% (30 episodes), whereas arterial thrombosis accounted for 31.8% (14 episodes). The most common venous thrombotic sites were deep veins of the lower extremities, mesenteric veins, cerebral veins and pulmonary arteries. The most frequently affected arterial sites was cerebral arteries. Comment: The identified genetic variants detected in the study cohort are summarized in Table 2. Sequencing analysis was performed for all participants in the study cohort. Table 2 summarizes the identified genetic variants and the corresponding frequencies of homozygous and heterozygous genotypes in the study population. Table 2Identified Genetic VariantsGeneVariantConsequanceChromosomal PositiondbSNPHomozygous Cases (n)Heterozygous Cases (n)F5NC_000001.10:g.169524538T>CMissenseChr1:169555300rs11820390502F7NC_000013.11:g.113118845G>AMissensechr13:113118845rs604601FGBNC_000004.12:g.154562863C>T5′UTR regionchr4:154562863rs18007871029NC_000004.12:g.154562556G>A5′UTR regionchr4:154562556rs18007901030THBDNC_000020.10:g.23028724G>AMissensechr20:23048087rs1042579427NC_000020.11:g.23049706C>T5′UTR regionchr20:23049706rs13306848019MTRNC_000001.10:g.237048500A>GMissensechr1:236885200rs180508707MTRRNC_000005.9:g.7870973A>GMissensechr5:7870860rs1801394330MTHFRNC_000001.10:g.11854476T>GMissensechr1:11794419rs1801131331NC_000001.10:g.11856378G>AMissensechr1:11796321rs1801133227VKORC1NC_000016.9:g.31104878G>AIntron variantchr16:31093557rs99344385410NC_000016.9:g.31107689C>T5′UTR regionchr16:31096368rs99232315410CYP2C9NC_000010.11:g.94981296A>CMissensechr10:94981296rs105791004P2Y12NC_000003.11:g.151056598A>CSynonymouschr3:151338810rs68096994513NC_000003.11:g.151057642G>AIntron variantchr3:151339854rs20469344610NC_000003.11:g.151099741A>GIntron variantchr3:151381953rs6787801633ABCG2NC_000004.11:g.89052323G>TStop gainedchr4:88131171rs223114214CYP2C19NC_000010.10:g.96540410G>AStop gainedchr10:94780653rs498689306NC_000010.10:g.96541616G>ASynonymouschr10:94781859rs4244285524ABCB1NC_000007.13:g.87138645A>GMissensechr7:87509329rs10456422233PEAR1NC_000001.10:g.156869714G>AIntron variantchr1:156899922rs120413312519ITGA2BNC_000017.10:g.42453065A>CMissensechr17:44375697rs59112019BMP2NC_000020.10:g.7106289T>C_chr20:7125642rs18843021232PROS1NC_000003.12:g.93886430G>TMissenseChr3: 93886430rs19946949502SERPINC1NC_000001.11:g.173914796G>TStop gainedchr1:g.173914796G>T–02 Identified Genetic Variants Comment: SHAP analysis identified rs1045642 as the most influential genetic variant in predicting thrombotic risk (Figure 1). The results indicate that rs1045642 has the strongest influence on thrombotic risk prediction, with higher allele values predominantly associated with negative SHAP values. Other variants, including rs12041331, rs1042579, and rs1801131, also contribute substantially to the model, with higher allele values of rs1042579 and rs1801131 tending to show positive SHAP values. Several additional variants exhibit moderate effects with a wide SHAP value distribution, whereas variants with lower importance show SHAP values clustered around zero, indicating a minimal contribution to the overall model performance. SHAP beeswarm plot illustrating the relative contribution of individual genetic variants to model-predicted thrombotic risk. Each point represents one individual; the horizontal position indicates the SHAP value (direction and magnitude of the contribution to the predicted risk) and the colour encodes the allele value of the variant. Variants are ordered by mean absolute SHAP value, with rs1045642 (ABCB1) showing the strongest overall contribution. -Comments: The SHAP interaction analysis between rs12041331 and rs1800790 is presented in Figure 2. Comments: The SHAP interaction analysis between rs12041331 and rs1800790 is presented in Figure 2. +Through SHAP interaction analysis between the two variants rs12041331 and rs1800790, the results indicate that the SHAP interaction values are strongly dependent on the combined genotypes of these two variants. Among individuals who are homozygous for the rs1800790 variant allele, SHAP interaction values are predominantly positive and of greater magnitude (≈ 0.025–0.032), particularly when rs12041331 is in the heterozygous or homozygous variant state. Through SHAP interaction analysis between the two variants rs12041331 and rs1800790, the results indicate that the SHAP interaction values are strongly dependent on the combined genotypes of these two variants. Among individuals who are homozygous for the rs1800790 variant allele, SHAP interaction values are predominantly positive and of greater magnitude (≈ 0.025–0.032), particularly when rs12041331 is in the heterozygous or homozygous variant state. +In contrast, among rs1800790 heterozygotes, when the variant allele of rs12041331 is absent, SHAP interaction values are clearly concentrated in the negative range (≈ −0.04). In individuals without the rs1800790 variant, SHAP interaction values generally fluctuate around zero; the combination lacking both variants shows a slight positive value (≈ 0.015), whereas the homozygous rs12041331 genotype is associated with lower interaction values (≈ −0.01). In contrast, among rs1800790 heterozygotes, when the variant allele of rs12041331 is absent, SHAP interaction values are clearly concentrated in the negative range (≈ −0.04). In individuals without the rs1800790 variant, SHAP interaction values generally fluctuate around zero; the combination lacking both variants shows a slight positive value (≈ 0.015), whereas the homozygous rs12041331 genotype is associated with lower interaction values (≈ −0.01). -Comments: The results of the SHAP feature importance analysis are presented in Table 3. Comments: The results of the SHAP feature importance analysis are presented in Table 3. +One-locus model (rs12041331): In the MDR analysis, the 1-locus model involving rs12041331 achieved a balanced accuracy of 0.6505 in the training set and 0.5926 in the testing set, with a cross-validation (CV) consistency of 9/10. These results indicate that rs12041331 provides a moderate and relatively stable classification performance. One-locus model (rs12041331): In the MDR analysis, the 1-locus model involving rs12041331 achieved a balanced accuracy of 0.6505 in the training set and 0.5926 in the testing set, with a cross-validation (CV) consistency of 9/10. These results indicate that rs12041331 provides a moderate and relatively stable classification performance. +Two-locus model (rs12041331 × rs1800790): When rs1800790 was incorporated into the model, the balanced accuracy increased substantially to 0.757 in the training set and 0.6897 in the testing set, while the CV consistency remained 9/10. Two-locus model (rs12041331 × rs1800790): When rs1800790 was incorporated into the model, the balanced accuracy increased substantially to 0.757 in the training set and 0.6897 in the testing set, while the CV consistency remained 9/10. +According to MDR-based risk stratification, in the 1-locus model involving rs12041331, the high-risk group included 17 cases of thrombosis, compared with 3 cases in the low-risk group, corresponding to an odds ratio (OR) of 3.73 (95% CI: 0.93–21.48; p = 0.0405). According to MDR-based risk stratification, in the 1-locus model involving rs12041331, the high-risk group included 17 cases of thrombosis, compared with 3 cases in the low-risk group, corresponding to an odds ratio (OR) of 3.73 (95% CI: 0.93–21.48; p = 0.0405). +When extending to the 2-locus interaction model combining rs12041331 and rs1800790, the odds ratio increased markedly to 8.0, with a more stable confidence interval (95% CI: 1.67–74.96; p = 0.0028). In this interaction model, the majority of thrombotic events (18 of 20 cases) were classified into the high-risk group, whereas only 2 events were observed in the low-risk group. Confidence intervals remained wide, reflecting limited precision due to sample size. The very small number of events in the low-risk stratum means that the point estimate of OR = 8.00 is statistically unstable, and its wide 95% confidence interval (1.67–74.96) should be interpreted as indicating the direction of the association rather than its precise magnitude. When extending to the 2-locus interaction model combining rs12041331 and rs1800790, the odds ratio increased markedly to 8.0, with a more stable confidence interval (95% CI: 1.67–74.96; p = 0.0028). In this interaction model, the majority of thrombotic events (18 of 20 cases) were classified into the high-risk group, whereas only 2 events were observed in the low-risk group. Confidence intervals remained wide, reflecting limited precision due to sample size. The very small number of events in the low-risk stratum means that the point estimate of OR = 8.00 is statistically unstable, and its wide 95% confidence interval (1.67–74.96) should be interpreted as indicating the direction of the association rather than its precise magnitude. SHAP interaction plot for rs12041331 (PEAR1) × rs1800790 (FGB). The vertical axis shows the SHAP interaction value, and points are stratified by the combined genotypes of the two variants. Positive, high-magnitude interaction values are concentrated among individuals homozygous for the rs1800790 variant allele who also carry the rs12041331 variant allele, indicating a context-dependent (non-additive) contribution to predicted thrombotic risk. Table 3MDR Analysis of Single-Locus and Two-Locus Genetic Models for Thrombotic RiskMDR ModelGenetic VariantsBal. Acc. (Training)Bal. Acc. (Testing)CV ConsistencyRisk GroupThrombosis (n)No Thrombosis (n)OR (95% CI)p1-locusrs120413310.65050.59269/10Low-risk3271 (Ref)–High-risk17413.73 (0.93–21.48)0.04052-locusrs12041331 × rs18007900.75700.68979/10Low-risk2321 (Ref)–High-risk18368.00 (1.67–74.96)0.0028 MDR Analysis of Single-Locus and Two-Locus Genetic Models for Thrombotic Risk Discussion General Characteristics off the Study Population The study population exhibited a wide age range (2–95 years), with a mean age of approximately 40 years. The highest proportion of participants was observed in the 40–60-year age group, which is consistent with the current demographic structure in Vietnam - the majority of the population is in middle adulthood.14 Although the overall sex distribution was relatively balanced, males predominated among patients with clinically manifest thrombosis. Previous studies have reported sex-related differences in thrombotic risk, with men experiencing a higher incidence of both venous and arterial thrombosis. In the case of venous thromboembolism, men also have and a greater risk of recurrence after exclusion of female-specific risk factors such as pregnancy and oral contraceptive use.15,16 This observation implies that, in our cohort, the age and sex distribution of the disease group are broadly consistent with the epidemiological characteristics of thrombosis events. Among the 66 participants included in the study, 19 individuals (28.8%) had a documented history of venous thrombosis, whereas 47 individuals (71.2%) remained clinically asymptomatic at the time of enrollment. Among affected patients, the mean age was higher than that of the overall population, which is consistent with previous studies showing that thrombosis predominantly affects older individuals.17,18 However, the median age at disease onset was below 30 years, suggesting that hereditary risk factors may be present long before clinical manifestations become apparent. This discrepancy indicates a prolonged latent phase preceding clinical manifestation and supports the concept that underlying risk factors, potentially inherited, may be present long before thrombosis becomes clinically evident. Previous studies on Factor V Leiden, the prothrombin G20210A mutation, and polymorphisms in genes related to IL-6 signaling, fibrinogen, and platelet activation have consistently demonstrated strong associations with thrombosis occurring at a young age.19 Regarding thrombotic localization, venous thrombosis was markedly more frequent than arterial thrombosis, accounting for more than two-thirds of documented events. This pattern is in line with established knowledge that inherited thrombophilic factors predominantly predispose to venous rather than arterial thrombosis.20 Arterial thrombosis is much more commonly associated with lifestyle and metabolic risk factors such as hypertension, smoking, and diabetes.18 The frequent involvement of deep veins of the lower extremities, pulmonary arteries, mesenteric veins, and cerebral veins reflects the systemic nature of the prothrombotic tendency. Although less common, arterial thrombosis was also observed, mainly affecting the cerebral arteries, suggesting that in certain individuals, complex interactions between genetic predisposition and inflammatory or atherosclerotic mechanisms may contribute to arterial events.21 SHAP-Based Importance of Genetic Variants in Thrombotic Risk Prediction The variant rs1045642 was identified as the largest contributor to the SHAP predictive model. This SNP is located in the ABCB1 (MDR1) gene, which encodes P-glycoprotein, a key membrane transporter involved in the absorption, distribution, and excretion of a wide range of drugs as well as endogenous substrates.11,22 The prominent contribution of rs1045642 raises the possibility that variability in membrane transport and substrate handling may indirectly influence thrombotic risk, potentially through modulation of drug exposure or treatment response, including direct oral anticoagulants such as rivaroxaban. Nevertheless, the present findings are derived from statistical modeling and do not establish a direct biological mechanism linking rs1045642 to venous thromboembolism (VTE). Further investigation in larger, independent cohorts, as well as functional studies, will be necessary to determine whether this association reflects a true biological effect or is influenced by sample size and cohort-specific factors. Ranked second, rs12041331 emerged as the most important variant in the 1-locus MDR model and also exhibited a high SHAP value. This SNP is located in the PEAR1 gene, which encodes a platelet transmembrane platelet receptor with tyrosine kinase activity that participates in platelet-platelet contact and aggregation and is expressed in endothelial cells, platelets, and several other tissues. Genetic variants in PEAR1 have long been hypothesized to contribute to thrombotic risk based on their physiological role in platelet aggregation.23,24 However, direct evidence supporting this hypothesis remains limited. In particular, the PEAR1 variant rs12041331 has not been consistently associated with thrombotic events in previous studies. Instead, its reported effects have primarily involved pharmacogenetic interactions with antiplatelet agents, whereby carriers of this variant exhibit lower PEAR1 protein expression and reduced platelet aggregation responses to various antiplatelet drugs.25,26 The variant rs1042579, located within the coding region of the THBD gene, affects the function of thrombomodulin (TM), an essential endogenous anticoagulant.27 In addition, rs6787801, located in the gene encoding the P2Y12 receptor, and rs5911 in ITGA2B, which encodes glycoprotein IIb-an important platelet receptor involved in von Willebrand factor-mediated adhesion-are both associated with platelet activation, adhesion, and aggregation.28,29 All three SNPs have been implicated in coagulation-related pathways, contributing to increased thrombotic risk and disease pathogenesis. The involvement of multiple SNPs affecting coagulation highlights that the disease model is not driven by a single determinant but rather reflects the synergistic effects of multiple genetic factors related to thrombosis. Moreover, the presence of PEAR1, platelet glycoprotein-encoding genes, and P2Y12 receptor-related genes among the SNPs with high SHAP values suggests platelet-related mechanisms in increasing thrombotic risk among affected individuals. Notably, the variant rs1801131 in the MTHFR gene, which encodes a key enzyme in folate and homocysteine metabolism, also showed a substantial contribution. Previous studies have demonstrated that elevated plasma homocysteine (Hcy) levels can induce endothelial damage, thereby increasing the risk of atherosclerosis and representing an established risk factor for vascular disease and thrombosis.30 Accordingly, the influence of this SNP within affected pedigrees provides supporting evidence for the link between metabolic factors and coagulation disorders within the predictive model. In addition, rs4244285 in the CYP2C19 gene, a major drug-metabolizing enzyme, exhibited a measurable impact on the model. CYP2C19 is involved in the metabolic activation of the antiplatelet agent clopidogrel, thereby influencing treatment efficacy, particularly in patients requiring coronary revascularization and stent implantation.31 This finding suggests that genetic variability in drug metabolism-especially for antiplatelet agents-may adversely affect clinical outcomes in variant carriers by reducing therapeutic effectiveness and compromising post-intervention prophylaxis. Finally, several variants located in regulatory regions or with currently unclear biological significance (RCV001724979, RCV001619021, RCV000017814) still demonstrated notable contributions to the model.32–34 However, given the limited sample size, the observed associations involving these genes may be due to random variation. This observation indicates the potential presence of additional, as-yet-uncharacterized genetic variants influencing disease susceptibility within the pedigree. Consequently, further studies are warranted to elucidate the functional roles of these genomic regions. Gene-Gene Interaction Between rs12041331 and rs1800790 Identified by SHAP Analysis The FGB gene encodes the β-chain of fibrinogen, a key component of the soluble fibrinogen complex. The variant rs1800790 (also known as −455G>A), located in the promoter region, has been identified as a functional polymorphism capable of modulating transcriptional activity and increasing circulating fibrinogen levels.35 Elevated fibrinogen concentrations have been hypothesized to increase the risk of thrombosis events, and the link between this variant and arterial thrombosis has been established; however, it is not yet consistent and large epidemiological studies and meta-analyses have yielded conflicting results: although this variant is strongly associated with plasma fibrinogen levels and has been linked to thrombogenesis, it has not been consistently demonstrated to be an independent predictor of venous thromboembolism or myocardial infarction across different populations.36–38 These findings suggest that rs1800790 may not act as a direct causal variant. Rather, its contribution to thrombotic risk may become apparent only in the presence of synergistic interactions with other genetic alterations in platelet activation pathways or with environmental and inflammatory triggers that disrupt the balance of the coagulation-hemostatic system. This also explains the absence of this variant in the above SHAP analysis for individual variant contributions to thrombotic events. Despite growing interest in the individual roles of rs12041331 and rs1800790, no previous studies have specifically examined the interaction between these two variants in relation to thrombotic risk. Most prior investigations have focused on single-variant analyses or linear multivariable models that assume additive and independent effects. In contrast, thrombosis is well recognized as a multifactorial and polygenic condition, in which disease onset may result from the convergence of transient triggers (such as pregnancy or surgery) with established risk factors including age, comorbidities, and inherited prothrombotic variants. Moreover, approximately 50–60% of the variability in thrombotic disease susceptibility has been attributed to genetic factors.39–42 Under such complex etiological conditions, linear analytical approaches may overlook biologically meaningful interactions between pathways involved in thrombus formation. In this context, SHAP interaction analysis provides a powerful framework for exploring context-dependent genetic effects within nonlinear machine-learning models. Examination of the interaction between rs12041331 and rs1800790 in the study population revealed that the contribution of rs12041331 to predicted thrombotic risk was strongly dependent on the rs1800790 genotype. Notably, individuals homozygous for the rs1800790 variant allele exhibited positive and high-magnitude SHAP interaction values, particularly when rs12041331 was present in the heterozygous or homozygous variant state, suggesting a potential synergistic effect that amplifies risk beyond simple additive contributions. In contrast, among individuals heterozygous for rs1800790 who did not carry the rs12041331 variant allele, SHAP interaction values were predominantly negative, reflecting a context-dependent reduction in predicted risk. In individuals lacking rs1800790, interaction values clustered around zero, which implies the modulatory role of rs1800790 in shaping the phenotypic impact of rs12041331. Collectively, these findings indicate that genetic risk is not uniform across genotype combinations. However, it is critical to distinguish between SHAP interaction values and experimental effect sizes. While SHAP enables a quantitative assessment of the direction and strength of interactions at the individual genotype level within a non-linear model, it does not provide conventional effect size estimates such as odds ratios (OR) or confidence intervals. Instead, SHAP interaction values represent the marginal contribution of a feature pair to the model’s prediction, serving as an exploratory tool to identify potential epistasis that may be “hidden” in standard additive models. From a mechanistic perspective, the synergy between PEAR1 (rs12041331) and FGB (rs1800790) highlights a potential convergence between platelet activation and the plasma coagulation cascade. While PEAR1 is traditionally associated with antiplatelet pharmacogenetics and arterial thrombosis, its role in venous thromboembolism is increasingly recognized through the “cell-based model” of coagulation. Biologically, the PEAR1 variant (rs12041331) modulates platelet signaling and surface receptor expression, which provides the essential phospholipid scaffold required for the assembly of tenase and prothrombinase complexes. This accelerated thrombin generation occurs in tandem with the effects of the FGB variant (rs1800790), which increases the availability of circulating fibrinogen-the primary structural building block of the thrombus. We propose that the co-presence of these variants may creates a feed-forward amplification loop: enhanced PEAR1-mediated platelet signaling facilitates a more robust procoagulant surface, which then rapidly converts the excess fibrinogen (driven by the FGB variant) into a dense, lysis-resistant fibrin matrix. This interaction may extend beyond simple additive effects, suggesting that the stabilization of the platelet-fibrin complex is a key driver of thrombosis risk in these pedigrees, particularly in the context of early-onset disease where genetic predisposition outweighs acquired triggers. Synthesizing these findings, we propose a hierarchical risk model to reconcile the high global importance of rs1045642 (ABCB1) with the clinical impact of the PEAR1-FGB interaction. In this framework, ABCB1-related variability in drug and substrate transport may represent a foundational genetic susceptibility to thrombosis. This baseline predisposition could be further accentuated in the presence of interacting prothrombotic pathways, such as the convergence of platelet signaling (PEAR1) and fibrinogen availability (FGB). This explains why SHAP analysis identifies ABCB1 as the most influential individual predictor across the entire cohort, while the PEAR1-FGB interaction identified by MDR serves as a critical amplifier that drives the highest observed odds ratios for clinical thrombosis manifestation. Importantly, these mechanistic interpretations remain hypothetical and are inferred from statistical modeling rather than direct biological evidence. Further investigation at both molecular and clinical levels will be required to validate the proposed interaction and clarify its pathophysiological relevance. Predictive Performance and Risk Stratification of MDR Models In the context of multifactorial diseases like thrombosis, genetic risk is rarely determined by a single variant but arises from complex combinations of multiple loci. Our study demonstrates that while the 1-locus model (rs12041331) achieved a balanced accuracy of 0.6505, its predictive power remained limited, with a modest odds ratio (OR = 3.73) and a 95% CI (0.93–21.48) that approached unity. This is consistent with prior research suggesting that individual PEAR1 variants typically exert only subtle effects on clinical outcomes unless considered within a broader genetic context.23 The transition to a 2-locus MDR model (rs12041331 × rs1800790) resulted in a substantial improvement in balanced accuracy (0.757) and a markedly higher risk discrimination (OR = 8.00; 95% CI: 1.67–74.96; p = 0.0028). By collapsing complex genotype combinations into high-risk and low-risk groups, the MDR approach effectively captured the non-linear interaction that traditional regression models often overlook. The high cross-validation consistency (9/10) further supports the stability of this interaction within our study cohort. While SHAP interaction analysis identified the existence of a synergistic relationship, the MDR-based OR provides the necessary experimental effect size to quantify the clinical impact of this interaction. A critical justification for employing this integrated MDR-SHAP framework over simpler statistical approaches, such as logistic regression, lies in the detection of “phantom heritability”. As demonstrated by Zuk et al, current estimates of heritability often implicitly assume that traits involve no genetic interactions, which can lead to a significant underestimation of the explained heritability. Detecting pairwise epistasis using conventional additive models often requires prohibitively large sample sizes-potentially in the range of 500,000 individuals-because individual interaction effects are frequently much smaller and harder to detect than marginal effects.43 Consequently, synergistic effects like the PEAR1–FGB interaction are often diluted or lost in large-scale studies of unrelated individuals. By utilizing SHAP as an exploratory lens, we can identify these subtle, context-dependent genetic signals that are typically hidden in standard additive frameworks. Furthermore, the non-parametric nature of MDR allows for the identification of high-risk genotype combinations without the rigid assumptions of linearity or independence of observations, which are often violated in pedigree-based designs. We acknowledge that the reported odds ratio indicates considerable uncertainty due to the wide confidence interval. This lack of precision reflects the inherent limitations of a pedigree-based study with a modest number of affected cases (N=19), where the clustering effect might influence the generalizability of the results. Despite these constraints, the use of a pedigree-based approach was intentional; it enhances the power to uncover these non-linear components by focusing on specific genetic backgrounds where such interactions are more likely to manifest. Clinical Implications and Limitations Risk grouping based on MDR therefore has potential value as a supportive tool for genetic risk stratification, particularly in settings where conventional risk factors do not fully explain thrombotic events. While not intended to replace established screening strategies, this approach may help identify individuals who warrant closer clinical monitoring, especially young patients with unexplained thrombosis or those with a family history of thrombotic disease. Identification of genetically defined high-risk groups could inform individualized preventive counseling, including heightened awareness and management of modifiable risk factors such as smoking, immobilization, hormone use, or perioperative exposure. In high-risk clinical contexts-such as surgery, pregnancy, or estrogen therapy-MDR-based genetic risk information could contribute to a more comprehensive benefit-risk assessment.44 Importantly, these findings do not support routine genotype-based prophylactic treatment but rather emphasize personalized risk evaluation. Because a large number of variants and pairwise interactions were screened, the risk of false-positive findings from multiple testing is a genuine concern. We did not apply a formal family-wise correction (such as Bonferroni or Benjamini–Hochberg control of the false discovery rate) to the exploratory SHAP rankings, because these methods are designed to identify candidate signals rather than to confirm hypotheses; instead, we controlled for spurious associations by requiring high cross-validation consistency (9/10) for any reported MDR model and by restricting formal effect-size estimation and permutation-based significance testing to the single pre-specified PEAR1–FGB interaction. The reported p-values for this interaction (p = 0.0028) were derived from MDR permutation testing and were not further adjusted; given the multiplicity of the initial screen, all associations should be regarded as hypothesis-generating and requiring replication. The primary limitation of this study is the relatively small sample size, which increases the risk of model overfitting and limits the precision of the interaction estimates. Although we employed robust techniques such as MDR and SHAP to mitigate these risks, the results remain exploratory in nature. Furthermore, while pedigree-based studies are powerful for identifying rare variants and familial interactions, they may not be fully representative of the general population. Consequently, our findings should be viewed as a hypothesis-generating framework rather than definitive causal evidence, necessitating validation in larger, independent longitudinal cohorts before clinical applications or screening could be implemented. Conclusion Our study identifies rs1045642 (ABCB1) as the most influential individual genetic contributor within the SHAP-based predictive model for venous thromboembolism (VTE)risk in the investigated pedigrees. Furthermore, variants such as rs12041331 (PEAR1), rs1042579 (THBD), and rs1801131 (MTHFR) also demonstrated substantial contributions to the thrombotic predictive model. Our study identifies rs1045642 (ABCB1) as the most influential individual genetic contributor within the SHAP-based predictive model for venous thromboembolism (VTE)risk in the investigated pedigrees. Furthermore, variants such as rs12041331 (PEAR1), rs1042579 (THBD), and rs1801131 (MTHFR) also demonstrated substantial contributions to the thrombotic predictive model. By integrating SHAP-based interpretation with MDR analysis, we observed a potential synergistic gene-gene interaction between rs12041331 and rs1800790, with the two-locus model identifying a high-risk genotype subgroup associated with an approximately 8-fold increased odds of thrombosis (OR = 8.00; 95% CI: 1.67–74.96), exceeding the risk estimate derived from single-locus analysis. Given the modest sample size and wide confidence interval, this interaction should be interpreted cautiously. By integrating SHAP-based interpretation with MDR analysis, we observed a potential synergistic gene-gene interaction between rs12041331 and rs1800790, with the two-locus model identifying a high-risk genotype subgroup associated with an approximately 8-fold increased odds of thrombosis (OR = 8.00; 95% CI: 1.67–74.96), exceeding the risk estimate derived from single-locus analysis. Given the modest sample size and wide confidence interval, this interaction should be interpreted cautiously. These findings support the importance of considering multi-locus genetic interactions in thrombotic risk stratification. Further validation in larger, independent longitudinal cohorts is warranted before these results can be translated into clinical screening or risk assessment strategies. These findings support the importance of considering multi-locus genetic interactions in thrombotic risk stratification. Further validation in larger, independent longitudinal cohorts is warranted before these results can be translated into clinical screening or risk assessment strategies. Funding Statement This work was supported by funding from Advanced Information Technologies Corporation (project number: 1316/CV-AIT). The funding organization had no role in the study design, data collection, analysis, interpretation of data, or writing of this manuscript. Disclosure The authors report no conflicts of interest in this work. References