Smoking Interacts With APOE ε4 to Influence Dementia: Evidence From the UK Biobank and Validation by Meta-Analysis of Longitudinal Studies Authors: Yu-Lin Cong, Yong-Chang Wang, Liang-Yu Huang, Ze-Lin Yang, Lan Tan, Xiao-Hui Chu, Wei Xu Published: 2026 ABSTRACT AbstractBackgroundGene-by-environment interaction plays important roles in influencing dementia occurrence. Smoking is a well-established modifiable risk factor for all-cause dementia (ACD). However, its interaction effects with APOE ε4 as well as the underlying peripheral mechanisms are unclear.MethodsPanoramic smoking burden (PSB) and smoking history score (SHS) were constructed for 173,366 participants without dementia from the UK Biobank. The associations of PSB and SHS with ACD risk (median follow-up = 13 years) and their interactions with APOE ε4 carrier status were explored in death-competing risk models. Next, the interactions between smoking and APOE ε4 carrier status were validated by meta-analyzing epidemiological evidence searched up to May 2025. Finally, plasma proteomics, bioinformatics, and causal mediation analyses were conducted to reveal the potential mechanisms.ResultsSmoking indices were associated with ACD risk, with significant multiplicative interactions of PSB (p < .001) and SHS (p = .031) by APOE ε4. The associations of smoking with ACD were more pronounced among ε4 noncarriers. Meta-regression of data from 207,887 participants validated that the relationship between smoking and ACD was modulated by APOE ε4 status (p = .011). Current smoking was associated with ACD risk exclusively among APOE ε4 noncarriers (p < .001). Inflammation, angiogenesis modulation, and regulation of endopeptidase activity could be mediating mechanisms.ConclusionsThese findings highlighted the interplay between smoking and APOE ε4 in influencing dementia, which should be considered in risk stratification and prediction models.Keywords: All-cause dementia, APOE ε4, Cohort, Interaction, Proteomics, Smoking Smoking is an established modifiable risk factor for all-cause dementia (ACD) and its subtypes including Alzheimer’s disease (AD) and vascular dementia (VaD) (1,2). Smoking cessation and a collection of smoking-related behaviors have been widely taken into consideration for prevention and risk prediction of dementia (3, 4, 5). Recently, gene-by-environment interactions are increasingly attracting attention for precise management of dementia (6,7). The APOE ε4 allele has been linked to an increased risk of dementia (8). However, due to conflicting results, little is known about whether the relationship between smoking and dementia risk is modulated by APOE ε4 carrier status (3,9, 10, 11). There are 2 types of interaction relationships, including multiplicative and additive interaction. An additive interaction occurs when the combined effect of 2 factors (smoking and APOE ε4) on an outcome (ACD) is different from the sum of their individual effects (12). A multiplicative interaction occurs when the effect of one factor (smoking) on an outcome is significantly different in the presence or absence of another factor (APOE ε4). Understanding the interaction relationships can help gain a more nuanced understanding of how variables interact, which can be important for deepening our knowledge of dementia etiology, the weight distribution of prediction models, and policy making for dementia prevention. Moreover, there are several limitations of previous studies, which hinder our interpretation of the interaction between smoking and genetic factors. First, the definition of smoking exposure mostly focused on a single smoking behavior, such as smoking status (ever, current, or past smoking) and pack years, but ignored the joint impacts of different smoking behaviors on health outcomes (3,4,13). This simple definition makes population grouping easier but could introduce measurement bias. Second, smokers tended to have cardiovascular diseases, cardiopulmonary dysfunction, and resulting elevated mortality rates (9). Thus, the attrition bias due to death during the follow-up could also lead to misestimation of the relationship between smoking and dementia risk (9,14). Third, the peripheral biological changes induced by smoking are related to multiple chronic stress conditions, which could promote the occurrence of neurodegeneration and cognitive decline (15, 16, 17, 18). However, the peripheral mechanism linking smoking exposure to dementia has not been fully uncovered. Therefore, our primary objective in this study was to investigate the interactions between smoking, defined by 2 comprehensive smoking indices, and APOE ε4 status on dementia risk, as well as the underlying mechanisms. We started by conducting an independent cohort study from the UK Biobank (UKB), followed by a validation meta-analysis to summarize the interaction between smoking status and APOE ε4 based on previous longitudinal studies. Finally, proteomic and bioinformatic analyses were combined to provide clues about the mechanisms underlying the observed associations. Methods and Materials Study Population The UKB (http://www.ukbiobank.ac.uk) is a large-scale prospective cohort study that collected baseline phenotypic and genetic data from more than 500,000 participants ages 37 to 73 years who were recruited from 22 centers across England, Scotland, and Wales during the period 2006 to 2010 (19). Participants were followed until the first diagnosis, death, loss to follow-up, or the most recent data update. Participants with complete smoking data were included. Baseline characteristics were compared in individuals with and without complete smoking data (Supplement 1). Individuals with baseline dementia or major neuropsychiatric disorders were excluded. Finally, a sample of 173,366 participants were included for smoking history score (SHS), and a subsample of 170,720 participants were included for panoramic smoking burden (PSB). A flowchart detailing the selection of participants is presented in Supplement 2. This study adhered to the principles of the Declaration of Helsinki. Ethical approval for the UKB was granted by the North West Multicenter Research Ethics Committee, and all participants had provided electronic informed consent. Constructions of Comprehensive Smoking Indices To comprehensively assess the smoking-related exposures, we developed 2 smoking indices, namely PSB and SHS. PSB covers a broader range of smoking behaviors, while SHS measures the cumulative dose and duration of smoking exposure. Comprehensive data on smoking behaviors were collected through a self-administered touchscreen questionnaire. According to previous literature (13,20), we calculated the 2 indices using the following algorithms: PSB = (β1 × SMK1 + β2 × SMK2 + β3 × SMK3 + β4 × SMK4) × (4/[β1 + 9 × β2 + 4 × β3 + β4]), where SMK1 to SMK4 correspond to the individual score of 4 smoking behaviors: smoking initiation (ever vs. never smoking), age of initiation (every 1 year earlier of smoking), heaviness of smoking (every 5 cigarettes per day heavier of smoking), and smoking cessation (current vs. past smoking) for each participant (13) (Supplement 3). Each smoking behavior was weighted by its β coefficient on dementia risk based on multivariable-adjusted Cox models (Supplement 5). On the other hand, SHS was calculated using the following algorithm: SHS = (7 × smokingi + PYi / 10) × (2 / [7 + MaxPY / 10]), where smokingi denotes the individual smoking status (current/past smoking, coded 1/0), PYi represents the individual value of pack-years, and MaxPY is the highest value of pack-years (20) (Supplement 4). The β coefficient for the association between current smoking and dementia risk was approximately 7 times higher than that for the association between a 10-year increment in pack-years and dementia risk (Supplement 6). The weighted PSB score ranges from 0 to 4, while the weighted SHS ranges from 0 to 2, with higher scores indicating greater smoking exposure. Never smokers were assigned a score of 0. PSB and SHS showed a significant moderate correlation (Spearman correlation coefficient = 0.58, p < .001). For PSB, we categorized participants into low- (0–1) and high- (3–4) dose groups (Figure 1A). For SHS, we utilized the “surv_cut-point” function from the “survminer” package in R to determine the cutoff value as 0.1156, using which we categorized the population into low- and high-dose groups (Figure 1B). In addition, because the β coefficient for smoking cessation is significantly larger than that for smoking initiation, age of initiation, and heaviness of smoking (Supplement 5), the high-dose group consists only of current smokers, while the low-dose group exclusively comprises past and never smokers. Considering that this categorization overlooked the effects of 3 other smoking behaviors on ACD risk, sensitivity analyses were conducted by utilizing the surv_cut-point function from the “survminer” package in R to determine an alternative cutoff value of 0.6552 for PSB that accounted for all 4 smoking behaviors (Supplement 7). Associations between PSB and SHS and their interactions with APOE ε4 status on ACD. (A) The density plot of PSB showing that participants were categorized into low- (0–1) and high- (3–4) dose groups. (B) The rank statistic plot of SHS showing that participants were categorized into low- (0–0.1156) and high- (0.1156–2) dose groups. (C, D) The forest plots showing the HRs for ACD associated with high vs. low PSB and SHS in the total population and APOE ε4–stratified subgroups, respectively. The circles denote HRs, and the error bars represent corresponding 95% CIs. Different colors represent different groups, with red and blue denoting total population and APOE ε4 subgroups, respectively. (E, F) The cumulative incidence plots comparing the incidence of ACD in the high PSB and SHS groups and their low-dose counterparts after controlling for mortality without a dementia diagnosis. The solid and dashed lines denote the high and low PSB and SHS groups, respectively. Red and blue represent the cumulative incidence of ACD and death, respectively. ACD, all-cause dementia; HR, hazard ratio; PSB, panoramic smoking burden; SHS, smoking history score. Dementia Diagnosis Dementia cases were identified through the integration of data from primary care information, hospital inpatient admissions, and death registries (21). This electronic linkage captured primary and secondary events across health care systems in England, Scotland, and Wales. According to the ICD, dementia diagnosis was defined as code 290.1 in ICD-9 and codes F00, F01, F02, F03, G30, G31.0, G31.1, and G31.8 in ICD-10 (22). The date of diagnosis was established based on the earliest record, regardless of the source. Covariates Covariates comprised sociodemographic characteristics, lifestyle factors, health and medical history, as well as APOE ε4 genotype. Sociodemographic factors included age, sex, educational levels, and the Townsend deprivation index. Data on lifestyle included alcohol dependence and physical activity, which was classified into low, moderate, and high levels based on the International Physical Activity Questionnaire guidelines (23). Health and medical history included body mass index (BMI), anxiety, cancer, depression, hypertension, diabetes, hyperlipidemia, stroke, and medication use of insulin, antihypertensive drugs, and cholesterol-lowering drugs. As the proportion of participants with missing covariate values in the eligible sample was relatively low (<1%), we conducted a complete case analysis (3). Assessments of APOE ε4 Genotypes Genetic data were obtained based on 2 similar Affymetrix arrays: the Applied Biosystems UK Biobank Lung Exome Variant Evaluation Axiom Array and the Applied Biosystems UK Biobank Axiom Array. Genotyping was conducted by the Affymetrix Research Services Laboratory, and quality control was managed by the UKB (24). APOE ε4 genotypes were determined through analyzing 2 single nucleotide polymorphisms: rs429358 and rs7412. Meta-Analysis We conducted a meta-analysis following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement (25). PubMed was searched up to May 2025 using terms related to smoking and dementia. Reference lists of relevant meta-analyses were also reviewed (1,26). We included prospective cohort studies that 1) tested associations of smoking exposure with incident all-cause dementia or AD, 2) included participants free of dementia at baseline, and 3) provided measures of effect size by APOE ε4 status. Two reviewers independently screened studies and extracted data. Study quality was assessed using the updated Newcastle-Ottawa Scale (27). Detailed methods, including the full search strategy and extracted variables, are provided in Supplement 8. Plasma Proteomics The UKB performed plasma proteomic profiling on approximately 55,000 participants using the Olink Explore 3072 proximity extension assay (28). This analysis quantified 2923 unique proteins for each sample, which were expressed as normalized protein expression values on a log2 scale (28). Any protein data that did not meet the quality control criteria were considered as missing values (28). Furthermore, we excluded 12 proteins with missing values >20%, resulting in a final pool of 2911 proteins. A total of 18,634 participants with proteomic data for SHS and a subsample of 18,352 for PSB were included. Statistical Analyses According to the smoking index scores, participants were categorized into low- and high-dose groups. Baseline characteristics in the groups were compared. For skewed continuous variables, data were expressed as median (IQR), and Mann-Whitney U tests were applied to assess group differences. For categorical variables, data were presented as n (%), and χ2 tests were utilized. Cox proportional hazards regression models were used to evaluate the associations between the 2 smoking indices and ACD, with multivariable-adjusted hazard ratios (HRs) and 95% CIs calculated. The proportional hazards assumption was tested using Schoenfeld residuals, with no violation detected. We performed sensitivity analyses by removing participants diagnosed with dementia within the initial 3 years. Moreover, the death-competing risk model lowered the risk of confounding bias due to mortality without a dementia diagnosis. The population attributable risk percentage (PAR%) was calculated for both smoking indices in the total population and in subgroups stratified by age and sex across different APOE ε4 genotypes using the following equation (29): PAR% = Pe (RR – 1) / (Pe [RR – 1] + 1) × 100%, where Pe denotes the proportion of individuals with high PSB or SHS values, and relative risk (RR) is the corresponding risk estimate from the Cox regression models. The multiplicative interaction effects between both PSB and SHS and APOE ε4 were estimated by adding the terms PSB × APOE ε4 and SHS × APOE ε4 to each Cox model. The additive interaction effect was also assessed, using the relative excess risk due to interaction, attributable proportion, and synergy index (30). Subgroup analyses by APOE ε4 status were performed, and the corresponding HRs, 95% CIs, and PARs% were calculated. To validate our findings, we meta-analyzed previous findings from longitudinal studies. The multivariable-adjusted risk estimates, including HRs or RRs, along with their 95% CIs, were log transformed and pooled using random-effects models (DerSimonian-Laird method) (31). To mitigate potential overestimation of effect size by odds ratios (ORs), we converted ORs to RRs using the following algorithm: RRadjusted = ORadjusted / ([1 − P0] + [P0 × ORadjusted]), where P0 represents the incidence of dementia in the nonexposed group, or in the total population if unavailable (32). To ascertain whether and how APOE ε4 modulated the relationship between smoking and dementia, we performed meta-regression analyses by APOE ε4 genotype, with subgroup analyses conducted subsequently. Heterogeneity was evaluated using the Q test, with its size determined by the I2 index (33). Publication bias was assessed using Egger’s regression test (34). Finally, to test the potential peripheral mechanisms underpinning the above-mentioned associations, proteomic and bioinformatic analyses were conducted. Multiple linear regression models were utilized to identify proteins associated with the constructed smoking indices, and Cox proportional hazards regression models were used to identify proteins associated with ACD. The overlapping proteins with consistently positive or negative β coefficients were selected, after Bonferroni correction (p < 1.72 × 10−5, number of proteins tested = 2911). Next, we conducted enrichment analyses using the STRING database (http://string-db.org). We selected biological pathways from the Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases, with strength values (log10 [observed protein/expected protein]) >1, which indicate a large enrichment effect. Causal mediation analyses were conducted to identify proteins that significantly mediated the relationships between smoking indices and ACD in the total population and in subgroups by APOE ε4 status. Specifically, the first equation (multiple linear model) regressed the mediator (overlapping proteins) on the independent variable (PSB or SHS). The second equation (parametric survival model) regressed the dependent variable (ACD) on the independent variable. The third equation (parametric survival model) regressed the dependent variable on both the independent variable and the mediator variable. The indirect effect was estimated, with the significance determined using 1000 bootstrapped iterations. Bonferroni correction was implemented to adjust for multiple testing, with significance thresholds tailored to the number of comparisons in each subgroup analysis. All models were adjusted for age, sex, education, Townsend deprivation index, anxiety, cancer, depression, hypertension, diabetes, hyperlipidemia, stroke, insulin, antihypertensive drugs, cholesterol-lowering drugs, alcohol dependence, BMI, physical activity, and APOE ε4 genotype. A 2-sided p value of <.05 was considered statistically significant unless stated otherwise. The “survival,” “survminer,” “cmprsk,” “ggplot2,” “interaction,” and “mediation” packages in R software version 4.3.1, Stata version 17.0, and GraphPad Prism 10.1.2 were used to conduct the above analyses and figure preparations. Results Associations of PSB and SHS With ACD Risk Varied by APOE ε4 Status in the UKB Among the 170,720 participants for PSB, 54.2% were female, with a median age of 57 years. During a median follow-up of 13 years, 12,004 deaths (10,432 without a dementia diagnosis) were recorded, and 3515 participants developed dementia. Among the 173,366 participants for SHS, 53.9% were female, with a median age of 57 years. During a median follow-up period of 13 years, 12,295 died, 10,704 of whom died without a dementia diagnosis, and 3558 participants were diagnosed with dementia. Compared with participants in the low-dose group, those with a high PSB or SHS were more likely to be male and less educated; tended to have a higher prevalence of depression, hyperlipidemia, and stroke; had lower socioeconomic status; were more likely to be users of cholesterol-lowering drugs and to have alcohol dependence; and were less physically active (Table 1). Table 1Baseline Characteristics of Participants for PSB and SHSIndividual CharacteristicsPSBSHSOverall, n = 170,720High PSB, n = 12,337Low PSB, n = 158,383p ValueOverall, n = 173,366High SHS, n = 35,106Low SHS, n = 138,260p ValueAge, Years57 (50, 63)54 (47, 61)57 (50, 63)<.00157 (50, 63)59 (51, 64)56 (49, 62)<.001Female92,554 (54.2%)6042 (49.0%)86,512 (54.6%)<.00193,400 (53.9%)14,098 (40.2%)79,302 (57.4%)<.001Education, Years13 (11, 22)11 (11, 21)13 (11, 22)<.00113 (11, 22)11 (10, 22)13 (11, 22)<.001Townsend Deprivation Index−2.30 (−3.71, 0.19)−0.31 (−2.73, 2.71)−2.39 (−3.77, −0.03)<.001−2.29 (−3.71, 0.22)−1.31 (−3.21, 1.78)−2.46 (−3.81, −0.20)<.001Anxiety2352 (1.4%)167 (1.4%)2185 (1.4%).8412389 (1.4%)493 (1.4%)1896 (1.4%).626Cancer926 (0.5%)28 (0.2%)898 (0.6%)<.001935 (0.5%)138 (0.4%)797 (0.6%)<.001Depression8944 (5.2%)1191 (9.7%)7753 (4.9%)<.0019120 (5.3%)2614 (7.4%)6506 (4.7%)<.001Hypertension42,851 (25.1%)2862 (23.2%)39,989 (25.2%)<.00143,481 (25.1%)11,224 (32.0%)32,257 (23.3%)<.001Hyperlipidemia19,845 (11.6%)1534 (12.4%)18,311 (11.6%).00420,219 (11.7%)6085 (17.3%)14,134 (10.2%)<.001Diabetes7706 (4.5%)580 (4.7%)7126 (4.5%).3007845 (4.5%)2682 (7.6%)5163 (3.7%)<.001Stroke1934 (1.1%)221 (1.8%)1713 (1.1%)<.0011985 (1.1%)802 (2.3%)1183 (0.9%)<.001Insulin1658 (1.0%)133 (1.1%)1525 (1.0%).2151695 (1.0%)533 (1.5%)1162 (0.8%)<.001Antihypertensive Drugs33,453 (19.6%)2093 (17.0%)31,360 (19.8%)<.00133,927 (19.6%)9309 (26.5%)24,618 (17.8%)<.001Cholesterol-Lowering Drugs28,081 (16.4%)2164 (17.5%)25,917 (16.4%).00128,615 (16.5%)9173 (26.1%)19,442 (14.1%)<.001Alcohol Dependence205 (0.1%)78 (0.6%)127 (0.1%)<.001212 (0.1%)134 (0.4%)78 (0.1%)<.001Body Mass Index26.57 (24.02, 29.64)26.16 (23.49, 29.31)26.60 (24.06, 29.67)<.00126.57 (24.02, 29.64)27.56 (24.84, 30.64)26.33 (23.85, 29.35)<.001Physical Activity Low31,673 (18.6%)2880 (23.3%)28,793 (18.2%)<.00132,209 (18.6%)7571 (21.6%)24,638 (17.8%)<.001 Moderate70,003 (41.0%)4810 (39.0%)65,193 (41.2%)71,079 (41.0%)13,801 (39.3%)57,278 (41.4%) High69,044 (40.4%)4647 (37.7%)64,397 (40.6%)70,078 (40.4%)13,734 (39.1%)56,344 (40.8%)APOE ԑ448,504 (28.4%)3404 (27.6%)45,100 (28.5%).03649,235 (28.4%)9861 (28.1%)39,374 (28.5%).150PSB0.000 (0.000, 0.517)3.517 (3.379, 3.690)0.000 (0.000, 0.414)<.001––––SHS––––0.000 (0.000, 0.085)0.323 (0.174, 0.529)0.000 (0.000, 0.009)<.001For continuous data, values are presented as median (lower quartile, upper quartile), and p values are from Mann-Whitney U tests because they do not follow normal distribution. For categorical data, values are presented as n (%), and p values are from χ2 tests.PSB, panoramic smoking burden; SHS, smoking history score. Baseline Characteristics of Participants for PSB and SHS For continuous data, values are presented as median (lower quartile, upper quartile), and p values are from Mann-Whitney U tests because they do not follow normal distribution. For categorical data, values are presented as n (%), and p values are from χ2 tests. PSB, panoramic smoking burden; SHS, smoking history score. PSB was significantly linked to over a 2-fold higher risk of ACD (HR = 2.24, 95% CI [2.00 to 2.52], p < .001) (Figure 1C). The association remained unchanged after excluding individuals developing ACD within the first 3 years (HR = 2.25, 95% CI [2.00 to 2.53], p < .001) and in the death-competing risk model (p < .001) (Figure 1E). Utilizing 0.6552 as the alternative cutoff value did not change the significance (HR = 1.36, 95% CI [1.25 to 1.48], p < .001). Individuals with a high SHS exhibited a 34% increased risk of ACD compared with their counterparts (HR = 1.34, 95% CI [1.24 to 1.44], p < .001) (Figure 1D). The correlation remained robust after excluding participants developing ACD during the initial 3 years (HR = 1.35, 95% CI [1.25 to 1.45], p < .001) and in the death-competing risk model (p < .001) (Figure 1F). There was a significant multiplicative interaction between PSB and APOE ε4 status (p < .001) (Figure 1C), while no additive interaction was detected (Supplement 10). High PSB was significantly associated with a 159% increased risk of ACD in the APOE ε4− subgroup (HR = 2.59, 95% CI [2.21 to 3.04], p < .001) (Figure 1C), while there was an 89% higher risk of ACD in the APOE ε4+ subgroup (HR = 1.89, 95% CI [1.59 to 2.25], p < .001) (Figure 1C). Subgroup analyses revealed significant multiplicative interactions in the male and late-life subgroups but not in the female or midlife subgroup (Supplement 11). We observed a higher contribution of PSB to ACD onset among APOE ε4 noncarriers compared with among carriers (PAR% = 10.41% vs. 5.88% in the total population, 13.14% vs. 5.89% in the male subgroup, and 9.18% vs. 2.68% in the late-life subgroup) (Supplement 11). Similarly, a significant multiplicative interaction between SHS and APOE ε4 status was detected (p = .031) (Figure 1D), while no additive interaction was detected (Supplement 9). Having a high SHS was significantly correlated with a 36% increased risk of ACD among APOE ε4 noncarriers (HR = 1.36, 95% CI [1.22 to 1.51], p < .001) (Figure 1D), while there was a 31% heightened risk of ACD among APOE ε4 carriers (HR = 1.31, 95% CI [1.18 to 1.45], p < .001) (Figure 1D). Nevertheless, we observed no significant interaction in subgroups stratified by age or sex (Supplement 11). In the total population, the influence of SHS on ACD was slightly amplified among APOE ε4 noncarriers compared with carriers (PAR% = 6.82% vs. 5.85%) (Supplement 11). Meta-Analyses Supported the Effect of the Interaction of Smoking and APOE ε4 Status on ACD Risk The literature screening process is illustrated in Figure 2A. A total of 14 association results from 7 longitudinal studies were available for meta-regression and subgroup analyses by APOE ε4 status. The characteristics are presented in Table 2, with additional characteristics reported in Supplement 12. The average quality score was 6.36 (Supplement 13). Current smoking was significantly associated with a 42% increased risk of ACD (RR = 1.42, 95% CI [1.12 to 1.79], p = .003) (Figure 2C) compared with never smoking. Meta-regression analyses indicated that APOE ε4 status significantly modulated the association between smoking and dementia (β = −0.50, 95% CI [−0.87 to −0.13], p = .011) (Figure 2B). Subgroup analyses found that among APOE ε4 noncarriers, current smokers exhibited a 75% higher risk of ACD (RR = 1.75, 95% CI [1.44 to 2.13], p < .001) (Figure 2C), whereas there was no association among APOE ε4 carriers (p = .613) (Figure 2C). No publication bias was revealed. Ever smoking (RR = 1.22, 95% CI [1.02 to 1.45], p = .032) but not past smoking (p = .360) was significantly linked to ACD risk (Supplement 14). Meta-regression analyses revealed no interaction effect by APOE ε4 status (Supplement 14). Current smoking was significantly associated with a 60% increased risk of AD (RR = 1.60, 95% CI [1.14 to 2.26], p = .007) (Supplement 14). No interaction with APOE ε4 was detected (p = .274) (Supplement 14). Meta-analyses on the associations between smoking status and all-cause dementia stratified by APOE ε4 genotype. (A) The search flowchart showing the screening process of literature. A total of 7 studies were finally included for meta-analysis. (B) The scatterplot with a fitted regression line illustrating that the APOE ε4 genotype contributed significantly to the observed heterogeneity. The x-axis represents the APOE ε4–negative and –positive subgroups. The y-axis represents the effect size (log-transformed RR). Each blue bubble denotes a subgroup from an individual study, with the size of the bubble reflecting its weight, which is calculated as 1/SE2. The red line represents the fitted regression line, suggesting a negative influence of APOE ε4 genotype on effect size. (C) Subgroup analyses of associations between current smoking and ACD, according to APOE ε4 genotype. ACD, all-cause dementia; DL, DerSimonian–Laird; MA, meta-analyses; MCI, mild cognitive impairment; RR, relative risk. Table 2Population Characteristics of the Seven Included CohortsFirst Author, YearCountryTotalAPOE ε4+APOE ε4−Type of ExposureType of OutcomeAge, YearsFemaleFollow-Up Duration, YearsSamples/CasesSamples/CasesSamples/CasesZhang et al., 2021 (3)United Kingdom64.152.4%Median 9.0193,198/178854,471/962137,807/812Current and past smokingACDRönnemaa et al., 2011 (42)Sweden71.00%Mean 131174/246274/96594/101Current smokingACDRusanen et al., 2010 (11)Finland50.662.65%Mean 211419/59 ACD; 1367/46 AD493/31 ACD; 477/25 AD887/26 ACD; 860/21 ADCurrent and past smokingACD, ADReitz et al., 2007 (10)The Netherlands69.559.9%Mean 7.16868/706 ACD; 6205/501 AD1761/271 ACD; 1705/215 AD4569/368 ACD; 4500/286 ADCurrent and past smokingACD, ADAggarwal et al., 2006 (41)United States73.861.9%Mean 4.11064/170NANACurrent, past smokingADTyas et al., 2003 (39)United States77.70%Maximum 313102/166NANACurrent smokingAD ± CVDMerchant et al., 1999 (40)United States75.468.74%Mean 2.041062/142262/45646/88Current and past smokingADThe samples/cases of the total population for each study were calculated based on the samples/cases from the APOE ε4+ and APOE ε4− subgroups. We directly utilized the primary samples/cases of the total population unless the corresponding data from each subgroup by APOE ε4 genotype were not provided. Additionally, we only considered types of exposure and outcomes that provided corresponding measures of effect size.ACD, all-cause dementia; AD, Alzheimer’s disease; CVD, cardiovascular disease; NA, not available. Population Characteristics of the Seven Included Cohorts The samples/cases of the total population for each study were calculated based on the samples/cases from the APOE ε4+ and APOE ε4− subgroups. We directly utilized the primary samples/cases of the total population unless the corresponding data from each subgroup by APOE ε4 genotype were not provided. Additionally, we only considered types of exposure and outcomes that provided corresponding measures of effect size. ACD, all-cause dementia; AD, Alzheimer’s disease; CVD, cardiovascular disease; NA, not available. Proteins and Enriched Pathways Potentially Mediate the Association Between Smoking and Dementia Baseline characteristics of participants with proteomic data are shown in Supplement 15. In the total population, 65 proteins (Bonferroni-adjusted p < 1.72 × 10−5) were significantly associated with PSB and ACD (Supplement 16), and 67 proteins were associated with SHS and ACD (Supplement 17). Among these, 57 overlapped (Figure 3A). GO and KEGG enrichment analyses identified 22 statistically significant pathways (strength value >1). The enriched pathways were categorized into 3 functional groups, including immune and inflammatory responses, angiogenesis modulation, and regulation of endopeptidase activity (Figure 3B and Supplement 18). Mediation analyses revealed significant mediation effects of these proteins in the association between PSB and ACD (p < .001, number of proteins tested = 57) (Figure 3C and Supplement 19). Remarkably, urokinase plasminogen activator surface receptor (PLAUR) emerged as the predominant mediator, exhibiting the highest mediation proportion. As for SHS, 38 proteins exhibited suggestive mediation effects (p < .05) (Supplement 20), but none survived Bonferroni correction. Proteomics, bioinformatics, and mediation analyses. (A) The Venn diagram showing the number of overlapping proteins associated with PSB, SHS, and ACD. Groups A and B denote proteins associated with PSB and ACD among participants for PSB calculation, respectively. Groups C and D denote proteins associated with SHS and ACD among participants for SHS calculation, respectively. (B) The Sankey bubble plot demonstrating the enriched pathways of overlapping proteins. The size of the bubble for each pathway represents the number of enriched proteins. The color gradient from blue to red denotes an increasing level of enrichment significance. (C) The Sankey plot showing proteins significantly mediating the associations between PSB and ACD. The size of each protein block represents its mediation proportion. The blank blocks represent the scale legends. Proteins denoted by ∗∗∗ are those with Bonferroni-adjusted p values < .001. (D, E) The 2 linked volcano plots showing the beta estimate (x-axis) and −log10 (p value) (y-axis) for the associations of PSB and ACD with proteins among APOE ε4 noncarriers and carriers, respectively. Differentially expressed proteins associated with both PSB and ACD were highlighted. (F, G) The 2 linked volcano plots showing the beta estimate (x-axis) and −log10 (p value) (y-axis) for the associations of SHS and ACD with proteins among APOE ε4 noncarriers and carriers, respectively. Differentially expressed proteins associated with both SHS and ACD are highlighted. (H, I) The proportional Venn diagrams showing the number of shared proteins among APOE ε4 noncarriers and carriers, respectively. (J, K) The dot plots showing the top 5 GO and KEGG pathways (strength value >1) for each of the identified proteins among APOE ε4 noncarriers and carriers. If a protein had <5 terms that met the above criteria, then all available pathways are displayed. The size of the bubble denotes the strength value of each pathway. The color gradient from red to blue indicates an increasing level of enrichment significance. (L) Categorization of proteins identified in APOE ε4 noncarriers and carriers according to the top 5 pathways of each protein. Shared proteins and pathways are in bold. (M, N) The flower plots showing proteins significantly mediating the associations of PSB and SHS with ACD among APOE ε4 noncarriers. The color gradient of the petal from yellow to blue indicates an increasing mediation proportion. ACD, all-cause dementia; BMP, bone morphogenetic protein; EGFR, epidermal growth factor receptor; FDR, false discovery rate; GO, Gene Ontology; HIF, hypoxia-inducible factor; JAK-STAT, janus kinase/signal transducer and activator of transcription; KEGG, Kyoto Encyclopedia of Genes and Genomes; NF, nuclear factor; PSB, panoramic smoking burden; SHS, smoking history score. Plasma Proteins Mediate the Amplified Effects of Smoking on ACD Among APOE ε4 Noncarriers We hypothesized that the plasma proteomic profiles and the enriched pathways linking smoking exposures to dementia would differ in the absence or presence of the APOE ε4 allele. Accordingly, candidate proteins were selected when they simultaneously showed significant associations (Bonferroni-adjusted p < 1.72 × 10−5) with smoking indices and ACD among APOE ε4 noncarriers (13 for PSB and 19 for SHS) (Figure 3D, F) and carriers (9 for PSB and 11 for SHS) (Supplements 21 and 22) (Figure 3E, G). Among APOE ε4 noncarriers, 12 were overlapping proteins (Figure 3H), while among APOE ε4 carriers, 8 were shared proteins (Figure 3I). We selected the top 5 GO and KEGG pathways for each protein, with ranking based on strength values (Figure 3J, K and Supplements 23 and 24). Accordingly, the 12 proteins identified in APOE ε4 noncarriers were classified into 5 functional groups: 1) immune and inflammatory responses, 2) cell growth and tissue development, 3) fibrinolysis, 4) regulation of the digestive system, and 5) phospholipid metabolism (Figure 3L). The 8 proteins identified in APOE ε4 carriers are shown in Figure 3L. Immune and inflammatory responses, cell growth and tissue development, and fibrinolysis were the overlapping mechanisms. Regulation of the digestive system and phospholipid metabolism were unique to APOE ε4 noncarriers. Mediation analyses were performed on the overlapping proteins in subgroups by APOE ε4 status. In the APOE ε4− subgroup, all 12 proteins significantly mediated the associations of smoking indices with ACD risk (p < .004, number of proteins tested = 12), with mediation proportions spanning from 3.7% to 37.5% for PSB (Figure 3M and Supplement 25) and from 5.1% to 37.3% for SHS (Figure 3N and Supplement 26). PLAUR consistently emerged as the predominant mediator among APOE ε4 noncarriers. No mediation effect was detected in APOE ε4 carriers (Supplements 27 and 28). Discussion Via a large population-based cohort and evidence-based methods, in this study, we uncovered that 1) significant multiplicative interaction effects on dementia risk were observed between comprehensive smoking indices and APOE ε4 status, with more pronounced effects among APOE ε4 noncarriers; 2) the interaction was validated by meta-analyzing previous evidence that current smoking was associated with dementia only among APOE ε4 noncarriers; and 3) immune and inflammatory responses may be the crucial mechanism linking smoking to dementia, while regulation of the digestive system and phospholipid metabolism may elucidate the enhanced effects among APOE ε4 noncarriers. We found that APOE ε4 modified the association between smoking and dementia. However, the lower HRs among ε4 carriers warrant further interpretation. First, a similar absolute risk increase leads to a smaller HR when added to a higher baseline risk (10,12). We observed no significant additive interaction, signifying that there was no difference in the actual risk increase from smoking between APOE ε4 carriers and noncarriers. Therefore, the lower HRs among APOE ε4 carriers may result from their higher baseline dementia risk. Second, APOE ε4 and smoking may contribute to dementia through differential mechanisms. APOE ε4 is potently linked to AD through amyloid-β (Aβ) accumulation (35) but weakly linked to VaD (36). In contrast, smoking is strongly linked to VaD via vascular damage but not to AD (37). These differential pathological mechanisms could contribute to the observed interaction. Finally, higher mortality among APOE ε4 carriers may mitigate the impact of smoking on dementia (38). However, Cox regression analyses identified no significant association between APOE ε4 and death, suggesting no survival bias. Previous studies examining the interaction between smoking and APOE ε4 status have yielded conflicting results. One study revealed a negative multiplicative interaction (3), while others identified positive (11) or no (10) interaction. When using pack-years to define smoking, the Honolulu-Asia Aging study identified no interaction with APOE ε4 on AD risk (39). Consistent with the UKB cohort, most studies demonstrated via subgroup analyses that current smoking predicted elevated risk of dementia only among APOE ε4 noncarriers (3,10,39, 40, 41, 42, 43), although one study reported the opposite (11). These inconsistencies may arise from the following methodological limitations: 1) reduced statistical power due to small sample size and insufficient covariate adjustments; 2) 3 studies utilized logistic regression instead of Cox regression (11,39,41), potentially leading to inaccuracies in longitudinal risk assessment (44); and 3) previous studies mostly tested single smoking behavior, thus overlooking the combined effects of diverse smoking behaviors. However, these limitations were well-addressed in the current study. Additionally, further age- and sex-stratified analyses suggested that smoking cessation may confer particular benefits in male and late-life APOE ε4 noncarriers. Potential mechanisms linking smoking to dementia include immune and inflammatory responses, angiogenesis modulation, and regulation of endopeptidase activity. Specifically, smoking could trigger neuroinflammation via microglia activation (9,45), disrupting blood-brain barrier integrity through vascular endothelial growth factor overexpression (46, 47, 48, 49), and reducing neprilysin levels, thereby impairing Aβ clearance and enhancing Aβ deposition (50,51). PLAUR is involved in intracerebral atherosclerosis (52), which may promote Aβ accumulation, cerebrovascular injury, and hypoxia of neurons (53, 54, 55). Notably, immune and inflammatory responses, cell growth and tissue development, and fibrinolysis were identified as overlapping mechanisms in both APOE ε4–stratified subgroups. Beyond overlapping mechanisms, we uncovered differential plasma proteomic profiles in each subgroup by APOE ε4 status. Among APOE ε4 noncarriers, proteins involved in digestive system regulation (GAST [big gastrin], TFF2 [trefoil factor 2]) and phospholipid metabolism (PLA2G15 [group XV phospholipase A2]) significantly mediated the smoking-dementia association. In contrast, no such effect was observed among APOE ε4 carriers. Therefore, we inferred that smoking could confer an additional risk of dementia among APOE ε4 noncarriers potentially via dysregulating digestive system and phospholipid metabolism. Smoking is associated with gut microbiota disturbance (56), which may impair the central nervous system through the gut-brain axis (57). Deficiency of vitamin B12, mediated by smoking-related chronic atrophic gastritis, contributes to dementia by disturbing myelin synthesis, inducing tau hyperphosphorylation and damaging cerebral vessels (58, 59, 60, 61). Smoking-induced oxidative stress oxidizes phosphatidylcholine into PAF (platelet-activating factor)–like lipids, which bind to PAF receptors to trigger inflammatory responses (62,63). However, future experimental studies are warranted to validate these proteins and pathways. The current study has several strengths. First, this study is the first to evaluate the interaction of comprehensive smoking behaviors by APOE ε4 status based on a large-scale prospective cohort study. The consistent interaction patterns that we observed across the 2 indices indicated the robustness of our findings. Second, the stratification effects were validated in independent populations using meta-analysis. Third, we performed sensitivity analyses by utilizing a death-competing risk model, excluding participants diagnosed with dementia within the initial 3 years, and using different methods for selecting cutoff values. Finally, the proteomic and bioinformatic analyses provided valuable insights to elucidate the peripheral mechanisms. The current study also has limitations. First, we utilized only baseline data on smoking behaviors but cannot account for changes in smoking behaviors over time. Second, PSB and SHS do not encompass all dementia-related smoking behaviors, such as years since smoking cessation (64). Third, because our study was observational, we cannot establish causal links, especially for the biological pathways. Finally, the UKB cohort primarily comprises European White participants, which may limit the applicability of our findings to other ethnic groups. Conclusions This study comprehensively demonstrated how APOE ε4 status modifies the relationship between smoking and dementia. Stronger effects were observed among APOE ε4 noncarriers, which is a relationship that may be mediated by regulation of the digestive system and phospholipid metabolism. This interaction should be emphasized in the risk stratification and prevention strategy-making process. Acknowledgments and Disclosures This work was supported by the Taishan Scholar Project (Grant No. tsqn202211375 [to WX]). This research has been conducted using the UK Biobank Resource under Application No. 108930. Y-LC, Y-CW, WX, and X-HC contributed to analysis of the data, drafting and revision of the article, and preparation of all the figures. Z-LY, L-YH, and LT contributed to revision of the article. WX and X-HC contributed to conceptualization and design of the study. UKB data are available upon reasonable request. The authors report no biomedical financial interests or potential conflicts of interest. Footnotes Supplementary material cited in this article is available online at https://doi.org/10.1016/j.bpsgos.2026.100800. Contributor Information Xiao-Hui Chu, Email: a13001677262@163.com. Wei Xu, Email: dr_xuwei@qdu.edu.cn. Supplementary Material Supplements 1–28 Key Resources Table References Associated Data Supplementary Materials Supplements 1–28 Key Resources Table