Validated inference of smoking habits from blood with a finite DNA methylation marker set

This study was embedded within the Biobank-based Integrative Omics Study (BIOS) Consortium [28], which consists of six Dutch cohorts (N = 3118), including the Rotterdam Study (RS) (N = 584) [29], Cohort on Diabetes and Atherosclerosis Maastricht (CODAM) (N = 156) [30], The Netherlands Twin Register (NTR) (N = 894) [31], Leiden Longevity Study (LLS) (N = 625) [32], Prospective ALS Study Netherlands (PAN) (N = 167) [33] and LifeLines (LL) (N = 692) [34]. Additionally, we included another 646 unrelated participants from the Rotterdam Study (RS-III-1) not included in BIOS. We externally validated our model in the Kooperative Gesundheitsforschung in der Region Augsburg (KORA) study (F4, N = 1608) [35], as well as in the Study of Health in Pomerania (SHIP)-Trend (N = 244) [36] cohort. Characteristics of all cohorts used can be found in Online Resource 1: Table S1. We additionally tested our model in samples from children included in the Generation R Study [37], in particular, we used data from children participating at birth (N = 1111), at the age of 6 years (N = 355), and at the age of 10 years (N = 309), of which 197 overlapped between all three time points, providing longitudinal data (Online Resource 1: Table S2). The smoking status information was obtained using questionnaires. The study characteristics are described in detail in Online Resource 2: Supplemental methods.

DNA was extracted from whole peripheral blood in all studies using standard procedures. All studies used the Illumina Infinium Human Methylation 450 K BeadChip (Illumina Inc, San Diego, CA, USA) for epigenome-wide DNA methylation measurements, except the SHIP-Trend study, which used the more recent Infinium MethylationEPIC BeadChip (Illumina Inc, San Diego, CA, USA). DNA methylation data pre-processing for cohorts included in the BIOS consortium were conducted together via the pipeline created by Tobi et al. [38, 39]. The DNA methylation data pre-processing in the external validation cohorts and the Generation R Study were done independently. The methylation proportion of a CpG site was reported as a methylation β-value in the range of 0 (representing completely non-methylated sites) to 1 (representing completely methylated sites). Further study-specific methods can be found in Online Resource 2: Supplemental methods.

EWASs using the Illumina Infinium Human Methylation 27 K or 450 K BeadChip investigating smoking-induced changes in DNA methylation patterns were reviewed [2, 21, 40‐50]. We excluded studies [11] that used cohorts included in our model-building dataset, to avoid over-estimation of our model. Envisioning future laboratory tool development, we only selected robust CpGs that were (1) highlighted in two or more studies, (2) with at least 10% difference in mean or median (depending on availability per EWAS) β-values between current smokers and never-smokers (or non-smokers when non-smoking data was available) in at least one of the studies, and (3) with the same direction in β-value difference between current smokers and never/non-smokers in all studies investigated.

Of the total participants considered for model building (N_total = 5178), we excluded those with (1) missing data for smoking habits (1206 participants), (2) missing β-values for the predictive CpGs (82 participants), or (3) extreme outliers for one or more CpGs (mean ± 4 SD) (126 participants). In the end, we included 3764 participants in the final model building set, who were then categorized based on their smoking habits as (1) current smokers or (2) former and never smokers combined. The association between the candidate CpGs and smoking habits (smokers vs. non-smokers) was replicated in our model building dataset using binomial regression analysis adjusted for age and sex using the “glm” function with “binomial” as family and “logit” as link. To identify the most informative set of DNA methylation predictors from the candidate CpGs, the association between the complete set of predictive CpGs and smoking habits was assessed in a binary logistic regression analysis, using the “glm” function with “binomial” as family and “logit” as link. Backward elimination procedures were used for the marker selection process. We excluded the CpGs one by one based on their absolute z-statistic per regression (calculated by dividing the regression coefficient by its standard error) assessed using the “VarImp” function (r-package “caret”). The predictive CpG with the lowest absolute z-statistic in the regression was removed. The model was applied to the dataset with the “predict” function (type = “response”) and the confusion matrix (r-package “caret”) was conducted using a probability threshold of 0.5. The prediction performance of the model was additionally assessed using “prediction” and “performance” (r-package “ROCR”), the Area Under the Curve (AUC) per model was calculated (r-package “ROCR”) and a cumulative AUC profile was conducted for each model to obtain a cumulative AUC profile. We selected the best-fit prediction model using a combination of the backward elimination approach and the Chi squared test. In particular, we compared the model including all CpGs (model_FULL) with the model excluding one CpGs, (model_FULL-1CpG), this model _FULL-1CpG was then compared with the model excluding another CpG (model_FULL-2CpGs), following the same order as conducted via the backward approach, and so on until we noticed a statistically significant difference between two models in the backward approach. Subsequently, we tested the inclusion of age, sex and cell counts to the final model.

Participants included in the model building dataset (N = 3764) without additional smoking data, including the age someone stopped smoking (former smokers) or the age someone started smoking or the number of cigarettes someone smokes per day (current smokers), were excluded, resulting in a dataset including 2939 participants. The association between the previously selected predictive CpGs and the three smoking categories was assessed in a multinomial regression analysis, using the “multinom” function (r-package “nnet”). We predicted the smoking categories using the “predict” function (type = “class”) and the confusion matrix (r-package “caret”) was conducted. The AUC per category was conducted using the “predict” function (type = “probs”) and “roc” function (r-package “pROC”).

In the former smokers (N = 1332), smoking cessation time was calculated as one’s age minus the age one stopped smoking. The participants were split into two categories for three models. For model 1, ≥ 5 years cessation time were coded as “1” and < 5 years smoking cessation were coded as “0”, for model 2, ≥ 10 years cessation time were coded as “1” and < 10 years smoking cessation were coded as “0”, and for model 3, ≥ 15 years cessation time were coded as “1” and < 15 years smoking cessation were coded as “0”. The predictions were conducted using the same method as described for the current versus non-smokers model. Probability thresholds were set to 0.8733, 0.7650 and 0.6397 respectively.

We combined the pack-year inference in current smokers with the cessation time in former and never smokers, resulting into five categories in two models (N = 2939) for inferring life-time smoking information. For model 1, the current smokers ≥ 15 pack-years were coded as “5”, with < 15 pack-years were coded as “4”, the former smokers ≤ 10 years smoking cessation were coded as “3”, with > 10 years smoking cessation were coded as “2” and never smokers were coded as “1”. In the second model the same categories were used except for the pack-years which were now divided in ≥ 10 pack-years (coded as “5”) and < 10 pack-years (coded as “4”). The predictions were conducted using the same method as described for the current vs former vs never smokers model.

For internal validation of the developed predictive models, we adopted a fivefold cross-validation scheme [51], in which the whole dataset is first randomly distributed into five equal and non-overlapping subsets. Four of the subsets (80% of the data) are combined to form a dataset used to train the logistic regression model which is then tested by inferring the smoking habits in the remaining dataset (20% of the data). This resulted in five different training (80%) and testing (20%) sets. The model was trained in the five training sets and applied to corresponding testing sets, resulting in five logistic regression models. Subsequently, we used the bootstrap method (r-packages “boot” and “parallel”) as additional internal validation to correct for potential overestimation of the prediction, since we use the same data for model building and predictions. We generated 1000 bootstrap samples, with replacement from the dataset for which we estimated the model and applied each fitted model to the original sample, resulting in 1000 AUC estimates. Thereafter, we recalculated the prediction accuracy by applying the fitted model to the bootstrap sample itself. The performance in the bootstrap sample represents an estimation of the apparent performance, and the performance in the original sample represents test performance. The difference between the average of the two conducted AUCs is a stable estimate of the optimism. We corrected for prediction overestimation by subtracting the optimism from the apparent AUC, to obtain an improved estimate of the prediction AUC [52, 53].

We externally validated our prediction models in two independent cohorts from German-European origin. The full models were validated in the KORA F4 study (N = 1608). Additionally, we externally validated our models in the SHIP-Trend study (N = 244). In this cohort, the EPIC methylation array was used which does not include all CpGs of the 450 K array. We therefore first generated the prediction models based on the overlapping CpGs in the model building dataset and subsequently externally validated them in the SHIP-Trend dataset.

We compared the outcomes of the CpG model to infer current vs non-smokers with the outcomes using a cotinine level cut-off of 50 ng/mL [54, 55] and applied smoking information from self-reports as reference. We employed a subset of our model building dataset (N = 488 participants included in NTR [56]) in which both DNA methylation levels and cotinine levels were available. First, participants were categorized as smokers when their plasma cotinine levels were > 50 ng/mL, or as non-smokers with cotinine levels ≤ 50 ng/mL, threshold according to previous studies including the used cotinine data [54, 55]. Second, the current versus non-smokers CpG model was applied to this subset, obtaining the inferred smoking status for the participants. Third, we compared the obtained smoking status for both models with the information obtained from the self-reported questionnaires and computed the sensitivity and specificity per model.

Studies have shown the impact of prenatal smoking exposure on the DNA methylation pattern of the offspring [57] and the ability of predicting maternal smoking status using these patterns [58]. In this context, we wanted to test the effect of prenatal exposure on model application in adults. Hence, when an adult does not smoke, but was exposed to prenatal smoking, do we predict this person indeed as a true non-smoker? To test for this putative impact of exposure to prenatal smoking on epigenetic inference of smoking habits using our model, we tested our model in umbilical cord blood of newborns (N = 1111), and in whole blood of children at the ages of six (N = 355) and 10 years (N = 309). We used five different analyses to evaluate the effects of active smoking of the mothers and passive smoking of the mothers (i.e. smoking of others in the mother’s home and work environment) during pregnancy on smoking habit inference using our model. In our first analysis, we did not take the smoking habits of the pregnant mothers or others in the pregnant mother’s home and work environment into account and all children were coded as non-smokers. The proportion of accurately predicted cases was calculated using a probability threshold of 0.5. In each of the following analyses, we coded the children “1” if their parents met the smoking habit criteria, otherwise they were coded as “0”. So, in the second analysis, only sustained maternal smoking throughout pregnancy was considered. Therefore, the children of mothers that smoked during the whole pregnancy were coded as “1”. In the third analysis, we additionally included the children of mothers who stopped smoking when they realized that they were pregnant by coding these children as “1”. In the fourth analysis, we additionally included smoking of the father and/or others in the mother’s household/at work (> 1 h per day) during pregnancy (i.e. passive smoking). In the fifth analysis, we assessed the sole effect of passive smoking i.e., where the mother did not smoke but the father or someone else in the house or at work (> 1 h per day) smoked during the pregnancy of the mother. For 197 children, DNA methylation levels were measured at all three time points, i.e. birth, 6 years of age and 10 years of age; hence, we repeated the previous models again in these children to allow a direct comparison of the findings at these three time points in the same individuals.

We inspected 14 published EWASs on tobacco smoking habits (N_total = 7015) [2, 21, 40‐50] to identify smoking-associated CpGs as candidate DNA methylation markers for prediction modeling of smoking habits. CpGs were selected as candidate prediction markers if they met three criteria as mentioned in the method section. This procedure highlighted 20 top smoking-associated CpGs as candidate markers used for further analyses (Table 1). The differences in β-values between smokers and never-/non-smokers reported previously for these 20 top smoking-associated CpGs are illustrated in Fig. 1.

Following the replication of the association between the CpGs and smoking habits (smokers vs. non-smokers) after adjusting for age and sex (Online source Table 3), we assessed the predictive effect of the selected 20 candidate markers in the model building dataset (N = 3764). Starting with a model including all 20 CpGs, the CpG with the lowest z-value per model was sequentially removed, and the AUC was calculated for each model to obtain a cumulative AUC profile (Table 1; Fig. 2).

To identify the minimal number of CpGs required to achieve maximum prediction accuracy, we additionally used Chi squared tests. Applying this backward approach, the first significant difference between two models was noted when we compared the model with and without cg09935388 (Table 1; Fig. 2). The combined marker elimination approach resulted in a finite set of DNA methylation markers comprising 13 CpGs (Table 1; Fig. 2). The AUC for the identified 13-CpG model was 0.901 for distinguishing between smokers versus non-smokers (for other prediction accuracy measures, see Table 2). The remaining 7 CpGs raised the cumulative AUC only on the 4th decimal i.e. from 0.9010 to 0.9016 (Table 1; Fig. 2). Hence, this finite set of 13 CpGs was used for subsequent prediction modeling. Using the 13-CpG model, we inferred the smoking status of the participants included in our model building dataset; the inferred probabilities are presented in a histogram in Fig. 3, where each probability bin is overlaid with the percentage of accurately inferred smoking habits in that probability range.

Adjusting the prediction model for age resulted in a minor AUC increase from 0.901 to 0.907, adjusting for sex from 0.901 to 0.903 and including both age and sex in the model increased the AUC slightly from 0.901 to 0.911 (Online Resource 1: Table S4). Additionally, we tested the influence of cell counts on the model accuracy. In the subset of participants for which cell count measures were available (N = 3402), our 13-CpG model without cell counts achieved an AUC of 0.906. Including the cell count measurements for monocytes, granulocytes and lymphocytes in our 13-CpG model, the AUC was almost identical at 0.907 (Online Resource 1: Table S5). Since age, sex and cell counts only had a minor impact on the prediction accuracy, these three non-epigenetic factors were not considered in the final model used in the subsequent analyses.

Next, we considered former smokers as an additional, separate category in the prediction model building based on the finite set of 13 CpGs, resulting in a three-category prediction model. To this end, we considered a subset of 2939 participants for which the relevant smoking habit information was available. We obtained for the current smokers (N = 364) an AUC of 0.928, for the former smokers (N = 1332) 0.772, and for the never smokers (N = 1243) 0.835 (for other accuracy measures, see Table 3). Additionally, we calculated smoking cessation time for the former smokers (N = 1332), and used the 13-CpGs to infer smoking cessation for ≥ 5 years (N = 1160) versus < 5 years (N = 172), which resulted in an AUC of 0.793, for ≥ 10 versus < 10 years smoking cessation time (N = 1028 and N = 304, respectively) an AUC of 0.778 was obtained and for ≥ 15 versus < 15 years smoking cessation time (N = 887 and N = 445, respectively) an AUC of 0.779 was obtained (Table 4).

Furthermore, for the current smokers (N = 364) we calculated the pack-years (see methods) and used the 13 CpG markers to infer pack-years for ≥ 15 pack-years (N = 210) versus < 15 pack-years (N = 154), which resulted in an AUC of 0.815. For ≥ 10 versus < 10 pack-years (N = 246 and N = 118, respectively) an AUC of 0.846 was obtained (Table 5).

Finally, we combined the pack-years in current smokers, smoking cessation in former smokers with the never smokers (N = 2939) into one model for life-time smoking information inferring. We obtained for the current smokers with ≥ 15 pack-years (N = 210) an AUC of 0.949, < 15 pack-years (N = 154) an AUC of 0.869, in former smokers with ≤ 10 years smoking cessation (N = 311) an AUC of 0.793, with > 10 years smoking cessation (N = 1021) an AUC of 0.739 and the never smokers (N = 1243) an AUC of 0.835 (Table 6). We obtained for the current smokers with ≥ 10 pack-years (N = 246) an AUC of 0.948, < 10 pack-years (N = 118) an AUC of 0.863, former smokers with ≤ 10 years smoking cessation (N = 311) an AUC of 0.794, with > 10 years smoking cessation (N = 1021) an AUC of 0.739, and the never smokers (N = 1243) an AUC of 0.835 (Table 6).

We validated the newly developed prediction models based on the 13 selected CpGs via both internal and external validation procedures. Internal validation was carried out in the model building set using fivefold cross-validation and bootstrapping. For the two-category model (smokers vs. non-smokers), the optimism from bootstrap internal validation was 0.0032, resulting in a bootstrap-adjusted AUC of 0.898 (0.901–0.0032), see Table 2 for other accuracy measures and cross-validation results. For the three-category model (smokers vs. former smokers vs. never smokers) the bootstrap conducted optimisms are 0.0032 for current smokers, 0.0063 for former smokers and 0.0036 for never smokers resulting in bootstrap adjusted AUCs of 0.925 (0.928–0.0032) for current smokers, 0.766 (0.772–0.0063) for former smokers and 0.831 (0.835–0.0036) for never smokers (Table 3). For the smoking cessation time inference in former smoker, (1) for ≥ 5 versus < 5 years smoking cessation the bootstrap optimism was 0.0170 resulting in a bootstrap-adjusted AUC of 0.776 (0.793–0.0170); (2) for ≥ 10 versus < 10 years smoking cessation the bootstrap resulted in an optimism of 0.0112, giving a bootstrap-adjusted AUC of 0.767 (0.778–0.0112); (3) ≥ 15 versus < 15 years smoking cessation the bootstrap resulted in an optimism of 0.0096, giving a bootstrap-adjusted AUC of 0.769 (0.779–0.0096) (Table 4). For the two pack-year models, (1) the bootstrap optimism for ≥ 15 versus < 15 pack—was 0.029 resulting in a bootstrap-adjusted AUC of 0.786 (0.815–0.029); and (2) for ≥ 10 versus < 10 pack-years the bootstrap resulted in an optimism of 0.026, giving a bootstrap-adjusted AUC of 0.820 (0.846–0.026) (Table 5). Finally, for the life-time smoking information inferring, we obtained for ≥ 15 pack-years a bootstrap optimism of 0.0034 resulting in a bootstrap-adjusted AUC of 0.946 (0.949–0.0034), for < 15 pack-years a bootstrap-adjusted AUC of 0.860 (0.869–0.0091), for ≤ 10 smoking cessation a bootstrap-adjusted AUC of 0.782 (0.793–0.0106), > 10 years smoking cessation a bootstrap optimism of 0.0075 resulting in a bootstrap-adjusted AUC of 0.732 (0.739–0.0075) and for never smokers a bootstrap-adjusted AUC of 0.831 (0.835–0.0037) (Table 6). For the second five-category model, very similar results were obtained (Table 6).

External validation was performed in independent samples of two population-based studies, KORA and SHIP-Trend. In KORA (F4, N = 1608), an AUC of 0.911 was achieved for the full 13-CpG two-category model (Table 2). In SHIP-Trend (N = 244), an AUC of 0.888 was obtained for the two-category model based on a subset of ten CpGs, since the EPIC-array applied for SHIP-Trend is missing three of the 13 CpGs (cg06126421, cg22132788 and cg05951221). This 10-CpG model in the model building set gave a cross-validated average AUC of 0.893 ± 0.012 (Table 2). External validation of the three-category model in the KORA study (F4, N = 1608) achieved an AUC of 0.914 for the current smokers (N = 226), 0.699 for the former smokers (N = 707), and 0.781 for the never smokers (N = 675) (Table 3). The three-category model validation in SHIP-Trend for the 10-CpG model resulted in an AUC of 0.882 for current smokers (N = 51), 0.654 for former smokers (N = 92), and 0.778 for never smokers (N = 101) (Table 3). For comparison, in the model building set, this three category 10-CpG model gave a cross-validated average AUC of 0.919 ± 0.019 for current smokers, 0.748 ± 0.023 for former smokers, and 0.823 ± 0.018 for never smokers (Table 3). External validation of smoking cessation time inference in former smokers in the KORA study (N = 652) resulted in an AUC of 0.760 for ≥ 5 versus < 5 years of smoking cessation time, an AUC of 0.764 for ≥ 10 versus < 10 years of smoking cessation time, and of 0.754 for ≥ 15 versus < 15 years of smoking cessation time (Table 4). Furthermore, we externally validated the prediction of pack-years in the current smokers of the KORA study (F4, N = 224) and obtained an AUC of 0.752 for inferring ≥ 15 versus < 15 pack-years and an AUC of 0.796 for ≥ 10 versus < 10 pack-years (Table 5). The pack-year validation in the current smokers of SHIP-Trend (N = 41) for the 10-CpG model resulted in an AUC of 0.779 for ≥ 15 versus < 15 pack-years (AUC of 0.757 ± 0.077 in the model building set) and an AUC of 0.837 for ≥ 10 versus < 10 pack-years (AUC of 0.809 ± 0.039 in the model building) (Table 5). The external validation of the five-category models in the KORA study resulted for the current smokers with ≥ 15 pack-years in an AUC of 0.955, for < 15 pack-years an AUC of 0.710, for ≤ 10 years smoking cessation an AUC of 0.791, > 10 years smoking cessation an AUC of 0.650 and for never smokers an AUC of 0.788. For the second five-category model, we obtained in the KORA study an AUC of 0.943 for ≥ 10 pack-years, of 0.694 for < 15 pack-years, an AUC of 0.791 for ≤ 10 years smoking cessation, of 0.651 ≥ 10 years smoking cessation and an AUC of 0.788 for never smokers (Table 6).

In a subset of 488 participants for which we had CpG, cotinine and smoking information available, we compared our validated CpG-based prediction model for current versus non-smokers with the use of a cotinine cut-off to determine current smoking, using smoking information from self-reported questionnaires as reference. Using our CpG-model, we accurately inferred 87 of the 140 smokers and 344 of the 348 non-smokers (sensitivity of 0.621 and specificity of 0.989) compared to 105 of the 140 smokers and 342 of the 348 non-smokers using the cotinine level cut-off of 50 ng/mL (sensitivity of 0.750 and specificity of 0.983). Out of the 87 accurately inferred smokers with our CpG model, 75 (86%) were also accurately selected as smokers based on cotinine, and out of the 105 participants correctly selected with cotinine as smokers, 75 (71%) were accurately inferred as smokers with our CpG model. For the non-smokers, out of the 344 accurately inferred with our CpG model, 340 (99%) were also selected with cotinine as non-smokers, and 340 (99%) out of the 342 accurately selected non-smokers with cotinine, were accurately inferred as non-smokers with our CpG model. Finally, when comparing all three methods(questionnaires/cotinine levels/DNA methylation prediction), 340 participants were highlighted as non-smokers and 75 as smokers with all three methods, 12 were selected as smokers based on questionnaires and DNA methylation inference, 30 as smokers with both questionnaires and cotinine, 2 were determined as smokers with both cotinine and DNA methylation inference, whereas 23 were determined as smokers with questionnaires only, 2 as smokers with DNA methylation inference only, and 4 as smokers with cotinine only.

Next, we investigated the putative effect of prenatal smoking exposure and passive smoking on the epigenetic inference of smoking habits achievable with our validated model. When applying our model to the DNA methylation data at time of birth collected from cord blood, the proportion of children accurately inferred as non-smokers was surprisingly low at 0.114 (N = 1111) (Online Resource 1: Table S6). We then classified children whose mothers smoked throughout pregnancy as “smokers”, and obtained an AUC of 0.773, with a high sensitivity of 0.981 and a low specificity of 0.131. The AUC decreased to 0.664 when additionally considering mothers who stopped smoking when they became aware of their pregnancy (generally in the first trimester), and decreased further to 0.591 when additionally considering passive smoking of the mother during pregnancy; assessing the latter solely, an AUC of 0.460 was obtained, reflecting random prediction.

Additionally, we applied our model to data of children from the Generation R Study obtained from blood collected at the ages of six (N = 355) and ten (N = 309) years. In contrast to the results for newborns obtained from cord blood, we found that the proportion of 6- and 10-year-old children accurately inferred as non-smokers with our model was very high at 0.994 for both age groups (Table 7). This suggests no impact of prenatal smoking exposure nor passive smoking exposure during early childhood on the model performance. Subsequently, we applied our model to those 197 children for which epigenetic data were available from serial samples collected at birth, 6, and 10 years of age. The proportion of children that with our model accurately inferred as non-smokers at birth was 0.112, whereas it was 0.994 at six and 0.995 at 10 years of age, which was highly similar to the results obtained from the total datasets available for these three time points. The β-values per CpG for the model building set and the three time points in Generation R are shown in Online Resource 3: Figures S1–15.