Proteomic signatures of smoking and their associations with risk of incident diseases and mortality in diverse populations

Fig. 1. Overview of the study design and analytic approach. A gradient boosting classification model was built in the 70% randomly selected UKB population, differentiating current smokers and never smokers. Boruta feature selection algorithm was then used to select only relevant features for downstream analysis. B The model trained in UKB training dataset was further validated externally in CKB male population. C proteomic Smoking INdex (pSIN) was calculated for the whole UKB cohort. Smoking behavioural, genomic, and exposome determinants of pSIN were studied, followed by associating pSIN with incident health outcomes.

Fig. 2. Protein profile differentiating current and never smokers and its determinants. A shows the performance of the model in UKB 30% left-out testing dataset and CKB external validation dataset. B shows a summary plot for the SHAP value of the top 20 selected proteins. Each dot denotes a participant, the colour of the dots denotes the protein expression level, and the X-axis denotes its contribution to the model decision. Proteins were ranked by means of the absolute SHAP value. C shows the distribution of pSIN in the whole UKB population. Dotted line denotes the cut-off value at FPR 0.05 when differentiating current and never smokers. D Four gradient boosting models were employed to assess the contribution of intrinsic and extrinsic factors to pSIN. The additional variance explained by each category was calculated by subtracting the variance explained by the previous model from the variance explained by the current model.

Fig. 3. Quartiles of pSIN lead to strongly diverging cumulative incidence of major incident diseases and mortality. Cumulative incidence plot of top, median and bottom 25% of the pSIN in the whole UKB population with 95% confidence interval shown as lighter shading. X-axis denotes the chronological age and Y-axis denotes the cumulative incidence. Cumulative incidence and number at risk at each age point are shown in Supplementary Data 15 and Supplementary Data 16.

Fig. 4. pSIN differentiates future risks of morbidities and mortality in current and previous smokers. A shows association between pSIN and health outcomes which has at least 80 cases during follow-up time using multi-variate cox proportional hazard model in current smokers. In model 1, the exposure was pSIN without adjusting for any covariate as age and sex already been regressed out in protein level. Model 2 was adjusted for recruitment centre, Townsend deprivation index, IPAQ physical activity group, ethnicity, alcohol frequency, BMI and education years. Model 3 further adjusted for smoking pack years and smoking cessation time. Hazard ratios with 95% confidence intervals were shown. P-values were corrected for FDR multiple testing and non-significant associations after corrections were shown as grey colour. B shows association between pSIN and health outcomes in previous smokers.

Fig. 5. Previous smokers with pSIN similar to never smokers (recovery) show lower morbidity and mortality risks. A shows the percentage of the previous smoker population in each year of smoking cessation bin who have a pSIN lower than the cutoff value used to differentiate current smokers. This percentage first reaches 80% after 10 years of smoking cessation. B shows the distribution of pSIN in previous smokers. Dotted line denotes the cut-off when differentiating current smokers from never smokers at FPR of 0.05, dividing previous smokers into two groups. Hashed part denotes the group of previous smokers with a similar pSIN as never smokers (recovery). C shows a cumulative incidence plot of low and high-risk groups defined by pSIN in previous smokers (orange and blue) with self-reported current smokers as positive control (yellow) and self-reported never smokers as negative control (purple). A 95% confidence interval was shown in shaded area. Cumulative incidence and number at risk charts are shown in Supplementary Data 24 and 25, respectively.