Exploring the link between walking and lung cancer risk: a two-stage Mendelian randomization analysis

To address the potential pleiotropic effects of genetic variants, we used 3 MR methods in this study to assess the causal effects of the exposure of interest on the target outcome. The inverse variance weighted (IVW) approach was used as the principal analysis, which combined the Wald ratio of each SNP to the outcome and gained summary causal estimates. This approach allows for over-dispersion. In addition, other MR methods supplementary to the IVW, such as MR-Egger regression and weighted median, can provide more reliable estimates in a wider range of situations. MR-Egger regression can provide tests for unbalanced pleiotropy and considerable heterogeneity, while it requires a larger sample size for same under-exposure variation [22]. The weighted median approach offers consistent effect estimates when the weighted variance provided by the horizontal pleiotropy is at least half valid [23].

Horizontal pleiotropy will occur when genetic variation associated with the exposure of interest directly affect the outcome through multiple channels other than the hypothesized exposure. Therefore, we further performed Cochran’s Q test, funnel plot, leave-one-out analysis, MR-PRESSO and MR-Egger intercept test to detect the pleiotropy and to assess the robustness of the results. Specifically, heterogeneity would be identified if the p-value of the Cochran Q test was less than 0.05, at which point a random effects model was required. The horizontal pleiotropy would be detected if the p-value of the MR-Egger intercept test and MR-PRESSO were less than 0.05. We also performed a leave-one-out analysis by which each exposure-related SNP was discarded in turn to repeat the IVW analysis, so as to determine whether causal estimates were driven by any single SNP.

The Bonferroni-adjusted p-values below 0.01 (0.05/5 exposures) were considered significant and showed a strong association when lung cancer was taken as an outcome. And p values of 0.01–0.05 was considered potentially significant.

All analyses were performed through the packages TwoSampleMR (version 0.5.6), gwasglue (version 0.0.0.9000) and VariantAnnotation (version 1.44.1) in R (version 4.2.2).

In our analysis of the relationship between these factors and lung cancer risk, we finally extracted 7, 3, 23, 13 and 100 SNPs associated with number of days/week walked for 10 + minutes, duration of walks, usual walking pace, fed-up feelings, and time spent watching TV as IVs, respectively. This selection was made following the exclusion of SNPs with F statistic below 10 and those associated with lung cancer's confounding factors (smoking, alcohol, family history of lung cancer, air pollution, IPF, COPD and occupational exposure). When analyzing the relationship between number of days/week walked for 10 + minutes, usual walking pace, fed-up feelings and lung cancer, we excluded rs3004179, rs13107325, rs1652376, rs34898535, rs7124682, rs7896518, rs9972653, and rs4630591, which were all associated with alcohol consumption. When analyzing the relationship between time spent watching TV and lung cancer, we excluded SNPs associated with smoking (rs71658797 and rs75499503) and alcohol (rs11714337, rs2352984, rs4339469, rs62199883, and rs872169). However, we did not find SNPs associated with other confounders.

The corresponding F-statistic ranges were 116–158, 30–38, 11–20, 10–16 and 16–46 respectively. These values exceeded the standard cutoff values (> 10), indicating a robust instrumental strength. In the analysis of the relationship between number of days/week walked for 10 + minutes and fed-up feelings, we extracted 6 relevant SNPs as IVs. Similarly, when analyzing the association between usual walking pace and time spent watching TV, we extracted 21 relevant SNPs as IVs. The corresponding F-statistic values were all higher than the cutoff values (> 10).

Details on the variants used to construct the above phenotypes’ IVs, are shown in Supplementary File 1.

The IVW results showed that number of days/week walked for 10 + minutes (OR = 0.993, 95% CI = 0.987–0.998, P = 0.009) and time spent watching TV (OR = 1.009, 95% CI = 1.005–1.014, P = 2.9e-5) affected the risk of lung cancer. Usual walking pace (OR = 0.989, 95% CI = 0.980–0.998, P = 0.015) and fed-up feelings (OR = 1.021, 95% CI = 1.005–1.037, P = 1.3e-7) were potentially significant in relation to lung cancer risk. In addition, number of days/week walked for 10 + minutes was significantly associated with fed-up feelings (OR = 1.095, 95% CI = 1.063–1.128, P = 2.5e-9), but fed-up feelings did not affect number of days/week walked for 10 + minutes (OR = 1.116, 95% CI = 0.722–1.725, P = 0.621). Usual walking pace was significantly associated with time spent watching TV (OR = 0.821, 95% CI = 0.771–0.875, P = 8.9e-10), and time spent watching TV may affect usual walking pace (OR = 0.714, 95% CI = 0.685–0.744, P = 3.4e-58). Details are shown in Table 2 and Fig. 1.

The total effect of the exposures to the outcome was β1, the effect of the exposures to the mediators was β2, and the effect of the mediators to the outcome was β3. Then the total effect of the exposures to the outcome was β1, the mediating effect was β2*β3, the direct effect was β1-β2*β3, and the share of the mediating effect share was β2*β3/β1.

When analyzing the effect of number of days/week walked for 10 + minutes on the risk of lung cancer, the mediating effects of fed-up feelings was 0.00186, accounting for 25.0%. But fed-up feelings could not affect number of days/week walked 10 + minutes. When analyzing the effect of usual walking pace on the risk of lung cancer, the mediating effect of time spent watching TV was 0.00192, accounting for 17.5%. Interestingly, time spent watching TV may affect usual walking pace and then lung cancer risk, in which the mediating effect of usual walking pace was 0.0037, accounting for 38.0%. When analyzing the impact of number of days/week walked for 10 + minutes, fed-up feelings, usual walking pace, and the time spent watching TV on lung cancer risk, there was no overlap of SNPs between IVs associated with the direct factors and those associated with the mediating factors. Details are shown in Fig. 2. In addition, all data in this study were obtained from UK Biobank. In order to assess the reliability of the study, we explored the impact of sample overlapping rate on the conclusions in Bias and Type 1 error rate for Mendelian randomization with sample overlap (https://​sb452.​shinyapps.​io/​overlap/​). However, previous studies haven’t showed the possible OR between exposures and outcomes, so we hypothesized the OR of possible protective factors was 0.3 or 0.9, and the OR of possible risk factors was 1.1 or 3.0 when lung cancer is the outcome. The same method was adopted when the outcome is other factors. The results suggested that, even though all data in this study are from UK Biobank, the sample overlapping rate may not affect the reliability of the conclusions, especially when compared with the OR obtained in this study. Moreover, when analyzing the relationship between number of days/week walked for 10 + minutes and fed-up feelings, the sample overlapping rate may also not affect the final conclusion. However, when analyzing the relationship between usual walking pace and time spent watching TV, the conclusion may be greatly affected by the sample overlapping rate, possibly because the sample overlapping rate and time spent watching TV are categorical ordinal data, rather than binary or continuous variables. The results are shown in Table 3.

A series of sensitivity analyses were performed to assess the robustness of the above results, including Cochran’s Q test, MR-Egger intercept test, leave-one-out analysis, MR-PRESSO and funnel plot, among them, MR-Egger intercept test had p-values > 0.05 and did not detect pleiotropy. However, when analyzing the relationship between usual walking pace and time spent watching TV, the P-value of MR-PRESSO was still less than 0.05 after removing the SNP with pleiotropy. Some of Cochran’s Q-tests showed heterogeneity, for which the data were restricted to a random effects model only (Table 4). The leave-one-out analysis showed no SNP-driven results (Fig. 3), and the funnel plot was symmetric (Fig. 4).