Review Papers Claim by Claim

Paper Review (Skeptical, Evidence-First): Epigenetic Signatures of Cigarette Smoking

What the paper claims (grounded in its reported results)

• In 16 cohorts (n=15,907), compared to never smokers, current smoking is associated with 2,623 CpGs Bonferroni-significant (p<1×10^-7), annotated to 1,405 genes; additionally, 18,760 CpGs show differential methylation at FDR<0.05 (annotated to 7,201 genes). Bonferroni/FDR counts (current vs never)
• Compared to never smokers, former smoking retains a signature: 2,568 CpGs (1,326 genes) at FDR<0.05 among CpGs significant in current vs never, with 185 CpGs also meeting Bonferroni correction. Former vs never persistence counts
• In Framingham (FHS) using long history of cessation, most smoking-associated CpGs return toward never-smoker levels within ~5 years; however, 36 CpGs across 19 genes do not return even after up to 30 years of cessation. Decade-scale persistence in FHS
• Functional interpretation in the paper relies on (i) gene-set enrichment of pathways associated with smoking and (ii) methylation-by-expression (MxE) cis-proximity integration identifying CpGs associated with expression at nearby genes (in cis windows). Pathway + MxE integration results

Visualization 1 — How many smoking-associated CpGs?

Visualization 2 — Persistence after cessation (FHS long follow-up)

Interpretation map — where the evidence supports the story, and where it doesn’t

1) Strengths that directly support the statistical signal

• Large multi-cohort sample size and harmonized meta-analysis improves power and reduces reliance on any single study’s idiosyncrasies. Large cohort meta-analysis framework
• Cell-type adjustment is explicitly addressed (in models where whole blood was used) and results are reported as comparable pre/post adjustment. Cell-type adjustment and coefficient correlation
• Random-effects meta-analysis is used, which is conservative under heterogeneity assumptions (the paper states the conservative nature when heterogeneity is present). Random-effects meta-analysis rationale

2) Key blind spots and failure modes (what could still produce false or overconfident claims)

• Observational confounding and tissue mismatch: methylation is measured in blood (mixed leukocytes or specific cell types in subsets). Even with cell-type adjustment, residual confounding by smoking-linked physiology (inflammation, metabolic changes, hematopoietic shifts) can remain. The paper itself flags blood as an accessible proxy but notes additional interest in target tissues. Stated limitations and tissue confounding concerns
• Smoking history measurement error: smoking status is self-reported with categorization rules. Misclassification can attenuate true effects or create apparent effects if misclassification correlates with covariates. (The paper uses specific definitions; it is still subject to measurement error.) Self-reported smoking phenotype definitions
• Selection bias across cohorts / platform heterogeneity: cohorts use different sample types (whole blood, CD4+ T cells, monocytes) and have different covariate availability (e.g., BMI added only in a subset). This can produce differential power and slightly different estimands. Cohort heterogeneity and conditional covariate inclusion
• Multiple testing & enrichment overinterpretation: the paper reports very large significant CpG sets and uses enrichment approaches. Enrichment can be real, but it can also be driven by biases in gene annotation, CpG-to-gene mapping, or correlated testing structures. The results therefore support “biological plausibility” more than definitive mechanism. Enrichment-based interpretation framework

What would most likely change the conclusion?

• If independent consortia with matched design and improved smoking exposure quantification (and ideally direct longitudinal sampling from relevant tissues) showed substantially fewer persistent CpGs after cessation, the “decades persistence” emphasis would weaken. What persistence is based on + limitations
• If methylation-by-expression (MxE) associations systematically fail to replicate in cell-type-matched expression data or cis-window definitions, the “functional impact” interpretation would become weaker (though biomarkers could remain plausible). MxE integration is associative and cis-window based
• If trait-enrichment is found to be strongly driven by correlated lifestyle variables (not fully captured by cohort covariates) or by residual cell-composition structure, the bridge from methylation to disease mechanisms would be less direct. Residual confounding and indirect-effect limitations

User-actionable next steps (no interventions recommended)

Author reviews (direct to BGPT)