Tobacco smoking is responsible for approximately 85% of lung cancer cases worldwide and nearly 90% of lung cancer deaths. While global smoking rates have declined from 32.7% of adults in 2000 to 21.7% in 2020, former smokers continue to face elevated lung cancer risk that can persist for decades after quitting.
The data on post-cessation risk is sobering. Analysis of the Framingham Heart Study found that 40.8% of lung cancers in former smokers occurred more than 15 years after quitting. Even after 25 years of cessation, lung cancer risk in former smokers remains over three times higher than in never-smokers. Excess risk declines slowly, plateauing at about 20% above baseline only after 20 years of abstinence.
Quitting earlier substantially helps. Stopping before age 40 reduces the risk of tobacco-related death by approximately 90%. Cessation at ages 60, 50, 40, or 30 years extends life expectancy by approximately 3, 6, 9, or 10 years, respectively. Despite these benefits, the large population of former smokers worldwide represents an ongoing, substantial public health burden.
The core scientific puzzle is why some smoking-induced molecular changes resolve after cessation while others become permanent molecular scars that sustain long-term cancer risk. This review synthesizes current knowledge about which changes are persistent versus reversible, and how artificial intelligence and multi-omics data integration can advance prevention in this high-risk population.
Tobacco smoke is a complex chemical mixture containing over 7,000 compounds, at least 69-80 of which are recognized carcinogens. These include polycyclic aromatic hydrocarbons (PAHs), tobacco-specific nitrosamines, heavy metals like arsenic and cadmium, and radioactive elements such as polonium-210. Together, these substances cause DNA damage, oxidative stress, and chronic inflammation across the respiratory tract.
The mechanisms of carcinogenesis are specific. PAHs form DNA adducts in the bronchial epithelium that generate characteristic mutations - particularly G to T transversions - at TP53 hotspots, consistent with squamous cell lung cancer. Tobacco-specific nitrosamine NNK acts as a systemic lung carcinogen that primarily induces adenocarcinoma. Nicotine and certain nitrosamines also activate signaling pathways that promote survival of damaged cells, favoring clonal expansion.
Field cancerization describes how these accumulated molecular insults spread across the entire respiratory tract. Large areas of airway cells harbor pre-neoplastic changes even when no visible tumor exists - representing a field of injury where cancer is more likely to arise. This concept explains why former smokers remain at elevated risk long after the direct chemical exposure has ended.
Even minimal exposure causes detectable molecular changes. As little as 0.1 pack-years of exposure produces measurable transcriptomic changes in small airway epithelium. Genes such as PLA2G10 and CXCL6 respond to nicotine metabolite levels below 2 ng/mL in urine, demonstrating that there is no truly safe threshold below which tobacco smoke causes no molecular damage.
Many smoking-induced molecular changes are reversible - particularly those in xenobiotic metabolism and acute stress response pathways. A core set of nine genes, including CYP1B1, ALDH3A1, AKR1B10, and MUC5AC, consistently return to normal expression after cessation. Xenobiotic metabolism genes like CYP1A1 and CYP1B1 begin reverting toward baseline within just four weeks of quitting.
The timeline of recovery varies. By eight weeks after cessation, 88.2% of smoking-upregulated gene expression changes show downregulation, with metabolic and antioxidant profiles resembling never-smokers after approximately two years. Short-term cessation (3-6 months) leads to global decreases in DNA methylation at 3,878 CpG sites, with these changes correlating with improved lung function and reduced inflammatory biomarkers.
Epigenetic and microRNA recovery is also substantial. Approximately 65% of smoking-altered microRNAs in small airway epithelium returned to baseline within three months of quitting. Analysis of blood DNA methylation identified 602 nonpersistent versus 149 persistently differentially methylated CpG sites, meaning recovery at most sites eventually occurs, though the pace varies dramatically.
The reversible changes have clinical significance beyond simply illustrating that cessation works. Identifying which molecular pathways revert reveals potential intervention targets - if a pro-cancer pathway normalizes after quitting, it may be targetable with chemoprevention drugs before permanent damage takes hold. The PI3K signaling pathway, which normalizes after months of cessation, represents one such candidate for chemoprevention intervention.
DNA mutations are the most permanent consequence of smoking. Approximately 62% of former smokers (average cessation of 27 months) harbor clonal genetic alterations in histologically normal lung tissue, including loss of heterozygosity at cancer-associated genes. Whole-genome sequencing of bronchial epithelial cells shows that tobacco exposure adds thousands to tens of thousands of mutations per cell, and these alterations persist in affected cell lineages indefinitely.
A subset of gene expression changes refuses to normalize. In one 12-month study, 53 (11%) of 475 smoking-dysregulated genes did not return to normal, with apoptosis and proliferation pathways most resistant. Strikingly, 13 genes remain persistently altered even 20-30 years after cessation, including decreased expression of potential tumor suppressors TU3A and CX3CL1 and increased expression of oncogenes HN1 and CEACAM6.
Epigenetic methylation scars persist for decades. Epigenome-wide studies using whole blood DNA identified 149 CpG sites that remain differentially methylated more than 35 years after cessation. The sentinel site cg05575921 in the AHRR gene shows approximately 24% hypomethylation in current smokers and only gradually approaches never-smoker levels over years to decades. Hypomethylation at AHRR and F2RL3 loci was significantly associated with future lung cancer risk in prospective cohort studies, even after adjusting for smoking status.
Epigenetic aging is accelerated by smoking and only partially reverses. Smoking increases the epigenetic age of airway cells by an average of 4.9 years and lung tissue by 4.3 years. While airway cells partially reverse this age acceleration after cessation, lung tissue does not fully reverse, suggesting that long-lived lung cells retain smoking-induced molecular damage that maintains a pro-oncogenic tissue environment even in long-term former smokers.
Immune dysfunction persists long after smoking cessation. Neutrophil-mediated immunity and interferon-gamma-related pathways remain dysregulated for over 10 years after quitting, contributing to sustained lung cancer risk. While innate immune responses may partially normalize, adaptive immune changes - particularly cytokine response patterns from T cells - remain altered in former smokers, potentially linked to persistent epigenetic memory in immune cells.
Structural lung changes become irreversible. Animal model studies demonstrate persistent alveolar enlargement, right ventricular hypertrophy, ongoing inflammation, macrophage accumulation, and neutrophilic inflammation that persist six months to a year after smoking cessation ends. Some measures of progressive alveolar damage lasted longer than the original exposure period itself, suggesting that tissue remodeling generates self-sustaining pathological processes.
A smoking-induced EGFR signaling loop was identified in airway epithelium that is absent in never-smokers. This autocrine EGFR-amphiregulin feedback loop drives basal cell hyperplasia and squamous metaplasia - early histopathological changes associated with cancer development. This self-amplifying loop may maintain epithelial remodeling long after direct chemical exposure ends, representing a mechanistic explanation for persistent risk.
MicroRNA dysregulation also persists. Of 34 microRNAs altered by smoking in small airway epithelium, 12 remained dysregulated after three months of cessation. These persistent microRNAs primarily regulate the Wnt/beta-catenin signaling pathway, which plays central roles in cell proliferation, differentiation, and cancer development - linking the epigenetic molecular scar to ongoing aberrant cell signaling.
Artificial intelligence offers capabilities that conventional statistical methods cannot match for this challenge. The molecular landscape of former smokers spans genomic, epigenomic, transcriptomic, proteomic, and metabolomic data layers with high inter-patient variability. Machine learning and deep learning models can integrate these fragmented multi-omics signals and identify coordinated patterns across numerous features rather than analyzing one marker at a time.
Transformer neural networks and graph neural networks are particularly suited for synthesizing multimodal data such as imaging combined with pathology or genomics. These architectures can link persistent epigenetic marks (such as DNA methylation at AHRR and F2RL3) with transcriptomic changes to model how they collectively sustain a pro-tumorigenic tissue environment. An AI model trained on longitudinal multi-omics profiles could predict which individuals follow a persistence-prone molecular trajectory requiring closer surveillance or chemoprevention.
AI enables what are called "virtual biopsies" - predicting molecular features from routine clinical data or images without requiring invasive sampling. By training on paired radiologic and molecular data, AI models can predict methylation status, mutation burden, or expression-based biomarkers from CT scan features. This concept has already demonstrated feasibility: deep learning models can predict clinically relevant mutations (EGFR, STK11, KRAS) from standard pathology slides in lung cancer, and can predict EGFR mutation status and PD-L1 expression from CT imaging.
Representative AI tools span the entire prevention workflow - from the Sybil deep learning model that predicts lung cancer risk from low-dose CT scans, to commercial clinical decision support platforms (Tempus Lens, FoundationOne CDx) that integrate genomic and clinical data for treatment decisions. Multi-omics integration tools like MOFA+ and WGCNA enable systematic analysis of coordinated patterns across data layers, while biomarker discovery tools provide pathway enrichment and network visualization of molecular interactions.
Despite rapid technological advances, clinical translation of AI-based multi-omics approaches for former smokers faces substantial challenges. Most existing datasets are ancestry- and geography-biased, with baseline methylation and expression profiles varying across populations in ways that limit how well models trained in one cohort generalize to others.
The definition of persistence itself is a methodological problem. Persistence labels typically require longitudinal follow-up data to demonstrate temporal stability - meaning that labeling a molecular change as persistent versus reversible requires studying the same patients over time, which is logistically difficult and reduces effective sample sizes. Studies have used variable follow-up durations ranging from months to over a decade, making direct comparison across studies difficult.
Statistical association does not establish causality. Most proposed molecular signatures remain observational - even strong associations between a biomarker (like AHRR methylation) and lung cancer risk do not prove that the biomarker mechanistically drives carcinogenesis rather than simply reflecting prior exposure. AI-derived composite signatures integrating numerous correlated features face an even greater causal inference challenge.
Missing data across multi-omics layers is a persistent technical challenge, since variable assay availability and quality control issues degrade model performance. Persistent smoking-associated alterations may also arise from clonal expansion of long-lived altered cell populations, meaning that ground-truth labels often reflect complex mixtures of cell states rather than discrete binary categories - complicating machine learning model training.
The evidence synthesized in this review establishes that persistent smoking-induced molecular alterations represent a distinct biological state that is not fully captured by smoking history alone. A former smoker decades after cessation carries a molecular landscape that differs meaningfully from both current smokers and never-smokers - with lasting DNA mutations, methylation scars, dysregulated microRNAs, and immune remodeling that collectively explain the persistent excess cancer risk.
Recognizing the persistent versus reversible distinction has direct clinical applications. Persistent biomarkers like AHRR and F2RL3 methylation can stratify former smokers by ongoing molecular cancer risk, independent of years since cessation. Reversible molecular alterations point toward intervention targets where chemoprevention - pharmaceutical or dietary - might intercept the carcinogenic process before permanent transformation occurs.
Future progress requires larger, diverse, longitudinal cohorts with standardized definitions of molecular persistence. Longitudinal sampling is essential for modeling molecular recovery as a dynamic process rather than a binary before-and-after comparison. Models must account for genetic susceptibility, cumulative exposure history, and tissue type when making predictions, as these factors modulate which alterations persist in which individuals.
The ultimate opportunity is to translate these biological and computational insights into practical tools that guide lung cancer screening, personalized surveillance schedules, and chemopreventive interventions for the millions of former smokers worldwide whose elevated risk has no specific medical management beyond standard low-dose CT screening criteria. AI-based precision prevention tools designed specifically for this population represent a meaningful near-term clinical priority.