Epigenetic fingerprints link early-onset colon and rectal cancer to pesticide exposure

The incidence of colorectal cancer (CRC) is rising rapidly in people younger than 50 years. Although this increase parallels shifts in lifestyle and environmental factors—collectively termed the exposome—whether these are indeed linked to the development of early-onset CRC (EOCRC) remains uninvestigated. Due to limited exposome data in most cancer cohorts, we constructed weighted methylation risk scores as proxies for exposome exposure to pinpoint specific risk factors associated with EOCRC compared to late-onset CRC (LOCRC) patients diagnosed at ≥70 years. Our analysis confirmed previously identified risk factors, including educational attainment, diet and smoking habits. Moreover, we identified exposure to the herbicide picloram as a new risk factor (adjusted P = 4.4 × 10 −4 ) in the discovery cohort (31 EOCRC versus 100 LOCRC), which was replicated in a meta-analysis comprising nine CRC cohorts ( P = 3.1 × 10 −3 ; adjusted P = 1.5 × 10 −2 ; 83 EOCRC versus 272 LOCRC). Subsequently, we analyzed population-based data from 94 US counties over 21 years and validated the association between picloram use and EOCRC incidence ( P = 4.52 × 10 −4 ), which remained significant after adjusting for socioeconomic factors and other pesticide use. These findings highlight the critical role of the exposome in EOCRC risk, underscoring the urgency for targeted personal and policy-level interventions.

Globally, CRC ranks as the third leading cancer type and the second most common cause of cancer-related death 1 . As CRC is an aging-associated disease, the rates of CRC incidence and death grow steadily with age, with an estimated 90% of worldwide cases and deaths occurring in people over 50 years old 2 . However, in recent years, worldwide cancer registries have reported a disproportionate increase in the incidence of EOCRC, generally defined as CRC diagnosed in people younger than 50 years of age 3 , 4

EOCRC presents unique clinical and pathological characteristics in comparison with CRC in older patients, with a predominance of rectal and left colon tumors, higher prevalence of synchronous and metachronous CRC, higher frequency of metastatic disease at diagnosis and more aggressive and less differentiated tumors 5 . However, genomic alterations are largely similar to those described in older patients, although several genes exhibit different alteration rates 6 .

As our understanding of risk factors for CRC expands, so does the evidence that lifestyle factors, such as diet and exercise, can substantially influence CRC risk. This has led to the hypothesis that changes in lifestyle and environmental exposures—collectively known as the exposome—may play a critical role in the rising incidence of CRC in young patients. Nonetheless, efforts to identify modifiable risk factors specific to EOCRC have met with only limited success 7 , as most studies compare EOCRC cases with controls and identify the same risk factors as those for LOCRC cases 8 . This limited success could be attributed to the lack of quantitative measurements of exposome traits in cancer cohorts, which consequently makes it challenging to establish clear associations with cancer risk. Emerging studies have identified exposome-induced changes in CpG site-specific DNA methylation levels 9 , providing a new avenue for investigation. As DNA methylation data is frequently available in cancer cohorts, exposome-related methylation changes could serve as a valuable proxy for direct exposome measurements. This approach may enhance our ability to pinpoint specific exposome risk factors and improve our understanding of their role in EOCRC development.

In this study we aimed to explore the exposome traits that may contribute to the development of EOCRC compared to LOCRC, herein defined as CRC diagnosed in people 70 years and older. To this end, we constructed methylation risk scores (MRSs) for lifestyle and environmental factors using DNA methylation data from The Cancer Genome Atlas (TCGA) as the discovery cohort and conducted a replication meta-analysis across nine independent cohorts. The resulting MRSs were then compared between EOCRC and LOCRC. The relationship between pesticide use intensity and EOCRC incidence in the USA was further investigated using population-based data.

For the discovery phase of this study, we utilized colon adenocarcinoma (COAD) samples obtained from TCGA 10 . The subsequent replication phase was carried out through a meta-analysis, incorporating data from studies on colon cancer ( GSE131013 , GSE42752 , E-MTAB- 7036 and GSE199057 ) 11 , 12 , 13 , 14 , rectal cancer (TCGA-READ and GSE39958 ) 15 and colorectal cancer ( GSE101764 , GSE77954 and E-MTAB- 3027 ) 16 , 17 , 18 . Patients were classified as EOCRC ( 1 , with a detailed discovery cohort characterization in Supplementary Table 1 .

The exposome’s association with EOCRC versus LOCRC was evaluated using 29 lifestyle and environmental factors, which we define here as the exposome, acknowledging its broader scope. The analyzed traits encompassed 11 lifestyle factors: the Alternative Healthy Eating Index (AHEI), alcohol consumption, birthweight, body mass index (BMI) (continuous variable in kg m −2 ), cannabis use, coffee consumption, education level, Mediterranean Diet Score (MDS), obesity (defined as ≥30 kg m −2 ), smoking habits and smoking inference model (smoking-Maas). Furthermore, we examined four air pollution particles: nitrogen dioxide (NO 2 ), polychlorinated biphenyls (PCBs) and particulate matter (PM) 2.5 ) and between 2.5 µm and 10 µm (PM 2.5–10 ). In addition, we included 14 pesticides encompassing 2,4-dichlorophenoxyacetic acid (2,4-D), atrazine, acetochlor, chlordane, dicamba, malathion, dichlorodiphenyltrichloroethane (DDT), heptachlor, lindane, glyphosate, mesotrione, metolachlor, picloram and toxaphene. For the marker selection, we selected for each trait significantly associated CpG sites (CpGs) from extensive epigenome-wide association studies (EWAS), employing various significance thresholds, namely genome-wide (GW: P −7 ), P −5 and false discovery rates (FDR) of 19 ; no additional selection threshold was applied in the current study.

Using available EWAS summary statistics across the five marker selection thresholds, we identified 63 exposome CpG sets across 29 exposome traits and computed 63 weighted MRSs from DNA methylation beta-values adjusted for Horvath epigenetic age 20 . Although termed ‘risk scores,’ these MRS serve as proxies for exposome exposure, enabling the assessment of exposure prevalence in early- versus late-onset tumors rather than estimation of individual cancer risk. The number of CpGs and their weights can be found in Supplementary Tables 2 and 3 .

The limited availability of exposome data in tumor samples hinders the evaluation of its impact on EOCRC and restricts direct validation of the constructed MRSs against measured exposures. To address this, we performed a comprehensive validation of the constructed MRSs across several datasets, with all detailed results summarized in Supplementary Table 5 .

Validation was performed by reconstructing the MRS in a dataset with available exposure measurements and assessing concordance between MRS values and the corresponding recorded exposure. By leveraging TCGA data, we observed a higher alcohol MRS in liver cancer patients with alcoholic liver disease (ALD), and a separation of current and former smokers from never smokers with the smoking-Maas across several cancer types (Supplementary Table 4 and Extended Data Fig. 1 ). We further validated the use of MRSs as proxies for cannabis dependence, higher PM 2.5 median exposure, and smokers (smoking-Maas) using publicly available datasets, including noncancer participants (Extended Data Fig. 1 and Supplementary Table 5 ).

Beyond the exposome traits included in our study, we examined additional traits using EWAS and publicly available datasets with DNA methylation and measured exposome data, including allergic asthma, major depressive disorder (MDD), polygenic risk scores (PRS) for depression, as well as prenatal exposures, including birthweight, gestational age, maternal education, maternal BMI and arsenic exposure. The constructed MRSs for these traits showed concordance with directly measured exposures, further strengthening the validity of MRSs as a proxy for exposure (Supplementary Table 5 and Extended Data Figs. 2 – 3 ).

To elucidate the exposome’s impact on early-onset colon and rectal cancer cases, we compared the 63 MRSs between early-onset and late-onset (reference group) patients using multivariate logistic regression models. Due to sex disparities in CRC incidence 21 , we adjusted the regressions for sex when feasible. Figure 1 summarizes the results for all 29 exposome traits, analyzed across the five selection thresholds used in the original EWAS studies. In the discovery dataset, positive associations were observed for MRSs related to PCB, PM 2.5 , smoking-Maas, heptachlor, metolachlor, picloram and toxaphene. In contrast, negative associations were found for MRSs corresponding to BMI, education level, MDS, obesity, atrazine, malathion and mesotrione (Fig. 1 and Supplementary Table 6 ). We highlight the results for four lifestyle factors previously linked to colon and rectal cancers, including the MDS 22 (Fig. 2a ) and education level 23 (Fig. 2b ), which are considered protective factors, as well as smoking habits 24 (Fig. 2c ) and obesity status 25 (Fig. 2d ), which are recognized as risk factors. To examine the directionality of our findings, the heatmaps in the left panels of Fig. 2 show the methylation level distributions across CpGs featured in each of the four MRSs, along with their direction in the original EWAS and after sorting by the derived MRSs. The heatmaps show that increased MRS correlates with higher beta-values in CpGs with positive associations in the EWAS and with lower beta-values in CpGs with negative associations (see Extended Data Fig. 4 for a more detailed explanation). These results suggest that an elevated MRS reflects greater exposure levels in the original EWAS. Specifically, for patients with early-onset colon cancer, this suggests a lower MDS ( P = 7.9 × 10 −3 ; adjusted P ( P adj. ) = 3.3 × 10 −2 ) (Fig. 2a ), lower education levels ( P = 4.0 × 10 −3 ; P adj. = 2.2 × 10 −2 ) (Fig. 2b ), increased smoking (smoking-Maas) exposure ( P = 1.1 × 10 −3 ; P adj. = 8.9 × 10 −3 ) (Fig. 2c ), and lower obesity rates ( P = 8.4 × 10 −4 ; P adj. = 7.5 × 10 −3 ) (Fig. 2d ) in comparison to those with late-onset, as illustrated in the middle panels of Fig. 2 . The association of lower obesity rates with early-onset cases was verified utilizing physical metrics from TCGA-COAD. Colon cancer patients with a BMI over 30 kg m − 2 , as measured in the clinic, were categorized as obese, resulting in 4 out of 24 early-onset and 18 out of 72 later-onset patients being marked as obese. This provides a relative risk (RR) of 0.67 (95% confidence interval (CI): 0.26–1.76) for obesity in early-onset colon cancer patients within the TCGA-COAD cohort (Extended Data Fig. 5 ), supporting the MRS results for obesity.

Summary overview of the association between the 63 exposome-related MRSs across 29 exposome traits ( y axis) and five selection thresholds ( x axis, as applied in the original EWAS), comparing early-onset (age 3027 , E-MTAB- 7036 , GSE199057 , GSE131013 , GSE42752 , TCGA-READ, GSE39958 , GSE101764 and GSE77954 ), including 83 EOCRC and 272 LOCRC patients. The selection thresholds ( x axis) indicate the significance threshold in the original EWAS used for CpG selection, namely P −7 (GW), P −5 (P1E5) and FDRs of <0.1 (F01), <0.05 (F005) and <00.1 (F001). The presence of CpGs across the five selection thresholds varies, with unavailable data indicated in white (NA), and nonsignificant findings in light gray.

a – d , Lifestyle-related MRS differed between EOCRC and LOCRC, including non-Mediterranean dietary patterns MRS-GW (marker threshold P −7 ) ( a ), lower education levels MRS-GW ( b ), higher smoking exposure (smoking-Maas) ( c ) and lower obesity rates MRS-P1E5 (marker threshold P −5 ) ( d ) in patients with EOCRC. The heatmaps (left) display epigenetic age-adjusted DNA methylation beta-values in TCGA-COAD, ordered by MRS; the top color bar indicates CpG effect direction in the original EWAS (red: positive; blue: negative). The boxplots (middle) show MRS distribution in patients with EOCRC ( N = 31) and LOCRC ( N = 100) in TCGA-COAD, stratified by sex (orange: female; purple: male). The forest plots (right) present sex-adjusted logistic regression results for each replication dataset, their meta-analysis (Replication), and combined with TCGA-COAD (Discovery and Replication), for colon (blue), rectal (purple) and CRC (red). Squares: ORs; horizontal lines: 95% CIs. Boxplots show median (center line), interquartile range (IQR) (box) and whiskers (1.5× IQR) and outliers.

Extending our investigation, we conducted a meta-analysis across the nine replication datasets (Figs. 1 and 2 and Supplementary Table 7 ). This analysis corroborated the initial findings, notably the associations of a lower MDS ( P = 1.5×10 −2 ; P adj. = 4.6×10 −2 ), lower educational levels ( P = 2.11 × 10 −5 ; P adj. = 3.59 × 10 −4 ) and higher smoking exposure ( P = 1.02 × 10 −5 ; P adj. = 8.6 × 10 −4 ) in EOCRC (Fig. 2 , right panel). To ensure platform consistency, the meta-analysis was repeated using eight datasets generated with the 450K array, excluding the EPIC-based cohort GSE199057 . This strengthened the association for lower MDS ( P = 3.06 × 10 −3 ; P adj. = 1.3 × 10 −2 ), whereas educational level and smoking-Maas remained significant (Supplementary Table 8 ). Finally, despite the considerably reduced sample size, an exploratory tissue-specific meta-analysis of rectal-only (TCGA-READ and GSE39958 ) or colon-only cancer datasets ( GSE131013 , GSE42752 , E-MTAB- 7036 and GSE199057 ) showed a consistent trend in colon for education and smoking (Extended Data Fig. 6 and Supplementary Table 9 ).

Our results highlight a new association between the MRSs for the herbicide picloram and EOCRC, in comparison to LOCRC cases, in both the discovery and meta-analysis (Fig. 1 ). We observed that a higher exposure level, as indicated by the original EWAS direction, is associated with an elevated MRS (Fig. 3a ). This association highlights increased exposure to picloram among patients with early-onset colon cancer ( P = 2.59 × 10 −5 ; P adj. = 4.41 × 10 −4 ) (Fig. 3b )—a finding consistently supported by our meta-analysis using all replication cohorts ( P = 3.07 × 10 −3 ; P adj. = 1.51 × 10 −2 ; odds ratio (OR): 1.56 [95% CI: 1.16–2.09]) (Fig. 3c ) as well as in the meta-analysis using only 450K array samples ( P = 7.17 × 10 −4 ; P adj. = 4.06 × 10 −3 ; OR: 1.71 [95% CI: 1.25–2.33]).

The figure shows the distribution and validation of the picloram MRS employing the genome-wide CpG selection threshold (MRS-GW). a , Heatmap of epigenetic age-adjusted DNA methylation beta-values in TCGA-COAD, ordered by MRS; the top color bar indicates CpG effect direction in the original EWAS. b , Boxplot showing MRS-GW distribution in early- ( N = 31) and late- ( N = 100) onset TCGA-COAD patients, stratified by sex (orange, female; purple, male). c , Forest plot showing sex-adjusted logistic regression results in replication cohorts, their meta-analysis (Replication), and combined discovery and replication, including colon (blue), rectal (purple) and colorectal (red) cancer. d , e , The associations were further adjusted for other MRSs (31 EOCRC versus 100 LOCRC) ( d ) and tumor purity, subdivision (left/right), first-degree family with cancer diagnosis (yes/no), race (White/non-White) and MSI status ( e ). The baseline model (red) includes sex adjustment only; subset analyses show sex-adjusted (M1, orange) and fully adjusted models (M2, blue square). f , Spearman correlation between the 37-gene picloram-specific ssGSEA and the picloram MRS in TCGA-COAD ( P = 2.127845 × 10 −12 ); solid line: linear fit with the shaded area indicating the 95% CI. g , h , The violin plot shows the permutation results, confirming the robustness of CpG selection, patient categorization ( g ) and ssGSEA association ( h ). In the forest plots, squares represent ORs, and lines 95% CIs. Boxplots show median (center line), IQR (box), whiskers (1.5× IQR) and outliers.

The model presented was initially adjusted only for sex to facilitate external validation in independent datasets lacking measurements for additional variables. To evaluate the robustness of the association between picloram MRS-GW and age at onset, we applied additional adjustments in TCGA-COAD patients. First, we accounted for the influence of other MRSs included in the GW marker selection threshold (Fig. 3d ). We observed only minor variations across most adjustments, and none resulted in the loss of the significant association between picloram MRS-GW and age at onset (Fig. 3d ). Second, we explored the distribution of the picloram MRS-GW between male and female participants ( P = 0.0024) and between microsatellite stable (MSS) and instable (MSI) tumors ( P = 0.01), tumor purity high (≥0.7) versus low ( P = 5.3 × 10 −15 ), and across the consensus molecular subtypes (Extended Data Fig. 7 ). The observed significant differences highlight the need for further adjustment when testing the association between picloram MRS-GW and age at onset. As not all variables were available for every patient, Fig. 3e presents Model 1 (adjusted only for sex) alongside Model 2, which incorporates each additional adjustment. This approach ensures consistency by analyzing the same subset of patients to account for sample size differences. Notably, no substantial changes in the association between picloram MRS-GW and age at onset were observed after adjustments.