Abstract

   Cancer incidence and mortality increase with increasing body mass index
   (BMI), but BMI-associated epigenetic alterations in cancer remain
   elusive. We hypothesized that BMI would be associated with DNA
   methylation alterations in cancers. To test this hypothesis, here, we
   estimated the associations between DNA methylation and BMI through two
   different methods across 15 cancer types, at approximately 485,000 CpG
   sites and 2415 samples using data from The Cancer Genome Atlas. After
   comparing the DNA methylation levels in control BMI and high BMI
   individuals, we found differentially methylated CpG sites (DMSs) in
   cholangiocarcinoma (CHOL), colon adenocarcinoma (COAD), and uterine
   corpus endometrial carcinoma (UCEC) (False Discovery Rate < 0.05). The
   DMSs of COAD or UCEC were enriched in several obesity-induced and
   cancer-related pathways. Next, when BMI was used as a continuous
   variable, we identified BMI-associated methylated CpG sites (BMS) (P
   (Bonferroni) < 0.05) in CHOL (BMS = 1), COAD (BMS = 1), and UCEC (BMS =
   4) using multivariable linear regression. In UCEC, three of the BMSs
   can predict the clinical outcomes and survival of patients with the
   tumors. Overall, we observed associations between DNA methylation and
   high BMI in CHOL, COAD, and UCEC. Furthermore, three BMI-associated
   CpGs were identified as potential biomarkers for UCEC prognosis.

   Keywords: Association analysis, Body mass index, Cancer, DNA
   methylation, Epigenetics

Introduction

   Body mass index (BMI) is the most widely used measurement of adiposity,
   which is an established risk factor for various diseases, including
   type 2 diabetes, cardiovascular disease, metabolic disturbances, and
   cancers [37]^1^-[38]^3. High BMI (25 kg/m^2 or greater), including
   overweight and obesity, contribute to cancer incidence and death across
   a broad range of cancers [39]^4^,[40]^5. High BMI affects the majority
   of the adult population in most developed countries, and its prevalence
   is increasing rapidly in developing countries. If the current worldwide
   trends continue, the number of people with obesity will reach nearly
   2.0 billion by 2025 [41]^6. In addition, the prevalence of cancer has
   increased in parallel with obesity [42]^7. In the period between 1982
   and 2012, a quarter of the cancer cases could be attributed to high BMI
   [43]^8. Overall, high BMI is likely one of the drivers for the
   increasing incidences of cancer.

   Numerous genome-wide association studies between single nucleotide
   polymorphisms (SNPs) and BMI have allowed for a better understanding of
   the biological basis of high BMI resulting in diseases ^9,10. A study
   on pancreatic cancer found that the inflammatory NF-κB-mediated
   chemokine signaling pathway was associated with obesity, and contribute
   to a risk for pancreatic cancer [44]^11. Another study revealed that 7
   BMI-associated SNPs located in 8 obesity-related genes were
   significantly associated with a risk for endometrial cancer [45]^12.
   Similarly, SNPs in genes known to be associated with metabolism and
   obesity are relevant to prostate cancer [46]^13. These findings suggest
   that high BMI associated with multiple genes and can influence various
   biological pathways, thus contributing to the susceptibility of complex
   human cancers.

   Further advances in genomic technologies have brought epigenome-wide
   association studies (EWAS) to focus on improving the understanding of
   epigenetic alterations that are associated with diseases [47]^14. DNA
   methylation modulation, a well-known and critical regulation level of
   epigenetic machinery, is easily susceptible to environmental factors
   [48]^15^-[49]^17. Unlike stable and conserved genetic variants,
   epigenetic processes are highly dynamic during an individual's
   lifetime, which has attracted much attention to the study of DNA
   methylation [50]^18. Increasing evidences for DNA methylation sites in
   relation to high BMI support the hypothesis that altered DNA
   methylation of metabolically important genes are mediated by adiposity
   in human blood [51]^3^,[52]^19^,[53]^20. Moreover, alterations in DNA
   methylation are consequences of adiposity rather than the trigger
   [54]^3.

   Changes in DNA methylation can indeed be implicated in the initiation
   and progression of neoplasm [55]^21. Sirtuin 1, a gene related to
   epigenetic silencing, links obesity to cancer, and also plays a role in
   colorectal cancer development, suggesting that obesity-related
   epigenetic changes may be an important determinant in colorectal cancer
   [56]^22. DNA methylation patterns that link high BMI to cancer
   development appear to be predictable, however, they are not well
   understood. Furthermore, a previous study that analyzed 345 breast
   tumor samples has also shown an association between BMI and DNA
   methylation in tumor tissues [57]^23. The publicly available Cancer
   Genome Atlas (TCGA) has provided large-scale genome sequencing data
   alongside clinical features for over 30 human tumors, which allows
   researchers to systematically study possible interactions between
   epigenetic changes induced by high BMI and cancer risk [58]^24. In this
   study, we examined genome-wide DNA methylation profiles in tumor
   tissues to investigate whether BMI levels change and whether these
   changes were associated with a high BMI. Understanding the epigenetic
   molecular links may provide an avenue for diagnostic and therapeutic
   strategies.

Materials and Methods

The data from The Cancer Genome Atlas

   We downloaded patients' clinical information from the Broad GDAC
   Firehose ([59]http://gdac.broadinstitute.org/) and selected 15 types of
   cancers with the patients' weight and height information at diagnosis
   to estimate individual BMI values as weight in kilograms divided by the
   square of height in meters. These cancers include bladder urothelial
   carcinoma (BLCA), cervical squamous cell carcinoma and endocervical
   adenocarcinoma (CESC), cholangiocarcinoma (CHOL), colon adenocarcinoma
   (COAD), lymphoid neoplasm diffuse large B-cell lymphoma (DLBC),
   esophageal adenocarcinoma (EAC), esophageal squamous cell carcinoma
   (ESCC), kidney renal papillary cell carcinoma (KIRP), liver
   hepatocellular carcinoma (LIHC), rectum adenocarcinoma (READ), skin
   cutaneous melanoma (SKCM), thymoma (THYM), uterine corpus endometrial
   carcinoma (UCEC), uterine carcinosarcoma (UCS), and uveal Melanoma
   (UVM). Patients without definitive information of age or race were
   excluded. The Human methylation450 BeadChip datasets (level three)
   [60]^25 were also downloaded from the Broad GDAC Firehose.

Identification of CpG probes associated with BMI

   For adults, BMI was classified according to World Health Organization
   (WHO) definitions as underweight (BMI < 18.5), normal weight (18.5 ≤
   BMI < 25), overweight (25 ≤ BMI < 30), and obese (BMI ≥ 30). In the
   initial analysis, BMI was used as a categorical variable by dividing
   the samples into two groups: control BMI subjects (BMI < 25) and high
   BMI subjects (25 ≥ BMI). Since the four clinical indicators, including
   age, gender, race and tissue source site, were all provided in the 15
   cancers, and three (age, gender and race) of them had been reported to
   be associated with BMI and DNA methylation [61]^26^-[62]^29, we
   included the four variables into linear models. After adjusting for
   age, gender, race, and tissue source site, differentially methylated
   CpG probes between two groups were identified using the linear models
   for microarray data (limma) package (default), which uses linear models
   to assess differential methylation, whereby information is shared
   across probes [63]^30. Differentially methylated CpG sites (DMSs) were
   identified if the resulting adjusted P-value was < 0.05. The
   Benjamini-Hochberg method [64]^31 was used to adjust the P-values and
   ensure that the False Discovery Rate (FDR) was < 0.05. In the second
   analysis, BMI was not normally distributed in any cancers and therefore
   was transformed by rank normalization, which was performed using the
   rntransform function in the R GenABEL package [65]^32. The Lillie test
   in nortest package, a modification of the Kolmogorov-Smirnov test, was
   used to determine whether the data were normally distributed.
   Multivariable linear regression of normalize-transformed BMI with
   methylation level at each CpG probe was adjusted for gender, age, race,
   and tissue source site.

Pathway enrichment analysis

   To evaluate the functional relevance of DMSs, a The Kyoto Encyclopedia
   of Genes and Genomes (KEGG) pathway enrichment analysis for genes
   involved in DMSs was performed using the Database for Annotation,
   Visualization and Integrated Discovery (DAVID) software [66]^33. The
   significant enrichment pathways were identified with P < 0.05.

Statistical analysis

   Statistical analyses were performed using the R software. The survival
   difference between patient groups was analyzed using the Kaplan-Meier
   and log-rank test function in survival package. The methylation level
   was compared between two groups by the Wilcoxon rank test. Over- or
   under-representation of the CpGs were determined using the chi-square
   test. P values less than 0.05 were considered statistically
   significant.

Results

The characteristics of the TCGA samples analyzed

   To characterize the specific genome-wide DNA methylation profile that
   is associated with BMI in cancers, we chose 15 publically available
   cancer datasets from TCGA (Table [67]1 and Table [68]S1). Totally,
   there are 2415 samples including 883 control BMI samples and 1532 high
   BMI samples. The average age ranged from 48.4 to 69.6 years old, and
   the average BMI ranged from 22.6 to 34.4 kg/m^2. After adjusting for
   age, race, gender, and tissue source site, we firstly treated BMI as a
   categorical measure, and performed DMSs analyses between high and
   control BMIs for each cancer through limma package. Secondly, we
   treated BMI as a continuous measure, and identified BMI-associated
   methylated sites (BMSs) using the multivariable linear regression
   models (Fig. [69]1).

Table 1.

   The characteristics of eligible TCGA samples
   Cancer types Age (years) Races Gender (% male) BMI (kg/m^2) No. Control
   BMI No. High BMI
   BLCA 67.8 (10.5) a(284); b(43);
   c(18) 75% 27.1 (6.3) 142 203
   CESC 48.4 (13.5) a(189);b(19);
   c(23); d(7); e(1) 0% 28.4 (7.8) 89 150
   CHOL 63.4 (12.8) a(30);b(3);
   c(2) 45.7% 28.0 (5.3) 10 25
   COAD 64.4 (13.1) a(164);b(8);
   c(51);d(1) 54% 29.4 (17.4) 75 149
   DLBC 56.3 (13.9) a(29);b(18);
   c(1) 45.8% 25.9 (5.9) 25 23
   ESCC 58.7 (10.5) a(41);b(44);
   c(3) 85.2% 22.6 (3.4) 70 18
   EAC 66.6 (11.8) a(66);b(1) 88.1% 28.6 (6.4) 16 51
   KIRP 61.7 (12.0) a(152);b(4);
   c(39);d(1) 74.5% 32.1 (32.8) 47 149
   LIHC 59.1 (12.9) a(159);b(157)
   c(14);d(1) 68.6% 26.1 (8.5) 175 156
   READ 61.1 (11.5) a(62);c(5) 59.7% 27.2 (5.8) 18 49
   SKCM 59.4 (15.3) a(227);b(11);
   c(1) 59.0% 28.0 (6.1) 79 160
   THYM 58.8 (12.9) a(83);b(12);
   c(3) 52.0% 27.2 (6.03) 38 60
   UCEC 64.0 (11.0) a(275);b(87);
   c(8);d(2);e(7) 0% 34.4 (13.0) 65 314
   UCS 69.6 (9.22) a(40);b(3);
   c(8) 0% 29.6 (9.03) 17 34
   UVM 61.3 (13.7) a(53) 52.8% 28.7 (8.9) 17 36
   [70]Open in a new tab

   Abbreviations: a, White; b, Asian; c, Black or African American; d,
   American indian or Alaska native; e, Native hawaiian or other pacific
   islander. For age and BMI, data are mean (Standard Error). For race,
   data are race names (number).

Figure 1.

   [71]Figure 1
   [72]Open in a new tab

   Flow chart indicating study design.

Genome-wide methylation differences in high and control BMI

   High BMI has an increased risk for cancers [73]^34. To gain insight
   into the association between high BMI and DNA methylation, we used 25
   kg/m^2 as a cutoff threshold to divide high BMI and control cases
   (Table [74]1). We separately examined DNA methylation levels between
   the two groups across the 15 types of cancers. In COAD, the average
   global DNA methylation levels in high BMI cases (β value = 0.46) were
   significantly lower than the control cases (β value = 0.47) (P =
   0.0024). None of other cancers displayed a statistical significance
   (Fig. [75]2a). Using limma method and a FDR value of < 0.05 as the
   significance cut-off value for downstream analyses, we identified 3,
   1169, and 394 DMSs in CHOL, COAD (Table [76]S2), and UCEC (Table
   [77]S3), respectively. There were no DMSs in the other 12 cancers. 3
   DMSs in CHOL included cg26987376 that was located at an intergenic
   region, cg21515384 that was located at 200 bp upstream of the
   transcription start site (TSS200) of the 3-oxoacid CoA-transferase 1
   (OXCT1), and cg02725055 that was located at 1500 bp upstream of the
   transcriptional start site (TSS1500) of the transcription factor 4
   (TCF4). Furthermore, OXCT1 plays important roles in ketone and fatty
   acid metabolism [78]^35. TCF4 (also known as TCF7L2) is also identified
   to be a possible determinant of type 2 diabetes mellitus and is a
   transcription factor of the Wnt signaling pathway, which also plays a
   crucial role in insulin secretion and proglucagon gene expression
   [79]^36^,[80]^37.

Figure 2.

   [81]Figure 2
   [82]Open in a new tab

   Analysis of differences in genomic DNA methylation levels in cases of
   high body mass index (BMI) and controls. (A) The differences in genomic
   DNA methylation levels in 15 cancer types. (B) Volcano plot showing the
   distribution of CpG sites from the site-level test assessed by
   methylation differences and FDR values. CpGs in blue indicate a FDR <
   0.05. (C) Significant differences in the genomic distributions of the
   differentially methylated CpGs regarding the CpG context and the gene
   region, compared with all analyzed sites on the Infinium
   HumanMethylation450 BeadChip, by chi-square test. (* p < 0.05; ** p <
   0.01; *** p < 0.001; N_Shelf, North Shelf; S_Shelf, South Shelf;
   N_Shore, North Shore; S_ Shore, South Shore).

   Among the 1169 DMSs identified in COAD, 1168 DMSs exhibited low
   methylation levels, and a single DMS exhibited high methylation level
   in patients with high BMIs (Fig. [83]2b). Compared to all of the
   analyzed sites, the DMSs were underrepresented in the promoter regions,
   including the TSS1500, TSS200, 5' untranslated regions (5' UTR), and
   1^st exons, but were overrepresented in the gene body and intergenic
   regions (Fig. [84]2c). Regarding the CpG context, the 1169 DMSs were
   significantly overrepresented in the CpG islands and DNA regions within
   2 kb of the CpG islands (North Shore and South Shore). Additionally,
   the genomic regions outside the CpG islands and their 2 kb flanking
   regions were underrepresented. The 394 DMSs identified in samples of
   high BMI in UCEC consisted of 128 hypomethylated CpGs and 266
   hypermethylated CpGs. These DMSs were significantly overrepresented in
   the gene body regions and underrepresented in the TSS200 and 1^st
   exons. Regarding the CpG context, the DNA regions 2 kb outside of the
   CpG islands were overrepresented and the North Shore and CpG island
   regions were observed to be underrepresented (Fig. [85]2c), suggesting
   that altered methylated regions for high BMI are clearly tumor- or
   tissue-specific.

Biological pathways analysis

   To understand the functions of DMSs-containing genes, we performed a
   functional enrichment analysis using the DAVID software [86]^33. In
   COAD, 1168 DMSs corresponded to 449 unique genes. KEGG pathway analysis
   provided significant enrichment in 20 pathways (P < 0.05), including
   numerous biological processes pertinent to cancer, such as the
   phosphatidylinositol 3-kinase (PI3K)/Akt and hippo signaling pathways
   (Fig. [87]3a). Interestingly, type II diabetes mellitus which has been
   reported to be associated with an excess risk of colorectal cancer
   [88]^38, was also included. Tenascin XB, encoding a member of the
   tenascin family of extracellular matrix (ECM) glycoproteins [89]^39,
   had 23 CpGs with high methylation and was involved in pathways of
   PI3K-Akt signaling, focal adhesion, and ECM-receptor interaction.
   Collagen type genes of the three pathways were also subject to
   methylation in certain CpGs. It is clear that the ECM plays an
   important role in adipocyte development and function and as such in
   lipid metabolism, weight regulation, and obesity [90]^39. 116 DMSs
   enriched in pathways mostly located at the gene body regions of their
   genes (Table [91]S4). We did not identify any significant correlation
   between these DMSs and their genes in COAD, probably due to that there
   were 19 overlapping samples between transcriptome and methylome. In
   UCEC, 394 DMSs corresponded to 253 unique genes, which were involved in
   focal adhesion, rap 1 signaling pathway, bladder cancer, platelet
   activation, GnRH signaling pathway, estrogen signaling pathway,
   proteoglycans in cancer, and endocytosis (P < 0.05) (Fig. [92]3b).
   Nearly all of these pathways included the epidermal growth factor
   receptor (EGFR), SRC proto-oncogene, and non-receptor tyrosine kinase,
   were reported to be drivers of tumorigenesis [93]^40^,[94]^41. 31 DMSs
   enriched in pathways mostly located at the gene body regions of their
   genes (Table [95]S5). Among them, 4 DMSs (cg01222719, cg17307474,
   cg22187246, cg11209289) in gene body regions and 4 DMSs (cg00130530,
   cg23410129, cg12726213, cg26033529) in promoter regions, were
   negatively correlated with their corresponding genes (Table [96]S5).

Figure 3.

   [97]Figure 3
   [98]Open in a new tab

   Bar graph showing significant KEGG pathways of genes involving with
   differentially methylated CpGs. (A) Biological pathways in COAD; (B)
   Biological pathways in UCEC.

Epigenome-wide association study through multivariable linear regression
model

   Next, to investigate whether or not there was a correlation between BMI
   and DNA methylation in cancers, BMI was used as a continuous measure,
   and EWASs were performed between CpG sites and BMI using FDR < 0.05 as
   the cutoff threshold. We identified 1 BMS in CHOL and COAD,
   respectively, whereas we found 113 BMSs in UCEC (Table [99]S6).
   However, we were unable to identify any significant BMSs in the other
   12 cancers that were investigated. The identified BMSs included
   cg04545963 (β = 0.680, FDR = 0.024) located in NF-κB inhibitor alpha
   (NFKBIA) in CHOL and cg15542880 (β = -0.082, FDR = 0.020) located in
   dual serine/threonine and tyrosine protein kinase (DSTYK) in COAD.
   Moreover, the identified 113 BMSs in UCEC were annotated to 84 unique
   genes, and approximately 32% of the BMSs showed a positive correlation
   to BMI (Table [100]S6). The genomic location revealed that 113 BMSs
   were distributed among 19 chromosomes except for chromosomes 9, 14, and
   20 (Fig. [101]4). Some BMSs were located at genes involved with
   metabolically relevant events. For example, UDP glucuronosyltransferase
   family 2 member B15 (UGT2B15) was known as a member of the
   UDP-glucuronosyltransferases superfamily catalyzing the conjugation of
   various lipid-soluble endogenous substances and xenobiotics with the
   glucuronic acid moiety UDP-glucuronic acid [102]^42. SRY-box 6 (SOX6),
   a transcription factor, has been linked to the regulation of the
   expressions of key adipogenic regulators as well as synergy with
   epigenetic pathways involving mesoderm-specific transcript genes that
   are associated with adipocyte size [103]^43. Meanwhile, a SNP in SOX6
   gene, rs297325, located 0.22 Mb downstream of cg21992400 (β = -0.088,
   FDR < 0.05), has been identified to associate with the risk of both BMI
   and endometrial cancer [104]^12. Among the 84 genes, tumor-related
   genes were also found, such as T-box 2 [105]^44, discoidin domain
   receptor tyrosine kinase 1 [106]^45, and lysine acetyltransferase 6A
   [107]^46. These results further supported the findings that BMI is
   associated with DNA methylation profiles in CHOL, COAD, and UCEC.

Figure 4.

   [108]Figure 4
   [109]Open in a new tab

   Circos plot of the epigenome-wide association of DNA methylation in
   UCEC with body mass index (BMI). Results are presented as CpG-specific
   association test results (-log[10] (FDR)) ordered by genomic position.
   Green symbols, CpGs positively associated with BMI; Red symbols, CpGs
   negatively associated with BMI; Chromosome numbers are shown on the
   inner ring. Tick marks on the outer ring identify the genomic locis
   reaching epigenome-wide significance. The genes located by the 113
   locis are listed around the circos plot.

BMSs in UCEC implicated to tumor progression and patient survival

   Using the more stringent Bonferroni correction for multiple testing
   (Bonferroni-corrected P < 0.05) in the analyses of multivariable linear
   regression models, cg04545963 in CHOL and cg15542880 in COAD were still
   significantly associated with BMI. However, in UCEC, the number of
   significant BMSs was markedly reduced, and only 4 BMSs remained,
   including cg12645852 (β = -0.107, Bonferroni-corrected P = 5.85E-04)
   located in an intergenic region, cg09621472 (β = -0.107,
   Bonferroni-corrected P = 0.018) located in NatE catalytic subunit of N
   (alpha)-acetyltransferase 50 (NAA50), cg27131667 (β = -0.198,
   Bonferroni-corrected P = 0.028) located in spastic paraplegia 7 (SPG7),
   and cg14044785 (β = -0.116, Bonferroni-corrected P = 0.041) located in
   thyroid adenoma associated (THADA). Of these 4 BMSs, cg12645852,
   cg09621472, and cg14044785 in patients with high BMI in UCEC had
   significant hypomethylation when compared to controls.

   BMI is a prognostic factor in patients with cancer. Bladder cancer and
   colorectal cancer show lower mortality in patients with higher BMI
   values [110]^47. However, the increase of BMI at an older age is
   associated with a lower overall survival time in patients with
   pancreatic cancer [111]^48. To explore whether BMS can also predict
   tumor survival outcome, based on these stringent BMSs, we therefore
   respectively categorized the study cases into low and high levels
   according to the median of methylation levels of these 6 BMSs and
   subsequently performed Kaplan-Meier and Log-rank test. With regards to
   methylation status, cg04545963 in CHOL, cg15542880 in COAD, and
   cg12645852 in UCEC did not significantly affect the patient survival
   rates (Fig. [112]5a). For cg09621472, cg27131667, cg14044785, patients
   with UCEC with low methylated levels had improved survival rates when
   compared with those who had high methylated levels (P = 0.0008, 0.0248,
   0.0109, respectively, Fig. [113]5a), suggesting that cg09621472,
   cg27131667, and cg14044785 could be used as potential biomarkers for
   predicting survival outcomes of patients with UCEC.

Figure 5.

   [114]Figure 5
   [115]Open in a new tab

   Analysis of body mass index (BMI)-associated CpGs during the survival
   initiation, stage, and progression of CHOL, COAD, and UCEC. (A)
   Kaplan-Meier survival analysis. The patients were divided into low and
   high groups using the median cutoff value of DNA methylated levels for
   the 6 BMI-associated CpGs, respectively. HR, hazard ratio. (B) The
   methylation levels of BMI-associated CpGs in the initiation and
   progression of the three cancers (* p < 0.05, ** p < 0.01, *** p <
   0.001, no represents statistically insignificant).

   Furthermore, analyses were performed to investigate whether these BMSs
   had implications in tumor initiation and progression. The methylation
   status of cg04545963 was not significantly associated with the
   initiation and grade of CHOL, and neither was cg15542880 in COAD (Fig.
   [116]5b). We found that, in adjacent tissue, the methylation levels of
   cg12645852 and cg27131667 were significantly higher than tumor tissue
   of the early stages (grade 1) from UCEC. However, for cg09621472,
   cg27131667, and cg14044785, the methylation levels in adjacent tissues
   were lower than those of tumor tissues (Fig. [117]5b). Interestingly,
   with advancing tumor grade, the methylation levels of the 4 BMSs in
   UCEC significantly increased and displayed the most notable increase
   from grade 2 to grade 3 (all P < 0.0001, the different β values between
   the two grades are 0.1598, 0.2135, 0.0883, 0.1447, respectively).
   Furthermore, their high methylation levels were also associated with
   other advanced clinical tumor parameters, including tumor invasion
   percentage, clinical stage, and histologic diagnosis type (Table
   [118]S7). The dataset of [119]GSE67116 [120]^49 from the GEO database
   showed higher methylation levels of cg09621472 (P < 0.001) and
   cg27131667 (P = 0.014) in 33 patients with endometrial metastasis
   cancer relative to 53 patients with endometrial primary cancer. These
   data suggested that cg12645852, cg09621472, cg27131667, and cg14044785
   may be associated with the effect of BMI on UCEC advancement.

Discussions

   In this study, we characterized the genome-wide DNA methylation
   profiles associated with BMI in 15 cancer types by modeling BMI first
   as a categorical variable and then as a continuous variable. The two
   analyses consistently showed that the dynamic features of DNA
   methylation in the three known obesity-related cancers (CHOL, COAD, and
   UCEC) [121]^50^,[122]^51 associated with BMI, but not in the other 12
   major cancers.

   The observed hypomethylation patterns in COAD with high BMI is
   predominant, indicating that high BMI cases with COAD may regulate gene
   expression through the attenuation of DNA methylation. These genes
   containing DMSs play roles in the pathogenesis of colorectal cancer
   (e.g. DCC netrin 1 receptor; netrin 1; Wnt family member 3A; Wnt family
   member 5A; receptor tyrosine kinase like orphan receptor 2; RAR related
   orphan receptor A) [123]^52^-[124]^57. Most genes with DMSs enrichment
   were in the PI3K-AKT signaling pathway, an intercellular signaling
   pathway responsible for carcinogenesis and metastasis of many cancers
   [125]^58. This finding is consistent with a previous report showing
   that the PI3K-AKT signaling pathway links obesity risk factors and
   colon carcinogenesis processes [126]^59. In obesity cases, increased
   levels of blood insulin, insulin-like growth factor I, leptin, tumor
   necrosis factor, and interleukin 6, and decreased levels of adiponectin
   activate the PI3K-AKT signal pathway, further activating downstream
   signals for cell survival, cell growth and cell cycle, and ultimately
   resulting in cancer development. Alterations in DNA methylation also
   respond to adiposity [127]^3, therefore, in COAD, high BMI inhibits
   methylation of PI3K-AKT signaling pathway genes, which in turn may
   activate the pathway and the development of COAD. Another noteworthy
   pathway is that for type II diabetes. Adipose tissue of obesity cases
   releases various factors, such as non-esterified fatty acids and
   pro-inflammatory cytokines, which could contribute to development of
   insulin resistance and perturb β-cell function, thus resulting in type
   II diabetes [128]^2. This disease plays an important role in the
   pathogenesis of colon adenocarcinoma. Moreover, epidemiologic studies
   were directed at providing a consistent association and several shared
   genetic factors between type II diabetes and subsequent risk of colon
   cancer [129]^38^,[130]^60^,[131]^61. We identified, in patients with
   COAD and high BMI, significant hypomethylation of type II diabetes
   pathway genes, thus supporting the possibility that an alteration in
   this pathway may result in the induction of DNA methylation, and that a
   high BMI may have an important role in some of the downstream effects
   of COAD development. Even though the other 18 biological pathways in
   COAD linking obesity to cancer have not been reported, these pathways
   allow us to examine whether DNA methylation alterations due to high BMI
   have effects on colon cancer, particularly in case of cancer-related
   pathways such as the Hippo signaling pathway and cAMP signaling pathway
   and in ECM-receptor interaction. High BMI cases in UCEC also present
   methylation differences in comparison with controls. Some DMSs were
   located at genes involved in the development of UCEC (e.g. EGFR; kinase
   suppressor of ras 1; metastasis associated 1; syndecan 1; secreted
   protein acidic and cysteine rich) [132]^62^-[133]^66. Increased
   platelet activity is mainly related to reactive oxygen species and
   inflammation adipokines as well as a reduced sensitivity to insulin
   [134]^67. Obesity drives adipose inflammation, leading to induction of
   aromatase and increased estrogen signaling in adipose depots [135]^68.
   These results indicated that high BMI may contributes to modulation of
   two pathways involved in platelet activation and estrogen signaling as
   a result of DNA methylation in UCEC. These pathways could provide
   guidance for further understanding the effects of high BMI on the
   development of cancer.

   When BMI was used as a continuous measure in the EWAS with a strict
   Bonferroni correction (P < 0.05), methylation alterations were
   identified in the genes including NFKBIA, DSTYK, NAA50, SPG7, and
   THADA. These genes are highly associated with human disease
   susceptibility [136]^69^-[137]^81, but associations of DNA methylation
   among them have not been reported. Additionally, BMI-associated CpGs
   located in NAA50, SPG7, and THADA among UCEC tumor tissues are
   significantly associated with the survival time and tumor progression
   of this cancer, indicating that BMI may encourage the role of tumor
   mechanism through controlling DNA methylation. Intensive epigenetic
   studies have demonstrated a causal relationship between high BMI or
   obesity and altered epigenetic regulation in a great deal of
   metabolically important genes or regions
   [138]^3^,[139]^19^,[140]^20^,[141]^82. Although results of this study
   reveal that there are no major changes in metabolic processes, the
   discordance may be interpreted by tissue heterogeneity and
   tissue-specific patterns of DNA methylation. This deduction is
   consistent with previous reports showing that DNA methylation in HIF3A
   was associated with BMI in white blood cells and adipose tissue, but
   not in breast tumor tissue [142]^20^,[143]^23^,[144]^83. Furthermore,
   most studies have previously focused on blood or adipose tissue that
   are metabolically active; however, this cannot be used to reflect
   epigenetic patterns of other tissues. In another study using breast
   tumor tissue, obesity was shown to influence the methylation of some
   cancer-relevant genes and pathways involved in carcinogenesis [145]^23.
   The use of human primary tumor tissues in our study from 15 tumor types
   exclusive of the breast cancer significantly revealed new information.
   Thus, we demonstrated that high BMI may induce metabolic perturbations,
   leading to alterations in DNA methylation in cancer development-related
   and obesity-induced pathways in primary tumor tissues of CHOL, COAD,
   and UCEC.

   Excluding the overweight group, we also analyzed the methylation
   differences between obese and control BMI of cancer patients (Table
   [146]S8). The result suggested that the cutoff values of BMI influence
   the methylation differences. Removal of the overweight group resulted
   in detections of DMSs (Table [147]S8). TCGA dataset is heterogeneous
   and lack of unified information parameters (e.g., the history of
   smoking and drinking, etc.) in the subjects examined. This may be the
   reason why BMI-associated methylation patterns were not determined in
   the other 12 tumors. We further analyzed regression association
   analysis between BMI and DNA methylation in adjacent tissues from CHOL,
   COAD, UCEC, but failed to identify any significant BMSs. Further
   investigations may be needed when more samples become available in the
   future. In addition, DMS outcomes might be inadequately estimated in
   gene expression because the number of the overlapping samples between
   transcriptome and methylome was low. BMI grouping is possible and it
   can minimize the possible bias due to an underestimation of weight.
   However, BMI grouping obscures detection of possible correlations at
   within-group level. Instead, the BMI is a continuous variable in TCGA
   dataset, allowing the identifications of significant CpGs with the
   considerations of common confounders (age, gender, race, tissue source
   site) via multiple linear regression. With larger datasets in the
   future, improvement may be made on specifying how independent variables
   are entered into a given linear regression modeling. More sophisticated
   models may be deployed to improve the analytical power too, because
   possibilities of lacking normal and/or random distributions of
   independent variables, as well as the linear validity between
   independent variables and BMI could reduce the statistical power of the
   linear regression model. Therefore, these BMI-associated methylation
   pattern still need to be validated by a large-scale clinical samples.

   Collectively, we systemically analyzed the association of BMI with
   genomic DNA methylation in tumor tissues. This study supports the view
   that changes in DNA methylation represent an intermediate step between
   the BMI and CHOL, COAD, or UCEC, which may in turn provide information
   for guiding treatment options for patients with CHOL, COAD, and UCEC
   with high BMIs.

Acknowledgments