Robust identification of driver genes via pathway models.

Before the investigation of pathway findings, we proposed a brief test to verify the reliability of the results derived from EntCDP and ModSDP models. The driver genes discovered by IntOGen provide insights into the mutational features of each cohort (Fig 1B and S1 Table) and provide a valuable benchmark for comparison with our findings. The main motivation is to examine, using the hypergeometric test, the degree of overlap between common (or specific) IntOGen genes of comparable cohorts and their common (or specific) driver gene sets discovered by EntCDP (or ModSDP) (Sect A in S1 Text). Since some cohorts lack IntOGen genes, we restricted the hypergeometric test to cohorts containing IntOGen genes (Sect B in S1 Text).

In spite of the incompleteness of IntOGen driver genes, those reported in the literature and associated with cohort characteristics, such as cancer type and geographic location, can serve as a credible reference for validating novel discoveries. In a sense, genes not assigned to any cohort by IntOGen and lacking annotation in the literature can be considered new candidate driver genes for a certain cancer.

Targeting novel driver genes in region-specific signaling pathways.

The information on nationality or sequencing location in the collected data enables us to geographically stratify samples with the same cancer type (Fig 1B). Here, we focus on cohorts of each cancer type and apply EntCDP and ModSDP to cohorts of the same type but different geographic origins.

There is some overlap between the genes discovered by our models and those cohort-sensitive genes (i.e., IntOGen driver genes). At least 20% of common genes identified by EntCDP in each cohort group (Fig 2A) are IntOGen driver genes, all except kidney cohorts KIRC and RECA_EU, which share only one IntOGen gene (S2 Table). In addtion, significant overlap between IntOGen and ModSDP can be seen in half (7/14) types of cancer (S3 Table).

While IntOGen overlapping driver genes can confirm the accuracy of our results, non-IntOGen driver genes can either be validated as drivers or predicted to serve in such a capacity. For example, two common genes, TP53 and PIK3CA, identified in breast adenocarcinoma cohorts from the US, EU and UK, are IntOGen genes shared by all three cohorts (colored by dark purple in Fig 2A), while GATA3 is a known IntOGen gene present in two of the cohorts (colored by light purple). The three genes function in the IGF1 receptor (IGF-1R) signaling pathway and are identified by EntCDP when the size of the search set (parameter K) is 3 (Fig 2C and Table S1.3 in S1 Data). It has been reported that the IGF-1R pathway plays an important role in gynecologic cancers such as breast tumors [20]. When K = 10, we identified other seven genes mutually exclusive with the three genes. Although H3F3A, CARS, GTF2I and SEPT9 are not IntOGen genes of the three cohorts (colored by yellow), they have been proved to play important roles in the development of breast adenocarcinoma [21–24]. For example, SEPT9, a septin gene encoding multiple isoform, has been confirmed as a promising tool in breast cancer detection [24]. However, those not previously reported in the literature (colored by pink), including TAF15, HIST1H4I and FAM46C, may facilitate the identification of novel targets for the treatment of breast adenocarcinoma. Moreover, there is a series of driver genes specific to the cohort BRCA relative to BRCA_UK and BRCA_EU, with gene set {CDH1, MAP2K4, PIK3R1, MAP3K1} detected by ModSDP when K = 4 (Table S1.6 in S1 Data). The four genes are members of the development HGF signaling pathway (Fig 2B and 2C), which can provide valuable diagnostic, prognostic and predictive biomarkers for personalized cancer therapy [25]. Targeting this pathway may provide clinical benefits for breast cancer patients from the United States.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Results of the reliability test and regional differences in signaling pathways.

(A) The hypergeometric test reveals significant results for six types of cancer with IntOGen genes in each cohort, which investigates the overlap between cohorts’ common genes identified by IntOGen and EntCDP. The size of the circles corresponds to the significance level. Different colors in the bar graph present different features of genes identified by EntCDP: dark (light) purple indicates common genes obtained from all (part of) selected cohorts by IntOGen; yellow (pink) denotes non-IntOGen genes of any selected cohort that have (not) been verified as drivers in the literature. (B) Specific pathways of patients with a certain cancer type coming from one country (left) relative to other counties (right, denoted by colorful circles). The gray-filled bar for each signaling pathway indicates the proportion of identified genes enriched in that pathway, as shown by the ratio on the right. (C, D) Common signaling pathways among, and pathways specific to, patients with breast adenocarcinoma from the United States versus those from European Union countries and Britain (C), or with bladder cancer from the United States versus China (D). The dashed line indicates no significant results. The bar above clearly lists the names of genes with different characteristics from (A). *1: COADREAD, colorectal cancer; *2: Transcriptional misregulation in cancer; *3: Apoptotic pathways in synovial fibroblasts.

https://doi.org/10.1371/journal.pcbi.1013349.g002

Significant overlap between common genes discovered by both IntOGen and EntCDP is also observed in bladder tumors . FGFR3, HRAS and KRAS, along with known oncogene TP53, are typical biomarkers for bladder cancer associated with relevant pathways, and are discovered by EntCDP when K = 4 (Table S1.1 in S1 Data). When K = 10, EntCDP detected another six genes mutually exclusive with these four genes. Among them, PIK3CB is an IntOGen gene of the bladder cohort from America but not China, and IDH2 is not present in any cohort but has been reported to be of potential therapeutic significance [26]. The remaining four genes, CRTC1, MB21D2, USP44 and FGFR2, may drive bladder cancer through other mechanisms of pathway crosstalk, which requires further investigation.

A gene set comprising 10 genes is the only significant gene set specific to BLCA_CN compared to BLCA, among which NTRK1, HLA-A and HLA-B are associated with the PI3K-Akt signaling pathway (Fig 2C and 2D and Table S1.2 in S1 Data). In contrast, American patients with bladder cancer tend to harbor mutations that disrupt signaling by GPCR, with CDKN1A, IRS4 and KMT2D identified in the 6-gene set. Apart from these genes in specific pathways, some light gray genes in Fig 2D, such as ARAF, ESRRA and PTMA (from the China cohort), as well as KDM6A, PABPC1 and TRIM49C (from the American cohort), are predicted by ModSDP as region-specific drivers of bladder cancer and require further experimental validation.

Common pathways shared by patients from different countries can provide novel targeted drivers for cancer treatment. Meanwhile, the therapeutic significance of pathway prediction is embodied within specific findings distinguished among areas, as shown in Fig 2B. There are also differences in the specific signaling pathways among different cancer types in the same region, which emphasizes the need to distinguish both regional and cancer-type differences in cancer treatments.

Signal in esophageal and gastric carcinoma.

The esophagus and stomach, two crucial organs of the digestive system, are closely related. However, there is a lack of studies on tumor biomarkers through genotype similarity analysis. Therefore, understanding the etiologies of ESCA and STAD at the pathway level will facilitate screening and prevention efforts in high-risk populations.

By applying EntCDP to cohorts ESCA and STAD, we identified a common driver gene set when K = 9 (Fig 3A top and Table S2.6 in S1 Data). Among them, ARID1A, TP53, and CDH1 are also IntOGen genes in both cohorts. Although not listed by IntOGen for these cohorts, ACVR1 and CXCR4 have been implicated in the tumorigenesis of both cancers [27,28]. The remaining genes, including LDB1, PCBP1, CDK4, and UBE2D2, are novel predictions by EntCDP and may inform future therapeutic strategies. Enrichment analysis further supports the relevance of these genes, since TP53, CDH1, CDK4, and CXCR4 are typical representations of pathways in cancer (Fig 3B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Pairwise comparison of cancer types similar in location or function.

(A, C, E) The top panels show the common gene sets between two cancers identified by EntCDP (A & B), and the bottom panels display the specific gene sets of cancer A relative to cancer B (A / B) or vice versa (B / A) by ModSDP. Only significant sets are shown. Numbers in the leftmost or rightmost columns denote the parameter K, which means the number of genes in each set. Genes in the highlighted rows are enriched in the signaling pathway at the bottom of each ladder table. (B, D, F) Regulation of genes involved in the signaling pathways shown in (A, C, E), respectively. Genes marked in purple are identified by EntCDP or ModSDP, while genes in green are not.

https://doi.org/10.1371/journal.pcbi.1013349.g003

To investigate specific signals, we further applied ModSDP to identify distinct driver genes between cohorts ESCA and STAD. We identified PRKAR1A, SETD1B and TP53 as ESCA-specific driver genes, which are essential for the metabolism pathway. In contrast, ARID1A, NFE2L2 and TP53 were recognized as STAD-specific genes and serve as major drivers in the gastric cancer pathway (Fig 3A bottom and Table S2.7 in S1 Data). Notably, TP53 plays a prominent role in the development of esophagogastric cancers, not only because of its high mutation frequency, but also due to its enrichment in multiple cancer-related pathways. Therefore, we hypothesize that TP53 may be key to resolving gastrointestinal disorders.

Pathways involved in non-small cell lung cancer.

LUSC and LUAD are two major subtypes of non-small cell lung cancer (NSCLC) but demonstrate distinct genetic drivers and divergent prognostic outcomes [29]. Despite some success in the conventional treatment for lung cancer, subtype-specific targeted therapies remain necessary to understand genetic heterogeneity among subtypes.

The biological relevance of EntCDP is supported by the identification of the common gene set {CTNNB1, KRAS, MET, PTPN6, TP53} in cohorts LUAD and LUSC (Fig 3C top). Among them, KRAS is an IntOGen gene of cohort LUAD, whose mutation is also a critical criterion in targeted therapy for patients diagnosed with LUSC [30]. Additionally, this gene set is enriched in the proteoglycans in cancer pathway (Fig 3D). In a combined experiment [31], the expression of proteoglycan genes significantly decreased in tumors compared to normal lung tissue, aligning with our findings and suggesting their potential as promising biomarkers.

As for the specificity, we identified a LUAD-specific gene set: {EGFR, STK11, PTPRD}, among which EGFR and STK11 are involved in the mTOR and PI3K-Akt signaling pathways (Fig 3C bottom). It has been reported that dioscin, a therapeutic agent for lung adenocarcinoma, can suppress the proliferation, invasion and EMT of lung cells via the inactivation of AKT/mTOR/GSK3β signaling pathway [32]. Notably, PTPRD is a novel candidate predicted by ModSDP within LUAD-specific gene set and may represent a promising target for future experimental validation. By contrast, the LUSC-specific gene set includes {KMT2D, NFE2L2, PTEN, SGK1}. Among these, KMT2D and NFE2L2 are key regulators of LUSC tumorigenesis [33,34] and well-established components of the FoxO pathway. Therefore, we predict that PTEN and SGK1 may serve as effective therapeutic targets for LUSC by mediating the FoxO signaling pathway. Collectively, these findings demonstrate the ability of ModSDP to recover biologically meaningful and therapeutically relevant gene sets specific to cancer subtypes.

Typical malignant tumors in males and females.

Sex differences in cancer arise from the effects of circulating sex hormones [35], which influence organs unique to or significantly different between males and females.

In the previous section, we focused on breast cancer due to its high prevalence among women in many countries. Here, we compared ovarian cancer and breast carcinoma given their close association [36]. We found no common gene sets between cohorts BRCA and OV; however, we surprisingly identified significant commonalities between cohorts BRCA and OV_AU . For K = 8, the ErbB signaling pathway emerged as a common pathway, characterized by TP53, PIK3CA, RAC1 and FGFR3 (Fig 3E top and 3F). Among these, TP53 and PIK3CA are IntOGen genes of both cohorts, while the roles of the other six genes, ACKR3, EIF1AX, FGFR3, LOX, RAC1, and GATA3, have also been highlighted in studies of both ovarian cancer and breast carcinoma [37–47]. These findings highlight the biological relevance and robustness of the identified gene set.

For the differences, BRCA-specific analysis highlighted the T cell receptor (TCR) signaling pathway, whereas OV was specifically characterized by the TNFR1 pathway activation (Fig 3F). In BRCA, we identified CARD11, PIK3CA, PTEN, and GATA3 as key genes in the TCR pathway when K = 6 (Fig 3E bottom). Gu-Trantien et al. examined the prognostic factor of CD4+Tfh cells and the function of the T cell receptor signaling pathway in human breast cancer [48]. Conversely, we detected {TP53, CBL, NFKBIE, RBM10} as OV-specific drivers, and the first three are components of the TNFR1 pathway that exerts indirect cytotoxic effects on ovarian cancer stem cells [49]. Furthermore, we identified novel predictions in these contexts, including MED12 and RABEP1 in BRCA, and MED12 as an OV-specific candidate for further functional validation.

It seems that pathway analysis for the two male urological cancers is less specific, as the specific pathways identified are the pathways in cancer for KIRC and the transcriptional misregulation in cancer for PRAD, respectively. See Sect C in S1 Text for more details.

Beyond the four cancer type pairs detailed in our study, other comparable groups may also be informative in future clinical trials (Tables S2.11–S2.17 in S1 Data).

Comparison of adult and pediatric cohorts in AML and GBM.

The prominence of GBM among brain tumors underscored by its high morbidity and mortality rates in both children and adults [50]. Meanwhile, AML, a representative liquid tumor, is the most frequently diagnosed leukemia in adults (25%) and accounts for 15–20% of cases in children [51].

By leveraging somatic mutation data from both adult and pediatric cohorts of GBM and AML, EntCDP successfully identifies biologically meaningful common gene sets between two age groups. Specifically, for GBM (adults) and GBM_DE (children), four genes (CASP8, PIK3CA, PTEN, LPAR4) related to pathways in cancer are identified (Fig 4A right A & C and Table S3.1 in S1 Data). For LAML (adults) and AML_TARGET (children), the common pathway is cytokine signaling in the immune system, comprising six genes: TP53, FLT3, KIT, KRAS, NRAS, and PTPN11 (Fig 4A left A & C and Table S3.6 in S1 Data). This finding is consistent with the fact that abnormalities in cytokine signaling pathways are characteristic of all types of leukemia [52].

Age-specific differences in signaling pathways can guide the discovery of therapeutic targets tailored to pediatric and adult cancers. In GBM, CEBPA and ID3 in the NF-B signaling pathway (Fig 4A right C/A) are discovered as specific genes of children compared to adults when K = 8 (Table S3.4 in S1 Data), which may serve as potential targets in childhood glioblastoma. The adult-specific signaling is the MAPK signaling pathway, regulated by TP53, FGFR4, EGFR and IRAK1 (Fig 4A right A/C), which has been reported to promote glioblastoma malignancy [53]. Treatment for AML children may rely on downstream effectors of receptor tyrosine kinase (RTK), such as the Ras and phospholipase D signaling pathways [54], since we discovered NRAS, KIT, PTPN11 and KRAS in the two pathways when comparing AML_TARGET (children) to LAML (adults), LAML_CN (adults) and LAML_KR (adults) at K = 4 (Fig 4A left C/A and Table S3.7 in S1 Data). When K = 8, we detected {ARHGAP35, CEBPA, DNMT3A, ASXL1, PHF6, SUZ12, TNC, MTOR} as an adult-specific gene set for cohort LAML (Table S3.10 in S1 Data), among which CEBPA, MTOR, SUZ12 and ARHGAP35 are core components of the developmental biology pathway (Fig 4A left A/C). All remaining genes except for TNC have been previously implicated in AML [55–57]. Thus, we predict TNC as a novel candidate gene identified by ModSDP.

AML and GBM comparison within pediatric and adult groups.

Next, we summarized some meaningful outcomes from the comparison between AML and GBM conducted separately in children and adults.

The results show that the neurotrophin signaling pathway [58] is the common factor driving tumor development in pediatric patients from the cohorts AML_TARGET and GBM_DE_K. However, no gene sets were detected from mutations in adults by EntCDP, which indicates greater cancer heterogeneity observed in adult patients (Fig 4A middle and Table S3.11 in S1 Data).

The distinct pathways identified between adult cancers demonstrate the biological significance of ModSDP. As illustrated in the upper part of Fig 4A, we successfully identified CEBPA, NRAS, KIT, KRAS and RUNX1 as specific genes for adults with AML compared to those with GBM, and these genes are typical members of the acute myeloid leukemia pathway. On the other hand, adults with GBM are specifically affected by the PI3K-Akt signaling pathway (Table S3.12 in S1 Data), inhibition of which has been confirmed to be effective in clinical trials [59]. For children (lower part in Fig 4A), Ras and PAK signaling pathways are specifically associated with AML and GBM, respectively (Table S3.13 in S1 Data). As a tumor suppressor gene in the PAK pathway, inactivated PTPRD may lead to increased cell growth and suppression of apoptosis in GBM [60]. Mutations in the PTPN11, which encodes a phosphatase and regulates the Ras signaling pathway, are well characterized in children with AML [61,62]. Therefore, alterations in these two signaling pathways may have important implications for potential anticancer treatments.

In summary, the PI3K-Akt signaling pathway (PIK3CA and PTEN) is critical in adults with GBM, while the PAK signaling (PTPRD and PDGFRB) is specific to children with GBM. In AML, KIT, NRAS, and KRAS constitute core drivers, with CEBPA and PTPN11 further distinguishing adult and pediatric cases, respectively. These genes represent key molecular drivers within their respective biological contexts. Other genes, such as TNC in adults with AML, and CEBPA and ID3 in children with GBM, are proposed in our study as predicted therapeutic molecular targets within their associated signaling pathways.

Biomarkers for subtypes of pediatric hematological and neural tumors.

The final section focuses on two subtypes of leukemia: acute lymphoid leukemia (ALL) and acute myeloid leukemia (AML), as well as four subtypes of neural tumors: glioblastoma (GBM), neuroblastoma (NBL), medulloblastoma (MBL) and pilocytic astrocytoma (PA).

Both ALL and AML are common malignancies in children and young adults, but they differ in biology characteristics and clinical responses to various chemotherapies [63]. When analyzing similarities between the two subtypes, we found a set of 10 common genes shared by ALL_ST and AML_TARGET (Table S3.19 in S1 Data), among which TP53, NRAS, KIT, NF1, PTPN11 and KRAS are involved in the RET signaling. It is consistent with previous reports that AML cells exhibit activation of the RET signaling pathway [64]. Although there are few related reports for ALL, our results predict novel therapeutic targets, such as WT1 and FLT3. Additionally, the two subtypes, ALL_TARGET and AML_TARGET, differ in signals dominated by specific genes shown in Fig 4B and 4C. Surprisingly, we noticed that JAK2 and JAK3, two members of Janus kinase (JAK) family, are distinguishing markers for ALL and AML, respectively (Fig 4C).

Central nervous system tumors are the leading cause of cancer-related death in children [65]. In addition to GBM mentioned above, MBL and PA are particularly prevalent in the posterior fossa, and NBL is the most common extracranial solid tumor in children [66,67]. Despite much success, more targeted and effective therapies are needed for children whose tumors cannot be completely cured.

The commonality among different neural tumors can be observed only in the comparison between GBM and MBL (Table S3.23 in S1 Data). It is also worth noting that GBM and MBL cohorts have specific gene sets relative to other three brain tumors (Fig 4D and 4E). For example, when compared to NBL cohorts (NBL_TARGET and NBL_BROAD), GBM_DE has specific genes, such as PDGFRB and PTPRD in the PAK pathway (K = 4), as well as ID3 and CEBPA (K = 8) in the NF-B pathway (Table S3.25 in S1 Data), while MBL_DE_K and MBL_CA feature ID3, GNA11, and PRF1 (K = 10) in circadian rhythm regulation (Table S3.28 in S1 Data). This suggested ID3 could serve as a common regulatory gene for GBM and MBL. Unexpectedly, their common genes PTPRD, PDGFRB, RBM38, and KLHL36 are also identified as overlapping specific genes distinguishing GBM and MBL from NBL. These results may provide insights into potential common targeted therapies.

Despite the similarities between GBM and MBL, we also focused on their molecular differences and successfully identified their specific therapeutic pathways using ModSDP. CASP8, LRP1B, PDGFRB, RAC1 and SMO are representative specific genes for GBM_DE (Table S3.24 in S1 Data) and are involved in the ERK signaling. This pathway can inhibit the degradation of programmed death-ligand 1 (PD-L1) and maintain PD-L1 stability in glioblastoma [68]. On the other hand, we detected a specific gene set for MBL_CA: {NUMA1, CLIP1, PTCH1, STK11, CLTC}. Among these, NUMA1 and CLIP1 act in the cell cycle pathway, while CLIP1, PTCH1, STK11, and CLTC affect signaling by GPCR. This suggests that signaling pathway crosstalk may contribute to the development of MBL.

Subtype similarity is also embodied in NBL and PA, since the results of GBM versus PA are the same as those versus NBL when K ranges from 2 to 4 (Tables S3.25 and S3.26 in S1 Data). Compared with two NBL cohorts (NBL_TARGET and NBL_BROAD) and PA_DE, the specific genes of GBM_DE are {CEBPA, SMO} and {FOXA1, RGL3}, respectively. However, NBL exhibited relatively more heterogeneity than PA, because NBL (NBL_BROAD) has specific gene sets to PA (PA_DE), while PA lacks specific gene sets relative to any other subtype (Fig 4D). This specificity is observed in IRS4, ATM, FBN2, and FLT4 (Fig 4E and Table S3.27 in S1 Data, K = 9), which are involved in the ERK signaling pathway that has been verified to impact NBL progression [69]. The two MBL cohorts (MBL_DE_K and MBL_CA) also have specific genes compared to PA_DE (Table S3.29 in S1 Data). When K = 5, we detected {FOXA1, KLHL36, PPP2R1A, PTPRD, RABEP1}, among which PPP2R1A and RABEP1 are involved in the regulation of CFTR activity pathway.

Overall, by comparing to PA, we identified subtype-specific pathways in neural tumors: PAK signaling (PTPRD and PDGFRB) in GBM, regulation of CFTR activity pathway (PPP2R1A, RABEP1) in MBL, and ERK signaling (IRS4 and ATM) in NBL. In addition, based on the specificity relative to NBL, we predicted RBM38, KLHL36, and ID3 as potential subtype-specific genes for GBM and MBL, while LRP1B and PTPN11 appear to be specific to NBL. These predicted genes are promising for follow-up studies on their roles in signaling pathways.

Treatment for pediatric tumors should differ from that for adults, especially in light of mutations specific to children. The specificity among different tumor types also suggests the extent of tumor heterogeneity and the importance of prioritizing for targeted therapy. This underscores the role of personalized therapy, which can be achieved by careful stratification of patient populations.

Risk factors differ between smokers and nonsmokers.

It is well established that lung cancer is closely associated with tobacco use [70,72]. The fact that smokers outnumber nonsmokers in cohort LUAD further demonstrates the higher risk of lung cancer among smokers (Fig 5A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Three environmental agents promote tumor progression. Specific signaling pathways in patients with risk exposures to tobacco (A), alcohol (B) and high BMI (C) are illustrated on the left side, while the control groups without harmful habits or abnormal indices are shown on the right.

The pie chart in the middle shows the proportions of the two groups. The heatmaps on both sides exhibit the gene mutation profiles of specific genes in the exposure group (the two maps on the left) and the control group (the two maps on the right) when parameter K is 5. Heatmaps in the same horizontal line share the same gene set. Genes enriched in identified pathways in the middle are filled in dark blue.

https://doi.org/10.1371/journal.pcbi.1013349.g005

To begin with, we analyzed the common driver pathways shared by all LUAD patients, regardless of their smoking history. Using EntCDP with K = 9, we detected {KRAS, RAF1, TP53, MET, CTNNB1, PTPN6, FOXA1, MAX, NT5C3A}. The first three genes are enriched in pathways related to non-small cell lung cancer, while the first four genes are enriched in the central carbon metabolism in cancer, a frequently altered pathway in lung adenocarcinoma [73]. These results demonstrate the effectiveness of EntCDP in capturing functionally meaningful genetic patterns.

However, what causes lung cancer in those patients who have never smoked remains unclear. Notably, when focusing on non-smoking LUAD patients, we identified signaling by GPCR when K = 9 (Fig 5 and Table S4.1 in S1 Data). There is reason to believe that EGFR plays a prominent role in inducing LUAD through this pathway, since it exhibits a high frequency of mutations and has been identified as a prognostic factor for lung cancer in never smokers (LCINS) [74]. In addition, we presume that the mutated CYSLTR2 may be involved in lung cancer development through the production of aberrant GPCRs, based on the evidence observed in other cancers [75].

Human life is seriously affected by another smoking-associated cancer: head and neck squamous cell carcinoma (HNSC) [76], the sixth most common cancer worldwide [77]. There are significant differences in signaling pathways between HNSC patients exposed to tobacco and those who are not. In patients without a history of smoking, RASA1, BRAF and CACNA1D may promote the proliferation of head and neck cells by regulating the MAPK signaling pathway (S3A Fig and Table S4.2 in S1 Data). For smokers, we identified COL1A1, PIK3CA, and LPAR4 as specific genes in the PI3K-Akt signaling pathway, which has been confirmed to be associated with HNSC [78,79]. We predicted the T-cell leukemia virus-1 (HTLV-1) infection pathway as a novel finding in HNSC because we identified XPO1 and EP300 as smoker-specific biomarkers involved in this pathway.

Shedding light on alcohol exposure.

Beyond tobacco use, alcohol consumption is another major risk factor for head and neck cancer [80]. We found that drinkers generally carry specific genes such as NOTCH1, CDKN2A and BRCA1 (S3C Fig and Table S4.4 in S1 Data). Their important role in the GPCR pathway is consistent with evidence that changes in GPCR function are implicated in ethanol exposure [81]. For nondrinkers, we detected {CASP8, FH, MAPK1, RUNX1, STAT3, ETV5, EXT2, NRP1, ZNF626} when K = 9. The first five genes are members of the pathways in cancer, among which CASP8, STAT3 and MAPK1 are related to viral carcinogenesis, including Kaposi’s sarcoma-associated herpesvirus and human cytomegalovirus (HCMV) infection. This highlights the need to explore virus-related mechanisms in non-alcohol-associated tumor development.

In addition to HNSC, drinkers make up almost three-quarters of the patients in the cohort ESCA (Fig 5B). In particular, NOTCH1, characterized as a drinker-specific gene in HNSC, is also identified in ESCA drinkers, which underlines its core status in alcohol-related cancers [71]. TP53 also attracts our attention because its high mutation rate exceeds 80% in both alcoholics and nonalcoholics. In addition, the remaining patients who never drink may be affected by a pathway called MHC class I antigen presentation (Fig 5B). The pathogenesis of eosinophilic esophagitis has been reported to depend on presenting antigens on the major histocompatibility complex (MHC) class II [82], which may imply the threat of MHC Class I antigen presentation in esophageal cancers.

Obesity as an underlying factor disrupting signaling pathways.

It is believed that being overweight may also increase the risk of many cancers, such as stomach adenocarcinoma, liver hepatocellular carcinoma, breast carcinoma, and colorectal cancer [83]. Therefore, we investigated whether signaling pathways can serve as a bridge from obesity to cancer.

It has been reported that the PI3K/Akt/mTOR cascade is a target of many obesity-associated factors that regulate cell proliferation and survival [84]. We verified this result when we focused on patients with hepatocellular carcinoma. We identified CDKN1A, ATM, SGK1, APC and CHD4 (Table S4.7 in S1 Data, K = 9) as specific genes of overweight individuals. Enrichment analysis further revealed that these genes can be mapped to the PI3K-Akt signaling pathway. For those of normal weight, the hepatitis B pathway, composed of RB1, EP300, PIK3CB, HSPG2 and JAK1 (K = 9), may serve as a key point for effective treatment and prevention (Fig 5C), because chronic infection with hepatitis viruses is a major causative factor for hepatocellular carcinoma [85].

Additionally, it is particularly common that people who are obese also tend to have colorectal cancer (S3E Fig). We conclude that mutation in CDKN1B, CDKN2C, TP53, ATR, and CCR7 may increase the risk of developing colorectal cancer through the ERK signaling pathway in overweight individuals. However, those with a normal BMI should be cautious to prevent Yersinia infection [86], since MAP2K1, PLCG1, and MYD88 exhibit mutual exclusivity and are significantly enriched in this pathway.

In summary, viral and bacterial infections as causes of human cancer have gained new significance because we identified the HCMV infection, HIV life cycle, hepatitis B, and Yersinia pathway in several control groups. Moreover, we have reason to believe that the Notch-mediated pathway may be responsible for bridging alcohol and cancers, while obesity would increase the risk of cancer through signals regulated by the CDKN gene family. For more details about risk factors for other cancers, please refer to Sect D in S1 Text.