The Cancer Spectrum Theory.

As described in previous sections, ample evidence exists that supports the cancer spectra of tumor characteristics. However, simple dichotomization approaches have prevailed in many cancer research studies. Why are dichotomized approaches so common? There are several possible reasons. First, such dichotomized approaches are embedded in most clinical decisions for patient management. In clinical settings, we need to make dichotomized clinical decisions regarding whether or not to administer a medication, to perform a specific surgical procedure, to give a radiotherapy, etc. Second, the dichotomized way is what we were educated in colleges and graduate schools. For example, scientists and researchers learn case–control studies (as a basic design of observational studies), which are based on a dichotomized approach (case vs. noncase; present vs. absent, etc.). Outcomes recorded are often binary such as response versus resistance to treatment, recurrence versus nonrecurrence, death versus survival, etc. Third, statistical analyses and interpretability are simpler for dichotomized approaches (e.g., χ 2 test or logistic regression model) compared with multicategorical and continuous analysis approaches. However, we must be aware of limitations associated with simple dichotomizations. The dichotomized approach may not be able to adequately reflect heterogeneous effects originating from cancer spectra. Moreover, dichotomizing or categorizing variables without prior knowledge of any biological cut-off points may result in loss of information and bias because of the arbitrary selection of cut-off points. Therefore, detailed information on cancer spectra should be integrated into large-scale population studies as much as possible.

Advances in biomedical sciences are revealing inherent heterogeneity of pathogenesis. These advancements have led to a shift from the conventional epidemiologic approach, in which a given disease is treated as a single uniform category, to the molecular pathological epidemiology approach, which can divide a single disease entity into distinct pathogenic tumor subtypes and examine specific etiologic associations with each subtype ( 13 ). Each exposure (sometimes a disease risk factor) influences systemic conditions, tissue microenvironment, and cells in tissues in exposure-specific fashions. As the exposome (i.e., the totality of exposures) is unique to each individual and cannot be shared exactly by any two individuals, disease processes in each individual must be distinct. This notion is conceptualized as the “unique disease principle” (“unique tumor principle” for neoplastic diseases), which makes the scientific basis for the “disease subtyping spectrum theory.” The uniqueness of each disease can be manifested as biomarker metrics measured in biospecimens. For instance, multi-omics studies revealed that there are no tumor pairs that share the exact same set of mutations and epigenetic changes. In addition, many tumor phenotypes are continuous in nature. Classifying continuous tumor phenotypes into binary (or a limited number of) categories may lead to loss of important biological information. Tumor subtyping based on somatic mutations in a certain oncogene or tumor suppressor gene or protein expression may be regarded as binary (mutated vs. nonmutated; expressed vs. nonexpressed). However, this also turns out to be an oversimplification. Downstream effects of various somatic mutations in a given gene are nearly always heterogeneous, forming a spectrum (from strong effect to neutral/no effect). In the current era of cancer immunotherapy, one of the important tumor biomarkers is a tumor mutation burden (TMB), which is the number of nonsynonymous mutations per a defined number of sequenced nucleotides (such as megabases). Currently, given the widespread use of sequencing assays of a panel of genes, TMB is a common biomarker, which may guide treatment decision-making. Nonetheless, TMB-based tumor subtyping is challenging because there is no clear cut-off point in any given tumor type. Tumor long interspersed nucleotide element 1 (LINE-1; a.k.a., long interspersed nuclear element 1) methylation level has been linearly associated with better colorectal cancer survival without any threshold of effect ( 2 ). Another study indicates a continuity of transcriptome profiles in colorectal cancer ( 14 ). Regarding tumor subtyping based on characteristics of the tumor–immune microenvironment, tumors exhibit considerably heterogeneous phenotypes according to immune cell infiltrates and microbial profiles in tumor tissue. We should be aware that any tumor subtyping system may have a spectrum of tumors with heterogeneous phenotypes rather than clear-cut binary or minimally categorical tumor subtypes.

Cancers are traditionally classified by the organ system. One may match one cancer type with one organ without considering detailed anatomic subsites. Nonetheless, even cancers that occur within the same organ often have different tumor characteristics depending on detailed anatomic subsites in which cancers arise. To account for heterogeneity in the local tissue microenvironment according to biogeographic location, detailed anatomic location is needed. However, information on detailed anatomic subsites in a given organ (which often have vague boundaries) has not typically been incorporated in cancer population studies. Traditionally, the colon has been divided into two segments: the proximal (right-sided) and distal (left-sided) anatomic segments, using the splenic flexure as a cut-off point. Numerous studies dichotomized colorectal carcinomas into proximal and distal tumors, possibly because of limited availability of detailed location data or insufficient statistical power; however, even with this limited categorization, there is evidence of differences in molecular pathologic features within the dichotomized sites. In particular, intestinal contents, including food debris, microbiota, and microbial metabolites, likely change continuously according to detailed colorectal sublocations (i.e., cecum, ascending colon, hepatic flexure, transverse colon, splenic flexure, splenic flexure, descending colon, sigmoid colon, rectosigmoid, rectum) without abrupt change at a specific colorectal site, questioning a scientific rationale of a dichotomous approach. For instance, a study has shown that the prevalence of tumors with high amounts of Fusobacterium nucleatum (a Gram-negative oral anaerobe and a significant contributor to colorectal cancer) gradually decreases from the cecum to rectum ( 5 ). Multiple studies have demonstrated the tumor molecular features (e.g., MSI, CpG island methylator phenotype, and BRAF and PIK3CA mutations) gradually decrease across several colorectal segments from ascending colon to rectum ( 3 , 6 ), providing a rationale for the colorectal anatomic location spectrum. Another epidemiologic study has shown differences in risk factor associations (such as smoking, alcohol, obesity, and family history of colorectal cancer) between detailed colorectal subsites ( 7 ). Furthermore, a recent study has also shown a trend for better survival in relation to tumor location from the cecum to sigmoid colon, especially in patients with non–MSI-high tumors ( 8 ). In contrast, patients with MSI-high cancer have shown a suggestive trend for worse survival from the cecum to sigmoid colon, implying the need to account for both anatomic location and tumor MSI status for better prognostication ( 8 ). The colorectum colorectal anatomic location spectrum is a biologically rational theory, considering gradual changes of intestinal luminal contents and the microbiome along the colorectum. A similar paradigm can be applied to other cancer types. Gastric cancer is typically classified as cardia and noncardia subtypes. Even within noncardia gastric cancer, tumors in multiple anatomic subsites may have different etiologies. For example, gastric cancer in the fundus and corpus is more autoimmune gastritis–related, while gastric cancer in other noncardia anatomic locations is more Helicobacter pylori –related ( 9 ). Primary tumor location in lung cancer is also relevant. Besides histologic types, evidence indicates that tumor location affects the prognosis of lung cancer ( 10 ). Non–small cell lung cancer in the lower lobe tends to have poor prognosis compared with that in the upper lobe ( 10 ). Further research that accounts for these anatomic location spectra in various organ sites is needed because such studies can provide novel insight into the architectural aspect of tumor development and progression by anatomical sites. A practical limitation is a lack of detailed information on anatomic locations (beyond organ systems) captured in clinical and pathology records for many cancer types. Future studies could potentially include digitized radiography images that could allow for research-level precise quantifications of three-dimensional tumor location as has been demonstrated by breast density studies using screening mammograms.

There are multiple rationales for applying the cancer spectrum theory to cancer research. However, challenges also exist. Figure 1 summarizes rationales and challenges in their applications. The preceding section briefly discusses challenges in conducting large-scale research incorporating the cancer spectrum theory. We acknowledge that research design accounting for cancer spectra demands bigger sample sizes to adequately analyze detailed variables (e.g., age, anatomic location) with reasonable statistical power. It necessitates effort for the collection of detailed clinical, epidemiologic, and tumor characteristics metadata in a large number of individuals. Therefore, pooled analyses and/or multicohort collaboration are crucial for research to address the cancer spectra. Statistical analyses that account for the cancer spectra are more complicated. There is an increasing need to devise statistical methods that can analyze cancer spectra in population research. For example, a new analytic method for evaluating how the exposure–disease association changes across continuous levels of a subtyping biomarker has been developed ( 15 ). Other analytic methods to deal with continuous, ordinal, and multicategorical measures need to be devised. Effective collaboration with experts in biostatistics and bioinformatics should be prioritized. Interdisciplinary approaches are also needed for cancer spectrum research. Experimental models using cell lines, organoids, animals, etc., can be used to study cancer spectra. However, even the best experimental model cannot exactly replicate the complexity of human tumors and their microenvironment. Findings from basic experimental studies need to be validated using human subjects, which is crucial in applying these findings to clinical settings. Therefore, we need collaborative efforts of diverse experts such as basic researchers, population scientists, and clinicians.

Emerging evidence suggests gradual differences in clinical and tumor characteristics, forming phenotypic spectra in terms of age of diagnosis, anatomic spatial location, tumor subtyping, and tumor neoplastic property. Therefore, there exists the need for a paradigm shift in research designs and analytic approaches accounting for various spectra in the era of precision cancer prevention and therapy. The integration of the cancer spectrum theory into large-scale population studies will contribute to a better understanding of the etiology of cancer, ultimately leading to effective prevention and precision oncology.

A neoplasm exists along a spectrum ranging from normal-appearing tissue with some cellular molecular alterations to premalignant lesions (benign neoplasms) to invasive nonmetastatic tumors to metastatic tumors. However, population studies often used a dichotomized variable of cancer versus noncancer, not taking this spectrum into account. In this spectrum, not all normal-appearing cells with molecular alterations progress to benign neoplasms. Not all benign neoplasms progress to invasive tumors. For instance, only a small portion (<5%) of colorectal adenomas are considered to progress to colorectal carcinomas, and the risk of progression differs by histologic subtypes, size, and location ( 11 ). Binary classification into lethal versus nonlethal cancer is sometimes used in cancer research for prostate, breast, ovarian, and colorectal cancer, etc. However, this can be problematic because some cancers are very lethal with aggressive behaviors and others are mildly lethal, forming a spectrum. Another issue is that exposures including treatment and lifestyle factors during cancer progression can modify the malignant property of tumors and then change their clinical course and outcomes. Importantly, cancer is not only a disease of neoplastic cells but also an environmental and microenvironmental disease that can be modified by the exposome ( 12 ). One caveat is that we know the true property of a given neoplasm only after we get information on clinical outcomes of the patient (usually many years after treatment). This means that we do not exactly know but can only infer without not knowing the exactly true property of any neoplasm at the time of diagnosis based on its clinical and tumor characteristics and our current knowledge. New developments, such as noninvasive diagnostic techniques, artificial intelligence, and machine learning, may be able to help better evaluate the true property at diagnosis. We need to study precursor–cancer spectra as they can provide clues as to how internal and external factors may contribute to tumor progression and potentially shed light on novel early detection strategies. At the same time, we need to recognize that not all precursors progress to invasive cancer and that any precursor or any malignancy at a specific stage or grade is not a homogeneous entity. In addition, time to progression from premalignancy to invasive cancer differs by each cancer. This is important because longer duration increases chances for interventions or early detection if there is any available intervention, but, at the same time, it increases the potential for overtreatment among overdiagnosed indolent tumors. Hence, there is a need for research accounting for the precursor–cancer spectrum theory to elucidate the heterogeneous process of tumor evolution from premalignancy to malignancy.

Cancer is a heterogeneous disease influenced by somatic molecular alterations, host immunity, genetics, diet, lifestyle, microbiota, and environment. Cancer exists along several spectra (or the continuum) in terms of several factors, such as age at onset, anatomical spatial localization, tumor subtyping, and the degree of tumor aggressiveness (encompassing a tumor neoplastic property spectrum). However, the concept of these cancer spectra has not been fully integrated into large-scale population research. For example, clinical and tumor characteristics often gradually change by age at onset, but dichotomized age cut-off points (e.g., 50 years) have been frequently used in early-onset cancer research, not accounting for the continuous cancer age spectrum. Another example is the cancer anatomic location spectrum. Certain clinical and tumor characteristics of colorectal cancer gradually change across colorectal subsites from cecum to rectum, but numerous studies dichotomize colorectal cancer into proximal (right-sided) and distal (left-sided) tumors using the splenic flexure as a cut-off point. Similar spatial localization spectra exist in tumors arising in other organs and body sites such as skin, lung, head and neck, esophagus, stomach, pancreas, biliary tract, bladder, etc. A third example is the degree of tumor aggressiveness, which encompasses but extends far beyond the precancer–cancer spectrum. Cancer exists along a spectrum ranging from normal-appearing cells with some molecular alterations to premalignant lesions to invasive but nonmetastatic tumors to metastatic tumors. However, typical cancer population studies use a dichotomous variable of cancer or not (noncancer) as an outcome or an inclusion criterion, without much consideration of the tumor aggressiveness spectrum. The objective of this article is to provide examples of cancer spectra and emphasize the importance of integrating the cancer spectrum theory into large-scale population cancer research. The cancer spectrum theory is a biologically-driven concept, considering the pathophysiologic continuum of neoplastic cells, microenvironmental constituents, and their interactions. Large-scale population research that reflects cancer spectra can contribute to a better understanding of the etiology of cancer, ultimately leading to effective prevention and precision oncology.