Oncology research produces data about a wide variety of entities such as tumor types, locations, pathology, and staging, patient treatments and outcomes, and experimental systems such as mouse models and cell lines. In order to conduct effective cancer research, terminologies, classification systems, and ontologies are needed that can integrate these various datasets and provide standards for consistently representing entities. In this paper, we discuss our ongoing efforts to address these difficulties by developing a realism-based ontology for representing instances of malignant neoplasms, disease progression, treatments, and outcomes. This ontology is being built using the principles of the OBO Foundry, and makes use of other OBO Foundry ontologies, such as the Ontology for General Medical Sciences, Uberon, and the Cell Ontology. As a result of our efforts, we have made worthwhile progress towards developing a robust ontological framework for representing malignant neoplasms.
Oncology research produces data about a wide variety of entities such as tumor types, locations, pathology, and staging, patient treatments and outcomes, and experimental systems such as mouse models and cell lines. In order to conduct effective cancer research, terminologies, classification systems, and ontologies are needed that can integrate these various datasets and provide standards for consistently representing entities. These standards facilitate the meaningful linking, sharing, and analysis of disparate datasets between researchers and across institutions. However, the incomplete and inconsistent representation of cancer-related data makes it difficult to perform these activities.
In this paper, we discuss our ongoing efforts to address
these difficulties by developing a realism-based ontology for
representing instances of malignant neoplasms, disease
progression, treatments, and outcomes. This ontology is being
built using the principles of the OBO Foundry
This research developed out of a number of interests. The first is that we recognized a need to connect cancer data from multiple sources with differing levels of granularity. Some important levels include: (1) diagnosis and treatment information about the patient and how the patient responses to treatment; (2) anatomical information about the organs in which the cancer originates; (3) pathology information about the tissues removed during procedures, such as tumor tissues and lymph nodes; (4) cellular information, such as data obtained from flow cytometry and immunohistochemistry; (5) and molecular information, such as genomic sequencing. Providing a framework for tying these kinds of data together is essential for cancer research by providing the basis for the use of advanced ontology-based querying and analytical methods that allow for data integration across multiple sources and scales.
A number of existing classification systems, ontologies,
and terminologies have terms for representing malignant
neoplasms. Prominent examples include the International
Statistical Classification of Diseases 10th Revision
(ICD10), the International Statistical Classification of Diseases
for Oncology (ICD-O), the National Cancer Institute
Thesaurus (NCIT), and the Systematized Nomenclature of
Medicine Clinical Terms (SNOMED CT). However, since many
of these information organization systems do not share a
common upper-level framework, it is not easy to leverage
information contained in other terminologies and
ontologies. For instance, SNOMED CT does not have terms for
checkpoint inhibitors, whereas the NCI Thesaurus does.
Ideally, we would like to use terms from each system (i.e.,
SNOMED CT and NCIT), but due to differences in their
relations and hierarchical structures, it is difficult to do so.
For example, the term ‘metastasis’ denotes a disorder in
SNOMED CT, but denotes the spread of cancer (i.e., a
process) in the NCIT. OBO Foundry ontologies, in contrast, are
generally designed using the Basic Formal Ontology (BFO)
as their upper-level framework, and this enables the creation
of domain specific ontologies whose terms can be reused by
other OBO Foundry ontologies. For instance, the Drug
Ontology
International Statistical Classification of Diseases
Due to its long history and widespread adoption, the International Statistical Classification of Diseases (ICD)1 is perhaps the most relevant system for classifying diseases. Maintained by the World Health Organization (WHO), ICD is a globally recognized healthcare classification system consisting of hierarchically structured codes that represent diseases, disorders, and other health related issues.2 In relation to the current topic, the International Statistical Classification of Diseases for Oncology (ICD-O)3 has codes for representing a number of pertinent characteristics of a malignant neoplasm, such as the anatomical site of the neoplasm, the neoplasm’s histology (e.g., small cell, clear cell), and behavior (e.g., if it has metastasized). For example, an ovarian adenocarcinoma is represented using the following combination of codes: • •
C56 – the site code for an ovary 8140/3 – 8140 is the code for a neoplasm arising from glandular epithelial tissue, and ‘/3’ represents that the neoplasm is malignant
The advantage of ICD’s coding system is that allows diseases to be easily grouped and counted for statistical and reporting purposes. For instance, to find all patients who have an adenocarcinoma, you only have to look for patients whose histological code begins with ‘814’ and has a behavior code greater than 3. However, there are two related noteworthy drawbacks to implementing ICD as an ontology. First, ICD does not contain codes for many of the important cancer related entities that need to be represented, such as treatments and molecular disorders. This shortcoming is compounded by ICD’s lack of formal relations that would allow codes to be linked to other information. Thus, even if we created code lists for the missing entities, we would still be faced with the task of creating well-defined relations that would allow this information to be linked to ICD codes. 3.2
The National Cancer Institute Thesaurus (NCIT) is a
reference terminology developed by the National Cancer
Institute
In our examination of the NCIT, we found that many of the definitions in the malignant neoplasms branch were sufficiently defined and the hierarchy was rich enough to suit our purposes. However, when we examined other branches of the NCIT, certain problems became apparent. In particular, we found the definitions for cell types related to cancer to be inadequate. Consider the following NCIT concepts and definitions: • • •
Abnormal Cell (C12913): An abnormal human cell type which can occur in either disease states or disease models.
Neoplastic Cell (C12922): Cells of, or derived from, a tumor.
Malignant Cell (C12917): Cells of, or derived from, a malignant tumor.
The definition for Abnormal Cell suffers from its being circular (i.e., an abnormal cell is defined as being an abnormal cell type), and thus the definition does not provide any new information. Furthermore, the definition specifically states that an abnormal cell is a human cell. This prevents the NCIT from consistently modeling data about abnormal cells from non-human species despite the fact that the NCIT does contain concepts for mouse diseases, such as Mouse Carcinoma (C24010). Given the importance of mouse models in cancer research, not being able to represent data from mouse studies correctly is a severe limitation.
The definitions for Neoplastic Cell and Malignant Cell do not provide much clarity about how these cells relate to neoplasms. Since a neoplasm may also contain normal cell types, more details are needed about what it means to be a neoplastic cell other than being derived from a tumor. Furthermore, while a metastasis may be said in some sense to derive from a tumor, this cannot be said of the originating neoplastic cells that first started proliferating during the tumor formation process. Lastly, it needs to be pointed out that these cell types form a hierarchy. A Malignant Cell is a type of Neoplastic Cell, and Neoplastic Cell is a type of Abnormal Cell. This information is not contained in the textual definitions in an Aristotelian fashion, although it is represented in NCIT’s taxonomic relations. 3.3
Systematized Nomenclature of Medicine Clinical Terms
Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT) is a comprehensive health terminology that provides a standardized way to represent clinical information in an electronic health record.5 Although SNOMED CT has a large number of terms for clinical findings and disorders, it does not have worked out terms for other terms 5 http://www.snomed.org/snomed-ct/what-is-snomed-ct, accessed 2017-0621. related to neoplasms. For example, the concept Tumor cell (SCTID 252987004) is defined as subtype of the concept Abnormal cell (SCTID 39266006), but this does not specify if the concept Tumor cell represents malignant cells.6 Furthermore, the concept Malignant tumor cells (SCTID: 88400008) is defined as being a subtype of the concept Malignant neoplasm, primary (SCTID: 86049000).7 This classification is incorrect for at least two reasons. First, although malignant tumor cells are often part of a malignant neoplasm, they are not a kind of malignant neoplasm. A malignant neoplasm (as stated above) will also include a number of non-cancer cells as part of its makeup. Second, even if we accept that a malignant tumor cell is a kind of malignant neoplasm, this definition is incorrect because a malignant tumor cell is also found in a metastasis (metastatic offshoot of a primary tumor). Finally, SNOMED CT classifies Malignant tumor cells as a kind of Morphologic abnormality (SCTID 4147007), and not a Disorder (SCTID 64572001). In SNOMED CT, the distinction between a morphologic abnormality and a disorder is that some underlying pathological process supports a disorder.8 However, the reason that cell becomes malignant is because of underlying pathological processes (resulting from dysregulation) occurring with it. 3.4
The Disease Ontology (DO) is an OBO Foundry
ontology built for the purposes of providing the biomedical
community with consistent, standardized, and reusable
definitions to represent the range of human diseases
First, the DO is not consistent in its use of the terms ‘cancer’ and ‘neoplasm’. In DO, cancer is defined as a kind of disposition. This means that cancer is not a material thing (i.e., does not have mass), but rather is a kind of latent potential that is actualized when cells start proliferating out of control. Malignant neoplasms, are material objects that come into being due to uncontrolled cell proliferation. In the DO, however, there are number of terms in the cancer branch that reference neoplasms as material things and not the disposition of cancer. For example, ovary neuroendocrine neoplasm is defined as a subtype of ovarian cancer. Because of DO’s inconsistent use of terms ‘cancer’ and ‘neoplasm’ and our remaining true to the OBO Foundry principles, we decided it would be beneficial to the development of our ontology to use the term ‘malignant neoplasm’ and avoid using the term ‘cancer’ when possible. 6 http://browser.ihtsdotools.org, accessed 2017-06-22. 7 Ibid. 8 https://confluence.ihtsdotools.org/display/DOCEG/6.1.1+Clinical+ +-+definition, accessed 2017-06-22. Second, the DO is missing needed formal axioms that relate entities having the disposition of cancer to the anatomical structures in which these entities are located. For instance, the DO term ovary epithelial cancer does not have axioms that formally relate the disposition to the epithelial cells that are part of the ovary. The lack of these axioms can make it difficult to query data modeled using the DO. For example, it is not possible to query for the most common anatomical structures in which malignant neoplasms are found. 3.5
On carcinomas and other pathological entities
In Smith et al. (2005b), the Ontology for Biomedical
Reality
An anatomical entity is defined as being a material
pathological entity when
It has come into being as a result of changes in some pre-existing canonical anatomical structure through processes other than the expression of the normal complement of genes of an organism of the given type.
It is predisposed to have health-related consequences for the organism in question manifested by symptoms and signs.
Material pathological entities include pathological structures and pathological bodily substances. These are anatomical structures and body substances, respectively, that host some kind of pathological formation, a formation being pathological when it affects an organism’s physiological processes to the degree that they give rise to signs and symptoms. For instance, a carcinoma is a pathological formation that arises within an anatomical structure, such as an ovary.
A high-level summary of the hierarchy for material anatomical, material pathological entities is depicted below: material anatomical entity o anatomical structure § canonical anatomical structure § variant anatomical structure portion of canonical body substance (e.g., portion of blood) material pathological entity o pathological structure (e.g., neoplasm) o Portion of pathological substance (e.g.,
portion of pus)
Pathological formations are then related to their hosts
and the entities out of they originate using the following
relations from the Open Biomedical Ontology
As an example, suppose a patient (patient1) has a carcinoma (carinoma1) that originated within her ovary (ovary1). We represent this using the axioms: ovary1 at t part of patient1 at t carcinoma1 at t part of ovary1 at t carcinoma1 at t instance of pathological structure at t
Because carcinomas arise from the epithelial tissue lining of organs, we can assert the following about the patient’s tumor: 9 Hereafter, relations are represented in bold. 10 In the referenced Open Biomedical Ontology relations, the name the relation is named ‘derives from’. However, to avoid confusion, we use the term as presented in the paper. carcinoma1 at t derived from some epithelial cell at some t’ prior to t
And, since the part of relation is transitive, we infer that:
carcincoma1 at t part of patient1 at t
Although this inference is trivial, the advantage of representing the patient’s tumor in this manner is that we are not required to explicitly state this within an information system using the ontology. Rather, we let the computer system handle this through automated inferencing.
The benefit of doing this becomes apparent when we consider the multiple ways we classify malignant neoplasms. A malignant neoplasm may be classified according to: • • • •
The cell type from which the neoplasm is originates, e.g., carcinomas arise from epithelial cells, and sarcomas arise from non-epithelial cells.
The organ in which the neoplasm develops, e.g., an ovarian carcinoma originates in the ovary.
The organ system to which the organ of origin belongs, e.g., an ovarian carcinoma is a kind of reproductive system cancer The anatomical site or region in which the organ of origin is found, e.g., a tongue carcinoma is a kind of head and neck cancer.
When such classification information is axiomatized, we can then query the information system along these multiple axes without have to maintain complex data structures that explicitly assert this information. For instance, we can now query an information system for all carcinomas (i.e., malignant neoplasms that are derived from epithelial cells) that belong to patients’ reproductive systems without having to explicitly link each kind of carcinoma (e.g., ovarian, uterine, testicular) to the organ and associated organ system. 4
While we consider the work of Smith et al. to be a
significant improvement over the aforementioned classification
systems and terminologies, a number of ontologies have
been developed after this work was published. We take
advantage of these more recent ontologies as follows. First, we
make use of the terms and relations from the Cell Ontology
(CL)
Second, an important criterion in Smith et al.’s definition
of an entity being pathological is that it is predisposed to
have health related consequences
Recall that above we criticized the DO for its inconsistent usage of the term ‘neoplasm’. However, given the need to represent the dispositional aspect of cancer, we find DO’s hierarchy appropriate for characterizing cancer as we are clear and consistent about which sense of ‘cancer’ we are using. 11 DO uses the OGMS term disease. Third, in order to account for Smith et al.’s distinction between material pathological entities and material anatomical entities, we adopt OGMS’ account of a disease (as a disposition) being based on a disorder: disorder: A material entity which is clinically abnormal and part of an extended organism. Disorders are the physical basis of disease.
Since OGMS uses the Basic Formal Ontology (BFO) as its upper-level framework and a disease in OGMS is type of BFO disposition, it cannot (like all dispositions) exist on its own. Rather, a disease must be borne by a disorder whose structural abnormalities serve as a disease’s basis. For example, a sprained ankle is a disorder in the sense that the physical structures are clinically abnormal, and these physiological abnormalities are the reason that a sprained ankle is disposed to swell.
Fourth, to relate a disease to the disorder upon which it is based, we define the has material basis in relation as follows: has material basis in: A primitive relation between an instance of a disease i and an instance of a disorder j at particular time t in which i exists because of the physical makeup of some part of j at time t.
In addition to relating a disease to its basis, we must also account for the processes that realize (or make manifest) an instance of a disease. For this we use OGMS’ term pathological bodily process: pathological bodily process: A bodily process that is clinically abnormal.
As observed in the definition, a pathological bodily process is a type of bodily process. However, the term bodily process is not defined in OGMS.
Fifth, in order to account for the temporal development
of malignant neoplasms, we make use of the Relations
Ontology’s derives from and develops from relations
abnormal cell derives from native cell entails that a particular native cell no longer exists once the abnormal cell derived from it comes into existence.
The develops from relation also represents new entities that arise from previously existing entities, but does not require that the originating entity cease to exist. This allows us to represent that an instance of a secondary neoplasm material anatomical entity: An anatomical entity that has mass. anatomical structure: A material anatomical entity that is a single connected structure with inherent 3D shape generated by coordinated expression of the organism's own genome. pathological anatomical structure: A material entity that comes into being as a result of changes in some pre-existing anatomical structure through processes other than the expression of the normal complement of genes of an organism of the given type, and is predisposed to have health-related consequences for the organism in question manifested by symptoms and signs.
We note here that although intuitively a pathological anatomical structure is a type of anatomical structure, for reasons that will be discussed below, we classify them in separate hierarchies. Moreover, we assert that a particular pathological anatomical structure (1) develops from an instance of a previously existing anatomical structure, and (2) has part an instance of a disorder. These two assertions define both necessary and sufficient conditions for an entity to be a pathological anatomical structure.
Lastly, with the above modifications in place, we define the following terms to necessary for an ontology of malignant neoplasms: develops from an instance of a primary neoplasm without having to commit the primary neoplasm’s ceasing to exist.
Sixth, given the importance of representing the
anatomical structures in which malignant neoplasm from, we
incorporate the Uberon’s anatomical structure and OGMS’
pathological anatomical structure terms, and define them as
follows
A summary of proposed ontology of malignant neoplasms is depicted in Figure 1. 5
We began our work in order to build an application ontology to assist us in analyzing data in an ovarian cancer patient registry (work in progress). Because of our commitment to OBO Foundry principles and ontological realism, we began our ontology development by considering existing ontologies, including OGMS and DO, and related ontologies such as Uberon and CL. Our aim has been to reuse ontology classes where possible and create new classes and hierarchies where existing ontologies either are missing classes or providing faulty modeling of the domain.
We have found the NCIT to be a very useful source of information about cancer related entities, their definitions, and their relationships to each other. Although the NCIT is very large and has been developed over many years, it really remains a terminology rather than an ontology. For example, the NCIT includes the term Disease or Disorder defined as:
Any abnormal condition of the body or mind that causes discomfort, dysfunction, or distress to the person affected or those in contact with the person. The term is often used broadly to include injuries, disabilities, syndromes, symptoms, deviant behaviors, and atypical variations of structure and function.
This definition does not adequately distinguish between the processes and material entities that result in abnormal conditions. This distinction is important for precisely representing the nature of a malady. If a cancer patient has difficulty breathing due to metastatic tumors spreading throughout the lungs, both the difficulty in breathing and the tumors are abnormal conditions, and hence, are would be classified using the term Disease or Disorder (C2991). But, in reality, the process of breathing is a distinct kind of entity than a tumor, which is a material entity. There are past and current efforts to redevelop NCIT or at least sections of it into a proper ontology. Our hope is these efforts will make the NCIT more aligned with OBO Foundry principles. One important result of our work was the addition the abnormal cell, neoplastic cell, and malignant cell types to CL. These CL classes parallel the naming and relationships of the NCIT concepts, but as discussed above, we chose to write new definitions that better define these cell types and do not limit their applicability unnecessarily.
In considering the Disease Ontology, we found it to be a useful catalog of cancer types, but as discussed above, we find that there is confusion as to whether neoplasms are dispositions or disorders. Because of our need to represent pathological findings, we need to reflect that these findings are about disorders (which are material entities) that are observed by pathologists, and not about dispositions, which are not directly observable.
An important finding of our work is that we found that OBO Foundry ontologies have difficulty representing abnormal or pathological entities. Two prominent examples are pathological anatomical structures and pathological processes. Intuitively, a pathological anatomical structure is a kind of anatomical structure. For instance, an ovary containing a carcinoma is still an instance of an ovary. However, the standard definition (with some variations) for anatomical structure found in Uberon, the Common Anatomy Reference Ontology, the Foundational Model of Anatomy, and the Anatomical Entity Ontology does not allow for this:12 Material anatomical entity that has inherent 3D shape and is generated by coordinated expression of the organism's own genome.
This issue is that disorders (such as neoplasms and
fractures) that arise with anatomical structures are not
necessarily generated by the organism’s genome. Thus, the definition
is too strong. Smith et al. are aware of this propose an
anatomical hierarchy consisting of top-level anatomical
structure term with subtypes of canonical anatomical structure,
variant anatomical structure, and pathological anatomical
structure
Anatomical structure o Canonical anatomical structure o Variant anatomical structure o Pathological anatomical structure
While we think this a reasonable proposal, the lack of a definition for anatomical structure makes is unclear as to what canonical, variant, and pathological structures have in common.
A similar problem exists for abnormal processes (such as dysregulation of cell proliferation). The OGMS, for its part, does provide the term pathological bodily process. But, this term is orphaned from other biological processes found in other OBO ontologies. For example, the Gene Ontology (GO) includes the term biological process:13
Any process specifically pertinent to the functioning of integrated living units: cells, tissues, organs, and organisms. A process is a collection of molecular events with a defined beginning and end.
Again, intuitively it makes senses that a pathological bodily process should be a subtype of biological process. However, the definition of biological process does not permit this.
Although we do not have any concrete solutions, at this point, for how to align pathological structures and processes 12 Definitions retrieved from www.ontobee.org, accessed 2017-07-06. 13 Definition retrieved from www.ontobee.org, accessed 2017-07-06. with these other OBO terms, we look forward to collaborating with the OBO Foundry community on creating a coherent structure for these upper level classes that is shared among all OBO Foundry ontologies. Thus, we simply leave pathological anatomical structure as a subtype of material anatomical entity, and pathological bodily process in its current OGMS hierarchy.
Our goal is to contribute to the oncology domain by creating a strong and consistent ontological foundation for providing metadata and data analysis of patient cancer data for both research and clinical applications including clinical decision support. The ontological framework described herein attempts to solve some continuing issues in the representation of cancer as a disease and the disorders (neoplasms) in which it presents. Our framework is intended to be useful for the description and classification of data used in cancer diagnosis and treatment. In future work, we will be adding classes to represent additional entities associated with cancer such as laboratory methods and results, treatments, and outcomes. We hope our ontology will support other oncology researchers in exploiting the full potential of patient data registries and other cancer-related datasets.
We gratefully acknowledge support as follows. William Duncan and Carmelo Gaudioso received support from the Clinical Data Network, a Roswell Park Cancer Institute Cancer Center Support Grant shared resource funded by NCI P30CA16056. Alexander Diehl received support from NCATS 5UL1TR001412. All three authors received support from NCI P50CA159981.