Conceptual models of drug-drug interactions: A summary of recent efforts

Highlights

• We present a review of drug-drug interactions knowledge representations.

• We have identified 29 relevant documents describing 15 models or resources.

• Most of the models were created for natural language processing or DDI inference.

• To compare the models, we represent them in a common framework using UML diagrams.

Abstract

Conceptual modeling elicits and describes general knowledge in a particular domain and is a fundamental step in the development of knowledge-based systems. However, different conceptual models (CMs) could represent the same domain because they result from human intellectual activity with different objectives. Analyzing previous related efforts is crucial when conceptualizing a domain to avoid duplication, increase interoperability and ensure scientific conformity. Our domain of interest is drug-drug interactions (DDIs), and here we review 15 studies that have attempted total or partial representation of the DDI domain. Direct comparison of these different conceptualizations is complex because CMs are usually not provided, differ considerably from each other or are described with diverse formalisms at different abstraction levels. Therefore, to compare these CMs, we represent all of them in a common representation framework. Here, we compare the scope, content, final implementation and applications of CMs of the DDI domain. We aim to identify which aspects of DDIs have been conceptualized, characterize how this information has been modeled by different research groups, describe how each CM has been translated and illustrate the applications generated from the final models.

Introduction

Knowledge representation is an essential activity in knowledge engineering. Particular knowledge about a domain (e.g., the patient presents a sudden rise in temperature (39 °C) and neck stiffness and is a suspected case of meningitis) requires prior general knowledge of how concrete objects are related in the world (e.g., A disease presents signs and symptoms. The identification of signs and symptoms is used to diagnose the disease. A suspected case of meningococcal meningitis is defined as any person with sudden onset of fever ( > 38.5 °C) and at least one of the following signs: neck stiffness, altered consciousness or other meningeal signs). Conceptual modeling elicits and describes the general knowledge of a particular domain. The sets of objects and facts in a particular domain constitute its conceptualization, and its formal description, which sometimes includes a graphical notation, is the conceptual model (CM) [1]. Usually, the design of a CM relies on the perspectives of experts in that specific domain. However, different CMs can represent the same domain because they result from human intellectual activity with different objectives. These CMs are abstract models that can be translated into different description languages and interpretable schemata such as ontologies, relational databases or XML schemata.

Because of the growing success of the Semantic Web, ontologies have become one of the most popular formalisms for knowledge representation. Indeed, the most comprehensive repository of biomedical ontologies, BioPortal,¹ doubled the number of collected ontologies from ∼200 to more than 400 in the last six years [2]. The enormous complexity of the biological, medical and pharmaceutical domains compels authors to create individual ontologies with more exhaustive descriptions of specific areas within a broader domain. Further, various applications such as the coding and indexing of medical records [3], semantic annotation of biomedical documents [4], data integration from the Semantic Web and Linked Data [5] or data analysis and discovery applications [6] may require different conceptualizations of the same domain.

Thus, some research groups initially develop their own independent conceptualizations de novo, which can lead to multiple isolated CMs that represent different or even overlapping aspects of the same domain. To avoid such duplication, the OBO Foundry,² a collaborative effort to develop and maintain biomedical ontologies, recommends collaboration to 1) avoid duplication of work, 2) increase interoperability and 3) ensure that ontology content is both scientifically sound and meets community needs [7].

The medical and pharmacological domains are active areas of knowledge-based systems research [8]. Representation of drug-drug interactions (DDIs), a serious type of adverse drug reaction (ADR) that occurs when one drug affects the levels or effects of another drug [9], is an important effort in these domains. DDIs pose serious risks to patients’ safety and increase healthcare costs [10], [11], so their early apprehension is vital in clinical settings [12]. Various research groups have proposed diverse computational approaches that rely on CMs or other formal representations of the domain to improve prediction or management of DDIs. Here, we review the aspects of DDIs that have been conceptualized, characterize how this information has been modeled by different research groups, describe how the different CMs have been finally implemented and illustrate the applications generated from the final CMs.