This section presents: (i) the modifications that have been made to the model of the first version of FIDEO, (ii) the enrichment process of FIDEO to include food-drug interactions from DrugBank and Hedrine, and (iii) the evaluation framework with regards to competency questions.

Evolution of the FIDEO model

The first version of FIDEO focused mainly on how to organize knowledge related to food-drug interactions, and in particular how to link together existing concepts (from external ontologies) and new ones (defined in FIDEO), but not on the actual interactions described in the literature between food and drugs. In this new version, we represent food-drug interactions listed in online resources but currently described in a non-formalized way.

Compared to the model proposed in Fig. 1 of our previous paper [6], we have greatly simplified the Data part, corresponding to information derived from scientific knowledge. We use only the newly created concept “information source for interaction” (child concept of “information content entity” (IAO:0000030)), which is the parent concept of all evidence describing a given food-drug interaction and is linked to the process of that interaction by the relationship is_about (Fig. 1a). Food-drug interactions have been represented in FIDEO in the form of precompiled concepts, each of which specifies both the food and the drug involved. More specifically, the process of the interaction is linked to the food concerned (material entity) by the relationship has_participant and to the drug affected (material entity) by the relationship has_input. The definition of these relations are as follows:

has_participant: “a relation between a process and a continuant, in which the continuant is somehow involved in the process” (source: http://purl.obolibrary.org/obo/RO_0000057),

has_input: “p has input c if and only if p is a process, c is a material entity, c is a participant in p, c is present at the start of p, and the state of c is modified during p” (source: http://purl.obolibrary.org/obo/RO_0002233).

In the food-drug interactions described in the Hedrine and DrugBank resources, the information we have (and therefore represent) concerns the change that occurs in the effect of a given drug in the presence of a given food. Thus, both the food and the drug are involved in the interaction, but we consider that only the state of the drug is altered during this interaction (implying the use of the has_input relationship, being more specific than has_participant). Finally, the Real world part, which describes information about biomedical entities, has also been simplified and the “chemical substance” (ChEBI:59999) concept has been replaced by its parent concept “chemical entity” (ChEBI:24431) because many drugs found in DrugBank are classified as “chemical entity” in ChEBI, not as “chemical substance”.

Fig. 1

a Organization of high-level concepts related to the food-drug interaction process in FIDEO, and b) Illustrative example of this framework with a food-drug interaction between garlic and warfarin described in DrugBank at the following URL: https://go.drugbank.com/drugs/DB00682. The ontologies from which the corresponding concepts and the relations existing between them originate, as well as the hierarchy to which these concepts belong are specified in the legend on the right of the figure

Figure 1 illustrates the main concepts involved in describing a food-drug interaction in the second version of the model of FIDEO (panel a), and how an interaction between garlic and warfarin listed in DrugBank is represented in the ontology (panel b).

FIDEO enrichment steps and design aspects

In this part, we present the general architecture of the approach implemented to enrich FIDEO, which consists of the following four tasks shown in Fig. 2: 1) term annotation of food and herb terms within the DrugBank food-drug interaction corpus, 2) ontology term reuse (XOD 1) of concept names from existing ontologies to be mapped to annotation terms, 3) ontology semantic alignment (XOD 2) of selected concepts within FIDEO, and 4) ontology design pattern usage for the generation of logical definitions for concepts (XOD 3) to be integrated in FIDEO.

Fig. 2

Complete FIDEO enrichment process. Knowledge was extracted from DrugBank and Hedrine, OLS was used to identify drugs, foods and herbs in existing ontologies and ROBOT to define patterns for creating logical definitions for concepts to be included in FIDEO

Task 1: term annotation

Textual descriptions provided in DrugBank were manually annotated to structure food-drug interactions. From the drugs with relevant information, we first annotated a subset of interactions involving 100 randomly selected drugs. This first corpus was annotated by three annotators (GB, RA, and FM) to check inter-annotator agreement for categories including Food, Food Component, Herb, Enzyme, Interaction Mechanism, Physiological Effect, and Meal Time. A total of 473 sentences was annotated in this way. In general, agreement was in the almost perfect range for all categories, with a Fleiss Kappa of 0.87 for the Food category and higher for the other categories. To ensure high coverage, food interactions with the remaining drugs were exhaustively annotated by two annotators (GB and RA). Each annotated half of these drugs and the third annotator (FM) finally reviewed all drugs to homogenize the annotation. Overall, a total of 1213 sentences have been annotated following this procedure. An illustration of the annotation of the five food interactions concerning “warfarin” is presented in Table 1.

Table 1 Annotations of the field Food interactions described in DrugBank for the drug “warfarin” (DrugBank:DB00682)

As food-drug interactions are fully structured in the Hedrine database, no additional annotation was required. Hedrine curators make use of dedicated fields for foods, plants, drugs, interaction mechanisms and clinical importance that can be directly integrated in FIDEO.

Task 2: ontology term reuse

The principle of this task was to find a correspondence between the food, herb and drug terms used to annotate DrugBank sentences or described in Hedrine and concept names coming from existing ontologies. We automated this task to facilitate further enrichment of the ontology in future versions.

To perform these matches, we used the OLS service (described in Tools for enriching FIDEO section) to search for concepts in reference biomedical ontologies.

Since the terms to be searched in our case were related to the drug, food and herb categories, we performed term matching as follows: (i) drug terms were searched in ChEBI and, if not found, in the Drug Ontology (DrOn); (ii) food terms were searched in FoodOn; (iii) herb terms were searched in FoodOn and, if not found, in DrOn.

For each category, we estimated this automatic term mapping using the ratio of the number of terms mapped to a concept from existing ontologies to the total number of terms. At the end of this step, we found that the matching step was as follows: (1) Drugs: 1026 out of 1177 (87%) and 702 (59%), respectively for ChEBI and DrOn; (2) Foods: 54 out of 121 (45%) for FoodOn; (3) Herbs: 3 (20%) and 6 (40%) out of 15, respectively for FoodOn and DrOn. Because the automatic process was insufficient, we performed term matching manually for unmatched terms.

Task 3: ontology semantic alignment

This task aimed to organize concepts coming from ChEBI, FoodOn, and DrOn in a coherent way within FIDEO.

All drugs were described as sub-concepts of “chemical entity” (ChEBI:24431). To provide a simple and widely used structuration of drugs, we extracted for each drug obtained from DrugBank and Hedrine the ATC classes to which it belongs. We chose to integrate only the drug classes of the first two levels of the ATC (as illustrated in Fig. 3a)). To comply with OBO principles, we have converted plural terms to singular and lower-cased drug class names. In addition, we have slightly modified names that do not correspond to a drug type, and we have included the ATC code via a hasDbXref annotation containing the link to the web page corresponding to this ATC code. For example, the ATC class A-ALIMENTARY TRACT AND METABOLISM has been converted to FIDEO:00000100 with the label “drug target of alimentary tract and metabolism” and https://www.whocc.no/atc_ddd_index/?code=A DbXref. Finally, the drugs without ATC code were placed under an intermediate concept “not elsewhere classified drug” (e.g. “ardeparin” (DrOn:00017208)).

Fig. 3

Integration of drugs and foods in FIDEO: a drugs are organized according to the ATC hierarchy (or as a “not elsewhere classified drug” for drugs not described in ATC), and b) foods are subclasses of the FoodOn concept “food product”

All foods for which a mapping in FoodOn was found ware integrated under the concept “food product” (FoodOn:00001002), which was positioned directly under the “material entity” (ignoring the intermediate concept “food material” (FoodOn:00002403)). Moreover, for the sake of simplification, we decided not to keep the sub-concepts described in the FoodOn hierarchy between “food product” and the foods actually involved in the food-drug interactions.

Herbs were integrated in the same way as foods for those that exist in FoodOn. For the other herbs (being associated with a DrOn concept), they were integrated under “processed material” as defined in DrOn.

Figure 3 illustrates the integration of the drug “warfarin” from ChEBI (panel a) and the food “garlic food product” from FoodOn (panel b).

Task 4: logical definition generation

For this task, we used a strategy based on ontology design patterns to assign logical definitions to concepts in FIDEO. This task complements the previous tasks by providing a specific, feasible and robust mechanism to obtain an easily maintainable ontology. For this purpose, we have implemented an approach based on the ROBOT tool, which allows generating an ontology in the Web Ontology Language (OWL) according to templates that can be used to develop modular ontologies. The advantage of this approach is that each module is associated with a spreadsheet, in which curators can update information about the ontology entities rather than directly editing the OWL ontology file. Then, ROBOT creates logical definitions and class annotations via the “template” command, which transforms the information described in a spreadsheet into OWL axioms.

Thus, we first built the template according to the model defined in Evolution of the FIDEO model section. In this template (Fig. 4), the first two rows represent the ROBOT header for specifying the entities and their properties (i.e. identifier, label, parent concept(s) if no equivalent axiom is provided, type - concept or relation - and equivalent axiom - or logical definition - of concepts). The rest of the file contains the concepts and relations of FIDEO. This approach allows to considerably reduce the time needed to integrate concepts, thus facilitating the ontology update process.

Fig. 4

Extract from the ROBOT file containing as many rows as concepts and relations in FIDEO for which the following characteristics are available in the five columns: 1) their identifier in the ontology from which they are derived, 2) their label, 3) their parent concept(s) or relation(s) in FIDEO, 4) their type (owl:Class or owl:ObjectProperty), and 5) the logical definition(s) of defined concepts. Note that the concept “FDI garlic food product–warfarin” does not have a parent concept specified in column 3 because its logical definition already states that it is a child of the “food drug interaction” concept

Evaluation according to competency questions

In order to assess the coverage of the ontology, the competency questions that were defined for the first version of FIDEO are listed below, focusing on the interaction between garlic and warfarin used as an illustration throughout this article.

CQ1

What foods potentially interact with warfarin?

CQ2

Which drugs potentially interact with garlic?

CQ3

Which antithrombotic agents may interact with garlic?

CQ4

What type of interaction mechanisms underlie the interaction between garlic and warfarin?

CQ5

What type of studies describe the interaction between garlic and warfarin?

CQ6

What is the level of clinical importance of the garlic - warfarin interaction?

CQ7

Which spices or herbs can be safely consumed by patients taking warfarin?

CQ8

What alternative drugs can be taken to avoid the interactions between warfarin and garlic?