OBO Foundry food ontology interconnectivity

1.1. The FoodOn food ontology

FoodOn is described in detail in its inaugural paper [3] and in the https://foodon.org website, so here we provide a brief introduction, followed by overviews of the other food related ontologies. Generally, we seek to show how the collective community of older and newer OBO ontology curators is supporting the harmonization effort required for cross-domain food systems analysis. FoodOn’s mandate is to describe and provide precomposed terms for generic (non-branded) food products that a food producer, food manufacturer / processor, or consumer can find in the supply chain, ranging from wild or farmed food, to processed, wholesale, retail, prepared, restaurant or home-cooked food. FoodOn now offers over 24,000 terms in its extensive food product hierarchy and description facets, such as applied cooking treatments, preservation methods, and packaging, along with another roughly 10,000 OBO ontology terms including chemistry, anatomy and food source organism (plant, animal, fungi or algae) taxonomy domains.

FoodOn entered the OBO Foundry in 2016, transforming and evolving the basic structure of LanguaL [5], a popular food composition database (FCD) vocabulary created by the U.S. Food and Drug Administration (FDA) Center for Food Safety and Applied Nutrition (CFSAN) in 1975. During FoodOn’s inception, it adopted hundreds of Environment Ontology (ENVO) [6] food terms, and it integrated with OBO’s Chemical Entities of Biological Interest (ChEBI) [7] ontology which organizes molecular entities – mainly ‘small’ chemical compounds (natural or synthetic) – pertinent to the processes of living organisms. FoodOn imports anatomical terms from the Plant Ontology (PO) [8] and UBERON multi-species anatomy ontology [9], taxonomy from NCBITaxon [10], and imports dietary and nutritional terms from the Compositional Dietary Nutrition Ontology (CDNO) [11] and Ontology for Nutritional Studies (ONS) [12]. Finally, FoodOn organises food related data items and processes under the Information Artifact Ontology (IAO) [13] and Ontology for Biomedical Investigation (OBI) [14] classes respectively, and draws mainly on existing Relation Ontology (RO) relations [15], with just a few others introduced, like “has ingredient” and “has defining ingredient”.

1.2. Food related ontology use cases

Use case examples of how FoodOn and other OBO ontologies are being used can be found in a number of standards and database projects expressed in knowledge graph or tabular formats.

Food sample composition, nutrition and bioactive analytics:

Foodborne disease investigation:

OPEN IN VIEWER

Fig. 1.

Key food-related ontologies within the OBO sphere, and some leading organizations and consortiums that curate them. The direction of ontology term import and export is indicated by the arrows of dotted lines; solid lines convey curation participation. Each ontology has a count of terms it introduces in proportion to all its terms including ones reused from other ontologies. Data provided by https://ontobee.org/ontostat as of October 25, 2022.

Figure 1 illustrates the emergence or expansion of food-related OBO ontologies to meet the needs of projects like these, describing food composition, agricultural production, food processing and animal and human health factors. The One Health [25] paradigm of interrelated human, animal, and environmental health that has emerged in the last decade also motivates interdisciplinary vocabulary, such as the Health Surveillance Ontology [26].

3.1. AgrO, CO, TO, and PECO: Agriculture related ontologies

The Agronomy Ontology (AgrO) [16], launched in 2017 in the context of the CGIAR Platform for Big Data in Agriculture, describes agronomic management practices, implements, and variables used during crop plot experiments. It includes over 200 plant morphology traits from the Plant Trait Ontology (TO) [46], parameters identified by agronomists and crop modelers from the pre-existing Crop Ontology (CO) [46,47], and a treatment menu of 15 terms from the Plant Experimental Conditions Ontology (PECO) [48]. AgrO contains over 600 geography and geology related terms from ENVO, as well as 550 ChEBI terms and 155 Phenotype And Trait Ontology (PATO) [49] terms. Along with harvested food product references, AgrO “crop residue” terms are imported from FoodOn to support research on reuse of such materials. Continuing the selenium example, AgrO references various forms of selenium for sampling measurements, TO provides a “selenium content” biochemical plant trait; and PECO provides a “selenium nutrient exposure” treatment.

AgrO is used in the CGIAR Agronomy Field Information Management System, a software tool that enables the digital collection of agronomic trial data [50]. AgrO, ENVO and other ontologies can be used within the MIxS package standard [51] for soil biosample metadata collection [52]. In the near future, to support the Periodic Table of Food Initiative, AgrO will import FoodOn food and nutrition terms which are important to measure at the post-harvest stage. As well, AgrO term language will be adjusted to cover commercial production crop treatments. There is a future opportunity to further align TO biochemical traits and CDNO nutrient chemical concentrations.

3.2. CDNO: Compositional dietary nutrition ontology

CDNO [53], launched in 2020, provides a hierarchy of over 1,000 terms for nutritional attributes – vitamins, carbohydrates, lipids, minerals, proteins, etc. – and matching bioactive chemical concentrations from crops, livestock, and fisheries that contribute to human diet and which are referenced in precision food commodity laboratory analytics. Rather than directly mirror ChEBI’s hierarchy of molecular entities and their roles, CDNO provides a more diet-and-nutrition community friendly arrangement (Fig. 2) by essentially flattening ChEBI hierarchies like “mineral nutrient” and “protein” to selected elements, and adding CDNO groupings like “plant secondary metabolite”. Initially, CDNO, FoodOn and Plant Ontology (PO) curators worked together in a collaborative effort to establish a set of over 58 food raw materials entities. They additionally defined sources of plant samples and crop production processes or stages from which the food raw material was taken.

OPEN IN VIEWER

Fig. 2.

The top level CDNO ‘dietary chemical component” class and its subclasses, most of which are imported from ChEBI.

OPEN IN VIEWER

Fig. 3.

Visualization of how CDNO and FOODON classes can be brought together in a graph database statement about selenium concentration in a whole wheat kernel sample.

CDNO’s “nutritional component concentration” classes have a design pattern that uses PATO’s “concentration of” term that references a particular analyte in proportion to the BFO “material entity” specimen that is being measured in. A specimen can be a plant or animal species’ anatomical part, yielding a comprehensive way to measure nutritional attributes of biosamples (like seeds) throughout the food production lifecycle. Following the selenium example, Fig. 3 shows the “concentration of dietary selenium in material entity” class defines a concentration characteristic of an analyte as part of some material entity biosample. The term’s OWL logical “Equivalent To” definition is shown, as well as CDNO’s dietary selenium and its ChEBI subclasses which may be measurement targets. To express selenium concentration in a whole wheat kernel, we need to reference “whole wheat kernel” a FoodOn reference that helpfully indicates (axiomatizes) the kind of plant anatomy part it is, and its taxonomic origin. These classes can combine in a graph database data structure (aka A-Box or instance data) as shown in the “Graph DB expression” box of Fig. 3, where each dashed box is an instance of a class, with its own identifier (not shown are data properties for specifying the concentration measurement itself).

“Concentration of dietary selenium in material entity” and “whole wheat kernel” are each an example of a “precomposed term” – a term that has an equivalency axiom describing a set of relations to other things that together are the criteria for recognizing something of its kind. In databases one may also encounter a “post-composed” term, one which doesn’t have a preassigned ontology id, but which has the same searchable sentence structure as a precomposed term, to match more detailed needs of researchers, for example comparing concentration of dietary selenium in the endosperm versus germ part of a Titicum kernel. Note that for projects such as CDNO’s CropStoreDB (https://www.cropstoredb.org/) that use tabular datasets rather than graph representations, a biosample table can simply have columns for biosample identifiers, dietary component measure, FoodOn product term or PO anatomical term and taxonomy, and any other agricultural context metadata. With the addition of some column semantics, tabular views can be converted back into appropriate graph representations.

Upcoming CDNO work is focused on a hierarchy of classes for physical and functional attributes and dietary functional roles, as well as nomenclature required to describe variation in analytic measures. This will allow harmonization of the nutrient measures used for international standards such as the FAO sponsored International Network of Food Data Systems (INFOODS) [54] system used in many food composition databases, as well as the commonly used USDA nutrient codes in FoodData Central [55] (where for example selenium is listed in the “Supporting data for CSV Downloads” archive nutrient.csv table with id 317 and a microgram unit).

ECTO [56], launched in 2016, provides language to describe experimental and environmental factors for a wide range of public health and environmental monitoring objectives, including modeling components of human disease and nutritional health, environmental toxin exposure, and alteration of biological function. It focuses on documenting precomposed experimental treatments, non-experimental exposures, and environmental conditions that may impact organisms. For modelling experimental designs, ECTO contains hundreds of “exposure to chemical” terms such as “exposure to chlorpyrifos” (a moderately toxic pesticide that has been banned since 2020 in the European Union, and since 2022 in the USA) which can be used to model human and other organism diseases and phenotypes that arise from exposure, and the underlying metabolic activity at work, as illustrated in Fig. 4. The RO relations “has exposure medium”, “has exposure stimulus”, “has exposure route” and “realized in response to” all connect processes which summarize a chain of exposure events, for example, exposure to an instance of an apple which itself has been exposed to chlorpyrifos.

OPEN IN VIEWER

Fig. 4.

Instance-level ECTO data representation of exposure to chlorpyrifos by way of food-ingestion, including the disease and phenotype associated with it.

ECTO currently has over 160 food ingestion terms such as “exposure to skim milk via ingestion”, and 17 vitamin and mineral ingestion terms including an “exposure to selenium via ingestion” process. An upcoming Dead Simple OWL Design Pattern (DOSDP) [57] will integrate exposures to specific diets that refer to terms found within the Ontology for Nutritional Studies.

PestOn [58], launched in 2022, aims to provide an ontological framework for organizing branded pesticide product regulatory information. Regulation of pesticides (mostly chemical agents used to protect crops from disease, pests, and weeds, and improve plant growth) and their reporting requirements and other publicly-mandated information is uneven among nations and geopolitical regions, leaving uncertainty about their responsible use, and challenges in evaluating health and environmental impacts. PestOn was developed by ontologizing The Italian Ministry of Health’s pesticide register dataset, comprising over 16,000 pesticide products used over several decades. Pesticide brand names, their active ingredients (including chemical name and concentration), production company information, pesticide by chemical role (e.g. fungicide, insecticide, herbicide, plant growth regulator, adjuvant, herbicide, etc. product classes linked to ChEBI chemical role where possible), type of associated hazards (e.g. dangerous for the environment, flammable, toxic, etc.) and regulatory events (issue date, expiry, revocation) and derived authorization status are all part of the PestOn model.

More than 99% of pesticide active ingredients were successfully matched against CHeBI ontology, with the remainder either chemicals that the authors then suggested to ChEBI curators, or were organisms like fungi and bacteria which PESTON does not currently model. At launch, PestOn involves 795 active ingredients contained in 16,458 pesticide products and their 532 production companies. PestOn is divided into a core ontology that can be used for describing pesticides in general, but it also comes with import files which supply an RDF graph of all the Italian instance data for the pesticide products, companies and regulatory events. The SPARQL querying language [59] can be used for example to retrieve the current available products in each pesticide role category, as well as a summary of products containing a given active ingredient. Taking the chlorpyrifos example in the ECTO overview, PestOn has it as an active ingredient in over 190 brand name pesticides offered in the Italian market, along with issue and revocation dates, enabling one to determine compliance with the European chlorpyrifos ban.

PestOn’s initial corpus of Italian regulatory data should be expandable to other jurisdictions although different regulatory events (besides application / revocation, etc.) may need to be added. PestOn is not yet a member of OBO, and does not reference foods or crops that are applicable to pesticide application, but it could be useful in research linking branded pesticide products to exposure effects as described in ECTO.

3.5. ONS: Ontology for nutritional studies

ONS [12], first published in 2018, focuses on vocabulary for modeling nutritional studies, including diet and dietary pattern variations. A central concept in ONS is “diet” (ONS:1000001, having 58 subclasses), a datum defined as “the sum of food consumed by a person or other organism”. Diet is currently distinguished from “dietary pattern” (ONS:0000094, having 34 subclasses), defined as “the quantity, proportion, variety and combination of different foods and drinks consumed in meals, and the frequency with which they are habitually consumed”. Dietary pattern was intended to represent a dataset typically resulting from assays in the context of nutritional epidemiology (for example, an ONS “Food Frequency Questionnaire”), containing a specification of foods consumed and, as a result, denoting the type of diet to which a subject has adhered. Curation and development revolving around the initial diet concept in ONS has greatly benefited from the collaboration with the Joint Food Ontology Workgroup. Thanks to this interaction, multiple different subclasses of the diet (and related dietary pattern) were defined.

Various flavours of diet and related dietary patterns currently have axioms that include or exclude FoodOn products by way of an RO “eats” relation. As an example, the “vegan diet” (ONS:1000021) and the “vegan dietary pattern” (ONS:2000021) are axiomatized as eating vegetables and excluding the consumption of animal products; however, these axioms need to be revised as “eats” can only connect from an organism to a relevant food, rather than from diet, which is currently a kind of IAO Information Content Entity (which is logically disjoint from material entities like food). Additionally, the distinction between diet and dietary pattern needs to be re-examined. While diet as a descriptive term (the ONS “descriptive diet” class) details what an organism has actually eaten, it appears that the remaining diet terms could be considered eating patterns, and thus be consolidated under dietary pattern.

3.6. PO²: Process and observation ontology

OPEN IN VIEWER

Fig. 5.

PO² uses similar classes for both production line actuation and observation.

PO² [60] details processing steps, characteristics and sampling regimes of foods during their industrial manufacturing process, as well as their food formulation (Fig. 5). The ontology contains a small core layer of 68 classes dedicated to the generic modeling of both transformation and characterization processes, while domain specific sub-ontologies specialize the PO² core model for different projects. PO² can represent a food transformation process described by a set of experimental observations available at different scales and evolving in time as a production process composed of unit operations. PO² is based partly on the Sensor, Observation, Sample, and Actuator (SOSA) ontology [61], but also has BFO, OWL-Time and IAO classes. It has been implemented in database-driven software that covers dairy, meat, and biorefinery production, to represent the unique characteristics of foods during their manufacturing process. The PO² team participates in the Food Process Ontology Workgroup which is helping to build FoodOn’s food transformation process branch.

Future directions for PO2 include the development of a PO2 specialization called TransformON [62] which will cover various bioresource transformation processes that produce food and biobased products. It includes three component subclasses, two for food and feed taken directly from FoodEx 2, as well as a branch of non-food substances, including organic wastes or residues (e.g. wastewater) and bio-based products. A future objective is to align the PO2 component branch with FoodOn, and in this way map between FoodEx 2 and FoodOn. Processing equipment (actuators) and measuring instruments (sensors), as well as branches for analytical procedures and measured properties will be mapped to or sourced from OBO ontologies. PO2, available at https://doi.org/10.15454/XSVVBW, is not registered with OBO as it is tightly integrated with several other non-OBO ontologies, but it has collaborated on a comparison of it and the OBO process model [41].

FOBI [63], launched in 2020, is aimed at describing the relationships between a food and its metabolome, the collection of all metabolites in the body directly derived from the digestion and biotransformation of foods and their constituents. These biomarkers are often measured in stool or urine samples. FOBI, currently with a total of 1197 terms, is composed of two interconnected branches: a food branch consisting of over 300 FoodOn raw single-ingredient and multi-component foods; and a food intake “Biomarkers” branch of about 968 chemicals (Figure 6), many from ChEBI but with some FOBI specific terms. Connections between the two are made with FOBI’s “BiomarkerOf” object property, with 590 connections made. Figure 6 illustrates two biomarkers that can be left in humans as a result of consuming apple directly, or in some processed form such as a baked apple pie. (Note that the FOBI paper [63] refers to its own “Food Ontology [ … ] created on the basis of dietary data obtained from self-reported surveys for dietary assessment [ … ]”, but this references FOBI’s own branch of food items that was later mainly converted to FoodOn terms).

OPEN IN VIEWER

Fig. 6.

Example class-level FOBI linkage between FoodOn “apple (whole)” and FOBI BiomarkerOf object properties.

Various tasks could advance FOBI from an inaugural ontology. There are some lingering inconsistent PURLs (e/g. for Benzodioxoles, http://purl.obolibrary.org/obo/fobi.owl#FOBI:050222 should be standardized to http://purl.obolibrary.org/obo/FOBI_050222). FoodOn and ChEBI could be reused more: FOBI has apple juice as a subclass of “apple (whole)” rather than more correctly as a sibling product; ChEBI has a “catechol” class that could be reused. Remaining FOBI chemical classes should migrate into ChEBI if they are missing from that resource. FOBI class names are often pluralized, which OBO usually discourages as it conveys an instance of a class could point to a plurality of a thing rather than a single countable thing. FOBI has a number of classes such as “Phenylpropanoids and polyketides” which have disjunct subclasses, and which may be desired in a user interface menu but the question arises whether such umbrella classes are needed within an ontology itself, or would a hierarchy containing separate (in this case ChEBI) phenylpropanoid and polyketide classes and their branches suffice?

FOBI contains some example multi-component foods, like apple pie, which not directly connected to biomarkers, but which, by itemizing ingredients with a “Contains” (and inverse “IsIngredientOf”) object property, convey the possibility of recipe-ingredient-biomarker or biomarker-ingredient-recipe deduction, and related nutritional implications. However, it is rare that a recipe can be defined as a class with an exhaustive list of ingredients, since non-essential ingredient substitution is usually allowed (for example, shortening for butter in apple pie crust). FOBI could convert its multi-component food examples to test-sample instance-data (a particular apple pie’s constituents) and employ FoodOn’s “has ingredient” and “has defining ingredient” object properties. Alternately FOBI could consider leaving multi-component food definitions to the domain of a recipe database. There is also the fidelity question of which part of a plant or animal is precisely responsible for a biomarker – perhaps it should be axiomatized that only an apple hypanthium is the source of a biomarker residue, then we leave it to a reasoner to infer that ingesting a peeled, cored apple could lead to the biomarker being present. Also, “BiomarkerOf” is a shortcut relation representing a more complex digestion model which would be helpful to elaborate on, so that models could be offered for describing how biomarkers’ concentrations trail off time-wise after digestion, thus enabling prediction of how much related food may have been consumed.

The Human Disease Ontology (DO, with term prefix DOID) [64] has 28 “nutrition disease” (DOID:374) categories devoted to nutritional deficiency, for example “Keshan disease” arising from selenium deficiency, as well as overnutrition-triggered diseases. DO also includes allergies arising from food product consumption. The RO object property “has allergic trigger” (RO:0001022) has been used to attach an allergic disease to the class of food that triggers it. Food allergy (positioned under “gastrointestinal allergy”, an “An allergic disease that is located in the gastrointestinal tract”) currently has 33 categories including fish and shellfish, fruit, milk, wheat and vegetable subclasses, many with corresponding relations to FoodOn food products that cause them. DO also supports the Immune Epitope Database (IEDB) [65] database by enabling a selection of DO terms – ones with an “oboInOwl#inSubset DO_IEDB_slim” annotation – to be filtered into a separate file for export, and this includes the entire food allergy branch.

The Mondo Disease Ontology (MONDO), which seeks to cover and standardize disease vocabulary found in a number of large disease thesauri) has a “nutritional disorder” (MONDO:0005137) term branch containing cross references to DO and other ontologies and thesauri, and under which 5 “food-related allergic diseases”, 68 categories of “overnutrition” diseases (mainly obesity variants), and 67 categories of “nutritional deficiency disease” (including Keshan disease) are listed.

For food-related diseases, symptoms such as obesity and allergic inflammation can be found in the Human Phenotype Ontology (HP). Although it is problematic in a disease ontology to directly link diseases to symptoms with object properties (since OWL logic can only express certainty rather than probability that a symptom might be associated to a disease) this could be accomplished in a more context-sensitive database or knowledge graph of fuzzy logic or probabilistic connections between the two.

3.9. MAxO: The medical action ontology

MAxO [66], launched in 2020, is a broad ontology that provides a structured vocabulary for medical procedures, interventions, therapies, treatments, or clinical recommendations, including nutritional recommendations. MAxO was designed to annotate diseases, particularly rare diseases where management of nutrients is often critical. MAxO provides a lexicon of 77 dietary intake avoidance behaviour terms (Fig. 7) such as “selenium intake avoidance”, and an “avoided food” object property in the case where FoodOn products are being avoided. For example if a person has a strong fish odor, and Trimethylaminuria (the increased presence of trimethylamine in urine) is detected, then a MAxO “broccoli intake avoidance” treatment may be recommended to reduce trimethylamine abundance [67]. Additionally, over 70 nutritional supplementation process terms are included, including “selenium supplementation”, with linkages to directly to ChEBI chemicals as inputs (and in one case, “multivitamin supplementation” which uses the FoodOn “vitamin supplement”).

OPEN IN VIEWER

Fig. 7.

A small portion of the nutrition intervention hierarchy within MAxO.

In order to capture the relationship between treatments and diseases, the Phenotypic Observation Explication Tool (POET, https://poet.jax.org/) was developed to establish a relationship between MAxO, HP, and MONDO terms. It allows researchers to actively participate in annotating diseases in their area of expertise.

FIDEO [68], launched in 2020, and providing over 170 FoodOn food categories, and over 260 ChEBI chemicals, is used in a database containing annotation and retrieval of scientific articles about food-drug interactions. Other domain-specific food categories such as “vitamin k-rich food product” are locally defined. While initial efforts were focused on the design of the ontology based on the Basic Formal Ontology (BFO) and the OBO Foundry principles, more recent efforts have resulted in a user-friendly visual interface that allows search and exploration of interactions [69] and currently focus on the automatic integration of food-drug interactions from DrugBank (https://go.drugbank.com) and Hedrine (https://hedrine.univ-grenoble-alpes.fr) to FIDEO using ROBOT’s [35] template command.

3.11. ONE: Ontology for nutritional epidemiology

ONE [70] details nutritional epidemiology manuscript and dataset characteristics, including the document structure of research manuscripts, dietary surveys, and food-based dietary guidelines (FBDGs). FBDGs represent a wealth of accumulated diet knowledge summarized from nutrition studies [71,72], and are important documents for policy makers, healthcare workers, educators and ultimately the general public to glean dietary policy and advice from. ONE brings into ontology form an established set of “Strengthening the Reporting of Observational Studies in Epidemiology” (STROBE) minimal requirements for the reporting of nutritional epidemiology research [73]. The first version of ONE extends IAO document parts to cover research paper structure, and as well description of food surveys which form the underlying datasets for many studies. Abstract, discussion, ethics, methods, results and supplementary methods specific to dietary studies are defined, as well as quality assessment classes of study designs and dietary recall methods (Fig. 8). FBDG terms were added into ONE in 2021 to support potential applications such as automated analysis of dietary trends, along with a graph database application for holding semi-automated assessments of nutritional epidemiology manuscripts [74].

OPEN IN VIEWER

Fig. 8.

ONE’s hierarchy of dietary survey quality indicators.

ONE curators began development of a Natural Language Processing (NLP) query language interface to enable plain language retrieval of relevant nutritional studies as encoded in a knowledge graph, as well as dashboard development to visualize nutritional knowledge contained in research manuscripts and population based recommendations [75].

HSO [76] is a knowledge model for data collection, collation and reporting of animal health, public health and food safety surveillance systems, including proactive surveillance and reactive investigation methods and objectives. While detailed surveillance data may not be easily harmonizable and/or shareable due to specific national agency reporting requirements, a top-down semantic harmonization of reporting elements is still required to pool summary data for international / multi-agency contexts. After engaging multilaterally about what the essential components of One Health surveillance data are, the HSO curation team developed vocabulary and a model centered around a planned process “surveillance activity” and related protocols, along with new vocabulary of over 200 classes, 40 object properties, and 20 data properties, for sampling strategies, host and pathogen taxonomy, specimen device type, and anatomical origin or substance. HSO is an outcome of the One Health suRveillance Initiative on harmOnization of data collection and interpretatioN (ORION) project and is currently interoperable with various catalogues in the European Food and Safety Administration’s (EFSA) Standard Sample Description. HSO can be naturally paired with the recently OBO-launched Vertebrate Breed Ontology (VBO) [77] of over 16,000 region-specific domestic animal breeds and wild lineages to pinpoint taxa beyond what NCBITaxon can offer.

Future HSO integration with OBI sample types would be beneficial. There may be opportunity to reduce the number of object properties, for example the HSO “Surveillance Activity” has “has surveillance actions” object properties which allow a “surveillance methods specification” – a diagnostic/reporting/sampling design/specimen collection protocol – in their range. Instead, in a Surveillance Activity’s plan specification could directly link to these protocols via “has part” without losing any reasoning or query power.