Chemicals.

The primary entity type in our study is milk oligosaccharides. Few existing corpora include annotations for chemical compounds. Krallinger et al. provide a survey of the publicly available corpora [21]. The CRAFT (Colorado Richly Annotated Full Text) corpus includes annotations of compounds with ChEBI (Chemical Entities of Biological Interest) concepts and biological taxa with the NCBI (National Center for Biotechnology Information) taxonomy [22]. However, the chemical type is described with a level of precision (e.g., subatomic particles, atoms, molecules) that is not suitable for our purposes, and the CRAFT sources were selected according to the Mouse Genome Informatics (MGI) group, which is not relevant to our milk analysis study. Other corpora are either unavailable or out of scope (e.g., GENIA [23], ChemEVAL [24]).

The ChemDNER corpus [21] was designed to provide extensive coverage of chemical compounds and drugs in the literature. The documents were selected from PubMed using general keywords. Its annotations include over 84,000 manually identified chemical mentions, corresponding to nearly 20,000 unique chemical name strings normalized by IUPAC (International Union of Pure and Applied Chemistry) nomenclature and InChI (International Chemical Identifier) for formulas. However, despite this broad coverage, the corpus contains very few mentions of oligosaccharides, and their sample types are not documented.

Species.

Most corpora with species annotations are related to biomedicine, focusing on human, animal models (e.g., mice, rats) and microbes. The S1000 corpus is dedicated to biodiversity surveys [25], with its protist subset containing 671 scientific articles and 1104 species annotations ranging from green plants to fungi. The low number of mammalian mentions and the absence of chemical annotations in S1000 render it irrelevant for our purposes. The Bacteria Biotope corpus [26] is dedicated to bacteria habitats and therefore contains many mentions of mammals related to positive or negative microbial flora. However, the mammalian mentions are not tagged with NCBI taxa, as required by our study.

Geographical locations.

The geographical location information holds minor importance for mammalian oligosaccharide literature. The relevance of geographical location stems from its linkage to mammalian data, a feature absents in any existing corpus. We consider ChemDNER, S1000, and Bacteria Biotope relevant for evaluating the generality of the named-entity methods but inappropriate for training in our specific study domain.

Relationships.

The relationships among chemicals, media, species, their physiological stages, and locations are specific to our research topic. Most corpora in the biology domain are heavily focused on human health, where such information has not been annotated. The absence of a relevant corpus for text mining studies on milk oligosaccharides motivated our work.

Annotation guidelines.

The annotation guidelines provide the detailed instructions on how to perform text annotations. The guidelines document gives a definition of each entity, entity boundary criteria, and annotation rules, i.e., what should be tagged or not, and how to assign tagged elements to semantic resources. A wide range of examples is also depicted in the document to illustrate how to handle ambiguities and inconsistencies that may lead to annotation errors and discrepancies.

To ensure accessibility for individuals with diverse backgrounds, the document includes detailed explanations of each entity type. These guidelines are intended not only for annotators of the MilkOligoCorpus but also for adaptation and use for annotating other corpora or designing instructions in prompting methods.

The annotation lead, skilled in biological science, worked alongside with text and data mining experts to develop these guidelines essential for ensuring consistent annotation. The initial draft of guidelines, authored by a single individual, was refined through an iterative annotation process after being independently reviewed by four others.

The entire guidelines, with detailed instructions and extensive list of examples are provided in the repository (https://doi.org/10.17180/94mt-v661).

Annotators.

Two annotators with a background in biology (a PhD student expert in milk oligosaccharides, and a BS graduate with experience in manual annotation) performed the annotation. Consistency checking was performed by an NLP researcher expert in manual annotation curation.

Annotation process.

Annotation was conducted in three stages. In the first stage, the annotator expert in milk oligosaccharides annotated all entities and binary relations from the corpus. In the second stage, the second annotator reviewed the annotations and added normalization annotations (linking entities to entries from the semantic references) as well as n-ary relations (grouping binary relations describing the same event). Finally, we applied automatic scripts to check annotation consistency and flag potential errors, which were then reviewed manually by the NLP expert. Double-blind annotation was not performed because of time constraints, so we did not measure inter-annotator agreement.

Detected errors included missing mentions (mentions annotated in some documents but not in others), type mismatches (identical mentions tagged with different entity types), and missing or erroneous normalization annotations (with no or unknown semantic reference identifiers).

Annotation tools.

The annotators used two different tools to perform the annotation. The first annotator used the tagtog tool [36] (https://docs.tagtog.com/), which is easy to set up and use, and so more suitable for first-time annotators. The second annotator and the NLP reviewer both had experience in manual annotation and used the AlvisAE tool [37], which provides advanced functionalities such as n-ary relation annotation and entity normalization.

Entity recognition.

Entity recognition aims to identify textual mentions of given types relevant to a given domain. Thus, here, entities are related to the topic of mammalian milk oligosaccharides. We designed two baseline models to recognize the entities defined in MilkOligoCorpus: a simple rule-based baseline and a machine-learning (ML) -based baseline.

The rule-based approach leverages semantic resources, such as lexicons and thesaurus, and a set of simple regular expressions. It is implemented using the AlvisNLP corpus processing engine (https://github.com/Bibliome/alvisnlp). A basic preprocessing pipeline based on the Stanza natural language analysis package (https://stanfordnlp.github.io/stanza/) is also used for sentence segmentation, tokenization, lemmatization and general domain entity recognition (e.g., times, cardinals...).

A common ML-based approach to recognizing entities is to fine-tune an existing transformer-based language model to classify tokens into the target entity types. Here, we use the BioBERT cased model [38], a model specific to the biomedical domain. We fine-tuned BioBERT on MilkOligoCorpus to classify all entity types at once, in a 5-fold cross-validation setting with 5 random initializations (random seeds) to take into account the variability due to the stochastic nature of neural language models. For geographical locations, which are common in both general-domain and specialized-domain corpora, we also evaluated an existing model from the Stanza package (the English-language NERProcessor, https://stanfordnlp.github.io/stanza/ner.html) trained to recognize general-domain entities (including locations), in addition to fine-tuning our own model.

The code for our experiments is released at https://forgemia.inra.fr/holooligo/rb-pipeline and https://forgemia.inra.fr/holooligo/ner_holooligo.

Entity normalization.

Entity normalization consists in aligning textual mentions of entities within a corpus with their corresponding entries (i.e., class, category) in a knowledge base or lexicon. Biomedical entities have the particularity that many of them can be referred to by different textual forms (e.g., both “Triose C” and “β6’-Galactosyllactose” refer to the same oligosaccharide) while similar textual forms can correspond to different entities. It can therefore be challenging to accurately map the textual mention to the correct class. We chose BioSyn as machine-learning baseline for entity normalization because it achieved good performances on similar tasks. The BioSyn framework [39] was developed to address this challenge. BioSyn relies on a pre-trained language model that is fine-tuned using synonym marginalization, a strategy that incorporates synonym sets during training to learn robust representations for entity linking. We also implemented a lexicon-based baseline that performs simple string matching between entity mentions in the text and the labels of the entries from the semantic references with which these entities have to be normalized. The code for our experiments is released in the following repositories: https://forgemia.inra.fr/holooligo/rb-pipeline, https://forgemia.inra.fr/bibliome/geonorm, and https://forgemia.inra.fr/holooligo/nen_holooligo.

Relation extraction.

Relation extraction is the task of detecting relationships between given entities in a text and classifying them. Relation types and argument types for these relations are predefined according to the domain. Relations may involve two entities (binary relations) or more than two entities (n-ary relations). From an information extraction viewpoint, binary relation extraction is a more common and less challenging task than n-ary relation extraction. For our baseline experiments, we only tackled the task of oriented binary relation extraction.

As is typical for this task, we look at the problem of relation extraction as a sentence classification task and use a machine-learning baseline. We augment the original text with special characters that delineate the boundaries of the reference named entities and predict whether the delineated entities are in relation with each other. We use REBERT (https://forgemia.inra.fr/bibliome/re-bert), a tool that fine-tunes various language models specifically for relation extraction. We use the BioBERT model as the foundation model, as for our entity recognition baseline. The code for our experiments is released at https://forgemia.inra.fr/holooligo/brel_holooligo.

The following sentence from Rostami et al. [28] gives an example of the input given to the relation extraction method. There are two entities, the oligosaccharide name mention “3’SL” and the sample type “milk”:

[Example 1] “<< 3’SL>> was found to be by far the most prominent dog @@ milk @@ oligosaccharide.”

The relation extraction task is to predict whether there is a relation between the two entities “3’SL” and “milk”. When the annotation schema only defines one possible relation type for every pair of compatible entities, then the relation type can be deduced from the argument types, and thus there is no need to specify the type of the relation. From Example 1, the only possible relation between an oligosaccharide name and a sample type is the relation composed of (see the Results section below for a list of all relation types). This means that if the method predicts the correct label, “yes, there is a relation”, then the system will be able to deduce from the entity types that the relation type is composed of.

Entities.

Studies on mammalian milk composition include experiment records such as the number of individuals sampled, the geographical location where the sampling took place, the sampling methods, and the procedures and devices used to analyze the samples. Some studies give detailed metadata on female samples, namely breed, stage of lactation and number of gestations (primiparous, multiparous). Initially, the information included only six entity types: animals, characteristics (i.e., lactation stage, parity or age), oligosaccharides, geographical location, sample, and methodology. Several entity types have been trimmed down by removal or merging, while others have been separated into distinct entities to cover a wide range of relevant information while avoiding ambiguities with clearly distinct entities.

Originally, species mentions were included in a broader entity type, “animals”, including breeds when specified. We decided to introduce a separate breed entity type. Breed is mainly referred to in articles dealing with domesticated species (farm or domestic animal). The Livestock Breed Ontology, presented in the methodology, was the only lexica for breed terms, however, it is still limited to certain species (buffalo, cattle, chicken, goat, horse, pig and sheep breeds). We thus restricted the annotation of the breed type to domesticated species.

Similarly, female metadata were initially gathered in a large entity type including all characteristics. A first split resulted in the creation of the entity type female physiological stage along with an entity type for the age of the female. The physiological stage indicates how many times the female has already been in gestation, and can significantly impact the MO concentration [13]. After several annotations we decided to create two distinct entity types, lactation stage and postpartum age instead of a unique type for the age to avoid ambiguity. While postpartum age indicates a clear period of time, the entity lactation stage is a specific step during lactation. The exact duration depends on the species studied; colostrum last only 48h for rabbits while it extends up to 30 days in giant pandas [40].

Finally, milk oligosaccharides are complex structures of core molecules composed of three monosaccharide building blocks that can be repeated up to 25 times but also decorated with two other monosaccharides. They are typically classified in three categories according to this structure, neutral if the structure is only composed of the core molecules and sialylated or fucosylated depending on the monosaccharide decorative. Two other categories are also used, based on the chemical linkages between the monosaccharides, type I or type II. The precision of milk oligosaccharide composition varies depending on the methodology used. Some analyses are used to identify as many structures as possible, e.g., shotgun-MS analysis, while others cannot distinguish between isomers – molecules with the same chemical composition and mass but different arrangements of atoms – such as MALDI-TOF-MS [17]. As a result, several studies have been unable to clearly identify specific molecules but have instead detected characteristics elements of these molecules, classifying them into broader categories. Given this, we have decided to add a distinct entity type for oligosaccharide types.

The methodology also influences quantitative characterization. Initially gathered in one entity, we decided to create three milk oligosaccharide quantification entities. The absolute quantification includes indication regarding the total concentration of oligosaccharide, while the relative quantification is the proportion of an oligosaccharide compared to the total milk oligosaccharide. Finally, the richness gives the number of an oligosaccharide or a group of milk oligosaccharides.

In the end, the annotation schema consists of fourteen entity types that we classified into three groups based on the aspect they refer to: (i) Individual-related; (ii) Sample-related; (iii) Oligosaccharide-related.

The complete list of entities is available in Table 2, along with the semantic resources described in the following paragraph.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Description of the 14 entities of the annotation schema. Semantic normalization refers to semantic resources used to normalize the entities. “Indiv.-related” stands for “individual-related” and “Oligo-related” stands for “Oligosaccharide-related”.

https://doi.org/10.1371/journal.pone.0319729.t002

Semantic resources.

We used semantic resources to normalize the relevant entities. As described in the Material and Method section, we identified existing semantic resources to normalize species, geographical locations, breeds, and milk oligosaccharide names. Additionally, we also developed our own resources to normalize the entities for which no suitable existing resource was found, i.e., female physiological stage, sample type, methodology of analysis and oligosaccharide type (Table 2). We created four glossaries, the Female physiological stage thesaurus, the Sample thesaurus, the MO methods thesaurus and the Oligo type concepts, to overcome the lack of reference sources for the female physiological stage, the type of sample studied, the methodology used for the analysis, and the oligosaccharide type respectively.

The Female physiological stage thesaurus gathers terms to designate the physiological stage of the female, i.e., the number of gestation of females. Primiparous and multiparous are common names for parity, but we also found phrasal forms (“Bred for the first time”) or the number of parity (“x”-parity). Besides, porcine species have specific words, synonyms for pig, that indicate the parity. Sows are multiparous pigs, whereas gilts are primiparous pigs. These two physiological stages, primiparous and multiparous, are compiled in a 6-column tabular file including an identifier for each stage, synonyms, and specific characteristics relevant to the pig species. The Female physiological stage thesaurus is accessible in the repository https://doi.org/10.57745/LFXGFO.

This study focuses on the analysis of milk in its raw form as any intervention can potentially alter its composition. However, in some experiments, it is necessary to combine samples from multiple individuals. This is usually specified in the article and must be considered when comparing results across different experimental designs. Given the variability in formulation (adverbs, verbs, nouns) we decided to standardize it in the Sample thesaurus. The two types of samples, milk or pooled milk, are compiled in a 7-column tabular file that includes the identifier of the type of sample, and synonyms. The Sample thesaurus is accessible in the repository https://doi.org/10.57745/LFXGFO.

A third thesaurus is used to normalize the textual mentions of analysis methodology. The Chemical Method Ontology (CHMO) lists instruments and methods to collect data and prepare material in chemical experiments. It is an accurate ontology but it is too broad for our field of interest. Milk oligosaccharides analysis relies on a few specific methodologies which we have compiled into a concise thesaurus called the MO methods thesaurus. The forty-two methodology types are compiled in a 6-column tabular file that associates to each methodology an identifier, abbreviation, synonyms, the doi from the article where the methodology was mentioned, and the identifier from the CHMO database when available to enable data linking. The MO methods thesaurus is accessible in the repository https://doi.org/10.57745/LFXGFO.

The oligosaccharides’ types refer to the usual classification used to categorize milk oligosaccharides. This classification is based on the structure of the molecules. The Oligosaccharide type concepts lists the terms used and sialylated is the only type that can be referred to with a different name, acidic. The Oligosaccharide type concepts is accessible in the repository https://doi.org/10.57745/LFXGFO.

Finally, entities referring to quantification, i.e., individuals’ number, absolute quantification, relative quantification, and oligosaccharide richness, do not require semantic resources for normalization. Those numerical entities must only have homogenized units of measure.

Relations.

Information extraction aims to extract relevant knowledge from unstructured textual data to make them available in a structured format. To associate the MO structure pattern, i.e., the abundance and richness of milk oligosaccharides in a specific species, isolated entities alone are insufficient. It is necessary to define the relationships between them to accurately describe their association. The following sentence from Gopal et al. [18] gives an example [Example 2] of entities in relations, where the relative quantification entity, 50%, is meaningful only when linked to the oligosaccharides it described, which in turn are relevant only when linked to the species in which they are found.

[Example 2]: “Together, 3’- and 6’-sialyllactose account for more than 50% of the total oligosaccharides present in bovine”

Relations can involve two entities (binary relations) or multiple ones (n-ary relations). In our annotation schema, we defined eleven types of binary relationships, that are listed in Table 3 with the types of the arguments involved. Examples for each relation type are available in the guidelines document. We have chosen to focus on positive relations and not to annotate negated relations (such as in “Isoglobotriose was not found in the polar bear colostrum”). Such negations will be treated as negative examples when training automatic methods.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Characteristics of the binary relations with the definition and argument entity types.

https://doi.org/10.1371/journal.pone.0319729.t003

Given the definitions of the relationships from, composed of and found in quantity, the argument types are interchangeable. The from relation can connect a geography entity either to species, sample type or lactation stage entities, depending on the context. For relations involving milk oligosaccharide entities (composed of and found in quantity), the variation arises from differences in the precision of the molecule details, i.e., whether the study specifies the exact molecule names or only the type, and in quantification, distinguishing between the exact number of molecules or a relative quantity.

Regarding the relations number of, has produced, collected x days after parturition and analyzed by we decided to select interchangeable argument types based on the distance between arguments. The arguments to include in the relations were chosen after consideration of the location of the entity mentioned in the text. Relation extraction (RE) methods typically give better performance when entities are close by in the text. However, critical relations may occur between entities mentioned in separate and sometimes very distant sentences. In fact, to avoid redundancy authors do not always repeat the information already stated. For instance, in Salcedo et al. [29] the relation composed of links “milk” and “fucosylated” located in two distinct sentences [Example 3]:

[Example 3] “Porcine milk had lower OS diversity […] In agreement with previous studies, only 3 fucosylated OS were identified.”

Similarly, in Urashima et al. [41] the relation analyzed by includes the methodology of analysis (“gel filtration and preparative thin layer chromatography”) and the sample studied (“milk”) which is mentioned in the previous sentence [Example 4]:

[Example 4] “Carbohydrates were extracted from high Arctic harbour seal milk, Phoca vitulina vitulina (family Phocidae). Free neutral oligosaccharides were separated by gel filtration and preparative thin layer chromatography […].”

We chose to annotate cross-sentence relations between entities occurring in the same paragraph and not further away. However, to simplify the task for very distant entities, we decided to include flexible relations where one argument type may be replaced by another one close in meaning, such as in the has produced relation where the entity type lactation stage can be used instead of the entity type sample type. This flexibility does not provoke disparities.

Furthermore, considering the number of entities to be extracted, binary relations may not be sufficient to capture the information unlike n-ary relations that link more than two entities. The sentence in example 5 gives an example of MO (referred to as OS in example 5) quantities measured in the milk of different species where four arguments must be linked to fully capture the meaning [42].

[Example 5]: “Approximately 29 different OS were identified in porcine milk, compared to 40 in bovine milk and ≥130 in human milk.”

In this sentence, we need to establish three 4-ary relations indicating the quantity of milk oligosaccharides in the species: (i) 29 MOs in porcine milk; (ii) 40 MOs in bovine milk; (iii) 130 MOs in human milk.

Since, “OS” is stated only once, a relation extraction method relying solely on binary relations cannot accurately determine the number of MOs present in the different species. It would link all numbers, 29, 40 and >130 to OS, the latter being related to milk, itself linked to porcine, bovine or human. We cannot simply merge the binary relations because it would not indicate the number of MOs associated with the respective species.

N-ary relations are thus an essential part of our annotation schema. Fig 1 is a simplification of our annotation schema with entities and relations organized in hierarchies based on the types of lines representing the relations. We have grouped entities by theme: (i) Individual-related entities are in blue; (ii) Sample-related entities are in orange; (iii) Oligosaccharide-related entities are in green.

As the number of arguments increases, automatically extracting relations becomes even more challenging. Therefore, we have prioritized the relations to be extracted, ensuring that efforts are primarily focused on the key relations of interest. These relations are framed in Fig 1 by “Main relations of interest”.

The entire schema represents the maximal n-ary relation and each subgraph is a potential relation. The main relations of interest contain four entities: species; sample; oligosaccharides and oligosaccharide quantities. This 4-ary relation contains the central biological information to describe MO patterns in mammalian species. Entities linked by dash lines bring a higher level of detail.

The two 5-ary relations in the sentence below [Example 6] from Rostami et al. [28] include detailed information on the species, postpartum age or lactation stage which indicate relevant biological information regarding the evolution of MOs during the course of lactation within a same individual. The extraction of this information is a key to understanding the biological significance of such variations in MO profiles.

[Example 6]: 3’SL was found to be by far the most prominent dog milk oligosaccharide. In the first days of lactation it was present at levels around 7.5 g/L (12 mM), which dropped rapidly to approximately 1.5 g/L at 10 days of lactation.

Finally, we estimated that geography and breed represent the least critical information to extract. While they provide relevant information, these entities appear less frequently in texts and add complexity to the n-ary relation extraction task. As shown in Example 7 below from [43], a 6-ary relation, including species, oligosaccharide name, geography, sample, postpartum age, and oligosaccharide absolute quantification, is necessary to address the differences in human MO (HMO in Example 7) concentration according to the postpartum age.

[Example 7]: “A decrease (P < 0.05) in HMO concentration was observed during the course of lactation for the US mothers, corresponding to 19.3 ±/- 2.9 g/L for milk collected on postnatal day 10, decreasing to 8.53 ±/- 1.18 g/L on day 120 (repeated measures; n = 14).”

In many cases, n-ary relations are necessary to resolve ambiguities and extract the most detailed information possible. The annotation schema takes into account the challenge of n-ary relation extraction by prioritizing information. This prioritization does not exclude entities but instead establishes a hierarchy in cases of complex extraction aiding decision-making when necessary. We identified the key milk oligosaccharides information of scientific relevance to support the development of automatic methods which will be the focus of development and training efforts.