This software demonstration paper presents Identifying Chemical Entities (ICE), a platform composed by algorithms for chemical entity recognition, entity resolution to a reference database, namely ChEBI, and validation using chemical semantic similarity. It aims to provide the users with an improved display of entity recognition results, exposing outliers which are possible recognition errors and displaying evidence that corroborates consistent chemical entities in the entity recognition and resolution process.
Chemical information in the scientific literature is increasing at a fast pace, making it difficult for researchers to keep up to date with what is being published. Chemical abstracting services addressed this issue by having chemistry experts manually extract the necessary information from the literature. However, with the exponential growth in publication rate, automatic methods for chemical entity identification are increasingly needed.
The lack of available chemical terminologies has been an
important aspect for the slow development of chemical text mining
systems, but the recent release of ChEBI
ICE (Identifying Chemical Entities) is a software platform that integrates algorithms for chemical entity recognition in biomedical literature, resolution of named entities to the ChEBI database, and validation of annotations using semantic similarity in the ChEBI ontology to gather annotation evidence. Figure 1 shows an outline of the ICE architecture.
The software platform uses the algorithms for chemical entity
recognition presented by
The first step in the entity recognition process is the splitting of the input text into a sequence of tokens, which are then classified Text Recognition Resolution
Annotated text mapped to ChEBI
Validation Annotated text with confidence scores Fig. 1. Architecture of ICE with the three modules for Recognition,
Resolution and Validation of chemical entities. according to the previous model. With this method chemical named entities are located in the input text, however there is no mapping to a reference database and thus an entity resolution module is required. 3
For entity resolution to the ChEBI database this software platform
uses the algorithms presented in
A lexical similarity method is used to compare the constituent words in the input string with the constituent words of each ChEBI term, to which different weights have been assigned according to its frequency in the database. A final score is provided with the mapping and a minimum score threshold can be used to allow for no mapping to be made in cases where the provided mapping score is too low, which might be an indication that the term is absent from ChEBI. 4
A novel algorithm was developed for this software platform to perform validation of named entities mapped to ChEBI. The underlying assumption is that most often a text fragment such as a paragraph has a limited scope, and therefore normally contains entities that are somehow related to each other. With this in mind, this algorithm takes as input entities mapped to ChEBI within a text fragment and searches for relationships between them. The output is for each input entity the most similar entity within the text fragment, with the corresponding similarity score.
Using the ontology structure of ChEBI we are able to
compare chemical entities according to both structural and
functional characteristics through several possible semantic
similarity measures
Based on the maximum similarity score of each entity we can filter outliers and corroborate consistent entities. This is performed using two thresholds. The algorithm has thus four parameters that can be tuned according to the user requirements: the text fragment window, which can be the full document, paragraph or sentence; the semantic similarity measure to be used for the comparison of the entities; and the two thresholds, which can be tuned to allow for more precision or recall.
The final result is advantageous in semi-automated tasks by providing an improved view over the entity recognition results, because the user will have an indication of which entities have increased consistency and are most probably correctly identified, as well as which entities are outliers in the sense that no similar entities could be find, which might be an evidence of a recognition error. 5
As an example lets consider the following sample sentence and follow the steps of ICE.
A mixture of ethanol, propanol and acetic acid with a small amount of sodium chloride.
In the entity recognition step four entities can be found are now highlighted.
A mixture of ethanol, propanol and acetic acid with a small amount of sodium chloride.
In entity resolution, ChEBI identifiers are assigned to the entities.
A mixture of ethanol [CHEBI:16236], propanol [CHEBI:28831] and acetic acid [CHEBI:15366] with a small amount of sodium chloride [CHEBI:26710].
Entities mapped to ChEBI are compared to each other, and those with high maximum similarity are considered consistent. That is the case of ethanol and propanol, which have high semantic similarity. Sodium chloride has a low similarity with the other entities, and is thus considered an outlier. Acetic acid in the example has reasonable similarity to both ethanol and propanol, but not high enough to be considered consistent. The final result would highlight differently the new three classes of entities.
In this example, the text fragment window is very small and thus there are few entities to be compared, which can provide misleading results. All four entities were in fact correct, but the indication of consistent and outlier entities still provide interesting and meaningful explanation that users can use in a semi-automated fashion. 6
This software demonstration paper presents ICE, a framework that performs chemical entity recognition and resolution to the ChEBI database. The entity recognition results are then further processed using the ChEBI ontology to identify outliers and consistent entities, providing the user with annotation evidence.
The goal is to provide the user, potentially a curator performing semi-automatic annotation tasks, different layers of certainty in the recognized entities for better analysis of automatic chemical entity recognition results, as well as providing evidence based on semantic similarity.
The framework is being extended to other ontologies in addition
to ChEBI using ontology matching techniques that can align shared
concepts between ontologies such as the Gene Ontology and ChEBI
Also, this software platform will be used in the context of the project SPNet. This project aims to uncover network motifs associated with virulence in Streptococcus pneumoniae. This requires an extensive analysis of the transcriptional and metabolic networks involved in virulence, and the integration of chemical data with genomic and proteomic data will be required.
The authors want to thank the Portuguese Fundação para a Ciência e Tecnologia through the financial support of the SPNet project (PTDC/EIA-EIA/119119/2010), the SOMER project (PTDC/EIAEIA/119119/2010) and the PhD grant SFRH/BD/36015/2007 and through the LASIGE multi-annual support.