Combining Evidence, Specificity, and Proximity towards the Normalization of Gene Ontology Terms in Text
© S. Gaudan et al. 2008
Received: 21 November 2007
Accepted: 15 February 2008
Published: 3 March 2008
Structured information provided by manual annotation of proteins with Gene Ontology concepts represents a high-quality reliable data source for the research community. However, a limited scope of proteins is annotated due to the amount of human resources required to fully annotate each individual gene product from the literature. We introduce a novel method for automatic identification of GO terms in natural language text. The method takes into consideration several features: (1) the evidence for a GO term given by the words occurring in text, (2) the proximity between the words, and (3) the specificity of the GO terms based on their information content. The method has been evaluated on the BioCreAtIvE corpus and has been compared to current state of the art methods. The precision reached 0.34 at a recall of 0.34 for the identified terms at rank 1. In our analysis, we observe that the identification of GO terms in the _cellular component_ subbranch of GO is more accurate than for terms from the other two subbranches. This observation is explained by the average number of words forming the terminology over the different subbranches.
Gene Ontology (GO) is a structured database of biological knowledge that provides a controlled vocabulary to describe concepts in the domains of molecular and cellular biology . Currently, the GO vocabulary consists of more than 23 000 terms and is distributed via the GO web pages at http://www.geneontology.org/. GO contains three ontologies: (1) molecular function that describes molecular activities, (2) biological process that describes ordered assemblies of events or a recognized series of molecular functions, and (3) cellular component that describes subcellular locations and macromolecular complexes. Gene products across all species are annotated with GO terms (called GO annotations) to support the description of protein molecular characteristics. These annotations support exchange and reuse of gene characteristics in a standardized way.
The GO annotation (GOA) data is generated by the GO Consortium members through a combination of electronic and manual techniques based on the literature. The manual annotations ensure high-quality reliable data set and have been frequently used by the biological and bioinformatics communities . Apart from manual annotation of proteins, the GO knowledge resource is used to annotate microarray data  and to automatically annotate proteins with GO terms amongst other semantic types .
However, generation of manual GO annotations based on the literature requires a team of trained biologists and is time consuming, which leads to a limited and thus insufficient coverage of proteins with manual annotations. As a consequence, more efficient means are required to increase the throughput of annotations. Text mining technology is a means to support the annotation process by efficiently identifying and extracting information from the literature and transforming it into GOA input. This approach can produce a significant increase in the number of GO annotations from the scientific publications and improve the benefit from reusing the data.
Automatic GO annotation methods using text mining have been evaluated in the BioCreAtIvE competition as part of task 2 . In the first subtask of task 2, the competition participants had to identify a passage in a full text document that supports the annotation of a given protein with a given GO term.
Different techniques have been applied to this task referenced in Blaschke et al.  that can be split into three groups. (1) The first group of techniques processed the GO lexicon to feed the content into pattern matching techniques. (2) The second group used the input information to train machine learning techniques for the identification of relevant passages. (3) Finally, the last group applied techniques that combine pattern matching and machine learning approaches (hybrid methods).
The pattern matching methods (group 1) delivered the highest number of correct annotations but produced many false positives. The hybrid methods returned a lower number of false positives but had the disadvantage of a lower number of true positives. Finally, the machine learning techniques showed intermediate performance. The outcome of the BioCreAtIvE contest is that automatic GO annotation remains a challenging task and that further research is required to obtain better results.
The work presented in this paper is best categorized into the pattern matching group and our results will be compared to two state-of-the-art methods from the same group: Ruch  and Couto et al. .
Both methods base their scoring functions on the word frequencies extracted from the GO lexicon. The more frequent a word is, the smaller the contribution of the word to the final score of the GO term. Furthermore, Ruch's method increments the score of a GO term for which the candidate matches one of the predefined syntactical patterns of the system while Couto's method limits the look-up of words to the sentences and filters out GO terms that are frequent in the Gene Ontology annotations.
Apart from the contributions to the BioCreAtIvE workshop, other solutions for the extraction of Gene Ontology terms from text have been proposed. GoPubMed  is a web-based tool that offers access to subgraphs from Gene Ontology that matches PubMed abstracts. The methods behind GoPubMed identifies GO terms in abstracts by matching first the suffix of the terms and then gathering one by one the more specific words of the left part of the terms. In addition, a selection of general terms' suffixes and some words from a stop list is ignored during the word matching process. The authors do not provide explicit measures of the recall.
Our approach is also based on the identification of weighted words that compose terms denoting GO concepts. The novelty of our method resides in the integration of two new aspects in the scoring method: the proximity between words in text and the amount of information carried by each individual words.
The paper is organized as follows. First, we describe the method used to score mentions of GO terms in text. Then we explain the design of the evaluation method and present the obtained results. Finally, we focus on the characteristics of the method's performance. The evaluation focuses on the identification of GO terms in text and not on the identification of a text passage containing the evidence for a given protein being related to a given GO term.
The identification of a term in text requires the localization of the evidence for its mention in the text. The occurrence of words constituting the terms is a reliable evidence. However, this evidence might introduce related terms that are hypernyms or hyponyms of each other. The selection of the most specific term avoids loss of detail. Evidences extracted from text are unlikely to have a relationship with each other if they are far from each other in text. As a result, such evidence is unlikely to be the result of the mention of the term they constitute. Finally, the described methods make abstraction of the order of the words occurring in the terms, thus allowing to identify syntactical variants of the terms in the text.
We describe in the following a method that integrates the concepts of evidence, specificity, and proximity to identify mentions of terms in text.
Before describing all three aspects of the method, the concept of a zone needs to be introduced. A zone is a continuous stretch of text that is composed of words. The decomposition of a document into zones can follow various definitions. The zones of a document can be its paragraphs, its sentences, or, for instance, the noun phrases contained in the document. The zone can also be the document itself.
2.1. Evidence and Specificity
Lin  illustrates the benefit from the similarity measure with a number of scenarios such as the assessment of string similarity based on -grams or measurement of the similarity between concepts derived from an ontology by considering the ontology's structure. But more generally, the proposed approach can be applied to any model for describing the entities to be compared. In our study, the units describing the entities are words.
A term consists of words where each word contributes to the syntactic and semantic representations of the term. Intuitively, the word " " is of lower importance than the word " " to convey the meaning of the term " ." Similarly, " " contributes less than " ."
Terms frequencies can nevertheless be nil, thus carrying the same undetermined amount of information. These nil frequencies can be observed for half of the Gene Ontology terms. Therefore, we will privilege the first estimation of the specificity.
The fact that the words of a term occur in a text zone does not necessarily mean that the term is mentioned in the zone, especially if the zone is relatively large (e.g., the whole document). The proximity between the words of in provides a clue on the likelihood that those words belong to a common term.
The proximity of the words forming a term and found in the text is a function of the positional index of the words in the text. The benefit of using proximity between keywords has been greatly shown in the field of information retrieval . Various approaches for measuring the notion of proximity or density have been proposed in order to score and rank more efficiently query results. In the following, we propose to use the proximity in the sense of density.
The relative distances between the individual tokens from the term and the minimum distance between the words in the optimal case where the words are consecutive are taken into consideration when measuring the proximity.
The proximity indicates whether the words constituting the terms and found in a zone are close to each other or dispersed over the zone. The proximity is when the words of are consecutive in and decreases towards , while the words move away from each other in .
In case a word occurs several times in a zone, the smallest distance from all combinations is considered. This solution gives the opportunity to keep a relatively high proximity for a term such as " " that occurs in a zone like " ."
Evidence, specificity, and proximity are the three parameters that are combined to score the mention of a term in a text zone . The three criteria may be of various importance and must be weighted accordingly. Their importance varies in function of the nature of the terminology and of the text.
2.4. Previous Work
We provide the details of Couto and Ruch's methods. Our approach is compared to these two systems in Section 3.
The approach of Ruch  is based on pattern matching and text categorization ranking. The pattern matching module measures the similarity between a term and a sliding window of five words from the passages. The text categorization module introduces a vector space that provides a similarity measure between a GO term, seen as a document, and a passage, seen as a query. The modules ranking the GO terms given a passage as a text categorizer would rank documents given a query. A fusion of the two modules is operated by using the result of the vector space as a reference list while the pattern matching result is used as a boosting factor. Finally, the candidates scores are reordered by considering the overlap between the candidates and the noun phrases extracted from the text. The noun phrases are identified by applying a set of predefined patterns on the text's part-of-speech tags.
We illustrate here the identification of terms derived from the Gene Ontology (GO). GO provides a controlled vocabulary to describe genes and proteins in any organism. Each entry in GO has, among other fields, a term name, a unique identifier, and a set of synonyms.
The evaluation has the goal to provide a measure of the accuracy of the system for identifying in text terms derived from the Gene Onology.
The system has been evaluated on the BioCreAtIvE corpus, the most suitable standard evaluation set for the annotation of text with Gene Ontology terms. The evaluation illustrates the performance in terms of precision and recall.
3.2. Statistical Significance
The precision and recall are well-established measurements for estimating the accuracy of text mining methods. However, the comparison of individual methods by only considering the precision/recall can often underestimate their differences . Therefore, while presenting the performance of the individual methods, we also provide a statistical significance of the differences between the precisions of the various methods at rank . To do so, we test the null hypothesis:
Method 1 and method 2 are not different.
where is the number of trial with a precision that is at least as good as the one of the so-called best method and where is the number of trials. It has been acknowledged that a probability below illustrates a statistical significance in the result differences .
3.3. Probabilistic Model
Three probabilistic models have been tested for computing . The first one is based on the word frequencies in Medline. The second one is also based on word frequencies in Medline, but only for abstracts that have been annotated by curators in mammalian annotation projects . The last probabilistic model is based on the word frequencies in the Gene Ontology itself, as in Ruch and Couto's methods.
3.3.2. Medline Abstracts from Gene Ontology Annotations (GOA)
3.3.3. Gene Ontology
where is the number of names in the Gene Ontology where occurs and is the total number of names in the Gene Ontology. The probability that a word occurs in the ontology is independent of the hierarchical structure of the ontology. Indeed, the structure does not contain information on the probability that a word occurs in a concept name. However, the opposite is not true; it can be observed that many GO terms contain words that also belong to their parents, thus the concept names provide some information on the structure of the ontology. Lin  uses the structural information contained in the ontology to measure the semantic similarity between two concepts. However, in the present context, this structural information is not available for the zone and thus cannot be considered. Furthermore, the aim of our approach is not to measure the semantic similarity but the similarity in terms of evidence.
The three models have been evaluated with the optimal parameters ( , ) described in the next section. The best performance is reached when the Gene Ontology is used as the model, with a global performance of . Then, when the Gene Ontology Annotation's corpus is used and decreases even more ( ) with the entire Medline as the model.
3.4. Evidence, Proximity, and Specificity
Zone = sentence.
Zone = noun phrase.
Setups (1), (2), and (3) illustrate the role of the zone while setups (4), (5), and (6) demonstrate the importance of the evidence, the proximity, and the specificity, respectively. For the sake of brevity, only the optimal parameter is provided. The optimal parameters have been determined by trying the different combinations of , and within with a step of on the second independent annotated corpus from BioCreAtIvE.
3.5. Comparison with Existing Methods
In order to provide a baseline in the comparison, the exact match method, in other words the identification of terms by the localization of the constituting words in their consecutive order, is evaluated. The method is compared to the exact match method as well as the ones described by Rush  and Couto et al. .
Global performance of the 4 methods.
3.6. Branches of GO
The role played by each individual aspect of the method is clearly illustrated in Figure 1. The evidence and the proximity are factors to decide whether the term is mentioned or not in the analyzed text whereas the specificity prioritizes the terms in the list of candidates. As expected, the evidence, in other words the proportion of information carried by words from found in , plays a major role in the identification of mentioned terms in text. But it also appears that the specificity of the terms is an important criterion when annotating text with GO terms. The proximity is essential when the zone's size increases and is a convenient technique to constrain words to occur in a common window.
The novelty of our method resides in the specificity and the proximity aspects. The integration of these two aspects justifies the better performance over Rush  and Couto et al. . In spite of the fact that Couto limits the look-up to sentences and that Ruch increases the score of a pattern match, the notion of a flexible proximity is absent from their methods.
Better performance for the identification of GO annotations has only been reported in two cases that clearly differ from our approach. In Couto et al.'s approach , prior knowledge from GOA for selected proteins was used to annotate orthologs based on the identification of GO terms from Medline abstracts with Couto's method. Shatkay et al.  transformed Medline abstracts into feature vectors of distinguishing terms that are predictive for a subcellular localization and these feature vectors were used to categorize proteins into a subcellular location. The method appears to be efficient in predicting cellular localizations of proteins when combined with other genomic and proteomic data. However, this method did not propose any evidence of a GO term from the text and is restricted to the cellular localization.
The overall precision of for a recall at at rank seems at first limited. In the context of BioCreAtIvE, the evaluation method judges as correct a GO term only if there exists an annotation between the GO term and the protein mentioned in the passage. However, many other GO terms appear in the passages and are classified as incorrect because they are not bound to a protein. Furthermore, GO terms are refereed with unambiguous names due to the descriptive nature of the names. As a result, if the evaluation focusses only on the mentions of the GO terms, independently of a potential relationship with a mentioned protein, then the recall gets close to 0.6 at rank 10 while the precision can be assumed to get close to 1 due to the unambiguous nature of the terminology. Indeed, the occurrence of a GO name in text has a high probability to refer to the GO concept instead of another concept sharing the same name. In contrast, this property is not observed with protein names due to their ambiguous nature. For instance, the mention of "why" can refer to the adverb or to a gene.
The presented evaluation provides nevertheless an order on the performances of the various methods. It also provides a lower boundary for the precision as well as an upper boundary for the recall on the BioCreAtIvE test set. In other words, the identification of GO terms, independently of a relationship with a protein, can be achieved with a precision close to one and with a recall of 0.6. Such performance provides satisfying results for assisting biologists in the task of curating manuscripts when the localization of Gene Ontology terms in text is needed. The remaining challenging task is then the association of the GO terms with the proteins they potentially describe. This last step is still performed by curators since automatic methods remain deceiving.
The automatic identification of GO terms using the biomedical literature can significantly increase the amount of information about proteins available as a structured knowledge resource. We composed a new method integrating the notions of evidence, specificity, and proximity to identify terms from text. The evidence is the proportion of statistically meaningful words from a term found in text. This criterion has been previously exploited to achieve the identification of GO terms in text. The specificity allows the selection of the most informative terms while the proximity provides a clue about the belonging of words found in text to a term.
Adding proximity and specificity into the scoring function drives the system to reach better performance in comparison to other state-of-the-art methods. All has been evaluated against the annotated BioCreAtIvE I corpus and has been benchmarked to a precision of 0.34 at a recall of 0.34 for the terms delivered at rank 1. The evidence parameter is the major criterion for scoring mentions of GO terms in text. However, the specificity and the proximity are important to improve the identification of the terms. The size of the zone selected to identify the words has an impact on the performance of the method with an optimal setup based on the passage for the evaluation. The optimal probabilistic model was used to compute the probability that the word occur is based on the Gene Ontology, model also chosen by Couto et al. .
Some variations in the performance regarding the individual branches of GO can be observed and might be correlated with the nature and the distribution of the number of words in their respective terms. The performance of the compared systems indicates that further research is needed to better identify terms from the biological process and molecular function branches. Disambiguation, synonym formation, and word variation could potentially improve the performance of the identification of GO terms.
The first author is supported by an "E-STAR" fellowship funded by the EC's FP6 Marie Curie Host Fellowship for Early Stage Research Training under Contract no. MEST-CT-2004-504640. The authors would like to acknowledge Julien Gagneur and Harald Kirsch for their wise advices. Last but not least, the authors would like to thank Julien Gobeill, Patrich Ruch, and Fransisco Couto for processing the passages with their respective systems.
- Harris MA, Clark J, Ireland A, et al.: The Gene Ontology (GO) database and informatics resource. Nucleic Acids Research 2004, 32: D258-D261. 10.1093/nar/gkh036View ArticleGoogle Scholar
- Lee V, Camon E, Dimmer E, Barrell D, Apweiler R: Who tangos with GOA? Use of Gene Ontology Annotation (GOA) for biological interpretation of '-omics' data and for validation of automatic annotation tools. Silico Biology 2005,5(1):5-8.Google Scholar
- Doniger SW, Salomonis N, Dahlquist KD, Vranizan K, Lawlor SC, Conklin BR: MAPPFinder: using gene ontology and GenMAPP to create a global gene-expression profile from microarray data. Genome Biology 2003, 4: R7. 10.1186/gb-2003-4-1-r7View ArticleGoogle Scholar
- Rebholz-Schuhmann D, Kirsch H, Arregui M, Gaudan S, Riethoven M, Stoehr P: EBIMed—text crunching to gather facts for proteins from medline. Bioinformatics 2007,23(2):e237-e244. 10.1093/bioinformatics/btl302View ArticleGoogle Scholar
- Blaschke C, Leon EA, Krallinger M, Valencia A: Evaluation of BioCreAtIvE assessment of task 2. BMC Bioinformatics 2005,6(supplement 1):S16.View ArticleGoogle Scholar
- Ruch P: Automatic assignment of biomedical categories: toward a generic approach. Bioinformatics 2006,22(6):658-664. 10.1093/bioinformatics/bti783View ArticleGoogle Scholar
- Couto FM, Silva MJ, Coutinho PM: Finding genomic ontology terms in text using evidence content. BMC Bioinformatics 2005,6(supplement 1):S21.View ArticleGoogle Scholar
- Doms A, Schroeder M: GoPubMed: exploring PubMed with the Gene Ontology. Nucleic Acids Research 2005,33(supplement 2):W783-W786.View ArticleGoogle Scholar
- Lin D: An information-theoretic definition of similarity. Proceedings of the 15th International Conference on Machine Learning, Madison, Wis, USA, July 1998 296-304.Google Scholar
- Keen EM: Some aspects of proximity searching in text retrieval systems. Journal of Information Science 1992,18(2):89-98. 10.1177/016555159201800202View ArticleGoogle Scholar
- Yeh A: More accurate tests for the statistical significance of result differences. In Proceedings of the 18th Conference on Computational Linguistics, Saarbrücken, Germany, July-August 2000, vol. 2. Association for Computational Linguistics; 947-953.View ArticleGoogle Scholar
- Camon E, Magrane M, Barrell D, et al.: The gene ontology annotation (goa) project: implementation of go in swissprot, trembl, and interpro. Genome Research 2003,13(4):662-672. 10.1101/gr.461403View ArticleGoogle Scholar
- Couto FM, Silva MJ, Lee V, et al.: GOAnnotator: linking protein GO annotations to evidence text. Journal of Biomedical Discovery and Collaboration 2006., 1, article 19:Google Scholar
- Shatkay H, Höglund A, Brady S, Blum T, Dönnes P, Kohlbacher O: SherLoc: high-accuracy prediction of protein subcellular localization by integrating text and protein sequence data. Bioinformatics 2007,23(11):1410-1417. 10.1093/bioinformatics/btm115View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.