Concept-based query expansion for retrieving gene related publications from MEDLINE
© Matos et al. 2010
Received: 1 July 2009
Accepted: 28 April 2010
Published: 28 April 2010
Skip to main content
© Matos et al. 2010
Received: 1 July 2009
Accepted: 28 April 2010
Published: 28 April 2010
Advances in biotechnology and in high-throughput methods for gene analysis have contributed to an exponential increase in the number of scientific publications in these fields of study. While much of the data and results described in these articles are entered and annotated in the various existing biomedical databases, the scientific literature is still the major source of information. There is, therefore, a growing need for text mining and information retrieval tools to help researchers find the relevant articles for their study. To tackle this, several tools have been proposed to provide alternative solutions for specific user requests.
This paper presents QuExT, a new PubMed-based document retrieval and prioritization tool that, from a given list of genes, searches for the most relevant results from the literature. QuExT follows a concept-oriented query expansion methodology to find documents containing concepts related to the genes in the user input, such as protein and pathway names. The retrieved documents are ranked according to user-definable weights assigned to each concept class. By changing these weights, users can modify the ranking of the results in order to focus on documents dealing with a specific concept. The method's performance was evaluated using data from the 2004 TREC genomics track, producing a mean average precision of 0.425, with an average of 4.8 and 31.3 relevant documents within the top 10 and 100 retrieved abstracts, respectively.
QuExT implements a concept-based query expansion scheme that leverages gene-related information available on a variety of biological resources. The main advantage of the system is to give the user control over the ranking of the results by means of a simple weighting scheme. Using this approach, researchers can effortlessly explore the literature regarding a group of genes and focus on the different aspects relating to these genes.
Advances in biotechnology, together with the widespread use of high-throughput methods for gene analysis, have helped shifting the focus of biological research from specific genes and proteins to a more systemic analysis of the underlying biological problem. Researchers now face the increasing need to plan their experiments and analyse the resulting datasets in view of the quickly expanding biomedical information available [1, 2]. Although much of this knowledge is being annotated in the various existing biomedical databases, keeping these up-to-date is a difficult task due to the rapid emergence of new results and the exponential increase of published articles. Therefore, many relevant research outcomes are still enclosed as free-text in the literature, which remains the major source of information for researchers . A significant challenge for researchers is how to identify the most relevant articles for their specific study, and how to integrate this with the scientific knowledge annotated in biological databases, in an efficient manner . This integrated view of the literature, in the framework of a more systematized and formalized knowledge extracted from databases and ontologies, and the ability to analyse large datasets (a list of genes from a microarray experiment, for example), are two important requisites for biological data analysis .
The most popular biomedical information retrieval system, PubMed, gives researchers access to over 17 million citations from a broad collection of scientific journals, indexed by the MEDLINE literature database. PubMed facilitates access to the biomedical literature by combining the Medical Subject Headings (MeSH) based indexing from MEDLINE, with Boolean and vector space models for document retrieval, offering a single interface from which these journals can be searched . However, and despite these strong points, there are some limitations in using PubMed or other similar tools. A first limitation comes from the fact that keyword-based searches usually lead to underspecified queries, which is a main problem in any information retrieval (IR) system . This usually means that users will have to perform various iterations and modifications to their queries in order to satisfy their information needs. This process is well described in  in the context of information-seeking behaviour patterns in biomedical information retrieval. Another drawback is that PubMed does not sort the retrieved documents in terms of how relevant they are for the user query. Instead, the documents satisfying the query are retrieved and presented in reverse date order. This approach is suitable for such cases in which the user is familiar with a particular field and wants to find the most recent publications. However, if the user is looking for articles associated with several query terms and possibly describing relations between those terms, the most relevant documents may appear too far down the result list to be easily retrieved by the user.
To address the issues mentioned above, several tools have been developed in the past years that combine information extraction, text mining and natural language processing techniques to help retrieve relevant articles from the biomedical literature . Most of these tools are based on the MEDLINE literature database and take advantage of the domain knowledge available in databases and resources like the Entrez Gene, UniProt, GO or UMLS to process the titles and abstracts of texts and present the extracted information in different forms: relevant sentences describing a biological process or linking two or more biological entities, networks of interrelations, or in terms of co-occurrence statistics between domain terms. One such example is the GoPubMed tool , which retrieves MEDLINE abstracts and categorizes them according to the Gene Ontology (GO) and MeSH terms. Another tool, iHOP , uses genes and proteins as links between sentences, allowing the navigation through sentences and abstracts. The AliBaba system  uses pattern matching and co-occurrence statistics to find associations between biological entities such as genes, proteins or diseases identified in MEDLINE abstracts, and presents the search results in the form of a graph. EBIMed  finds protein/gene names, GO annotations, drugs and species in PubMed abstracts showing the results in a table with links to the sentences and abstracts that support the corresponding associations. FACTA  retrieves abstracts from PubMed and identifies biomedical concepts (e.g. genes/proteins, diseases, enzymes and chemical compounds) co-occurring with the terms in the user's query. The concepts are presented to the user in a tabular format and are ranked based on the co-occurrence statistics or on pointwise mutual information. More recently, there has been some focus on applying more detailed linguistic processing in order to improve information retrieval and extraction. Chilibot  retrieves sentences from MEDLINE abstracts relating to a pair (or a list) of proteins, genes, or keywords, and applies shallow parsing to classify these sentences as interactive, non-interactive or simple abstract co-occurrence. The identified relationships between entities or keywords are then displayed as a graph. Another tool, MEDIE , uses a deep-parser and a term recognizer to index abstracts based on pre-computed semantic annotations, allowing for real-time retrieval of sentences containing biological concepts that are related to the user query terms.
Despite the availability of several specific tools, such as the ones presented above, we feel that the demand for finding references relevant for a large set of is still not fully addressed. This constitutes an important query type, as it is a typical outcome of many experimental techniques. An example is a gene expression study, in which, after measuring the relative mRNA expression levels of thousands of genes, one usually obtains a subset of differentially expressed genes that are then considered for further analysis [16, 17]. The ability to rapidly identify the literature describing relations between these differentially expressed genes is crucial for the success of data analysis. In such cases, the problem of obtaining the documents which are more relevant for the user becomes even more critical because of the large number of genes being studied, the high degree of synonymy and term variability, and the ambiguity in gene names.
While it is possible to perform a composite query in PubMed, or use a list of genes as input to some of the IR tools described above, these systems do not offer a retrieval and ranking strategy which ensures that the obtained results are sorted according to the relevance for the entire input list. A tool more oriented to analysing a set of genes is microGENIE , which accepts a set of genes as input and combines information from the UniGene and SwissProt databases to create an expanded query string that is submitted to PubMed. A more recently proposed tool, GeneE , follows a similar approach. In this tool, gene names in the user input are expanded to include known synonyms, which are obtained from four reference databases and filtered to eliminate ambiguous terms. The expanded query can then be submitted to different search engines, including PubMed. In this paper, we propose QuExT (Query Expansion Tool), a document indexing and retrieval application that obtains, from the MEDLINE database, a ranked list of publications that are most significant to a particular set of genes. Document retrieval and ranking are based on a concept-based methodology that broadens the resulting set of documents to include documents focusing on these gene-related concepts. Each gene in the input list is expanded to its various synonyms and to a network of biologically associated terms, namely proteins, metabolic pathways and diseases. Furthermore, the retrieved documents are ranked according to user-defined weights for each of these concept classes. By simply changing these weights, users can alter the order of the documents, allowing them to obtain for example, documents that are more focused on the metabolic pathways in which the initial genes are involved.
Gene name mapping and normalization, to deal with the different identifiers used in different databases;
Query expansion, to deal with the high degree of synonymy in gene names and also to obtain documents that not only refer to the genes but also to related concepts;
Retrieval and ranking of the results, so that the top documents are the most relevant to the entire set or subsets of the input genes;
Term re-weighting, which allows users to change the influence of the different concept types in document ranking, therefore focusing on specific aspects related to the gene list.
The term mapping scheme implemented in the database works as a thesaurus through which each gene identifier (or symbol) in the user input is validated. The input genes that are present in the database are mapped to an internal identifier which is then used in the following operations. This mapping is performed considering the organism selected by the user, therefore reducing the problem of homonymy between genes in different species.
In terms of query expansion, we have followed two different perspectives. The first one deals with the different synonyms that can be found in the literature for a given gene [22–24]. For this, all know synonyms for each gene in the user input are obtained from the data integration database and included in the query.
The second perspective, concept-based query expansion [25, 28], allows to obtain other documents that deal with concepts related to the input genes, like proteins or pathways. Our approach is based on exploiting the direct links between genes and other biological concepts obtained from public biological databases. These networks of associations are implemented through direct relations in the integration database mentioned above (see 'Gene mapping and normalization'). The following entries from this database are used currently: gene names and symbols (obtained from Entrez Gene and UniProt), protein names (from UniProt), pathway names (from KEGG) and diseases (from OMIM).
Query expansion is performed as follows: for each gene in the query, the algorithm obtains, from the term expansion table, all the alternative gene and protein names corresponding to that gene's ID, and the associated pathways and diseases. Since the internal identifier for each gene is used at this stage (see above), only the synonyms and concepts referring to the organism selected by the user are considered. The full list of terms from all input genes is then accumulated in four separate query strings (one for each concept type). Pathway and disease terms occurring for more than one gene in the input are given an increased weight, which is used in constructing the queries. These query strings are then used in document retrieval as described next. The reason for adding the term weights is that we want to give greater relevance to terms which are linked to two or more genes in the input. For example, if three genes in the user input have been linked to the same pathway or disease, abstracts containing that pathway or disease name will have their score increased.
For efficient document retrieval, QuExT uses a local index of the PubMed database, created with Lucene . During indexing, both the document title and abstract are processed using the Lucene standard analyzer. These are tokenized and indexed as one single field, but not stored. Each indexed document is associated with the corresponding PubMed document identifier (PMID). This allows retrieving the corresponding document from PubMed once the query results are obtained, and eliminates the need to store the abstracts locally, therefore greatly reducing the storage requirements of the application.
Document retrieval is performed through Lucene. We run four index searches using the four query strings obtained in the query expansion stage (one for each concept type). For each search, we obtain the documents that match the query and the corresponding Lucene scores. These results are then organized in order to have, for each PubMed ID (PMID), a list of four scores: gene, protein, pathway and disease.
where W j is the weight attributed to the concept type j and s ij represents the score for document i in terms of the jth concept type. This score is obtained from Lucene, as described above. In the example of Figure 1, the document with PMID 2142119 has a score of 5.00 for the concept type 'Genes', 3.47 for the concept type 'Proteins', and 3.00 for the concept type 'Diseases'. According to Equation 1, these scores are multiplied by the corresponding weight and added to produce the final score for this document (5.0 × 0.5 + 3.47 × 0.3 + 0.0 × 0.1 + 3.0 × 0.1 = 3.84). Scores are then normalized in terms of the highest one to obtain a relative score for each document, as shown in Figure 1.
This weighting scheme allows users to alter the order of the retrieved documents by simply changing the concept weights. This allows, for example, focusing on documents describing the metabolic pathways in which the initial genes are involved. From a user perspective, this is achieved by changing the value of the slide bars on the application's user interface. When the new weight values are set, the documents are re-ranked according to Equation 1, without the need to perform the query expansion or the document retrieval stages.
A wide range of web-based information retrieval tools offer optimized or specialized versions of the search functionalities provided by the PubMed database. Each of the available web-based systems has distinct features, and may be preferred by particular users or to address specific problems . Despite that, the tool presented here offers three features that, to our knowledge, are not found in combination in any other tool. First, the automatic expansion of each gene identifier in the initial query to a set of other biologically related concepts. Second, the document retrieval and ranking method was designed to give higher relevance to terms that relate to more than one gene in the input list. This way, documents containing such terms will be ranked higher in the final result. Third, by allowing users to define different weights for the classes of concepts used in the query expansion, the tool allows users to explore the scientific literature from different perspectives, by focusing on particular aspects of the basic gene biology. Together, these characteristics make QuExT a unique tool for addressing the issue of selecting the most relevant papers that combine information relating to a set of genes, such as the outcome of a gene expression study.
Four classes of concepts are currently used for query expansion, but the flexibility of the concept-based expansion and weighting scheme allows the inclusion of more concepts in a straightforward manner. For example, new concepts such as Biological Process terms (from the Gene Ontology), may be included to enrich the expansion. Another possibility is the inclusion of MeSH terms and resource identifiers for the concepts that appear in the document, such as UMLS concept IDs or UniProt accession numbers, in order to categorize the documents and offer links to the primary data sources describing each concept. Likewise, seven reference organisms are supported at present: Homo sapiens, Mus musculus, Rattus norvegicus, Candida albicans, Saccharomyces cerevisiae, Drosophila melanogaster and Apis mellifera. Inclusion of new organisms only requires a straightforward update of the information in the database and can be easily accomplished if there is demand by users.
In order to assess the retrieval performance of the method, we performed an evaluation using data from the ad-hoc retrieval task in the 2004 TREC genomics track . This task is based on the information needs of real biologists, and consists of 50 queries. However, since QuExT is a gene-centred application, not all queries could be used for performing this evaluation: general ones such as "Gene expression profiles for kidney in mice" or "Cause of scleroderma" had to be excluded since these cannot be addressed by QuExT. From the 50 available topics, 18 that were found as being more centred in a particular gene (or set of genes) were selected. Furthermore, since the approach in QuExT is to use as input a gene or a set of genes, rather than a more specified query such as the ones used in TREC, these queries had to be pre-processed to select just the gene names. For example, for the queries "Role of TGFB in angiogenesis in skin" and "Role of p63 and p73 in relation to DNA damage", only the gene name(s) are passed to QuExT ("TGFB" and "p63, p73", respectively). After this step, two queries that explore different roles of the same gene, namely "TGFB", became the same query for QuExT. These two queries, and the corresponding relevant documents, were therefore merged, producing a final set of 17 queries for the evaluation.
The evaluation test produced a mean average precision (MAP) of 0.425, with an average of 4.8 and 31.3 relevant documents within the top 10 and 100 retrieved abstracts, respectively. These results are comparable with the best systems in this task . The best evaluation results were obtained with concept weights set at 100 for gene names/symbols, 10 for protein names and 0 for pathway names (the concept type "disease" was not used in this evaluation). Using pathway names produced a decrease in performance for this set of queries (MAP of 0.405 with pathways weight set at 5). Comparing to executing the search with no query expansion, we obtain an increase of 26% in the MAP statistic (from 0.337 when using just the gene names).
As another measure for comparison, we also ran the same evaluation using PubMed as the retrieval engine. For this, we used the Entrez e-utilities to obtain the first 10.000 documents for each gene query, limited to the date range in the 2004 TREC collection (from 1994 to 2003, inclusive). This experiment yielded a MAP value of 0.239, with 1.5 and 17.7 relevant documents within the top 10 and 100 retrieved abstracts, respectively.
These evaluation results indicate good overall results, especially when comparing to the use of PubMed. Although PubMed uses query expansion through Automatic Term Mapping (ATM), and includes manually annotated information in the search, through MeSH terms , our evaluation results show a 78% increase in the MAP statistic when using QuExT (0.425), as compared to PubMed (0.239). The reduced performance obtained for this task with PubMed (as measured by MAP) is related to the lack of any relevance ranking of the resulting documents, leading to relevant documents appearing lower in the returned list.
Although performance measures based on a manually annotated test set gives some indication of the method's retrieval performance, the 2004 TREC genomics evaluation data and methodology may not be entirely suited for testing our method. First of all, QuExT is oriented to a very specific query type. In contrary, the systems used in a task such as the ad-hoc retrieval in TREC are usually more generic IR tools that can be tuned to match the requirements of the task. Another limitation, which is related to the later, is the fact that not all query types in the TREC task could be used to evaluate QuExT. Since this is a gene-centred application, we had to exclude the majority of the queries, ending up with 17, which were considered to be more focused on a particular gene (or set of genes). These limitations imply that the results obtained do not allow a completely valid comparison with other systems that use the same evaluation, including the ones that participated in the TREC task and other more recent ones. Also, even if the same evaluation set could be used, it is not straightforward to compare QuExT to these tools, as the philosophy and approaches are very dissimilar. Finally, although our evaluation tests with the TREC data show that the inclusion of gene-related concepts helps improving retrieval performance and that optimal results can be obtained with different weight settings, this latter feature is still the most difficult to evaluate in an objective manner. It is important to evaluate this feature from the point of view of users, since it gives them a simple way to influence the ranking of the results in order to obtain the documents that better reflect their information need. In order to do so, we have asked a group of researchers to use the tool and compare it to the use of PubMed. The initial observations obtained indicate that this tool, and specially the concept of using weights to modify the relevance of the results, may be very useful to researchers studying a set of genes. Although this small study gave some good indications about the validity of the approach, the results are very subjective and a more structured evaluation, using a set of well-defined and validated information requests, is required and will be the subject of future research.
An important difficulty in information retrieval systems in biomedicine has to do with matching entity names, such as gene, protein and pathway names, to their occurrence in the texts. At this stage, we rely upon the use of well established terminological resources extracted from reference biomedical databases. Namely, we use the Entrez Gene database for gene names and symbols, UniProt for protein names, KEGG for pathway names, and the OMIM database for disease names. However, these resources do not cover all possible textual variations for a given term. For instance the gene "ATP-binding cassette sub-family A member 1" can appear in different ways in the literature. This includes lexical variation from the full gene name (e.g. word suppression or different word order) as well as morphological variations in gene symbols (ABC1, ABC-1 or ABC 1). Our indexing method uses the standard document analyzer and tokenizer implementations in Lucene, and does not deal with these variations entirely. Our ongoing research includes investigating the use of string similarity algorithms and rule-based strategies during indexing time in order to deal with this limitation and improve the retrieval performance. Namely, we are working on specific tokenization rules to normalize gene/protein symbols to a canonical form in order to address these morphological variations. In this regard, techniques similar to the methods that have been proposed in the literature [22, 23] are being considered.
Another major difficulty is ambiguity, that is, the same symbol identifying different genes and/or proteins. In QuExT, this problem is reduced when performing the mapping and query expansion, since the user is requested to select an organism for the input genes. Furthermore, the inclusion of related terms in the query also addresses this problem, as documents containing more terms associated to the input genes will have a higher ranking. For example, an abstract containing the terms 'CAT' and 'catalase' will have higher relevance than an abstract containing just the term 'CAT'. This aspect can be further explored by introducing reliability (or ambiguity) scores to the terms in our thesaurus.
This paper presents QuExT, a document retrieval and prioritization tool specially designed to obtain from the literature the most relevant articles relating to a given list of genes. QuExT follows a concept-oriented query expansion methodology to find documents containing concepts related to the genes in the user input, such as protein and pathway names. The main innovation of the proposed application is the possibility to modify the weights for the four concept classes used in the query expansion: gene names and symbols, protein names, metabolic pathways and diseases. This gives the user control on how the expanded search terms affect the final ranking of the documents. The user can, for example, select to give more strength to the pathway names in the search query, therefore bringing to the top of the results those documents in which pathway names relevant to the initial gene list appear. Although a more detailed evaluation and comparison with other tools is required, initial evaluation based on data from the TREC genomics track indicates good retrieval performance. The application of concept weights to modify the order of the returned documents represents a significant advantage when compared with the available methods.
Project name: QuExT - Query Expansion Tool
Project home page: http://bioinformatics.ua.pt/quext
Operating system(s): Platform independent
Programming language: Python
Other requirements: n/a
Any restrictions to use by non-academics: no restrictions
The research leading to these results has received funding from the European Community's Seventh Framework Programme (FP7/2007-2013) under grant agreement n° 215847 - the EU-ADR project. J. Arrais is funded by Fundação para a Ciência e Tecnologia (FCT) grant SFRH/BD/23837/2005. S. Matos is funded by FCT under the Ciência2007 programme.
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.