There is an increasing interest in using literature mining techniques to complement information extracted from annotation databases or generated by bioinformatics applications. Here we present PLAN2L, a web-based online search system that integrates text mining and information extraction techniques to access systematically information useful for analyzing genetic, cellular and molecular aspects of the plant model organism Arabidopsis thaliana. Our system facilitates a more efficient retrieval of information relevant to heterogeneous biological topics, from implications in biological relationships at the level of protein interactions and gene regulation, to sub-cellular locations of gene products and associations to cellular and developmental processes, i.e. cell cycle, flowering, root, leaf and seed development. Beyond single entities, also predefined pairs of entities can be provided as queries for which literature-derived relations together with textual evidences are returned. PLAN2L does not require registration and is freely accessible at http://zope.bioinfo.cnio.es/plan2l.

Gene regulatory mechanisms and protein interactions are studied in detail to understand how complex developmental processes are controlled. Biological annotation databases provide functional descriptions of gene products, the basic components of such biological processes, through manual literature inspection, resulting generally in associations of biological entities to a set of controlled vocabulary terms contained in structured database records (1). Despite the obvious strength of controlled vocabularies for annotation consistency, information exchange and data analysis, functional annotations of proteins do not provide a straightforward way to trace back the biological evidence supporting each annotation, making it sometimes cumbersome for human domain experts to directly interpret under which biological context and experimental conditions a given annotation applies. Considering the growing amount of published articles, to manually annotate newly described functional gene product characterizations as well as maintenance and update of already annotated entities is a challenging task. This motivated recent attempts to enable a more systematic access to relevant information hidden in large literature repositories using text mining and information extraction (IE) technologies, with the aim of not only supporting the literature curation process, but especially for providing suitable information retrieval systems useful for life sciences (2). Results generated by text mining systems have the general advantage to be directly interpretable by the end users, the human domain experts, in case direct links to the textual evidences are supported. Currently existing online literature mining systems mostly focus on very particular biological aspects, such as the extraction of protein–protein interactions (3), protein–keyword co-mentions or gene regulation events without directly integrating all these heterogeneous relation types into a single application. Relevance, not only of individual bio-entities but also of their interactions and regulation events for developmental processes studied in model organisms such as the plant Arabidopsis thaliana have not been addressed previously using text mining approaches. Arabidopsis thaliana, the first plant to be completely sequenced, is being used not only for experimental research, but also increasingly by systems biology and bioinformatics approaches to understand and model central biological processes (4). Databases like TAIR [The Arabidopsis Information Resource, (5)] or UniProt (6) are providing plant biologists with valuable infrastructures of manually curated information, but only few attempts have been made to implement literature mining applications for this model organism. The Dragon Plant Biology Explorer (DPBE) was an online text mining application for plant biology based on integration and combination of collections of manually curated vocabularies compiled for several topics to facilitate a more targeted literature search (7). PubSearch constitutes another system for semi-automated retrieval of literature that can be used to curate articles to extract manually annotations for gene products with Gene Ontology terms (8). It is primarily based on simple term matching and does not explore machine learning techniques to provide more sophisticated retrieval capabilities. Another system primarily used for literature curation by model organism databases is Textpresso (9), now integrated at TAIR to facilitate a better access to information relevant for Arabidopsis.

PLAN2L is an online text mining application dedicated to improve retrieval of knowledge by integrating and scoring information extracted from textual sources for various biological topics related with the description interactome associations in Arabidopsis. Figure 1 shows a general flow chart of the PLAN2L system 1 (a more complete description of the technical aspects can be found in the documentation of the web system).

A document retrieval pipeline that takes into account several sources of evidence for the determining whether a given article is associated to A. thaliana was implemented exploiting: (i) external references derived from multiple databases providing annotations for Arabidopsis proteins. (ii) Organism and taxonomic name tagging using dictionary lookup based on a species lexicon derived from the NCBI Taxonomy that was automatically extended using a rule-based approach to account for typographical variants and abbreviations of species names. (iii) Keyword-based retrieval from PubMed and PubMed Central. Additionally, a full text collection of Arabidopsis-related articles was constructed from a local repository of open access full-text articles as well as using customized article collection tools. Plain text conversion was carried out through a combination of systems including pdftotext and PDFlib. The detection of links between the literature and protein or genes of PLAN2L is based on the construction and lookup of a gene lexicon. This gene dictionary integrated A. thaliana gene names and symbols derived from multiple databases, including TAIR, SwissProt and from a collection of gene and protein names identified by a machine learning-named entity recognition program (ABNER) as well as based a rule-based approach considering morphological cues and name length to identify potential Arabidopsis gene symbols. Lexicon expansion using manually crafted rules was carried out. To detect gene regulatory relations, we adapted an IE architecture relying on a pipeline of semantic/syntactic rules. We applied part-of-speech tagging of each word using a GENIA-trained version of Treetager (10). Some of the POS tags were automatically substituted with more semantically oriented labels (e.g. organisms, protein/gene names and activation verbs). The text with mixed syntactic and semantic tags was fed into a SCOL parser (11), which generated a tree-like structure by applying a modified CASS grammar. These rules constitute cascades of finite-state automata, and use patterns that combine both grammatical- and biological-meaning features in the linguistic structure. We implemented extensions of the rules to handle frequent phrase coordination and prepositional anaphora. The extraction of protein interaction evidence associations was addressed using a machine learning sentence classifier approach relying on manually selected interaction evidence sentences (12). The used sentence classifier relies on a Support Vector Machines algorithm trained on set of manually classified interaction evidence passages derived from a collection used at the second BioCreative challenge (12), and obtained a performance of 89.75 for precision and 92.62 for recall using a radial basis kernel function on a balanced test set. For retrieving protein localization descriptions, we explored both the use of semantic–syntactic frames for extracting a fine-grained association between proteins and subcellular location mentions together with a machine learning sentence classifier for retrieving protein localization description sentences in general. The initial step followed, consisted in the construction of a sub-cellular location dictionary that integrates location keywords and synonyms derived from SwissProt together with cellular component terms from Gene Ontology. Location term mentioning sentences were manually revised to derive hand-crafted location frames. Additionally, a location sentence classifier was constructed using a collection of 2264 protein location description sentences. A central component of PLAN2L is the scoring of each evidence sentence according to its relevance for complex temporal biological events (topics), at the cellular level (cell cycle) as well as at the level of developmental processes. We therefore implemented a classifier for scoring cell-cycle relevant abstracts and document passages. The SVM text classifier was trained on a collection of cell-cycle relevant abstracts and nonrelevant abstracts and then applied to a literature collection of abstracts and full-text articles mentioning A. thaliana genes. Additionally, four specific sentence classifiers for the most relevant developmental processes in higher plants, namely (i) flowering, (ii) leaf development, (iii) root development and (iv) seed development/germination have been developed. The tool provides a comprehensive approach to assist in the selection and ranking of genes, proteins, documents and terms relevant to a specific biological process for this model organism. Additional details on the different modules, their characteristics and assessment are provided at the PLAN2L web.

The PLAN2L interface handles user-provided plain text keywords, protein/gene names or symbols. Some of the components allow additionally searching with gene or protein identifiers. The currently supported identifiers include TAIR gene identifiers and UniProt primary accession numbers. Based on a user survey that we carried out to get feedback from biologists on the PLAN2L system, aspects that were positively rated included its easy to use and intuitive query interface and the direct retrieval of evidence sentences for multiple topics. Aspects that were improved according to the user comments covered additional documentation on the system and the sentence scoring mechanism. PLAN2L supports six types of search strategies, each with its own query page, to avoid introducing unnecessary complexity through advanced search interfaces with complicated menu options. We will briefly describe each of these six search types and provide a case study to illustrate the type of results generated by PLAN2L.

• The basic search type allows retrieval of multiple biological topic associations for a given user query, consisting in ranked lists of evidence sentences. The default ranking is based on the association strength of the article for the organism source (i.e. A. thaliana), sorted in descending order. The structured result table page provides the document identifier (PMID) from which the evidence sentence has been derived, the extracted sentence text showing the color highlighted user query term as well as the relevance of the sentence for multiple biological topics: cell cycle, gene regulation, protein interaction, cellular location, flowering, leaf development, root development and seed development. Additional external links to other resources are also included, namely the BioCreative metaserver, iHOP, WikiGene and TAIR. In the results page, color-coded scores are given for each category represented visually by a unique informative glyph. Scores above classifier cutoff are displayed in green, otherwise in red. The evidence sentences can be dynamically reranked for each of the categories, and to facilitate rapid visual cueing of the association between the scores and their class, a background-coloring schema was used.
• PLAN2L also facilitates retrieval of physical interaction associations for a given query protein or protein identifier. Each of the co-mentioned proteins is returned together with its corresponding identifier, interaction evidence sentence and experimental interaction detection method keywords. Ranking of the interaction association evidence sentences is based on the corresponding interaction sentence classifier score.
• In case of regulatory associations, the typical user-specified query consists in a protein/gene name or in which the system returns a tabular regulation evidence summary. To facilitate the interpretation of the directionality of a given regulatory event, a color-coding strategy was used to highlight the regulator and regulated gene, together with a key sentence display. To qualify the type of regulation, PLAN2L labels each regulation evidence as activation, repression or undefined (for unclear cases).
• The sub-cellular location evidence sentences for a given protein can also be retrieved by PLAN2L. Complementing the location sentence scoring of the basic search option, more accurate location sentence retrieval, integrating also location term co-occurrence and location association words is available through PLAN2L.
• As A. thaliana is one of the most important model organisms for studying the plant cell-cycle, PLAN2L integrates a cell-cycle relevance ranking of PubMed abstracts using a document classification tool. The Arabidopsis literature can be searched using gene names or keywords and the resulting abstracts are ranked according to their cell-cycle relevance (cell-cycle score from the document classifier). Also cell-cycle relevant terms are tagged. The species ambiguity score of abstracts can be used to determine how specific a document is for Arabidopsis, something useful when several species are mentioned in the same article.
• In order to enable more effective searches of association evidences existing between two biological entities of interest, PLAN2L provide the association retrieval interface, where for two entities entered by the user, the application returns a set of sentence where both entities are co-mentioned. Each evidence sentence is scored according to the biological topics supported for the basic search query. The default ranking is based on the interaction sentence classifier scores. The typical time taken for text processing is 2–3 min, but in case of heavy server load and the actual query type, some jobs might take between 4 and 5 min, upon completion of the search task PLAN2L displays its results together with hyperlinks to the resources providing links to the original article.

We described a text mining system, PLAN2L that enables exploration of literature information at different levels of granularity, from retrieval of gene description sentences derived from multiple documents, to qualified biological relations important to understand the sub-cellular context and both physical as well as regulatory interaction networks of bio-entities. PLAN2L extracts biological information from both abstracts as well as full-text articles and integrates different language processing strategies from simple co-occurrence to syntactic/semantic rule-based algorithms and supervised machine learning methods. PLAN2L is intended to be useful for general retrieval, topic-specific retrieval as well as for finding association evidences for user-specified entities and for knowledge and hypothesis confirmation. A similar strategy as used for PLAN2L can be easily adapted to other model organisms as well as specific biological topics with minimal additional manual data preparation.