Ranked retrieval of Computational Biology models

Abstract

Background: The study of biological systems demands computational support. If targeting a biological problem, the reuse of existing computational models can save time and effort. Deciding for potentially suitable models, however, becomes more challenging with the increasing number of computational models available, and even more when considering the models' growing complexity. Firstly, among a set of potential model candidates it is difficult to decide for the model that best suits ones needs. Secondly, it is hard to grasp the nature of an unknown model listed in a search result set, and to judge how well it fits for the particular problem one has in mind.

Results: Here we present an improved search approach for computational models of biological processes. It is based on existing retrieval and ranking methods from Information Retrieval. The approach incorporates annotations suggested by MIRIAM, and additional meta-information. It is now part of the search engine of BioModels Database, a standard repository for computational models.

Conclusions: The introduced concept and implementation are, to our knowledge, the first application of Information Retrieval techniques on model search in Computational Systems Biology. Using the example of BioModels Database, it was shown that the approach is feasible and extends the current possibilities to search for relevant models. The advantages of our system over existing solutions are that we incorporate a rich set of meta-information, and that we provide the user with a relevance ranking of the models found for a query. Better search capabilities in model databases are expected to have a positive effect on the reuse of existing models.

Figure 1

Growing number of Computational Biology models and their components in BioModels Database. Upper chart: Numbers of models in BioModels Database, as of January 2010. Lower chart: Number of species (black bar) and reactions (gray bar) in BioModels Database, as of January 2010. BioModels Database started with 20 models and a total of 322 species when it was launched in April 2005. In 2007 it already reached almost 200 models and 10482 species. The release in January 2010 recorded 453 models with 33702 species and 41069 reactions.

Figure 2

Conceptual architecture. Overview of the conceptual architecture of the proposed ranking- and retrieval system. A version has been implemented in BioModels Database. The architecture shows the process of transforming a user given query by creating sub-queries, which are then assembled by enrichment of structural information and semantic indexing (see also Figure 3). The re-assembled query is then sent to the retrieval and ranking module, which makes use of the Extended Boolean Model to retrieve a list of matching models, and the Vector Space Model to rank the list of retrieved models. To determine the ranking, different weight information is used. Those are, however, not shown in the given Figure.

Figure 3

Sample query on the new BioModels Database search interface. Screenshot of a part of the new search interface of BioModels Database. The interface allows to search for Persons, SBML elements, Resources, and allows to restrict the search terms to particular features using a single qualifier. Models may only be considered for a certain range of dates. The sample search correspond to recent models by non-bogus authors describing the effect of caffeine in human's digestive tract when drinking coffee.

Figure 4

Ranked results. Search result obtained on BioModels Database with the given sample query (see Figure 3). The upper panel shows the enriched query. Due to the precise formulation of the query, and the requirement that caffeine must occur and additionally must be qualified with is, the result contains only three hits. (1) This model matches the top two constituents resolved by the semantic index, and additionally the term gut in the compartment feature. (2) The model matches the constituent ranked third by the semantic index. (3) The lowest ranked model only matches one constituent ranked eight by the semantic index - this is a very weak relation resulting in a very low rank.