Research in the Life Sciences depends on the integration of large, distributed and heterogeneous data sources and web services. The discovery of which of these resources are the most appropriate to solve a given task is a complex research question, since there is a large amount of plausible candidates and there is little, mostly unstructured, metadata to be able to decide among them. We contribute a semi-automatic approach, based on semantic techniques, to assist researchers in the discovery of the most appropriate web services to fulfill a set of given requirements.
Introduction Contemporary research in the Life Sciences depends on the sophisticated integration of large amounts of data obtained by in-house experiments with reference databases available on the web. This is followed by analysis workflows that rely on highly specific algorithms, often available as web services. The amount of data produced and consumed by this process is prodigious; however, the sheer amount of available resources is a source of severe difficulties.
Within this huge set of resources, one of the main problems to the user is
finding the right web services for a given research task. The landscape of Life
Sciences web services is large and complex: there are thousands of resources [
Given this context, it is a pressing question how to help researchers to discover the best possible mapping between their requirements and the available tools. We present a semi-automatic approach to assist the researcher in web service discovery, looking for web services that are appropriate to fulfill the information requirements in the Life Sciences domain. The whole process is driven by wellcaptured requirements, in order to avoid the high costs associated with nondisciplined, non-reusable, ad-hoc development of integration applications. The matching between the requirements and web services is based on a semantic normalization of both the requirements and the web services metadata.
Approach overview
The overall approach we propose to assist the selection of web services based
on users requirements consists of three main phases: (i) Requirements elicitation
and specification, (ii) Normalization, and (iii) Web service selection phase. Here,
we focus mainly on the normalization and the web service selection phases.
1. Requirements elicitation and specification. The user’s information
requirements are the information that drives the discovery of web services.
Therefore, the requirements are gathered and formally described in the
requirements model using the i* formalism [
sim(C, T ) = maxS∈lex(C)(ratio(S, T )) ratio(S, T ) = idf (cw(S, T )) − missing(S, T )
idf (S) missing(S, T ) = (idf (S) − idf (cw(S, T ))) The function ratio(S, T ) defines the ratio between the achieved information evidence for the text T , in our case the task description, and the information encoded in the lexicon form S. The function missing(S, T ) is the amount of information contained in S that have not been covered with T . idf (S) measures the relevance of the terms in the string S, and cw(S, T ) is the set of terms in common between the concept string S and the text T . The function idf is defined as follows: idf (S) = −
log(P (w|U M LS))
w∈S
The relevance of word is measured by means of its estimated
probability within the whole UMLS lexicon (i.e. P (w|U M LS)). After applying
the semantic annotator, each task is represented with a semantic vector
with the tf ∗ idf values of each concept. For example, the task
“Analyze domains in protein sequences” is represented with the semantic
vector {’C1513868’:8, ’D9000419’:15} where C1513868 is retrieved by
“domains” and D9000419 is retrieved by “protein sequences”.
2.2. Application specific annotations. The next step is to determine the
functionality the task is describing. For that purpose, we use the
taxonomy of categories defined by BioCatalogue to classify the user-defined
tasks. The matching between the tasks and the categories is made by
using the ISub metric [
The discovery process is made by two independent searches: (i) search the web services that are annotated with the same category as the task, and (ii) search the web services that have concepts in common with the user-defined task. The results of both searches are combined and scored based on the following linear combination:
score = C-score ∗ w1 + S-score ∗ w2 where w1 and w2 are weights that depend on the relevance of each search and fulfill the condition w1 + w2 = 1, C-score is the score calculated by the ISub metric, that is, the score of the string matching between the categories of the taxonomy and the user-defined task, and S-score is the cosine similarity between the vector of the task and the vector of the service.
Table 1 shows the highest ranked services for the task “Analyze domains in protein sequences”.
At the end, the user obtains a set of ranked lists of web services that are supposed to provide the functionality required by the tasks. In case the results are not those expected by the user, she can refine the process at the three phases of the guide. 3
We have presented a semi-automatic approach that guides researchers in the Life Sciences to the discovery of web services that respond to their informational requirements. Due to the importance of the semantic normalization, we have annotated the available information of the services registered in BioCatalogue by using a biomedical semantic annotator. In turn, the user requirements are also annotated with the same ontologies, thus allowing the application of a semantic search technique to find mappings between requirements and services. The result of the process is that the user is provided with a set of ranked lists of web services that are appropriate for her stated information needs.
Some direct follow-ups of this work are the refinement of some details of the semantic techniques, the exploitation of other sources of metadata (bibliographical information, referenced web pages), and the creation of a GUI to facilitate its application.
This research has been supported by the Spanish Ministry of Education and Science (grant TIN2008-01825/TIN) and by Universitat Jaume I – Fundaci´o Bancaixa (grant P11B2008-43). Mar´ıa P´erez has been supported by Universitat Jaume I predoctoral grant PREDOC/2007/41.