Methods for the automatic composition of services into executable workflows need detailed knowledge about the application domain, in particular about the available services and their behavior in terms of input/output data descriptions. In this paper we discuss how the EMBRACE data and methods ontology (EDAM) can be used as background knowledge for the composition of bioinformatics workflows. We show by means of a small example domain that the EDAM knowledge facilitates finding possible workflows, but that additional knowledge is required to guide the search towards actually adequate solutions. We illustrate how the ability to flexibly formulate domain-specific and problem-specific constraints supports the workflow development process.
The challenge of automatic workflow composition, in particular the automatic
composition of bioinformatics web services into executable workflows, has been
addressed by several projects in the past years (see, e.g., [
In this paper we show how the domain knowledge that EDAM provides can be used as basis for the automation of service composition. After an introduction to the principles of automatic workflow composition and to the software framework that we used for this study (Section 2) we describe the EDAM ontology (Section 3), especially focussing on the parts that are relevant for automatic workflow composition. In Section 4 (Results and Discussion) we present an example domain and workflow composition problem, showing that the EDAM knowledge facilitates finding possible workflows, but that additional knowledge is required to limit the search to the actually desired or adequate solutions. We use the PROPHETS synthesis framework, which facilitates a very flexible way of expressing additional knowledge: it can either be specified during domain modeling (especially suitable for domain-specific constraints) or during the actual synthesis (especially suitable for problem-specific constraints). In this way, users can flexibly interact with the workflow development framework, collecting possible solutions and continuously refining the constraints according to any arising requirements. The paper ends with a conclusion in Section 5. 2
Automatic Workflow Composition
The automatic composition of (small) software units into (large) runnable pieces
of software has been subject to research for several years (see, e.g. [
The commonly known methods can roughly be distinguished into
synthesisbased and planning-based approaches: Synthesis methods usually are rather
behavior-oriented (e.g., the temporal-logic-based methods of [
Accordingly, the quality of the solutions crucially depends on the quality of the available information about the services of the domain (more than on the actual search strategy): Merely syntactic information (as provided by most programming language APIs or standard web service interfaces) is not sufficient. Proper semantic descriptions of services and data types are required to obtain meaningful results. For instance, assume a web service that returns a job ID, and another service that consumes a nucleotide sequence. From a programming language point of view, both would be classified as character sequences (strings) and would thus be assumed to match. While this is syntactically correct, submitting a job ID to a sequence analysis service would of course not lead to the desired results, so a more precise - semantic - interface description is required. Automatic Workflow Composition in Bio-jETI For the work presented in this paper, we used the PROPHETS extension of the Bio-jETI framework [21]. PROPHETS seamlessly integrates automatic service composition methodology into Bio-jETI, in particular by supporting visualized/graphical semantic domain modeling, loose specification within the workflow model, non-formal specification of constraints using natural language templates, and automatic generation of model checking formulas (to check global properties of workflows). For a detailed introduction into the framework and the underlying ideas the reader is referred to [22, 23]. In a nutshell, working with PROPHETS incorporates two phases: 1. During domain modeling meta-information about the available services is collected and provided in appropriate form, and semantic classifications of the services and their input and output types are defined. The service and type taxonomies are stored in OWL format, where the OWL classes represent abstract classifications, and the actual types and services are then represented as individuals that are related to one or more of those classifications by instance-of relations. 2. In the actual workflow design phase this domain is used to (automatically) create the workflow models. Central is the idea of loose specification: branches in the model can be marked as loosely specified, which are then automatically replaced by adequate composite services.
The algorithm [24] that is then applied to complete a loosely specified workflow to be fully executable takes two orthogonal aspects into account: On the one hand, the workflow must be type-consistent to be executable, on the other hand, the constraints specified by the workflow designer must be met: – The configuration universe constitutes the algorithm’s basic search space.
It contains all valid execution sequences and is implicitly defined by the domain model. As this configuration universe is usually very large, it is not explicitly generated from the domain definition, but on the fly during the synthesis process. – The specification formula is the second aspect. It describes all sequences of services that meet the individual workflow specification, but without taking care of actual executability concerns. As the explicit representation of all these sequences would be extremely large, the formula is not explicitly built, but given as a declarative formula in SLTL (Semantic Linear Time Logic) [24], a variant of the commonly known propositional linear-time logic (PLTL). The formula is a conjunction of all available constraints, comprising: • technical constraints (configurations of the synthesis algorithm), • domain-specific constraints (globally specified by the domain modeler), and • problem-specific constraints (derived from the loosely specified branch or specified by the workflow designer).
To start the search for solutions, the synthesis algorithm requires an initial state (i.e. a set of start types). They are determined automatically according to the output specification of the service at the beginning of the loose specification. Based on this the synthesis algorithm performs a parallel evaluation of the configuration universe and the specification formula to search for paths that are consistent with both the configuration universe and the SLTL formula. Each of those paths is a valid concretization of the loosely specified branch. 3
EDAM as Background Knowledge
The EDAM (EMBRACE Data And Methods) ontology3 has been developed
in the scope of the EMBRACE (European Model for Bioinformatics Research
and Community Education) project4 as an ontology for describing life science
web services [
Strictly speaking, EDAM is not a single, large ontology, but consists of six separate (sub-) ontologies: 1. Biological entity: physically existing (parts of) things, such as SNPs, alleles, or protein domains. 2. Topic: fields of bioinformatics study, such as nucleic acid sequence analysis, model organisms, or visualization and rendering. 3. Operation: particular functions of tools or services (e.g., web service operations), such as annotation, codon usage analysis, or sequence database search by motif or pattern. 4. Data resource: various kinds of data resources, such as Biological pathways data resource, Literature data resource, or Ontology data resource. 5. Data: semantic descriptions of data entities that are commonly used in bioinformatics, such as BioCyc enzyme ID, phylogenetic consensus tree, or sequence alignment. 6. Data format: references to (syntactic) data format specifications, such as Clustal sequence format, PubMed Central article format, or HMMER hidden Markov model format. 3 http://edamontology.sourceforge.net/ 4 http://www.embracegrid.info/
Importantly, EDAM is not a catalogue of concrete services, data, resources etc., but a provider of terms for classification of such entities – exactly what is meant by “background knowledge”. Thus, using EDAM is simply taking the (relevant parts of the) ontology and sorting the application-specific resources into this skeletal domain structure. In the following we use the Operation (sub-) ontology as vocabulary for service classifications, and the Data (sub-) ontology for data type classifications. Employing also the Data format part in order to take into account the more technical interface specifications is subject of ongoing work. (For this study we used EDAM version beta08.) 4
Results and Discussion Using EDAM as background knowledge in the domain model for synthesis with PROPHETS, the domain setup involves three major steps: 1. Converting EDAM from OBO (Open Biomedical Ontologies) format into
OWL format (using the Prot´eg´e OWL API). 2. Generating the service taxonomy from the Operation term and (transitively) all its subclasses, and the type taxonomy from the Data term and (transitively) all its subclasses. 3. Sorting the available services and their input/output types into the service and type taxonomy, respectively.
Steps 1 and 2 can be executed fully automatically. Step 3 can be automated if EDAM annotations are available for the types and services, as it is the case, for instance, for EMBOSS [28], which provides EDAM relations in its Ajax Command Definition (ACD) files. In the following we will use a smaller example domain that comprises a set of bioinformatics tools for which the taxonomic classifications have been defined manually. It uses only a small part of EDAM, but it is already large enough to illustrate our results.
Table 1 lists the services in the example domain, along with their input and output data types. Concrete service and data type names are given in normal font, while (abstract) services and types in terms of the EDAM ontology are given in italics. Note that the service interface descriptions in this domain are quite simple: each one has at most one input and one output, whereby only user data (i.e. input and output files) are considered, while other parameters (primarily used for configurations that are not actually inputs) are not taken into account.
Figures 1 and 2 show the service and type taxonomies (screenshots from PROPHETS’ built-in ontology editor). The classes (the blue squares) correspond to the Operations and Data (sub-) ontologies of EDAM, respectively, but have been cut down to the classes that are relevant for the services and data that are used in the domain model. The purple rhombs represent the concrete services and types of the domain that have been added to the domain model as instances of the respective EDAM classes. Fig. 1. Service taxonomy.
Fig. 2. Type taxonomy.
Figure 3 shows the simple loosely specified workflow that we used as basis for experimentation with the described domain model, along with five possible concretizations. The workflow begins with reading a DNA sequence file (ReadDNASequence) and ends with displaying a result (Viewer). These components are connected by a loosely specified branch (colored in red). Accordingly, the basic synthesis problem is to find a workflow that takes a ReadDNASequence’s output (a DNA sequence) as input, and produces Viewer’s input (some data) as output. Note that this setup is in particular appropriate for experimentation with the domain, exploring what workflows are generally possible with certain input data. In many realistic applications, the target of the loosely specified branch is a more specific component, for instance an alignment viewer, so that the synthesis problem as such becomes more specific, too.
The five concretizations shown in Figure 3 are of course not the only possible ones. As we will detail in the following, thousands if not millions of solutions are easily possible with the described domain model, but they are not necessarily desired or adequate. We will show in the following how playing with configurations and constraints may help mastering this enormous potential by excluding inadequate or bundling equivalent solutions.
In a first experiment, we executed the synthesis with no further constraints. That is, only the input/output specification defined by the loosely specified branch was considered and a “naive” search was performed on the synthesis universe. When we used the basic configuration of the synthesis algorithm, we obtained 264,118 possible solutions for the synthesis problem already for a search depth of 5. Additionally using a solution filtering mechanism that removes service sequences that are mere permutations of others decreased the number of solutions to 5,325. As the example domain is characterized by consisting of services that have at most one input and one output data type, it turned out to be most efficient, however, to use a configuration in which the synthesis also mimics a “pipelining” behavior. This means that output data is only transferred to the direct successor of a service in the workflow and in particular not available for any other subsequent services. This configuration led to 2269 solutions in the unconstrained case.
In the following we will develop a small set of constraints that is already sufficient to drastically reduce the solution space further by excluding clearly inadequate solutions. The constraints are based on a few general observations about the solutions obtained so far: 1. Workflows that contain services that make no contribution to solving the synthesis problem. Such services are, for example, ReadFile (requiring no input but producing arbitrary new data that is planted into the workflow and can distract from the actual synthesis problem) or WriteFile (consuming data but without producing a new analysis result). 2. Workflows that contain services that make no progress within the workflow. Particular examples are redundant service calls, such as a the multiple invocation of Gblocks on a multiple sequence alignment where only the first call does make sense. 3. Workflows that contain “dead” functionality in the sense that certain service outputs are not adequately used. For instance, a BLAST result is usually not used as such. Rather, the sequence identifiers of the BLAST hits are of interest in order to retrieve the corresponding entries from a database. Thus, to make sense, workflows that call BLAST should also contain a call to a BLAST parser that gets the IDs from the BLAST result. 4. Workflows that contain several services that are useful when seen individually or in parts of the solution, but the overall workflow is not the envisaged analysis. For example, a solution where an alignment service (such as ClustalW) is called with the input data is certainly a possible and useful workflow, but if the workflow developers’ intention was actually to do the alignment with homologous sequences of the input sequences, or to obtain a phylogenetic tree from the input, this solution is not adequate.
In order to address the first observation (“services that make no contribution”) we defined a constraint that excludes all corresponding services of our example domain (i.e. ReadFile, ReadDNASequence, WriteFile, and Viewer) from the synthesis. This can easily be realized via the SLTL formula
G(¬ S true) which prohibits any occurrence of S in a solution. Since the exclusion of particular services is a quite common constraint, it is provided as a template in the PROPHETS Synthesis Wizard, which only requires the user to list all ‘unwanted’ services. Applying this template to all four services named above yields a constraint that in itself reduced the number of solutions from 2,269 to 55.
Adding an SLTL constraint for the second observation (“services that make no progress”) in a similar fashion, specifying that Gblocks should not never be called twice (or more), decreases the number of solutions 49. This number can be further reduced by an SLTL constraint exploiting the third observation (“dead functionality”): requiring that a call of WUBlast should be followed by a call to the corresponding parser eliminates another 18 solutions.
In order to address the last observation (“workflow is not the envisaged analysis”), we provided the synthesis with constraints that express general ideas about the desired workflow. As one example, a frequently performed analysis for molecular sequences is to search for similar sequences in a database and compare the new sequences to the known ones by means of a sequence alignment. Accordingly, we formulated a constraint that enforces a sequence database search by sequence and after that a multiple sequence alignment as part of the solution. This decreased the number of solutions further down to 24. The remaining solutions are similar to the two workflows at the bottom of Figure 3: they start with WUBlast and the WUBlastParser, followed by either DBFetch FetchBatch or DBFetch FetchData, after which one of the alignment algorithms is called. Half of the solutions appended Gblocks as final service.
Another analysis that is often applied to molecular sequences is the construction of a phylogenetic tree, which represents the evolutionary relationship between the sequences. As an alternative to the above constraint for the fourth observation, we formulated a constraint that enforces the use of a phylogenetic tree construction service. The only remaining result when this constraint is applied together with constraints 1, 2 and 3 is a one-step solution consisting of the PhyML DNA service, which computes a phylogenetic tree from a set of DNA sequences.
Table 2 summarizes our findings, containing also all other combinations of the constraints developed above: The first constraint (excluding services that make no contribution towards solving the synthesis problem) has the strongest impact on the number of solutions, whereas the second constraint leads only to an exclusion of around 80 solutions. Notably, already constraints 1 and 4 together decrease the number of solutions to 24, i.e. to the set of solutions that was also obtained with the first four constraints together. Likewise, constraints 1 and 4’ together already lead to the single-step solution that was initially obtained by a combination of four constraints.
As the table furthermore shows, some combinations of constraints, in fact most combinations involving 4 and 4’, leave no solutions at all. This may indicate that the constraints are inconsistent (i.e. do not have any solution), or, as in this case, that they require solutions of length greater than the current search depth of 5. In the latter case, increasing the search depth may be the solution of choice, but there are two options which apply in both cases: – loosening the constraints, in order to resolve the inconsistencies or to allow shorter solutions, and – revise/extend the domain model in order to increase the solution space. Whereas playing with the first option may well be in the competence of the workflow designer, the second option requires domain modeling expertise.
The second column of the table gives the number of nodes that are visited by the synthesis algorithm during its iterative deepening depth-first search. The numbers reflect that the synthesis search space is not only constituted by the (static) service descriptions, but also by the additionally provided logical constraints (cf. Section 2). In fact, the search space may grow with the product of the sizes of the formula and the constraints. As the table shows, in our case constraints can both decrease and increase the search space: In the unconstrained case, 34,026 nodes are visited, the least number of nodes (1,103) is visited when constraints 1 and 2 are used, and with constraints 2, 3, 4, and 4’ a total of 1,764,843 nodes are visited until a search depth of 5.
Note that not all constraints we defined here are useful in every setting. For instance, services that generate data “from scratch” may well be useful when there are other services involved that need this additional input. Thus, in cases where producing entirely new data is considered a useful or even necessary feature, the first constraint needs to be relaxed. Thus changing the considered problem/purpose may in addition to adapting the purpose-specific constraints (here 4 and 4’) also require to reconsider all the other constraint classes.
Conclusion Methods for the automatic composition of services into executable workflows need detailed knowledge about the application domain. In this paper we discussed how the EMBRACE data and methods ontology (EDAM) can be used as background knowledge for the composition of bioinformatics workflows. We found that the EDAM knowledge facilitates finding possible workflows, but that additional knowledge is required to limit the search to the actually desired/adequate solutions: – EDAM provides a controlled vocabulary for data and methods in the bioinformatics domain, whereby it covers in particular the service and data type description terminology that is needed for the automatic composition of services into workflows. Given that the services in the domain and their input/output types are properly annotated in terms of EDAM, this knowledge ensures that the synthesis algorithms find adequate workflows. – However, finding the actually desired workflows requires more knowledge. In this paper we provided simple examples of additional constraints, such as exclusion of particular services, dependencies between services, and general patterns for a desired solution.
Larger service collections like EMBOSS [28] or the BioCatalogue [29], which provide hundreds or even thousands of services, are not manageable without systematic discovery or service composition techniques. When dealing with small domains, human experts may be unbeatable in composing tailored workflows, but firstly not every human is an expert in bioinformatics services and data types, and secondly also experts cannot always keep track of all changes in large domain libraries. Thus both experts and average users may profit from tools that automatically exploit arising domain-specific and problem-specific knowledge beyond the usual “static” domain model (i.e. service interface descriptions).
The PROPHETS synthesis framework enables a very flexible way of
expressing additional knowledge: it can either be specified during domain modeling
(especially suitable for domain-specific constraints) or during the actual
synthesis (especially suitable for problem-specific constraints). Thus, the expression of
(additional) domain knowledge is, in particular, cleanly separated from the
implementation of the synthesis algorithm. This is in contrast to other approaches
to automatic composition of bioinformatics services that we are aware of (such
as [
Currently we are exploring the bioinformatics domain further in order to identify general domain-specific constraints, and problem-specific constraints especially in the shape of often recurring workflow patterns. Such a library of constraints that can be added and removed dynamically during the workflow development process will enable users to work and experiment with the domain in a very flexible manner by tailoring their solution space on demand. This kind of experimentation provides users with an easy entry into this complex landscape of tools and technologies, and later with means for scalability: experimenting with options and constraints it is possible to tailor the setting in a way that at the same time improves the adequacy of the specification as well as the search depths for solutions of the synthesis procedure. We plan to support this approach by extending the domain knowledge with domain-specific search heuristics enhancing the synthesis algorithms. 15. Steffen, B., Margaria, T., Beeck, M.: Automatic synthesis of linear process models from temporal constraints: An incremental approach. In ACM/SIGPLAN Int.
Workshop on Automated Analysis of Software (AAS’97) (1997) 16. Kupfermann, O., Vardi, M.Y.: mu-calculus Synthesis. In: Mathematical Foundations of Computer Science 2000. (2000) 17. Mayer, M.C., Orlandini, A., Balestreri, G., Limongelli, C.: A Planner Fully Based on Linear Time Logic. In: AIPS. (2000) 347–354 18. Margaria, T., Steffen, B.: LTL Guided Planning: Revisiting Automatic Tool Composition in ETI. In: Proceedings of the 31st IEEE Software Engineering Workshop, IEEE Computer Society (2007) 214–226 19. Russel, S., Norvig, P.: Artificial Intelligence: A Modern Approach. 3 edn. Prentice
Hall (12 2009) 20. Bacchus, F., Kabanza, F.: Using temporal logics to express search control knowledge for planning. Artif. Intell. 116(1-2) (2000) 123–191 21. Margaria, T., Kubczak, C., Steffen, B.: Bio-jETI: a service integration, design, and provisioning platform for orchestrated bioinformatics processes. BMC Bioinformatics 9 Suppl 4 (2008) S12 PMID: 18460173 PMCID: 2367639. 22. Lamprecht, A.L., Naujokat, S., Margaria, T., Steffen, B.: Synthesis-Based Loose Programming. In: Proceedings of the 7th International Conference on the Quality of Information and Communications Technology (QUATIC). (September 2010) 23. Lamprecht, A.L., Naujokat, S., Margaria, T., Steffen, B.: Semantics-based composition of EMBOSS services. BMC Bioinformatics (2010) to appear. 24. Steffen, B., Margaria, T., Freitag, B.: Module Configuration by Minimal Model Construction. Technical report, Fakulta¨t fu¨r Mathematik und Informatik, Universit¨at Passau (1993) 25. Ashburner, M., Ball, C.A., Blake, J.A., Botstein, D., Butler, H., Cherry, J.M., Davis, A.P., Dolinski, K., Dwight, S.S., Eppig, J.T., Harris, M.A., Hill, D.P., IsselTarver, L., Kasarskis, A., Lewis, S., Matese, J.C., Richardson, J.E., Ringwald, M., Rubin, G.M., Sherlock, G.: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nature Genetics 25(1) (May 2000) 25–9 PMID: 10802651. 26. Smith, B., Ashburner, M., Rosse, C., Bard, J., Bug, W., Ceusters, W., Goldberg, L.J., Eilbeck, K., Ireland, A., Mungall, C.J., Leontis, N., Rocca-Serra, P., Ruttenberg, A., Sansone, S., Scheuermann, R.H., Shah, N., Whetzel, P.L., Lewis, S.: The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration. Nat Biotech 25(11) (November 2007) 1251–1255 27. Farrell, J., Lausen, H.: Semantic Annotations for WSDL and XML Schema.
http://www.w3.org/TR/sawsdl/ (8 2007) W3C Recommendation. 28. Rice, P., Longden, I., Bleasby, A.: EMBOSS: the European Molecular Biology Open Software Suite. Trends in Genetics: TIG 16(6) (June 2000) 276–7 PMID: 10827456. 29. Bhagat, J., Tanoh, F., Nzuobontane, E., Laurent, T., Orlowski, J., Roos, M., Wolstencroft, K., Aleksejevs, S., Stevens, R., Pettifer, S., Lopez, R., Goble, C.A.: BioCatalogue: a universal catalogue of web services for the life sciences. Nucl. Acids Res. 38(suppl 2) (July 2010) W689–694