Quantitative Structure-Activity Relationship (QSAR) modeling is a ligand-based approach to quantitatively correlate chemical structure with a response, such as biological activity or chemical reactivity. The process is widely adopted and has for example been used to model carcinogenecity 12, toxicity 34, and solubility 56. Further, the literature is replete with QSAR studies covering problems in lead optimization 7, fragrance design, and detection of doping in sports 8. In QSAR, chemical structures are expressed as descriptors, which are numerical representations such as calculated properties or enumerated fragments. Descriptors and response values are concatenated into a dataset, and statistical methods are commonly used to build predictive models of these.

There exist many examples of investigations regarding the resulting statistical models with respect to validity and applicability in QSAR and similar fields 910. However, most of these investigations consider the dataset as fixed, and the choice of descriptors and implementations is left outside the analysis.

Part of the problem is the lack of a controlled vocabulary regarding descriptors; there is no easy way of defining what descriptors were used, which the underlying algorithms were, and how these were implemented. It is common to use several different software packages with results manually glued together in spreadsheets, sometimes with custom in-house calculated descriptors. The lack of a unifying standard and an exchange format means that QSAR datasets are published in articles without clear rules, usually as data matrices of precalculated descriptors, with chemical structures in a separate file.

The field of bioinformatics has acknowledged the standardization problem to a much larger extent than cheminformatics. Numerous standards, ontologies, and exchange formats have been proposed and agreed upon in various domains. The Minimum Information standards are examples that specify the minimum amount of meta data and data required to meet a specific aim. The MGED consortium pioneered this in bioinformatics with Minimum Information About Microarray Experiements (MIAME) 11, and it has now become a requirement that data from microarray experiments must be deposited in MIAME-compliant public repositories in the MAGE-ML exchange format 12, in order to be published in most journals. Standardization initiatives in cheminformatics are not as common, even though the problem of incompatible file formats and standards has been frequently discussed 13. Grammatica 14 has addressed the issue of QSAR model validation and notes that descriptor versioning as well as precisely defined algorithmic specifications are vital for developing QSAR models that can be considered reliable, robust, and reproducible (in addition to the usual issues of statistical rigor).

Initiatives that work towards standardizing cheminformatis in general include the Blue Obelisk, an internet group which promotes open data, open source, and open standards in cheminformatics 15, which has proposed dictionaries for algorithms and implementations suitable for QSAR. Distributed Structure-Searchable Toxicity (DSSTox) Database Network has proposed standardized structure-data files (SDF) as a file format for exchanging raw data in toxicological SAR analyses 16. This approach does however not include any information regarding descriptors, and SDF is a legacy text format which has many variants. OECD has established rules and formats for how to report QSAR models and QSAR predictions 17, but its intended use is communication, not complete technical coverage. It also lacks an ontology, which makes interpretation and reasoning around results much more complicated and subjective. Public repositories of QSAR datasets are limited to a few internet resources (e.g. 18 and 19) where they are usually not deposited but reproduced from articles by others than the original authors, and due to the lack of an established exchange format and missing raw data, structures are sometimes redrawn, data manually copied from articles, and educated guesses are made in some cases. QSAR DataBank 20 is a proposal for the electronic organization and archiving of QSAR model information. It is an interesting initiative that builds on other standards, but also lacks an ontology for descriptors. The OpenTox project is another project developing a framework to share QSAR datasets using REST services 21.

In general it is not uncommon that information about what software package that was used for descriptor calculation (and its version) is unavailable, and that custom descriptors have been added manually or results preprocessed. To further complicate matters, many QSAR software packages are proprietary, closed source, and it is a non-trivial task (sometimes impossible) to get insights into how algorithms are implemented. Due to these impracticalities, journals are limited to establishing simple rules for QSAR publications such as to state that structures should be publicly available 22.

A well-defined standard with a corresponding exchange format will have problems getting accepted in the scientific community if user-friendly tools supporting them are not available. This paper introduces a file format for exchanging QSAR datasets, together with tools implemented in the graphical workbench Bioclipse 2324 to facilitate working with QSAR according to the standard.

The QSAR-ML exchange format together with the Blue Obelisk Descriptor Ontology has many implications. To the best of our knowledge it is the first initiative which encompasses completely reproducible definition of QSAR datasets, including descriptor definitions and implementations. QSAR-ML is equipped with built-in properties to ensure integrity and consistency of included resources. For example are molecular resources appended with generated InChI, which can be used to verify the integrity of the chemical structures such as accidental changes or errors when transmitting data over networks. That descriptors in QSAR-ML are defined in Blue Obelisk Descriptor Ontology means that they have a formalized and clear meaning, and are uniquely referenced. Defining descriptor implementations by software name, version and identifier, and connecting this information with an entry in the descriptor ontology, uniquely defines, and makes it possible to accurately reproduce descriptor calculations. Open standards, a defined terminology, and reproducible results allows people to have trust in publicly available datasets and reconstruction of such datasets, and hence improves the reliability of the subsequent statistical analysis. Much research has been done on various aspects of QSAR modeling, such as validation, robustness, and domain applicability of models 14. This is not covered here as it is a research topic of its own, but we stress that the handling of the original chemical structures as well as the choice and implementation of descriptors are of great importance 37. This is a neglected topic, and QSAR-ML sets new standards for the field. A reproducible dataset setup enables validation not only of the resulting datasets, but allows for inclusion of e.g. chemical variability and descriptor selections with respect to model robustness and performance inside a cross validation loop 10. There is a large amount of descriptors available, and people continuously improve existing descriptors and develop new ones. An exchange format capable of harmonizing this requires an extensible architecture in order to be successful, and also intuitive tools that make this easily available for scientists. QSAR-ML, implemented as an XML Schema, and BODO, implemented in the Web Ontology Language 38, fulfills this demand of extensibility. We would like to point out that this is a proposal for an open standard and that we welcome suggestions to improve the specification further.

The Bioclipse-QSAR feature turns Bioclipse into a workbench which greatly simplifies the setup of QSAR datasets, with full support for the QSAR-ML exchange format. Bioclipse also supports many other features which are common in QSAR projects, such as conversion, editing, and visualization of chemical structures. Rich clients are software applications that take full advantage of today's modern desktop computers, but also leverages on the new e-Science tools such as online (Web) services. The Bioclipse-QSAR is a formidable example of this; Descriptors can be calculated on the local computer while, if connected to a network, remote services can provide additional descriptors or offer high performance computers for speeding up demanding calculations.

There would be great rewards if QSAR-ML is widely adopted by the scientific community. For example, users could download entire QSAR datasets and reuse it together with in-house data, extend existing models, join different models, search for overlap between datasets, collaborate, reproduce, and validate results. Further, QSAR-ML enables the establishment of public repositories of QSAR datasets. We envision that deposition of QSAR models in such repositories will become a standard operation procedure prior to future publication of QSAR models and results, similar to microarray experiments in bioinformatics, and that QSAR-ML is a strong candidate for such a format.

We describe a new exchange format for QSAR datasets, named QSAR-ML, which relies on the Blue Obelisk Descriptor Ontology for uniquely defining descriptors, and supports any implementation of these. QSAR-ML comprises all data and metadata required to reproduce the setup of QSAR datasets, enabling validation of chemical structures and descriptor calculations. Sharing QSAR datasets in an open, standardized format has profound implications for collaboration and information validity and reuse. We also describe a QSAR plugin for Bioclipse with full support for QSAR-ML, which greatly simplifies setting up QSAR datasets using graphical user interfaces. The implementation integrates with other cheminformatics component that are valuable in dataset preparation, such as database searching as well as editing and visualization of chemical structures.

Future plans include addition of subsequent statistical analysis into the QSAR-ML format and hence not only support dataset setup but also model fitting and prediction. We also aim at setting up a public repository with means for publishing QSAR-ML datasets, which is a first step towards public repositories for sharing QSAR on a global level, and could provide the basis for supplemental data in future QSAR publications.

The authors declare that they have no competing interests.

OS and ME designed the QSAR-ML. EW and RG designed and implemented the BODO. OS implemented the Bioclipse-QSAR plugins. EW implemented the JOELib plugin. JW supervised the project. All authors read and approved the final manuscript.

The authors would like to thank all the people who have contributed to the Blue Obelisk Descriptor Ontology and the Bioclipse project, as well as Anders Lövgren at the computing department at Uppsala Biomedical Center (BMC) for hosting the CDK REST and XMPP services.

This work was supported by the Swedish VR (04X-05957) and Uppsala University (KoF 07).