A prominent characteristic of modern chemistry is the increasing automation of processes such as synthetic combinatorial chemistry, or of data capture by instruments associated with high-throughput screening. These developments presage the era of intelligent robotic chemistry where components of the system are programmed to make decisions from structured information and to issue instructions to other components using structured messages. An example of this might be a robotic chemical synthesiser automatically identifying and scanning a newly published journal article for interesting molecules reported there, and either ordering them for testing or deciding to synthesise them from existing starting materials.

Part of this vision stems from that of the "Semantic Web" proposed by Berners-Lee, where he conceives of simple robots or agents which can accurately gather and act upon high-quality data found on the Web or other networks. Web-based resources ("URLs") would contain "metadata" describing the resource, e.g. the authorship, ownership and access rights of the resource. Metadata can also be used to make assertions about other forms of metadata such as "resource X conforms to protocol Y", or "I confirm that resource Z was authored by organisation Q". Part of the necessary process to define and identify what we mean by the term "high-quality" data is to establish a so-called "web of trust", which would be based on certifiable metadata and information objects. In other words, can we, or the agents that act on our behalf, trust the authorship and content of metadata statements?

Chemistry as a discipline is well suited to this sort of reasoning. Chemical entities are well understood and described by nomenclature that is in the public domain. It is meaningful to make a series of related assertions including, but not limited to, the following;

1. "Agent H identifies material K from an article in a journal authored by F and published by G"
2. "Agent H works/acts for organisation E"
3. "Agent H decides that material K, which corresponds to a molecular connection table X, will be required"
4. "To be useful, K must have 1% impurity"
5. "Acetylsalicylic acid matches connection table X"
6. "Authority M states that aspirin is a common name for acetylsalicylic acid"
7. "Aspirin is entry A-312 in the catalogue from supplier J"
8. "Supplier J asserts that compound A-312 is 99% pure"
9. "Y represents the NMR spectrum of compound A-312 in chloroform as solvent"
10. "Computer program Z analyses Y as having 1% impurity"
11. "Program Z was written by organisation Q"
12. "Q is on E's list of approved NMR software suppliers"
13. "J is on E's list of approved compound suppliers"
14. "M is on E's list of approved chemical metadata suppliers"
15. "G is on E's list of approved chemical publishers"
16. "F is on H's list of known/trusted authors"

From this, a semantic chain can be constructed which automatically leads to Agent H proposing an order of compound A-312 from supplier J. Whilst in most chemical organisations Agent H is currently likely to be a research scientist, delegation of such purchasing decisions to a robot agent is already happening in other sectors of business-to-business (B2B) electronic commerce. We argue that such a "chemical web of trust" is a realistic vision, but is mainly dependent on the ability of suppliers to use a uniform web-based approach to data and metadata. In this context, we recently proposed one such universal approach to Web-based Chemistry using XML (eXtensible Markup Language), and a specific implementation of XML termed CML (Chemical Markup Language). Here we describe a mechanism for the certification of chemical data and metadata which could be used to implement this vision.

We require the following components to authenticate information:

1. A list of authorities whom we are prepared to accept as making assertions we can act upon. These assertions may be about "facts" or may authenticate assertions made by authorities that we do not as yet accept. For example if we assert "The absolute conformation of (-)-labetalol is (R,R)" the reader should not accept this without confirmation. However much greater confidence is generated by accepting it from the assertions:
   1. In Acta Cryst (1984), C40, 825, the authors state that the absolute configuration of (-)-labetalol hydrochloride is (R,R).
   2. The International Union of Crystallography requires that the reporting of absolute configurations by X-ray crystallography in Acta Cryst conforms to its guidelines.
   3. The editorial office of Acta Cryst use algorithms to check the data consistency of every published structure.
   4. The editorial office of Acta Cryst have the power to reject publications of absolute configuration which do not meet its guidelines
   The reader may now feel able to accept the assertion with a high degree of confidence (e.g. 95%).
2. The ability for each authority to certify their assertions. The assertions made above cannot be automatically validated without referring back to (printed) material, which itself provides the authentication. When assertions are made in electronic form it is essential that their origin is verifiable and that their content is shown to be uncorrupted.

Thus we see authors and publishers making authenticable assertions about the validity of their facts and assertions. It is important to distinguish between these concepts:

• Authenticity is the ability to verify that a document/assertion has been created by the authority to whom it is attributed and that it is uncorrupted after its creation.
• Validity is the ability to show that a specified validation process has been correctly carried out.
• Certification is the association of a specified validation or authentication procedure with an identified authority (normally a named person or organisation) or an automated process certified by a specific authority. Certification is accomplished by incorporating a digital signature into the document (signing).

In general, "facts" (primary data) are not validatable other than verifying their authenticity from an authority. However secondary data derived from primary data by a process (often algorithmic), can be validated. Thus "I mounted crystal X on an X-ray diffractometer and collected diffraction data Y" is not normally validatable electronically. However "From diffraction data Y, I used program Z to determine the crystal structure of A as B" is validatable. It is possible to support or refute this claim by running the same or equivalent computation.

We indicate below some of the electronic components that will need authentication and/or validation:

1. Chemical structures, with both two and three dimensional co-ordinates. Validation includes the following hierarchy:
   1. Syntactic conformance. We have described earlier how Markup Languages can be used to ensure syntactic validity (i.e. that the components of a chemical structure obey a given CML DTD or Schema). Schemas have greater power than DTDs and we shall describe later how they enhance validation of CML document instances.
   2. Self-consistency. CML-based representations of molecules contain some required internal self-consistency (e.g. that the atomRefs in bonds must reference valid atoms).
   3. Regularisation, normalisation and canonicalisation. We assume algorithms to determine "chemical consistency" such as valence checking, aromaticity and tautomerisation. There is no single approach to this and a variety of protocols will exist.
   These all involve agorithmic "checking of validity" through certified procedures. For example, XML-based validation requires an authentic schema and authenticated output from a validating XML parser. We assume that the appropriate DTD or Schema is authenticated before use and that the output includes this authentication. If the output itself is authenticated, it represents a trustable object for use in the semantic chemical web.
2. Validation of computations. Much chemical information now originates in the output of algorithmic-based modelling procedures, which may themselves include access to databases. Such programs will need to be authenticatable and their output certified.
3. Instrumentation. As with modelling programs, the parameters defining instrumental procedures need to be declared and the output authenticated.

For information to be easily authenticable and certifiable, it must be structured into precisely defined information components or objects, and these definitions must be openly declared and accessible. One such approach has been developed, called XML (eXtensible Markup Language). Documents marked-up according to XML specifications have several characteristic features. They comprise data contained within tags known as elements, which may themselves be hierarchical. For example in the CML (Chemical Markup Language) specification, a molecule element may contain many atom sub-elements or children. Elements can have associated attributes and attribute values and such documents are said to be 'well-formed' if the syntax for expressing the elements, their attributes and their values all follow generic XML specifications.

A precise specification can be created of the allowed elements, their characteristics and the boundaries of the attribute values. This is referred to as a Document Type Description (DTD), and it can itself be expressed as an XML-conforming document called a schema. Any XML document conforming precisely to a given schema is said to be "valid". Because the structure of the document is precisely specified, this allows transformations of the data contained within it, using eXtensible Style Sheets (XSL) or other tools. An illustration of this process as applied to chemistry is the ChiMeraL Project, where CML based documents are used to express molecules and their properties, and where the data can be visualised using conventional Web browsers and appropriate XSL transformations.

As noted above, XML-based validation will require an authentic schema. For example, CML version 1.0 is defined by a published DTD and only validation against this can certify that documents or components of documents contain authentic CML 1.0

We propose here that authenticity can be ensured by a procedure known as XML signing, which is based on the generation and use of key-pairs contained in so-called X.509 certificates. A description of X.509 certificates and their use for establishing the authenticity of Java applets has been previously described by us in some detail. In the next section, we describe how X.509 certificates can be used to create authenticated XML documents or document components, and how such procedures can in turn be used to create certified documents and components contained within them.

Recent legislation in the UK and elsewhere has granted digital documents similar legal status to printed ones. It is self evident that someone receiving such documents must be able to establish that they are uncorrupted or unaltered since they were created by their originator (i.e. they are authentic), and that their source is itself authentic and verifiable. The first document format to become accepted in this context is Adobe Acrobat PDF (Portable Document Format). For example, virtually all electronic journals now offer articles in this format, and other important applications of PDF include patent submissions and compound safety sheets. An Acrobat PDF file can be digitally signed using X.509 certificates which can serve to ensure that the documents, like the Java applets noted above, are authentic. However, an important limitation of the Acrobat 4.0 format is that the digital signature can only be applied to the entire PDF document. Although the Acrobat 5.0 specification allows such documents to be further annotated by others, it is not possible in general to associate specific components of the documents with multiple and/or different certificates.

We propose that these and other limitations of Acrobat-format files can be overcome by using XML documents containing XML signatures. As examples we will use two sources of information. The first is an article describing the ChiMeraL project and available in XML form as supplemental information associated with the journal. This article contains not merely components authored by different people, but has well defined data components originating from different sources, and which we therefore use as a model for thechemical publication of the future. The trust associated with such an article originates at least in part from the original authors. In this example this is individually signed by the authors, and hence is not an automated procedure. The second example derives from the use of an Web-based agent or robot which has been programmed to automatically traverse a collection of documents from a remote site and on the basis of pre-defined procedures and associated algorithms, to express the documnets in valid XML form and to create specific metadata for the collection. The procedures include automatically validating any specific occurrences of e.g. an MDL Molfile and converting these files to valid CML. These components are then automatically signed by the robot agent as certifying that the metadata and converted CML files were produced by this procedure and no other. The authors of the robot and the procedures it implements have in effect granted the robot a proxy to use their certificates.

The signing procedures above were applied manually to a single document and its components by the authors of that document. An alternative approach is to delegate the signing process to an agent. This agent can perform a task pre-programmed by its author/creator. In our case, we illustrate this by using the ChemDig, JChemTidy and JChemMeta agents described previously. This involves using a tree traversing robot (ChemDig) to retrieve the HTML content of a remote site and to perform three operations on each document located in the tree:

1. Each HTML document is normalised to conform to XHTML syntax (JChemTidy), producing a well formed and valid XML document.
2. Any chemical data files transcluded into the original HTML document are then parsed, and key chemical meta-information (JChemMeta) is either extracted from the original file or derived using suitable software (e.g. a canonical SMILES string is produced using a specified algorithm). This meta-information can then be expressed into the XHTML document via an appropriately defined schema.
3. Finally, selected types of transcluded chemical files such as the MDL Molfile are converted to a CML representation according to published Molfile specifications using a conversion tool written by us.

The next stage involves establishing a level of trust that the conversions and transformations of the chemical data indeed involve authentic processes as described by us. To achieve this, we have now added two further steps in the sequence indicated above:

1. The derived meta-information as contained in the two elements meta.../meta and link.../link and the converted XHTML is signed as described in .
2. Each of the XML/CML files converted from a Molfile is similarly signed, resulting in a collection of XHTML and XML documents each of which can be authenticated against a formal digital signature. This authentication signifies that a specific Molfile to CML conversion process has been conducted, and that meta-information has been added according to our specifications. These components, if added to a larger database of chemical information would automatically carry their provenance and authenticity.

To illustrate the complete process, we have taken another recently published article. The supplemental information for this article contains a collection of HTML documents, together with 13 MDL Molfiles which have been linked (transcluded) to the main document. We have now charged our ChemDig agent/robot to produce from this a well formed and valid XHTML document from each original HTML file, to add specific chemical metadata to the XHTML, and to transform the 13 original Molfiles to valid CML form. This process creates in effect a set of assertions about the operations conducted on the original information. It does of course, not validate the original chemical claims made. The resulting collection of documents is deposited as supplemental information with the present article..

The techniques described in this article will allow an entire document expressed in XML, or individual components of such a document, or both, to be authenticated against a unique digital signature encapsulated in what is referred to as an X.509 certificate. This represents a considerable advance over the use of Acrobat PDF files for producing authenticatable documents. The various ways of implementing signatures allow for a diverse set of applications of this technology. Most importantly, the signature can be used to authenticate the specific schema and hence XML language used to encapsulate the information in the document, in the form of a signature attached to the document schema itself. In this manner, the recipient can be assured that the structure of the document is valid.

Chemical information and data could be assembled from a variety of sources, including instruments, computational procedures, databases, e-journals and e-books, and the authenticity of each source individually signed as being authentic. Because each signature is individually a well formed and valid component of the overall XML document, it can be retained as an audit trail of the source of the information it refers to. As such, the signature might authenticate that component, or it may indicate that the information has been deliberately or otherwise either altered or superseded. Signatures can be also regarded as envelopes. Thus small components of a CML document which contain one or more molecules and associated properties can be wrapped in a large envelope containing in effect a database of such molecules. At the other extreme, individual properties such as e.g a melting point could be signed as authentic if deemed necessary.

The signatures can be included within the document itself if the nature of the document is such that it is not intended to be further edited. This would represent an archive of the information. Such documents are extensively used in areas such as patent submissions and drug submissions. If the document is intended to be extensively revised, then the signatures can be deposited in an external database, with URI links to the components they refer to. This approach might be useful for establishing a clear audit trail of the various information components in an organisation, or of the process representing the eventual publication of a journal article. The signing procedures could be extended to two other areas, namely access control lists for individual document components, and encryption of these components to ensure that only the intended recipient can make use of them.

The preceding discussions makes it apparent that a context or role must be placed on the signature, which would include certifying the authenticity of a document, its XML/CML syntactic validity and conformance and most challengingly its chemical validity. Similar roles can in fact be declared when an Acrobat PDF file is signed, but only in the context of the entire document and not its components. Declaration of signature roles in XML is currently less well developed, and we have not applied it here. The chemical community faces a major challenge in making available communally agreed procedures for evaluating chemical validity, an example of which might be validating a declaration that one compound is a tautomer of another. We hope that in the future, globally available resources for establishing chemical validity will become available.

We suggest that the concept of information component authentication and delivery to intended recipients represents a radical departure from current methods of document transmission and processing. We anticipate the techniques describe here will have widespread importance in the industrial chemical and publishing industries, and could ultimately serve as an infra-structure on which to base the creation of a chemical and semantic web of trust associated with on-line information.

One of us (GVG) thanks Merck Sharpe and Dohm and the EPSRC for the award of a studentship.