Chapter 5.5. The use of mmCIF architecture for PDB data management

For the past 30 years, the PDB has served as the single central repository for macromolecular structure data. The data format used to store archival entries in the PDB is a column-oriented data format resembling many data formats developed to accommodate the limitations of paper punched-card technology (see Chapter 1.1 ). An example of the data format is shown in Fig. 5.5.2.1.

Figure 5.5.2.1 | top | pdf | Excerpt of records from a PDB data file.

Many of the data records in this format are prefixed with a record tag (e.g. CRYST1, ATOM) followed by individual items of data. The specifications for the records in this data format are described informally by Callaway et al. (1996). In addition to the labelled records as in Fig. 5.5.2.1, many data records in the PDB format are presented as unstructured or only semi-structured remark records.

In 1998, the Research Collaboratory for Structural Bioinformatics (RCSB) assumed the management responsibilities for the PDB. One important outcome was the change in the underlying data representation used to process PDB data. The PDB now collects and processes data using a data representation based on a comprehensive ontology of macromolecular structure and experiment: the PDB exchange data dictionary. This representation is an extension of the mmCIF data dictionary, now the standard data representation for experimentally determined three-dimensional macromolecular structures. The dictionary and data files based on this data ontology (Westbrook & Bourne, 2000) are expressed using Self-defining Text Archival and Retrieval (STAR) syntax (Chapter 2.1 ).

Although the mmCIF dictionary was developed within the crystallographic community, the metadata model employed by mmCIF is quite general and has been adopted by other application domains including NMR, molecular modelling and molecular recognition (dictionaries are available at http://mmcif.pdb.org/ ). Within the crystallographic community, metadata dictionaries have also been developed for other types of diffraction experiments, electron-microscopy data and for the general description of image data. The metadata concepts and tools that have been developed to support mmCIF are sufficiently general that they may be applied to the description of data in virtually any application.

The demands of structural genomics projects have driven the development of extensions to capture an increased level of experimental detail. These are available at http://mmcif.pdb.org/ . Extensions have also been introduced to describe NMR, cryo-electron microscopy and all aspects of protein production. The ability to rapidly add extensions and incorporate these into the PDB data-processing system is an important feature for supporting the rapidly evolving technologies associated with high-throughput structure determinations.

The mmCIF metadata architecture is built from three levels as illustrated in Fig. 5.5.2.2 (see also Chapter 2.6 ). Individual data files are described at the top level (e.g. Fig. 5.5.2.2a). The contents of these data files are defined by a data dictionary (e.g. Fig. 5.5.2.2b) in the next lower level (see Chapters 3.6 and 4.5 ). The attributes used in this data dictionary to build data definitions are in turn defined in the dictionary description language (DDL) (e.g. Fig. 5.5.2.2c) in the lowest level (see Chapters 2.6 and 4.10 ).

Figure 5.5.2.2 | top | pdf | Files at different levels of the mmCIF metadata architecture. (a) mmCIF data file excerpt. (b) Example mmCIF data dictionary definition. (c) Example DDL dictionary attribute definition.

The major syntactical constructs used by mmCIF are illustrated in the data file example of Fig. 5.5.2.2(a). Each data item or group of data items is preceded by an identifying keyword. Groups of related data items are organized into data categories. Two categories, CELL and ENTITY_POLY_SEQ, are shown in the example. CELL contains an individual instance describing a single set of crystallographic cell constants. ENTITY_POLY_SEQ contains a loop_ (i.e. table) of instances describing a polymer residue sequence. Essentially all mmCIF data are described as a set of tabular data structures.

Each mmCIF data item is defined in a data dictionary. Data definitions are given between save-frame delimiters (i.e. save_); apart from this, the data definitions share the same simple syntax as used in data files. An example definition for a crystallographic cell constant is shown in Fig. 5.5.2.2(b). Many features of the cell constant are described in this definition, including data type, range restrictions, units of expression, dependent quantities, related definitions, necessity and related precision estimate. Although not shown in this example, dictionary definitions can also include parent–child relationships that have important consequences in maintaining data consistency.

The attributes of each data definition are defined in the DDL dictionary. Fig. 5.5.2.2(c) shows example DDL definitions describing data types. DDL definitions have the same syntax as definitions used in the data dictionary. Because the attributes of the DDL are also used in DDL definitions, this metadata architecture is described as self-defining.

The RCSB PDB distributes parsing tools that support all three levels of this metadata architecture (http://sw-tools.pdb.org/ ). The CIFPARSE_OBJ package (Tosic & Westbrook, 2000) provides high-level methods to read, write, validate and manage data from data files, dictionaries and DDLs. Data files can be validated relative to an input data dictionary, and dictionary files can be validated relative to an input DDL. CIFPARSE_OBJ stores information in a collection of table objects. Access methods are provided to search and manipulate the table objects. A companion package, CIFOBJ (Schirripa & Westbrook, 1996), provides an alternative representation of dictionary and DDL data. CIFOBJ organizes dictionary information into a collection of category and item-level objects. Access methods are provided for all dictionary attributes.

One of the greatest benefits of a dictionary-based informatics infrastructure is the flexibility that it provides in supporting alternative data formats and delivery methods. Because the data and all of their defining attributes are electronically encoded, translation between data and dictionary formats can be achieved using light-weight software filters without loss of any information.

XML provides a particularly good example of the ease with which data can be converted to and from the mmCIF format. XML translations of mmCIF data files are currently provided on the Worldwide PDB ftp site (ftp://ftp.wwpdb.org/pub/pdb/data/structures/divided/XML/ ). These XML files use mmCIF dictionary data-item names as XML tags. These files were created by a translation tool (http://sw-tools.pdb.org/apps/MMCIF-XML-UTIL/ ) that translates mmCIF data files to XML in compliance with an XML schema. The XML schema is similarly software-translated from the PDB exchange data dictionary.

Other delivery methods such as Corba (http://www.omg.org/cgi-bin/doc?lifesci/00–02-02 ) do not require a data format, as data are exchanged using an application program interface (API). A Corba API for macromolecular structure (Greer et al., 2002) based on the content of the mmCIF data dictionary has been approved by the Object Management Group (OMG). Software tools supporting this Corba API (OpenMMS, http://openmms.sdsc.edu , and FILM, http://sw-tools.pdb.org/apps/FILM ) take full advantage of the data dictionary in building the interface definitions and supporting server on which the API is based (see also Section 5.3.8.2 ).

The basic functions of the ADIT deposition system are shown in Fig. 5.5.3.2. Users interact with the ADIT system through a web server. The CGI components of the ADIT system (that is, functional software components interacting with web input data through the Common Gateway Interface protocol) dynamically build the HTML that provides the system user interface. These CGI components are currently implemented as compiled binaries from C++ source code.

Figure 5.5.3.2 | top | pdf | Schematic diagram of ADIT editing, format translation and validation functions.

User data can be provided in the form of data files or as keyboard input. Input files can be accepted in a variety of formats. ADIT uses a collection of format filters to convert input data to the data specification defined in a persistent data dictionary. Data in the form of data files are typically loaded first. Any input data that are not included in uploaded files can be keyed in by the user. ADIT builds a set of HTML forms for each category of data to be input. At any point during an input session, a user may choose to view or deposit the input data. Users who are depositing data may also use the data-validation services through the ADIT interface.

Comprehensive data ontologies like the PDB exchange dictionary contain vast numbers of data definitions. A data-input application may only need to access a small fraction of these definitions at any point. To address the problem of selecting only the relevant set of input data items from a data dictionary ADIT uses a view database. In addition to defining the scope of the data items to be edited by the ADIT application, an ADIT data view also stores presentation details that are used in building the HTML input forms. An important use of the data view is to provide a simple and intuitive presentation of information for novice users which disguises the complex details of a data dictionary.

Fig. 5.5.3.3 shows an example ADIT editing screen for the crystallographic unit cell. The data dictionary category containing this information is named CELL, and the length of the first cell axis is defined in the dictionary as _cell.length_a (Fig. 5.5.2.2b). In this case, the data view has substituted Unit Cell and Length a for the dictionary data names. Although this example is simple, some dictionary data names are as long as 75 characters, and in these instances the ability to display a simpler name is essential.

Figure 5.5.3.3 | top | pdf | Example ADIT data-input screen.

Precise dictionary definitions and examples obtained from the data dictionary are accessible from the ADIT interface through buttons next to each data item. ADIT makes full use of the dictionary specification in data-input operations. Data items defined to assume only specific values have pulldown menus or selection boxes. Data type and range restrictions are checked when data are input and diagnostics are displayed to the user if errors are detected.

For performance reasons, the data dictionary is converted from its tabular text structure to an object representation using CIFOBJ. The class supporting the object representation provides efficient access functions to all of the data dictionary attributes. A dictionary loader is used to check the consistency of the data dictionary and to load the object representation from the text form of the data dictionary.

Any dictionary that complies with the dictionary description language (DDL2) can be loaded and used by ADIT. All ADIT software components gain their knowledge of the input data from the data dictionary and any associated data views. Consequently, ADIT can be tailored for use in virtually any data-input and data-processing application.

In addition to the data editing and processing functions, ADIT also supports a versatile database loader (mmCIF Loader; http://sw-tools.pdb.org/apps/MMCIF-LOADER ) that builds data­base schemata and extracts the processed data required to load database instances. The relation of the database loader to the central components of the ADIT system is shown in Fig. 5.5.3.4.

Figure 5.5.3.4 | top | pdf | Schematic diagram of ADIT database loading functions.

Schemata are defined in a metadata repository that is accessed by the loader application. In the simplest case, a schema can be constructed that is modelled directly from the data dictionary. Since the data model underlying the dictionary description language used to build ADIT data dictionaries is essentially relational, mapping a data dictionary specification to a relational schema is straightforward.

In other cases, a mapping is required between the target schema and the data dictionary specification. This mapping is encoded in the schema metadata repository. The database loader uses this mapping information to extract items from data files and translate these data into a form that can be loaded into the target database schema. The definition of the mapping operation can include: selection operations with equijoin constraints (e.g. the value of _entity.type where _entity.id = 1), aggregation (e.g. count, sum, average), collapse (e.g. vector to string), type conversions and existence tests.

Schema definitions are converted by the database loader into SQL instructions that create the defined tables and indices. Loadable data are produced either as SQL insert/update instructions or in the more efficient table copy formats used by popular database engines (i.e. DB2, Sybase, Oracle and MySQL). Loadable data can also be produced in XML.

One goal of high-throughput structural genomics is the automatic capture of all the details of each step in the process of structure determination. Fig. 5.5.3.5 shows a simplified structure-determination data pipeline. The essential details of each pipeline step are extracted and later assembled to make a data file for PDB deposition. The RCSB PDB data-processing infrastructure has been developed in anticipation of a data pipeline in which automated deposition would be the terminal step. The dictionary technology and software tools developed by the RCSB PDB to process and manage mmCIF data can be reused to provide the data-handling operations required to build the pipeline.

Figure 5.5.3.5 | top | pdf | Schematic diagram of a structure-determination data pipeline.

Dictionary definitions have been carefully developed to describe the details of each step in the structure-determination pipeline. These data items are typically accessible in electronic form after each program step. The information is either exported directly in mmCIF format or is printed in a program output file. To deal with the latter case, a utility program, PDB_EXTRACT (http://sw-tools.pdb.org/apps/PDB_EXTRACT ), has been developed to parse program output files and extract key data values. In either case, the results of this incremental extraction of data from each program step must be merged to build a complete mmCIF data file ready for deposition. The PDB_EXTRACT program also carrys out this merging operation.

Some steps in the structure-determination pipeline may not be driven by software. For instance, the details of protein production may be held in laboratory databases or within laboratory notebooks. A version of ADIT with a data view including all of the structural genomics data extensions has been created for entering these data. This ADIT tool can also be used to validate and check the completeness of the final data file.