1.1. Databases 101

Why do we need structure databases? In simple terms, a database is software that permits you to organize the data to efficiently answer a specific set of questions of that data. For example, in finding all structures in the PDB that belong to the protein kinase family it does not make sense to search the more than 3 Gbytes of text information in the current PDB looking for the text string `protein kinase'. It is more efficient to have previously partitioned the data, so for example, you only search the equivalent of all PDB COMPND records searching for references to the text string `protein kinase.' However, once you have found an instance of `protein kinase', how do you reference other features of that structure, for example, the resolution? In any database associated with the instance of `protein kinase' is a unique identifier for that protein. That same identifier is associated with other features of the same protein, including the resolution and is used to locate that feature. Similarly, by grouping the resolutions of all protein structures together makes returning all structures of better than, or less than, a certain resolution very efficient. There are three major types of database that implement this generic idea of partitioning data and all have been used for storing and querying macromolecular structure data. They are introduced in increasing order of complexity. The indexed file uses the basic file structure of the computer operating system along with an index as a unique identifier to partition the data. The relational database partitions data into tables where a tuple is a row of the table having a unique identifier (the primary key) for one or more attributes. So the PDB identifier could be a primary key and the compound name an attribute. The object-oriented database partitions data into classes where each tuple is an object instance from that class with an object identifier.

Each of these three database types has positive and negative features. Object-oriented databases can contain complex objects (e.g., the surface of the active-site region of a protein) and have methods associated with each object instance. The latter implies that a query can also perform a calculation on data without being passed to an external program. Conversely, object-oriented databases are relatively inefficient and difficult to develop. Relational databases are more efficient but more limiting in the type of data they can contain. Indexed files are simple to implement, but have none of the characteristics associated with a formal relational database, such as concurrency control, enhanced security and formal query language. It is beyond the scope of this paper to discuss these issues in detail. Interested readers can refer to Bourne et al. (1998) for a discussion aimed at crystallographers wishing to access data from multiple databases, and Date (1994) for a detailed treatment of databases.

3.1. Parsing PDB entries

A major task in loading all PDB structures into any database is consistently parsing the complete PDB. Problems in parsing may not be apparent to a user of a single PDB file interested in only the atomic coordinates, but there are major problems when trying to consistently parse the complete corpus and extract information from the so-called `header records', the information preceding the atomic coordinates. Morris et al. (1992) highlighted such problems and significant efforts have been made by the PDB and others to rectify these problems, but many problems remain. Problems such as undocumented inconsistencies between SEQRES records and the sequence found on ATOM records, the failure to use IUPAC notation, syntax errors, and inconsistent atom and group labeling. Note that these errors are distinct from the stereochemical quality of the structure but refer to the ancillary information so important for classifying structures according to name, taxonomy and experimental detail.

The problem with the current corpus begins with the format definition itself. The PDB format is defined in the Contents Guide provided by the PDB. However, the corpus contains entries in at least four distinct formats v1.0, v2.0, v2.1, and v2.2, with only those entries released from December 1996 indicating the format to which they comply through use of a REMARK 4 record. Early versions of the Contents Guide to which many structures conform are no longer available. Further, there is significant undocumented variation within a given format. For example, using PDB+ to analyze the change in the number of polypeptide chains from one version of a structure to another produces a number of surprises. While the application or removal of non-crystallographic restraints satisfactorily explained most of these surprises, several anomalies remained. Close inspection revealed that the definition of what constituted a polypeptide chain had changed from one version of a structure to another and single residues or non-polymer chemical components were assigned, or had removed, chain identifiers. This is obvious to an experienced human inspecting the entry, but not obvious to a computer program. While such a parser can be written to handle anomalies, it becomes a major software maintenance effort when the input format continues to change. Such anomalies are to be expected in an evolving discipline and are not unique to the representation of structure data, but are more frequent given the complexity of the information being represented. The use of a robust structure deposition tool (e.g., Autodep) is a vital step in improving the consistency of entries. Further, current efforts by several groups to `clean-up' existing PDB data by both removing anomalies and improving the level of annotation are most welcome, provided they lead to a single community-accessible version of each structure.

3.2. Content of PDB+ and lack of experimental detail

The analysis of PDB+ summarized subsequently raises a number of questions that remain unanswered because of missing or unparsable information relating to the experiment performed in determining the structure. Consider the case of resolution. Resolution is often used as the single yardstick for defining the accuracy of a structure and hence how that structure is used in a modeling experiment such as drug deign. However, there is variation in how the overall resolution is defined by different depositors. In principle 100% of the data in a given resolution range should be available for resolution to be reported at the limit of that range. This is often not the case. Version 2.0 and above of the PDB format provides REMARK 3 for including the completeness of the high-resolution shells, but it is not obligatory to provide this information. It should be possible to determine this information from the structure factors, but these are not available in many cases. It is particularly disappointing that experimental information as vital as that relating to resolution does not find its way into the final PDB deposition in a seamless and consistent manner, since it is available in electronic form at some point during the experiment. Closer interaction between software developers and the PDB could lead to better automated capture of experimental detail.

Consider another example of the lack of machine-readable experimental detail, which has important consequences when accessing structure information from a database, rather than simply reading it from PDB REMARK records. Interrogating the PDB+ database reveals that human deoxy hemoglobin (1HHB) was replaced by three entries (2HHB, 3HHB and 4HHB) in 1984, and these entries reside in the PDB today. A review of their fold deviation scores (FDS) color-coded and plotted with Rasmol (Sayle & Milner-White, 1995) on a per residue basis indicates three distinct situations (refer to http://pdbobs.sdsc.edu to review this case). The FDS is a useful first indicator when reviewing deviations from ideal geometry at each residue position. The FDS is the mean deviation from the ideal values of Engh & Huber (1991) summed over all bond lengths, bond angles, and dihedral angles in each amino-acid residue, including the side chain. The FDS plots reveal 2HHB to be a highly restrained model close to ideal geometry and 4HHB to be a loosely restrained model. 3HHB is presented as a highly restrained dimer to which non-crystallographic symmetry needs to be applied to generate the tetramer. Close inspection indicates that these three entries are derived from the same data set taken at 1.74 Å resolution and all three interpretations are correct within the accuracy of the experiment.

The following describes these observations in each of the PDB entries, 2HHB, 3HHB, and 4HHB

While this is adequate for a human to decipher it is impossible to have a computer program, including a parser attempting to load these structures into a database, take this information into account when the guide in writing such a program states

Non-standard remark annotations or those with no clearly defined topic or assigned remark number appear with remark number 6 or greater, but less than remark number 100.

Yet the distinction between these different atomic models can be important when using the data. For example, application of the Kabsch and Sander algorithm (Kabsch & Sander, 1983) for determining secondary structure provides slightly different answers using data from the three structures. While it can be argued that all determinations are correct within the accuracy of the original experiment and the user should be responsible for making themselves aware of such distinctions, this is just not practical when examining a large corpus. The level of machine-readable annotation that can be queried must be improved such that cases like that described above can be distinguished. Only then will future structure databases realise their full potential.

3.3. Analysis of the complete corpus

Given the inconsistencies and lack of machine-readable annotation, what can be ascertained from the complete corpus? This question is addressed by queries of the PDB+ database, the results of which will be discussed. A more detailed set of queries and results is found in Weissig & Bourne (1998). It should be noted that we purposely examined the complete corpus, since this is what most users do, even though it is biased by a disproportionate numbers of certain types of structures. For example, there are currently over 600 lysozyme structures in the complete corpus, many determined at high resolution in the same laboratory.

4.1. The corpus

The content of PDB+, given as the total number of depositions per year for the past 26 years since the PDB's inception, is shown in Fig. 1(a). Rapid growth began at the beginning of the decade, brought about with the insistence by many journals that structures be deposited with the PDB prior to publication and by the many advances in the field that have been documented previously (e.g., Kleywegt & Jones, 1995). Throughout this 26-year history, the absolute number of replaced structures has remained relatively constant, but as a percentage of the total depositions in a given year it has decreased markedly over time. The majority of structures deposited before 1982 have been replaced at least once (Fig. 1b). All figures use data available as of 1 January 1998, however, it should be noted that data for 1997 are not complete since many depositions for that year were either still being processed or are on hold at the PDB. It is estimated that at least 600 structures deposited in 1997 were not available as of 1 January 1998.

4.2. Temporal trends

Figs. 2(a) and 2(b) plot the mean resolution and mean size of structure, respectively, for all depositions made in a given year. This data varies widely in the early years when a small number of depositions were made per year, but since 1991, given a better statistical sample, both resolution and size show a slight upward trend. Since 1991 the mean resolution has increased from approximately 2.1 to 2.3 Å, while in the same time period the mean number of non-H atoms per structure has increased from approximately 2400 to 3000. This relationship is not surprising since complex structures with larger unit cells tend to diffract to a poorer resolution (see below). Thus, while the overall number of high-resolution structures deposited in a given year has increased over time (not shown), there is a predominance of large structures at slightly lower resolutions being deposited in successive years. This agrees with earlier findings using a set of 462 proteins (Morris et al., 1992).

Fig. 3(a) plots the mean FDS for all structures deposited in a given year against the year of deposition. While the FDS is a crude measure, it is nonetheless clear that the overall stereochemical quality of structures deposited since 1990 has continued to converge towards ideal values. At the same time χ₁ values, which are frequently not restrained during refinement (Fig. 3b), have shown little change. One possible hypothesis that could be drawn from this analysis is that with the recent focus (beginning in 1990) on checking the stereochemical quality of models with programs like PROCHECK and Whatif, models are being over restrained during refinement to reach ideal values even when the data may dictate otherwise. Thus, rather than having a stereochemical model that matches the quality of the experimental data, the model matches an ideal representation of the structure. In principle this should lead to higher mean R values over time, but the reporting of R value is itself open to question (see below).

4.3. Global trends

The time of deposition is never a consideration when selecting a structure for further study. The most common yardsticks when seeking accurate data are the R value and resolution. Yet R value and resolution are poorly correlated (correlation coefficient 0.36) in the current PDB as shown by the clustering in Fig. 4. That is, there are a significant numbers of high-resolution structures with poor R values and vice versa. As expected superceding structures show stronger clustering and correlation than the complete corpus and obsolete structures show less clustering. Outliers in these scatter plots can be explained in one of two ways. Either they are justifiable given the experimental conditions. For example, a high-resolution data set was collected but the quality of the diffraction spots was poor leading to a high R_merge and hence high R value. Or the outlier represents a potentially problematic structure. For example, an over-restrained model refined against a high-resolution data set. In either situation a user of that structure data should first be aware of this apparent anomaly and secondly should be able to probe the cause further by having experimental data available. Again χ₁ data are useful since they are less affected by the refinement. This is shown in Fig. 5 where there is a stronger correlation (correlation coefficient 0.58) between χ₁ and resolution. A more fundamental question to address is, what criteria are used in reporting data at a particular resolution? Early structures deposited with the PDB did not report the amount of data available in the high-resolution shells, or if they did it was buried in a non-parsable REMARK record. As of v2.0 of the PDB format this information is parsable, although the depositor has the option of whether to include it or not. For structures where the information is available, Fig. 6 plots the completeness of data in the high-resolution shell. There are a significant number of structures where less than 75% of the data have been collected at the reported resolution. A strong argument could be made that 100% of the data should be available at the reported resolution.

There is evidence of inconsistent reporting of water content across the corpus, particularly in early versions of a structure. Looking at structures that have been replaced indicates that 73% reported increased water content, 12% no change, but 15% reported decreased water content, often from higher resolution data. This indicates over zealous reporting of water in the first structure determination. Brändén & Jones (1990) have concluded that for structures of 2 Å resolution or better one water molecule per residue is appropriate provided there is evidence of hydrogen bonding. Fig. 7 plots the average number of atoms and water molecules for all structures in PDB+ within 0.2 Å resolution increments. The correlation between number of atoms and resolution across the whole corpus is striking up to 2.6 Å. The overall trend in water content is as anticipated – less water per residue at low resolutions, more water per residue at higher resolutions. Specifically, between 3.0 and 3.2 Å there is approximately one water per 12 residues, between 2.2 and 2.4 Å one water per two residues, between 1.6 and 1.8 Å resolution, one water per residue, and between 1.0 and 1.2 Å resolution, three waters per residue. Thus, except at very high resolution, water is conservatively reported relative to the guidelines suggested by Brändén & Jones (1990).

The plateau in the mean number of waters per structure for structures solved between 1.6 and 2.4 Å resolution is noteworthy. Overall a constant number of waters have been reported per structure regardless of resolution (and number of polymer atoms). While the number of polymer atoms ranged from 1700 to just over 3000, the mean amount of water remained constant at approximately 180 water molecules per structure. These findings need further study since they may be biased by a large number of very similar structures, for example the lysozymes, within this resolution range. Moreover, to be rigorous, water content should be plotted against the volume of exposed surface.