Visualizing ligand molecules in twilight electron density

A software script is presented for facilitating the analysis and visual inspection of ligand molecules in the context of the electron-density maps calculated from experimental data associated with protein structures determined by X-ray crystallography.

Abstract

Three-dimensional models of protein structures determined by X-ray crystallo­graphy are based on the interpretation of experimentally derived electron-density maps. The real-space correlation coefficient (RSCC) provides an easily comprehensible, objective measure of the residue-based fit of atom coordinates to electron density. Among protein structure models, protein–ligand complexes are of special interest, given their contribution to understanding the molecular underpinnings of biological activity and to drug design. For consumers of such models, it is not trivial to determine the degree to which ligand-structure modelling is biased by subjective electron-density interpretation. A standalone script, Twilight, is presented for the analysis, visualization and annotation of a pre-filtered set of 2815 protein–ligand complexes deposited with the PDB as of 15 January 2012 with ligand RSCC values that are below a threshold of 0.6. It also provides simplified access to the visualization of any protein–ligand complex available from the PDB and annotated by the Uppsala Electron Density Server. The script runs on various platforms and is available for download at http://www.ruppweb.org/twilight/.

1. Introduction  

The Protein Data Bank (PDB; Berman et al., 2000 ▶) contains almost 10⁵ protein-structure models. Among these, one class of models determined by means of X-ray crystallography is of particular interest, namely those of protein–ligand complexes. While extensive validation performed for the protein component of each crystal structure almost assures that gross inaccuracies are avoided, validation of the small-molecular-weight ligands is more challenging. An attempt to address this issue through geometry validation, for example, is provided by the ValLigURL server (Kleywegt & Harris, 2007 ▶) for comparing the conformation of a ligand across PDB entries containing the same ligand, highlighting instances of ligands with deviating geometries.

In practical terms, however, the primary evidence in support of both the presence and the specific pose (position and conformation) of a ligand molecule is the corresponding electron density. User-friendly molecular-visualization programs such as Chimera (Pettersen et al., 2004 ▶) or PyMOL (http://www.pymol.org) provide direct access to the structure models deposited in the PDB. However, outside of the context of experimental data in the form of electron density, one has to accept the deposited atomic coordinates of protein–ligand complex structure models without qualification.

In reality, crystal structure models are to a large extent interpretations of experimental evidence (that is, electron density) by their originators. Given the lack of universally accepted guidelines for ligand validation, a broad range of interpretations exist in which authors may push the boundaries of justification despite sparse evidence. The average ‘consumer’ of the structural information is often not an expert crystallographer and may be unfamiliar with the procedures of obtaining electron-density maps or may even be unaware of this essential presentation of the experimental electron-density evidence in support of protein–ligand complexes. The program we describe, Twilight, is designed to fill this gap and provide easy access to the visualization of protein–ligand complexes based on experimental data in form of electron-density maps. Twilight is linked to the ligand structures deposited in the PDB, which are filtered to include only entries which are in low agreement with electron density as indicated by low real-space correlation coefficients (RSCCs; Brändén & Jones, 1990 ▶) provided by the Uppsala Electron Density Server (EDS; Kleywegt et al., 2004 ▶). The applications of the tool extend beyond the inspection of the pre-compiled list of potentially problematic PDB entries. It also allows the interested researcher to easily access the corresponding information for any PDB entry with available density maps at the EDS for the purposes of analysis, verification, review and education.

2. Design and implementation  

2.1. Ligand database  

The PDB online Advanced Search Interface was used to retrieve 37 419 database entries with protein chains determined by X-ray crystallography as of 15 January 2012 which were cocrystallized with a ligand (including covalently bound sugars from glycosylations) and which were also listed in the EDS database, excluding entries containing additional chains of either DNA or RNA. For each entry, the corresponding real-space fit MAPMAN (Kleywegt & Jones, 1996 ▶) output file was downloaded with permission from the EDS. All values displayed in the final table were retrieved from the PDB, except for the residue number, the RSCC value and the occupancy-weighted average B factor (OWAB) of the atoms of the residue, which were parsed from the real-space fit files from EDS. The OWAB factor is computed as the sum of the B factors of the atoms of the residue multiplied by their occupancy, divided by the sum of the occupancies of the affected atoms. The ligand charge and the total number of atoms per residue, split into the number of H and non-H atoms, were derived from the chemical formula as provided by the PDB. We included only PDB heterogens which met the following criteria: (i) the RSCC is between zero and 0.6, (ii) the ligand consists of at least six non-H atoms, (iii) the ligand is not a glycerol molecule and (iv) the ligand does not include a peptide bond, as is frequently encountered in modified residues contained in protein chains. The list of peptide-bond-containing residues is provided as Supplementary Table 11. Furthermore, the choice of RSCC threshold (0.6) reflects our attempt to balance the length of the resulting list and the detection of cases with ligands in high disagreement with electron density: higher RSCC threshold values will result in a notably broader selection, while a lower real-space correlation threshold excludes interesting cases. The entries were ordered by a score S, which provides an intuitive ranking for the degree of misfit of the ligand in the electron density dependent on the resolution of the protein structure,

where RSCC is the real-space correlation coefficient and RESOL is the resolution of the crystal structure. This score becomes S = 1.0 for RSCC = 0.6 and RESOL = 1.3, and penalizes residues in crystal structures with a resolution better than 1.3 Å by assigning values greater than one. By the definition of this score, ligands in high-resolution structures with low correlation to electron density are ranked high. However, the score S is solely used to provide an ordering of the list of ligands in the graphical user interface and is not used as a threshold in the identification and annotation of problem­atic ligands. The program also allows sorting of the entries based on RSCC only. We have collected a list of 63 commonly encountered buffer molecules (Supplementary Table 2) and any ligand encountered in this list is assigned classification code ‘B’, which is displayed in the Class column in the tabular view (Table 1 ▶) for convenient elimination of ligands of limited interest.

Table 1. Description of column names in the Twilight script.

This table contains a reference sheet for all columns shown in the Twilight main table, supplying remarks on their origin. If a datum is unavailable for a particular PDB entry, the description mentions its default value for the corresponding column.

Column name	Description
Rank†	Initial sort order upon loading, which is a combination of descending score S and grouping by PDB four-character code.
PDBID‡	RCSB Protein Data Bank (PDB) four-character accession code.
LigNm‡	Three-letter PDB ligand name.
ResNr‡	Unique residue identifier. This has the general format CNNNNI, where C is the chain name, NNNN is the four-digit residue number and I is the insertion code. The residue number NNNN is right-justified and the insertion code I is mostly the space character. Residue number 66 of chain A is therefore represented as ‘A  66 ’. Notice that the RSCC values are not reported by the EDS separately when alternate conformers are present and therefore the alternate conformation indicators are not included.
RSCC§	Real-space correlation coefficient for the ligand retrieved from EDS. Used in the computation of the score S.
OWAB§	Occupancy-weighted average B factor of the ligand, also retrieved from EDS.
MolWt‡	Molecular weight of the ligand in units of Da.
Charge¶	Ligand charge, computed from its chemical formula.
#Heavy¶	Number of non-H atoms that the ligand is composed of.
#Hydrog¶	Number of H atoms for the ligand.
#Atoms¶	The overall number of atoms; that is, #Heavy + #Hydrog.
DepDate‡	Protein structure deposition date according to the HEADER record from the PDB.
Resol‡	Experimental resolution of the crystal structure in units of Å. Used in the calculation of the score S.
Rwork‡	Working-set R value. A negative number indicates its absence.
Rfree‡	Test-set R value (R free). A negative number indicates its absence.
Robs‡	Observed-set (working set and test set taken together) R value. A negative number indicates its absence.
Softwre†	Software used for crystallographic refinement as stated in REMARK 3 of the PDB entry. In the case that multiple refinement programs are stated, this becomes a comma-separated list of program names.
Jrnl‡	Abbreviated journal name if an article describing the structure has been published. Contains ‘To be Published’ if no publication is associated with the PDB entry.
PMID‡	NCBI PubMed identifier of the primary citation, if the structure has been published; otherwise, the cell is empty.
Class†	Ligand type. Frequently encountered crystallization buffer molecules are conveniently flagged with the letter ‘B’, otherwise the hyphen character ‘-’ is printed.
Valid†	This column contains the letter ‘Y’ (RSCC ≤ 0.6) or ‘G’ (RSCC ≥ 0.95) to indicate that the ligand has passed all of the filter criteria described in §2. The letter ‘A’ is used for heterogroups from a manually imported structure.
Score†	Combination of the real-space correlation coefficient and the experimental resolution into a single score S, which is used for ranking the list of ligands.
Comment	Author comment if the particular ligand has been analyzed; otherwise, it is empty. We have blanked all our in-house-generated annotations except for false positives. However, users are especially encouraged to add their own annotations in this column.

^†Value resulting from the filtering procedure mentioned in §2.

^‡Value directly taken from the result list returned by the PDB Advanced Search Interface query described in §2.

^§Value retrieved from the EDS server.

^¶Value derived from the chemical formula returned by the PDB Advanced Search Interface query described in §2.

While the primary use of Twilight is to identify and review problematic ligands in protein crystal structures, we also provide an option for inspecting a ‘daylight’ list of entries, which are characterized by an RSCC of ≥0.95. This option may be used as an educational tool to highlight cases of very clear electron density confirming the presence of a ligand.

2.2. Graphical user-interface components of the Twilight script  

The Twilight script aims to simplify electron-density-map-based visualization, analysis, validation, annotation and correction of ligand molecules found in protein structures determined by X-ray crystallography. The GUI as shown in Fig. 1 ▶ is composed of a File menu for loading and saving pre-calculated data files containing ligand data, with the possibility of directly importing structure data from the PDB and EDS, a large main table for displaying these data and a second, smaller table for providing detailed ligand information from a particular protein structure.

Figure 1.

The Twilight application window. This screenshot shows different components of the graphical user interface. The majority of window space is dedicated to the list of ligands with low correlation to electron density. In this example, the second line with PDB entry 1zjy (Schlieben et al., 2005 ▶) is highlighted. In the view below the main table, all of its heterogeneous group names are displayed in various colours. The NAI ligand of the selected row is printed on a blue background, whereas another ligand SS2 from this PDB entry is shown in orange on a grey background, indicating another ligand that is not in agreement with electron density. The magnesium ion (MG) is compatible with the electron density and since it has an RSCC of ≥0.95 it is printed in green on a white background. A status bar located at the lower end of the main view provides information about time-consuming procedures such as downloading files from EDS through the Internet. For the currently selected row, various websites can be accessed by means of the context-sensitive menu shown in the middle of the window, and ultimately it can be used to assign an annotation from a list of predefined classes, which is written to the Comment column. On the right-hand side the Preferences window allows the configuration of user-specific parameters, which are stored across sessions.

2.2.1. Table elements  

The main table, with its 23 sortable columns, represents the most dominant part of the application window presented in Fig. 1 ▶. The following paragraph explains each column in detail and Table 1 ▶ provides a brief and compact summary for fast lookup.

The initial sorting order (that is, the highest combined score S at the top), grouped by PDB entry, is given in the Rank column. (If the Twilight script is invoked with the --sort-rscc command-line option, initial sorting will be based on RSCC only.) The PDB four-character accession code is displayed in the column termed PDBID, and the ligand three-letter code is shown in the LigNm column. In the ResNr column, a unique ligand identifier is displayed as a concatenation of chain identifier, residue integer number and, if present, insertion code. The column labelled RSCC shows the real-space correlation coefficient which, together with the experimental resolution displayed in the Resol column, forms the basis for the combined score S (1) given in the Score column. The occupancy-weighted average B factor of the ligand is shown as OWAB, and the molecular weight (in Da) is presented in the MolWt column. The overall charge of the ligand, its number of non-H and H atoms, and the sum of the latter two are given in the columns Charge, #Heavy, #Hydrog and #Atoms, respectively. The DepDate column stores the deposition date of the protein structures as provided by the PDB HEADER record. The R values for the working, test and observed (i.e. both the working and test) sets are shown in the corresponding columns Rwork, Rfree and Robs. The software used for refinement presented in the PDB REMARK 3 record is shown in the Softwre column. We found that 82% of the analysed structures were associated with a peer-reviewed publication, and abbreviated journal names as well as NCBI PubMed identifiers are shown in the Jrnl and PMID columns. The Class column contains the letter ‘B’ for ligands defined as buffer molecules, as explained in §2. The Valid column discriminates between manually imported ligand data (letter ‘A’) and our set of ligands that have passed the filter criteria described in §2 (letters ‘Y’ and ‘G’, respectively). The last column, Comments, is reserved for user-defined annotations. These can either be entered as free text or can conveniently be assigned from a set of predefined strings which cover all frequently observed cases. When the PDB entries did not contain data for one or more of the above-described parameters, unrealistic default values that do not interfere with the sorting scheme of the affected column were assigned (e.g. −66.0 for a missing R value). A pop-up explanation of the content of the column is given when placing the mouse pointer on the respective column header text. Sorting is achieved by clicking on the header; a second click reverses the sort order.

A table can be loaded with the File menu and the initially provided table is contained in a file called ligands-2012-01-15.tsv. This menu also offers the import of ligand data from an arbitrary PDB entry, provided that a corresponding entry exists at the EDS. Newly imported ligands are added on top of the pre-calculated table or into an empty table and are highlighted in light blue. For unrestricted investigation of these ligands, no filtering is applied to these data. After adding or amending comments in the corresponding column, it is possible to save the file under the same or a different name, with distinct and increasing version numbers automatically appended to the name in the former case.

2.2.2. Customization  

The Preferences menu allows customization of the application. We utilize the macromolecular model-building tool Coot (Emsley et al., 2010 ▶) for displaying protein structures and their associated electron-density maps, so that the Coot window size on startup, the initial contouring of the 2mF _o − DF _c and mF _o − DF _c electron-density maps and the location of the Coot binary can be specified in the Preferences menu. A local PDB-file mirror may be entered, which will be used to load coordinate data prior to accessing the remote EDS server. Any three-dimensional coordinate file or electron-density file loaded from the web server is stored locally in a configurable location for faster access in the future. It is possible to choose an identification code for the currently active annotator which will automatically be used as a prefix for any annotation text entered in the Comment column. Input fields in the Preferences window are furnished with help texts that appear on mouse-over. Coot instructions are sent either as Scheme or as Python scripts and our script supports them both, which is especially helpful for some Coot versions which lack either interpreter. Finally, a system-dependent configuration file is used to store preferences between sessions. This file can be edited to alter or extend the set of predefined annotation texts, electron-density levels or Coot window sizes according to individual needs.

The Twilight tabular viewer is written in the programming language Python and utilizes the Python wrapper PyGTK to the GIMP Toolkit for the graphical user interface. Both Python and PyGTK are highly portable software packages and the tabular viewer has been tested in combination with Coot on Mac OS X, Windows XP and several commonly used Linux distributions. The software package can be downloaded from http://www.ruppweb.org/twilight/ and the distribution package contains further instructions for invocation and platform-dependent installation of additional packages required by Twilight.

2.3. Twilight for review  

The overarching objective of our work is to raise the awareness that electron density is the primary evidence to support the assessment of local model quality. In an accompanying publication (Pozharski et al., 2013 ▶), we provide technical details of why electron-density inspection is decisively important for ligand validation. Twilight therefore allows the coordinates and electron density of any given PDB code to be fetched, provided that an EDS entry exists.

Specifically, an entry can be added to the existing database or to an empty database generated by the user. If for some reason the electron-density maps are not available from the EDS, the program issues a warning. It also reports if a specific entry does not have any ligands, if the EDS server is inaccessible or if the PDB code does not exist at all.

We envision that the following process would be extremely simple and easy to use for review, requiring almost no additional effort on the parts of the PDB, journal editors and reviewers. Following nearly universally required submission of model coordinates and structure factors and the issue of a corresponding PDB code, the PDB creates the electron density via the EDS or other appropriate means, which can be password-protected if desired by the deposition authors. The password is revealed to the corresponding editor of the associated primary publication and provided through the editor to appropriate reviewers. Reviewers can then utilize the Twilight review script, which automatically displays electron density and model, beginning at the first heterogen that complies with our (or any agreed-upon) definition of bona fide ligands. Identification of the appropriate ligand and inspection of its electron density and environment should be intuitive even for non-expert users.

While many variants of the above scheme are conceivable, we believe that the inspection of electron density in a simple and accessible way for non-expert reviewers and editors will greatly enhance the quality of ligand structures in the PDB. We show in the accompanying publication (Pozharski et al., 2013 ▶) that this is indeed a pressing need and a serious matter for the credibility of protein crystallography.

3. Results and discussion  

The standalone Python script Twilight implements a tool for highlighting ligands which are insufficiently supported by the electron density reconstructed from the data and model, and is tightly connected to the three-dimensional macromolecular-graphics model-building program Coot. The program visualizes its central data element, a table of flagged ligands enriched with additional data and manual annotations by users, and offers a variety of utilities for sorting, searching, editing and viewing three-dimensional structures, and provides links to the most important web resources. As of 15 January 2012, the PDB includes 2850 entries with 5154 heterogens assigned an RSCC value of 0.6 or less according to EDS, encompassing the full range of possible ligands, which are subject to further filtering as described below. This represents 7.6% of the 37 419 PDB entries investigated; that is, entries with protein chains but no DNA or RNA chains, including a free ligand and determined by X-ray crystallography with a corresponding entry available at the EDS.

The choice of the RSCC cutoff of ≤0.6 was simply driven by the desire to arrive at a list of PDB entries that would not be overwhelmingly long, while at the same time assuring that a significant mismatch exists between the electron density and ligand model. However, it should be noted that the real-space correlation coefficient is sensitive to the B factors of the ligand atoms, the ligand size and the resolution of the experimental data used to calculate the electron-density maps (Kleywegt et al., 2004 ▶) and also depends on both the precision of the data and the accuracy of the model (Tickle, 2012 ▶). Consequently, we designed the Twilight tool specifically to facilitate the important actual inspection of the electron-density maps.

A certain number of the flagged entries are related to molecules from the solvent and therefore these are of only secondary interest. This was addressed by excluding 1555 ligands with fewer than six non-H atoms. Of the remaining 3599 ligands, we removed 516 glycerol molecules (PDB ligand name GOL), an abundant cryoprotectant molecule with six non-H atoms. Another 267 ligands were eliminated since they contained a peptide-bond linkage; the majority of these were unknown residues (listed with residue name UNK in the PDB entries). Whether the remaining 77 heterogeneous groups are true unbound ligands or appear as a linked residue in a polypeptide chain cannot be determined using the PDB web interface and requires the analysis of the surrounding atoms in three-dimensional space or the inspection of individual LINK/CONECT records, a task that we have omitted owing to the relatively small number of expected ligands.

After the application of the described filtering steps, the final list contained 2815 ligands distributed across 1464 PDB entries. Fig. 2 ▶ presents a graphical overview of the filtering steps divided into experimental resolutions, which range from 0.85 to 4.3 Å, with a median resolution of 2.1 Å computed over all 2815 ligands. In this figure, a large portion of filtered and eliminated nonstandard peptide residues is found in the resolution range of 3 Å or higher, and manual inspection has shown that almost all are unknown peptides covalently bound to a polypeptide chain, which could not be assigned an amino acid owing to weak, or a complete lack of, electron density.

Figure 2.

Resolution-dependent discrete distribution of ligands with RSCC ≤ 0.6. The number of ligand structure models containing ligands with RSCC values less than or equal to 0.6 is plotted versus the experimental resolution of the respective PDB entry. The distribution of ligands which have been excluded by individual filtering steps is proportionally displayed: (i) diagonally hatched light grey bars indicate the fraction of both unknown residues (190) and other nonstandard residues with peptide linkages (77), (ii) the 516 glycerol molecules are represented by light grey bars and (iii) ligands with fewer than six non-H atoms (1555) are rendered as dark grey bars. In total, the filtering steps decreased the initial set of 5153 ligands with RSCC ≤ 0.6 to our final set of 2815 ligands of interest displayed as black bars in the histogram.

3.1. Program use  

After launching the tabular viewer program Twilight, a file containing a list of ligands incompatible with experimental electron density can be opened. Initially, this is a precompiled and partially annotated list (Pozharski et al., 2013 ▶) packaged with the software. Upon loading, the ligands are grouped by PDB code and sorted by the combined score S, which links experimental resolution and ligand RSCC value (1), with higher values indicating a worse fit to electron density with respect to crystallographic resolution. For each PDB entry, the ligand with the highest score S is presented first, succeeded by all remaining ligands from this entry in descending order. This defines a monolithic block of ligands from the same PDB entry. All blocks are then sorted by their highest scoring ligand in descending order, resulting in distinct blocks of ligands from a single PDB entry with the highest scoring ligand at the top of each block. This is illustrated by taking an example from our data set. The second row in our sorted table refers to PDB entry 1zjy (Schlieben et al., 2005 ▶), an oxidoreductase from Lactobacillus brevis complexed with 1,4-di­hydronicotinamide adenine dinucleotide (NADH) and 1-phenyl­ethanol (Fig. 1 ▶). The NADH ligand has the second highest score S, and 1-phenylethanol has been added to this block in order to group all investigated ligands from PDB entry 1zjy into a single block. The next block starts with the fourth line, an N-acetyl-d-glucosamine residue from PDB entry 2hox (Shimon et al., 2007 ▶), which succeeds the NADH ligand by combined score S. Another five ligands from this PDB entry were added to complete the block.

In addition, the Twilight script supports viewing ligands that have very high agreement with electron density (defined as ligands with an RSCC value of 0.95 or greater), which is achieved by starting the program with the command-line option --show-hq-rscc. The table thus retrieved contains more than 30 000 entries and features the same filter criteria as described above, except that instead of RSCC ≤ 0.6 we here apply RSCC ≥ 0.95.

A middle-button click on a row retrieves the PDB entry from the RCSB PDB server and displays all heterogeneous groups in a separate table below the main table. These groups are conveniently colour-coded. The ligand from the currently selected row is displayed on a blue background. Any other ligands contained in the main table are shown in orange with a grey background. Ligands that have not passed the filter criteria but still have an RSCC value of below 0.6 are shown in orange on a white background. Conversely, ligands in excellent agreement with the electron density (i.e. RSCC ≥ 0.95) are highlighted in green. All other remaining heterogeneous groups are shown in black on a white background.

Double clicking a row in the table loads the structure and its related electron density from the PDB and EDS servers and displays the corresponding ligand using Coot. By default, the Sigma-A weighted 2mF _o − DF _c and mF _o − DF _c maps (Read, 1986 ▶) are contoured at 1σ and ±3σ density level, respectively. If there are multiple flagged ligands in a PDB entry, pressing the ‘g’ key in Coot advances to the next ligand listed in the main table, whereas pressing the ‘v’ key goes back to the previous ligand, such that all affected ligands can be investigated without reloading the structure or manually locating them in the structure model.

A right click on a row in the main table presents a popup menu which provides links to the PDB, EDS and PDB_REDO (Joosten et al., 2009 ▶) websites for the associated entry. In cases where there is a publication connected with the PDB entry, an NCBI PubMed link is also available. Choosing an entry from the popup menu will load the respective website in the default browser of the operating system. Finally, after inspection and analysis, a ligand can be assigned to a predefined category by selecting one from the Assign submenu in the popup menu. This assignment is written as a self-explanatory text to the Comment column in the main table, which is located at the rightmost end of the table.

3.2. Example  

In Fig. 3 ▶, we present human dihydrofolate reductase (hDHFR) bound to the inhibitor molecule D2Q in the ternary complex with the cofactor NADPH refined at 1.8 Å resolution and deposited as PDB entry 3nzd (Cody & Pace, 2011 ▶). The inhibitor molecule is well modelled (RSCC = 0.853), but NADPH was flagged by EDS with RSCC = 0.493, indicating a mismatch of the model with experimental evidence in the form of electron density. Inspection of difference electron-density maps revealed that NADPH has been suboptimally positioned. This is easily adjusted by re-refinement with REFMAC (Murshudov et al., 2011 ▶) or downloading the corresponding fully optimized entry from PDB_REDO (Joosten et al., 2009 ▶). With either method, the RSCC value of NADPH increases to 0.98 and is now in almost perfect agreement with the electron density computed from the deposited experimental structure factors.

Figure 3.

Misplaced NADPH cofactor close to its density in a dihydrofolate reductase. Example of human dihydrofolate reductase (DHFR) in a ternary complex with NADPH and the inhibitor 2,4-diamino-6-{5-[3-(ethoxycarbonyl)butoxy]-2-methoxybenzyl}-5-methylpyrido[2,3-d]pyrimidine (D2Q) determined to a resolution of 1.8 Å (PDB entry 3nzd; Cody & Pace, 2011 ▶). The inhibitor D2Q is well placed in the electron density, which is reflected by its RSCC value of 0.853. The electron donor NADPH is listed with an RSCC of 0.493 and is in inferior agreement with the electron density. (a) presents mF _o − DF _c difference electron density for the NADPH ligand and its vicinity. Positive electron density contoured at +3σ is rendered in green and negative difference density contoured at −3σ is coloured red. Atoms are coloured by element: oxygen, red; nitrogen, blue; phosphorus, orange. C atoms from hDHFR are coloured grey and C atoms belonging to the NADPH ligand are shown in yellow. (b) Electron density from the 2mF _o − DF _c map contoured at 1σ is shown as a light blue mesh. The NADPH ligand is clearly present in this map, but the difference density (a) indicates that it is slightly displaced. (c) provides a comparison of the NADPH ligand in the deposited structure (RSCC = 0.493) and after manual rebuilding in Coot followed by ten cycles of positional refinement using default parameters of REFMAC (RSCC = 0.982) in the same view as in (a). Here, C atoms from the re-refined ligand are shown in dark grey; the remaining atoms are rendered in paler colours. There is still high agreement for the nicotinamide moiety, which is close to inhibitor D2Q, but the differences are becoming more compelling towards the adenine moiety, where the displacement between equivalent original and re-refined atom pairs reaches values of higher than 1 Å. This figure was generated with PyMOL (http://www.pymol.org).

4. Conclusions  

The Twilight script was developed for viewing, analyzing, annotating, validating and correcting ligands that appear to have low correlation with electron-density maps. In an accompanying manuscript (Pozharski et al., 2013 ▶), we discuss common patterns that we have recognized during an extensive analysis; we supply specific examples and suggest possible solutions to correct the problems. The underlying data distributed with the program in fact encourage a variety of additional queries and analyses to be performed, as they include descriptions of all ligands deposited in the PDB prior to 15 January 2012 regardless of the filtering steps that we have applied. Interested readers are invited to download the software package and to critically review and annotate the protein–ligand structure models. The automated analysis described in this work will be repeated at regular time intervals and we will publish updated versions of the resulting data on our website: http://www.ruppweb.org/twilight/.