1. Introduction 1.1. Validation A crystallographic experiment and the ensuing process of phasing, model building and refinement (hopefully) culminates in a three-dimensional structure model that fits both the experimental X-ray data and our prior knowledge of macromolecular chemistry. Validation helps to guide this process and enables crystallographers to produce high-quality structure models suitable for biological interpretation. Validation software routines have been available since the early 1990s (Laskowski et al., 1993 ▶; Hooft et al., 1996 ▶; Davis et al., 2004 ▶). These check many aspects of crystallographic (or other) structures, typically with a strong focus on macromolecules and specifically on proteins. Validation of nucleic acid structures is also available, for example in MolProbity (Chen et al., 2010 ▶; Richardson et al., 2008 ▶). Validation software for other chemical entities is available as well, e.g. PDB-care (Lütteke & von der Lieth, 2004 ▶) for carbohydrates and WASP (Nayal & Di Cera, 1996 ▶) for metal ions. Various validation tools are also available directly through graphical model-building software such as O (Jones et al., 1991 ▶) and Coot (Emsley & Cowtan, 2004 ▶; Emsley et al. 2010 ▶). While finalizing a crystallographic structure, validation can be used constructively to detect anomalies in the model and the crystallographer can either remove the anomaly or confirm that it is real, i.e. by assuring that there is sufficient (experimental) evidence for a model with this anomaly. This process has many complications. For example, large overall bond-length deviations typically arise from wrong or erroneous restraint settings; however, they can also be the result of errors in the experimental determination of unit-cell parameters and should not be resolved by tightening the restraints but rather by updating the unit-cell parameters. Most validation tools focus on the identification of outliers: residues which are, based on statistical measures of our growing empirical knowledge of macromolecular structure, not similar to other residues in known structures. Focusing on outliers is sensible because these residues are most likely to be either interesting or wrongly modelled, but it takes away attention from the rest of the model. This becomes a problem when the absence of outliers is seen as a confirmation that the model is optimal. Side-chain rotamers are a good example: having no (or few) outliers does not mean all the rotamers are optimal with respect to the experimental data. 1.2. Validation after structure deposition After a crystallographic structure has been finalized, deposited in the PDB and released, the atomic coordinates are set in stone. The Protein Data Bank (Bernstein et al., 1977 ▶; Berman et al., 2003 ▶) is a historical archive and does not change the atomic coordinates, although their annotation may be updated (Henrick et al., 2008 ▶) to deal with the changing needs of depositors and users. The static nature of structure models in the PDB has many implications. The obvious consequence is that models that are several decades old often lack the accuracy and precision that modern crystallographic software offers, while they also lack the benefit of constructive validation which was either lacking or less ‘mature’ at the time these older models were constructed. This way, the users of the PDB often have access to models that are suboptimal by modern-day criteria. In addition, a less obvious consequence is that even when a user validates an existing PDB entry or uses ready-made validation reports as provided by, for example, the PDBREPORT data bank (Hooft et al., 1996 ▶; Joosten, te Beek et al., 2011 ▶), validation is no longer constructive: any anomalies found do not lead to an improved structure model because most PDB users do not have the crystallographic expertise to make the structure better before using it. In this context, validation may lead to the outright rejection of a structure model when an alternative with better validation results is available. This situation not only affects the older models in the PDB, which were accumulating at a slower pace (more crystallographic structures have entered the PDB in the last three years alone than in its first thirty years): a model submitted now will simply be an old model in a few years time, as crystallographic methods are still improving at an appreciable pace. 1.3. PDB_REDO To keep PDB models up to date, we decided to apply some of the latest crystallographic methods to all PDB entries for which experimental X-ray data are available (Joosten, Salzemann et al., 2009 ▶) and created the PDB_REDO software pipeline that takes atom coordinates and X-ray data from the PDB and re-refines the structure model in REFMAC (Murshudov et al., 1997 ▶, 2011 ▶). A strong focus was placed on automation in order to deal with the tens of thousands of PDB entries, which meant dealing with the problems the PDB still had (at the time) with inconsistent annotation of coordinate and reflection files. The most important strength of automation is that it allowed us to optimize the weight between the X-ray data and the geometric restraints, something that is a lot of work to do by hand even for a single structure model. In our re-refinement we consistently used TLS models for anisotropic atom movement (Schomaker & Trueblood, 1968 ▶; Winn et al., 2001 ▶), which has only recently been made accessible to macromolecular crystallography, thus using the latest developments in refinement for all PDB entries. The resulting structure models showed an improvement in terms of R free (Brünger, 1992 ▶), and model validation showed a substantial improvement in overall model-quality estimators such as the Ramachandran plot (Ramachandran et al., 1963 ▶) Z score, which compares the combination of backbone torsion angles for each residue with a residue-type and secondary-structure specific distribution in high-quality protein structure models (Hooft et al., 1997 ▶), and the number of atomic clashes or bumps. A notable result was that very recent PDB entries also improved, although in most cases the improvement was not as great or as common as in older PDB entries. Making the improved structure models available to the PDB user community through the PDB_REDO data bank (Joosten & Vriend, 2007 ▶) was an initial step towards a constructive form of structure validation of deposited structure models. However, it was clear that local fitting errors and other problems in structure models could not be resolved by the approach we used (Joosten, Womack et al., 2009 ▶). A much more comprehensive approach that incorporates real-space model rebuilding was needed. 1.4. Real-space rebuilding Algorithms for real-space fitting and rebuilding parts of the structure model have existed for decades (Diamond, 1971 ▶; Jones, 1978 ▶). The problem lies in deciding how they should be applied and where in the structure. The typical approach has been to manually update the structure model using molecular graphics, with a validation report to hand. The program OOPS (Kleywegt & Jones, 1996 ▶) provided a significant speed-up of this process by automating the where part of the problem: it used validation results from O (Jones et al., 1991 ▶) and WHAT_CHECK (Hooft et al., 1996 ▶) and turned them into a ‘macro’ for O that takes a user through the flagged parts of the structure model automatically. Similar implementations are available to interface MolProbity (Chen et al., 2010 ▶) with Coot (Emsley et al., 2010 ▶). Solving the how part of the problem still remains a challenge in structure validation: finding an anomaly is often easier than understanding what caused it. We recently described two programs (pepflip and SideAide) that rebuild structure models using a strategy that incorporates both the where and the how in a single decision-making framework (Joosten, Joosten et al., 2011 ▶). Instead of improving crystallographic structures just by rebuilding the parts of the model flagged by validation routines, the main chain and side chain of every residue is rebuilt and validated; if rebuilding leads to an improved fit to the crystallographic maps, then the model is updated with the new conformation. Instead of specifically looking for errors, this approach focuses on improvable features of the structure model and therefore has much greater coverage of the structure model than validation-based methods. The algorithms are fully automated and not particularly computationally intensive, enabling us to incorporate them in the PDB_REDO pipeline. The main challenge in automation lies in decision making. Decisions, and the priority they are given, are often taken based on statistic measures, but also on personal preference, experience or gut feeling, or even on dogmatic principles. Many decisions are taken without even realising it, e.g. by applying default values to many parameters. Fully automated software pipelines need to formalize all these decisions to a closed set of rules. Here, we discuss PDB_REDO as a decision-making framework, showing the choices we face and how we resolved them, and also present an extensive test on 12 000 PDB entries. 2. Methods The different model-quality metrics used in the decision-making process of model optimization are described in Table 1 ▶. The software programs used in the procedure are discussed in the text and summarized in Table 2 ▶. 2.1. The PDB_REDO pipeline The PDB_REDO software pipeline uses the PDB file with the deposited crystallographic structure coordinates and the associated reflection file containing the X-ray diffraction data. 2.1.1. Preparation of the diffraction data (CIF file) The reflection file is standardized with Cif2cif (Joosten, Salzemann et al., 2009 ▶), which writes out Miller indices and amplitudes (or intensities if amplitudes are not given) for each reflection. σ values and R free-set flags are written out if available. Only the first data set in the reflection file is used. Some basic sanity tests are performed as follows. (i) Reflections with negative values are rejected if they are marked as amplitudes, but are kept if they are intensities. (ii) If σ values are given the set is tested for information content. If all values are the same, the σ values cannot be used for scaling purposes in refinement. Individual σ values of 0.0 are reset to the highest σ value in the data set. (iii) The R free set is validated, as discussed below (§2.2.1). The standardized data set is then completed and converted to an MTZ file using tools from the CCP4 suite (Winn et al., 2011 ▶). Intensities are converted to amplitudes using CTRUNCATE, which deals with negative intensities (French & Wilson, 1978 ▶). The B Wilson is calculated by SFCHECK (Vaguine et al., 1999 ▶). SFCHECK is also used to check the data completeness and to check for the presence of twinning. 2.1.2. Preparation of the coordinates (PDB file) The PDB file is parsed with Extractor to extract cell dimensions, space group, waters and special residues that are involved in chemical interactions (as marked in LINK records in the PDB header). Parameters about refinement are also extracted at this stage: the R factor and the R free from the PDB header (R head, R free,head), TLS-group selections and tensors, and the type of solvent model. The following records are removed from the PDB file by the program Stripper, mostly to ensure proper restraint generation for refinement: explicit H atoms, atoms with occupancy 5%) and a twin fraction of >5% was detected by SFCHECK, then REFMAC is run with twinning. (iii) If the above difference persists, rigid-body refinement is tried. (iv) If a difference of above 5% still persists, then five cycles of TLS refinement are tried (only if TLS tensors were extracted from the PDB file header) in an attempt to deal with possible corruption of the TLS tensors. (v) If after these steps the difference between the reported and the calculated metrics is more than 10%, it is decided that something is inherently wrong with this PDB entry and the PDB_REDO pipelines stops prematurely. New restraint files for compounds not yet in the REFMAC dictionary (Vagin et al., 2004 ▶) are generated automatically. The final values of R calc and R free,calc plus those of R ratio, R free,unb,calc, Z(R free,calc), r.m.s.Z(bond)calc and r.m.s.Z(angle)calc are used as baseline values for further refinement. 2.1.4. Re-refinement To get the most out of refinement in REFMAC, many parameters should be optimized. In the original PDB_REDO pipeline we only optimized the geometric restraint weight. Despite the presence of a robust and reliable option to automatically weight X-ray data and geometry in newer versions of REFMAC, we find that our search method can have advantages. We use the geometry-weight optimization and also systematically explore the usage of many other refinement strategies, as follows. (i) If SFCHECK finds a twin fraction of >5%, twinning is evaluated in REFMAC. If REFMAC finds twin operators with R merge 7%, refinement is performed with twin target functions. (ii) Four different B-factor models are evaluated: anistropic B factors, isotropic B factors with TLS, isotropic B factors without TLS and one overall B factor with TLS. The selection of the optimal algorithm is discussed below. (1) If it is decided that anisotropic B factors cannot be used, the possibility of TLS refinement is tested by resetting all atomic B factors to the B Wilson (or the average B factor at resolutions 4 Å or worse) and calculating R TLS and R free,TLS, followed by TLS refinement alone. Multiple TLS models are tested whenever possible: a simple model with one TLS group per chain, the TLS model extracted from the PDB header and any additional user-provided TLS models (for instance, from TLSMD; Painter & Merritt, 2006 ▶). The selection algorithm for the best TLS model is also discussed in §3. If TLS refinement decreases R free with respect to R free,TLS, the output TLS model will be used in further refinements as is. (2) If individual B factors are used, the weight of the B-factor restraints (which keep neighbouring B factors similar) is optimized by performing a grid search of up to seven different values. For each weight a short refinement is performed with automatic geometric restraint weighting. The selection algorithm for the best weight is discussed in §2.2.4. After all these choices have been made, the actual re-refinement is performed. A riding hydrogen model is always used to optimally benefit from van der Waals restraints. Local NCS restraints (Murshudov et al., 2011 ▶) are used when applicable, regardless of the data resolution. In this re-refinement, up to seven different geometric restraint weights are used, meaning that a final model will be selected from a number of candidates; the selection algorithm is also discussed in §2.2.4. Our previous PDB_REDO results showed that re-refinement could benefit from making the search space for the geometric restraint weight resolution-dependent. We have performed this by assigning each input model to one of six different categories (Table 3 ▶) based on the resolution of the data and the number of X-ray reflections per atom. The latter is important because at a given data resolution the number of reflections per atom varies greatly depending on the solvent content: for example, at 3.0 Å resolution values of between 1.9 and 7.8 reflections per atom are observed. In addition to assigning the geometric restraint-weight search space, the ‘resolution’ categories are also used for determining the B-factor restraint-weight search space and the application of ‘jelly-body’ restraints (Murshudov et al., 2011 ▶), which we use to stabilize the re-refinement of structures belonging to the two lowest resolution categories (‘vlow’ and ‘xlow’). 2.1.5. Rebuilding The re-refinement results in a structure model with new (difference) electron-density maps. These maps are used to rebuild the structure model in four steps. (i) The program Centrifuge removes all waters with low weighted mean density fit to the 2mF o − DF c map [fit(ρ) 10.0]. (4) The difference between R free,calc and R calc is much smaller than expected [R free,calc − R calc 18). (ii) If there are still 50% more reflections than parameters including anisotropic B factors (18 > RPA > 13.5) the atomic B factors are set to B Wilson and two refinements using isotropic and anisotropic B factors are run using automatic geometric restraint weighting and default B-factor restraints. The program bselect (see below) is then used to pick the best B-factor model. (iii) If there are more than three reflections per atom (13.5 > RPA > 3) isotropic B factors are used. (iv) If there are fewer than three reflections per atom (RPA R free,w,simple) the simpler model is used. (ii) All possible combinations of w 1 and w 2 are used for Hamilton tests. The percentage of tests that gave an acceptable result (where ‘acceptable’ means that the more complex model is appropriate) is then calculated. If less than 30% of the Hamilton tests were acceptable, the simpler model is used. If more than 95% of the Hamilton tests were acceptable, the more complex model is used. (iii) If the step above is inconclusive (acceptable tests between 30 and 95%), the complex model is examined by looking for signs of overrefinement and the simpler model is used if the Z score for R free is too low [Z(R free)complex cutoff, where cutoff is 4% for anisotropic B factors and 6% for isotropic B factors) then the simpler model is used. Finally, if the difference between the R free and the R factor for the complex model is more than two times the difference for the simpler model [(R free, complex − R complex) > 2.0 × (R free, simple − R simple)] then the simpler model is used. Otherwise, we consider that there are no signs of overrefinement and the more complex model is used. 2.2.3. Selecting the number of refinement cycles A large number of REFMAC refinement runs are performed to try different parameters; thus, the number of internal refinement cycles becomes an important contributor to the time needed to optimize a model. Based on our experience with REFMAC, we use a set of empirical rules to assign the number of refinement cycles. (i) For rigid-body refinement, we use ten cycles when attempting to reproduce the R factor reported in the PDB header, R head (see §2.2.2). Rigid-body refinement is also used for ‘legacy’ structure models predating 1990; in this case, 15 cycles are used since older structure models may be further away from convergence. (ii) For TLS refinement, we use five cycles for reproducing R head (see §2.2.2), ten cycles for optimizing the TLS model (15 cycles for ‘legacy’ models) and five cycles in the final model refinement. (iii) During re-refinement, 20 cycles are used by default. If no TLS models are used, five additional cycles are introduced. In the cases where a new R free was created 30 cycles are used, and finally when using anisotropic B factors 40 cycles are used since anisotropic refinement takes longer to converge. For ‘legacy’ models, 50 cycles are used by default and 60 cycles are used when using anisotropic B factors. (iv) For choosing the B-factor model and weights, we use the same number of cycles as in re-refinement when deciding between individual isotropic B factors or one constrained B factor. For deciding between isotropic and anisotropic B factors 50 cycles are used. Finally, ten cycles are used for optimizing the B-factor restraint weight with TLS models and 15 cycles when TLS models are not used. 2.2.4. Selection of the optimal model from a set of refinement results The TLS-model optimization, B-factor restraint-weight optimization, re-refinement and final refinement require the selection of the best model from a set of refinement results. The program picker selects the model with the best fit to the experimental data while minimizing the risk of overrefinement. To do this, the quality of the starting structure is also taken into account. Picker uses the following procedure. (i) Firstly, we establish R co and R free,co as cutoff values depending on the source of the models in question. R TLS and R free,TLS are used if the models originate from TLS-model optimization refinement; R calc and R free,calc are used in all other cases. If R free,calc has been considered to be ‘biased’ then R free,unb,calc is used instead of R free,calc. (ii) Similarly, r.m.s.Z(bond)co and r.m.s.Z(angle)co are established as geometry cutoff values. The default cutoff values are 1.0, but the values are increased to r.m.s.Z(bond)calc and r.m.s.Z(angle)calc, respectively, if these were greater than 1.0. (iii) For each refinement condition i, Z(R free) i and σ(R free) i are calculated. From these, a maximum value for R free is calculated [R free,max,i = R i × R ratio + 3.5 × σ(R free) i ]. R free,max,i is then set to the lower of R free,max,i and (R i + 6%) to make sure that the difference between R and R free is not too large. Finally, R free,max,i is set to the higher of R free,max,i and (R free,co/R co × R i ) to deal with structures that had a high initial R free/R ratio. (iv) After the cutoff values have been established, any model is rejected if a model metric exceeds the preset limits [r.m.s.Z(bond) i > r.m.s.Z(bond)co, r.m.s.Z(angle) i > r.m.s.Z(angle)co, R free,i > R free,max,i , R free,i > R free,co). For structures in the ‘vlow’ and ‘xlow’ categories the difference between R free and the R factor is also taken into account for model rejection [(R free,i − R i ) > 2.0 × (R free,co − R co)]. (v) Finally, the optimal refinement is selected from the remaining conditions. The models with the lowest free likelihood (LLfree) and the lowest R free are selected. If these two models are different, that with the highest Z(R free) is finally selected. The algorithm in picker rejects all refinements if none pass the established criteria and the PDB_REDO pipeline adapts accordingly. Depending on the set of refinements evaluated by picker, TLS might not be used further in refinement or the B-factor restraint weight can be set to the REFMAC default. If the evaluation after re-refinement does not produce a better model, the original structure is used in the subsequent rebuilding steps and the final refinement is performed with auto weighting for the geometric restraints. 2.2.5. Selection of atoms for exclusion from rebuilding The first three steps of rebuilding are meant to be comprehensive, so by default all residues are considered. Negative selection is applied to ‘special’ residues for which automated unsupervised rebuilding is too risky or not needed. (i) Centrifuge ignores all waters involved in LINKs (see §2.1.2). (ii) Pepflip ignores all residues in which the backbone N or the backbone O atom is involved in a LINK. Residues in the middle of secondary-structure elements, as designated by DSSP (Kabsch & Sander, 1983 ▶), are very unlikely candidates for flips. They are ignored as well to speed up peptide flipping. (iii) SideAide ignores all side chains involved in LINK records and all side chains with multiple conformations. The final rebuilding step uses positive selection of candidates based on a WHAT_CHECK validation report. However, the negative selectors above still apply. 2.3. Evaluation data set We attempted to optimize all of the PDB entries with deposited X-ray diffraction data (>53 000) for inclusion in the PDB_REDO data bank. Fewer than 900 PDB entries ( 5σ(R free)calc + R free,head], while for the total data set this was 5%. Reproducing R factors is known to be a challenging problem (Kleywegt et al., 2004 ▶; Afonine et al., 2010 ▶) and many problems can be reduced to a lack of knowledge of the original refinement parameters, e.g. about the treatment of bulk solvent. Because the annotation of new PDB entries has improved substantially, we can now adapt Extractor to obtain a more detailed description of the original refinement settings. These collectively indicate that part of the problem with worsened R free is partly artificial. Deterioration in R free does not mean that the model becomes worse in terms of all other quality metrics. In fact, this is very rare and only occurs in 0.5% (58 structures) of the test set (Fig. 4 ▶); deterioration of three or more metrics is still rare and occurs in 6% of the test set. The opposite, improvement of all model-quality metrics, occurs in 16% (1934 structures) of the structures (Fig. 4 ▶), while improvement of three or more quality metrics occurs in 85% of structures in the evaluation set (Fig. 5 ▶). 3.2. More constructive validation During high-throughput testing of PDB_REDO on existing PDB entries, more than 800 fixable errors were encountered and were reported back to the PDB. Although mostly trivial annotation problems, these errors can be devastating when structure models are used in automated computational studies. Most issues were resolved at short notice, fixing the problem for all PDB users rather than just for PDB_REDO. 3.3. Outlook 3.3.1. Ongoing development The results shown here are encouraging, but also show that there is still ample room for improvement. For instance, the sulfate ion in Fig. 1 ▶(d) was distorted, which was traced back to a problem with chiral volume restraints for the S atom. In real-life chemistry the O atoms are equivalent and the sulfur is not chiral. However, the O atoms are different computationally (they have different names), which makes the sulfur chiral during refinement. This means that a swap of any two O atoms in sulfate inverts the chirality of the sulfur. The chiral volume restraints now erroneously push the refinement towards improving the chiral volume, resulting in a distorted molecule. We are currently testing a new tool, called chiron, that fixes these computational chirality problems. In the long term, chemical chirality problems (where the atomic coordinates do not match the residue name) should also be fixed by either renaming the compound or rebuilding it. However, this can only be performed when reliable information about the chemical nature of the compound can be obtained automatically, and touches on the issue of constructive validation of ligand entries in the PDB, arguably a more important, but also a significantly more complex, task. The refinement of a structure model with NCS constraints is available in REFMAC and implemented in PDB_REDO. Unfortunately, using this option often requires manual intervention because the so-called MTRIX records that describe strict NCS in PDB files frequently have annotation errors. We are working on a decision-making algorithm that can properly deal with such cases. Our current rebuilding tools can be improved to allow support for noncanonical amino acids such as methylated lysines. Also, more substantial backbone rebuilding, for instance by building missing or poorly defined stretches of many residues, is a target for further improvement of PDB_REDO. Adding multiple conformations of side chains and of stretches of main chain presents an additional challenge. Representation of the optimization results has become much more important now that we are actively changing structure models. We currently use molecular scenes from YASARA to highlight changes in the model, but this requires that installation of software and can only show one thing at a time. An approach such as Proteopedia (Hodis et al., 2008 ▶; Prilusky et al., 2011 ▶) could be more flexible and could allow the visualization of results directly in a web browser. Such a tool can allow us to highlight peptides that have been flipped or waters that have been removed. The current model-quality Z scores are robust metrics, but are somewhat less comprehensible by statistically unaware users than, for example, the percentile scores used in MolProbity (Chen et al., 2010 ▶); the percentile scores have, besides obvious advantages, some caveats for asymmetric distributions. In addition, the absolute number of bumps that we use in our reports can cause misinterpretation, simply because it does not account for the severity of the bumps. This can lead to misleading results: in absolute terms ten bumps with atomic overlaps of 0.05 Å are worse than a single bump with overlap 0.5 Å even though the former is likely the result of suboptimal restraints and the latter of a fitting error. A new metric for bumps should be developed. Local measures for the fit to the experimental data are not yet included in the PDB_REDO pipeline and should be added. In the near future, the brief validation reports proposed by the PDB Validation Task Force (Read et al., 2012 ▶) are very likely to become the preferred way of presenting PDB_REDO results. The relative metrics that are recommended for adoption in this report (and are under development in the PDB) can and should be implemented comparing PDB_REDO structures both against average PDB structures but also against average PDB_REDO structures. 3.3.2. Using PDB_REDO The PDB_REDO pipeline can also be used to refine structures in the process of finalizing a structure in any laboratory. We now regularly employ PDB_REDO in our laboratory, usually to optimize near-complete models, but sometimes as early as straight after molecular replacement. While the software is available for download at http://www.cmbi.ru.nl/pdb_redo for use in one’s own laboratory (see, for example, Peng et al., 2010 ▶; Guan et al., 2011 ▶), it is admittedly not straightforward to install and has several system dependencies at the moment. We are also working on a PDB_REDO web server that will provide a more user-friendly way to use the PDB_REDO pipeline in the near future. We encourage the usage of PDB_REDO prior to structure deposition because it can improve the structure model and its interpretation. It must also be noted that PDB_REDO chases a moving target: unlike the PDB models, the PDB_REDO data bank models are not static and must all eventually be replaced or supplemented by a new version incorporating new methodological advances. Individual entries in the PDB_REDO data bank (http://www.cmbi.ru.nl/pdb_redo) can be used for any structural biology study, e.g. for homology modelling (van der Wijst et al., 2010 ▶; Flueck et al., 2011 ▶). Importantly, the collection of models can also be used as a homogeneously treated data set for statistical analysis of structure models. For example, the Ramachandran plot quality (Fig. 3 ▶) or average r.m.s.Z(angle) (Fig. 6 ▶) distributions in the PDB_REDO differ distinctly from those in the PDB. These distributions can be used to define new refinement targets and new criteria for choosing reliable water molecules or to construct (stricter) validation criteria (Kota et al., 2011 ▶). Further development to alleviate all cases in which some models deteriorate according at least to some criteria will be required before the full potential of PDB_REDO can be unleashed for the community. 4. Conclusions The natural and constructive follow-up to structure model validation is to improve the model based on the validation results, as is now common practice by competent X-ray crystallographers. The process of improving the model still requires many decisions to be taken by crystallographers. The PDB_REDO pipeline combines refinement and model rebuilding with a decision-making framework that can autonomously optimize structure models. It makes ‘constructive validation’ possible without the need for manual intervention. This is particularly important for structure models in the PDB, which are never updated otherwise. Applying the PDB_REDO pipeline to 12 000 random PDB entries showed that the majority of PDB entries can be improved according to commonly accepted quality criteria. This improvement is resolution-dependent: greater improvement occurs at lower resolution. Real-space model rebuilding has substantial added value to re-refinement, especially for improving geometrical targets. The limiting factors to improving models appear to be in finding optimal refinement parameters, but also in extending the scope of rebuilding to larger portions of the main chain and adding multiple conformations. The final hurdles are likely to be uniform and reliable water modelling, and last (but by no means least) the PDB-wide rebuilding of additional macromolecules (nucleic acids and carbohydrates) and the various hetero-compounds (ligands) bound to protein structures.