Application of the AMPLE cluster-and-truncate approach to NMR structures for molecular replacement

Processing of NMR structures for molecular replacement by AMPLE works well.


Introduction
Molecular replacement (MR) is an increasingly common route to solving the phase problem for protein crystal structures. In 2012, for example, 77% of protein structures submitted to the Protein Data Bank (PDB; Rose et al., 2012) were solved using MR. MR involves the placement of an existing structure (the search model) in the new unit cell of the target structure in such a way as to reproduce its crystallographic lattice. This provides approximate phasing information allowing the initial calculation of electron-density maps (Rossmann & Blow, 1962). Typically, the search model is derived from an experimental structure. The structure itself, a processed version of it or a homology model of a related protein may all be used, but the existence of a similar structure to the target is ultimately required.
Historically, NMR structures have been considered to be more problematic than crystal structures for use as search models in MR: it is not uncommon for a crystal structure to be insoluble even with an NMR structure of the same protein (Chen et al., 2000). While genuine conformational differences may exist between the crystalline and solution states of a protein, more frequently the problem stems from the intrinsic variability within NMR ensembles and the fact that NMR structures generally score more poorly by protein structurequality measures than their crystal structure counterparts (Bhattacharya et al., 2007). In NMR, spectral overlap and peak broadening are factors that reduce the number of experimental restraints that can be assigned to specific parts of the molecule and hence employed in model calculation. The reduced experimental definition of the affected areas translates into their increased conformational variability during structure calculation and thus results in local divergence within the NMR ensemble (Doreleijers, Sousa da Silva et al., 2012).
Nevertheless, the introduction of residual dipolar couplings (RDCs; Tjandra & Bax, 1997) has provided valuable longrange information that helps to define large-scale features of the protein structure, and general progress in NMR methods has led to a corresponding improvement in the quality of NMR structures (Mao et al., 2011).
Irrespective of methodology, some protein regions may be truly mobile, lacking defined structure, and hence more variable in calculated NMR ensembles. Such locally divergent regions in NMR ensembles tend to deviate more strongly from the crystal structure counterpart. Thus, in order to reduce noise in the MR search model, variable regions in NMR ensembles are normally excluded. Since the most variable regions in an NMR ensemble are commonly surface loops and termini that broadly correlate with regions of higher B factors in crystallographic structures, their elimination often has the additional advantage of avoiding some less well defined parts in the crystal structure, i.e. those that contribute least to the scattering. Recently, a specialized tool, FindCore (Snyder & Montelione, 2005), has been applied to processing NMR ensembles for MR (Mao et al., 2011). FindCore reduces variability within the NMR ensemble by calculating an atomic pseudo-B factor based on structural variance and eliminating any atoms (plus trailing side-chain atoms) with a pseudo-B factor of over 60 Å 2 . In a benchmarking exercise using Phaser for MR and using ARP/wARP for tracing, FindCore-derived ensembles solved 22 of 25 cases in which the NMR ensemble and the target crystal structure were 100% sequence-identical, and a further two cases were successful when Rosetta refinement was employed (Mao et al., 2011). In cases of homologous proteins a sequence-identity threshold was observed: above 40% identity success was assured, but below 30% identity only one of four cases yielded a correct solution.
As stated, among the predicted structures used for MR, homology models predominate. However, in recent times there has been rapid development in the area of ab initio protein modelling (also known as de novo or template-free modelling; Gajda et al., 2012). This aims to predict protein structures without relying on evolutionary relationships and so can address novel folds that are inaccessible to homology modelling. Although the combinatorial nature of the ab initio folding algorithm limits the accessible target size to around 120 residues for soluble proteins or 145 residues for membrane proteins (Yarov-Yarovoy et al., 2006;Barth et al., 2007), ab initio models have successfully been employed for MR. This was first with performed compute-intensive all-atom models (Qian et al., 2007;Das & Baker, 2009 truncate approach combined with different modes of sidechain treatment. Now implemented as the CCP4 program AMPLE, the pipeline produces many search models for each target (up to around 500) and succeeds over a range of searchmodel size from very small, generally accurately modelled structures to larger more approximate representations. As mentioned above, locally divergent regions in NMR ensembles often differ most from the corresponding crystal structures and are often eliminated prior to MR attempts. This is conceptually similar to the rational elimination of divergent and likely inaccurate regions by AMPLE in ensembles derived from ab initio modelling (Bibby et al., 2012). We therefore explored the application of AMPLE to NMR ensembles, proposing too that its sampling of both large and small search models, combined with different side-chain treatments, could improve performance compared with the approach of finding a single core structure (Mao et al., 2011). Here, we describe the results, demonstrating the successful application of AMPLE to solve crystal structures using search models derived from NMR structures. Furthermore, we find that a protocol including Rosetta (Leaver-Fay et al., 2011) remodelling of NMR structures can lead to successful structure solution where simple editing does not. Detailed hands-on guidance for running AMPLE is also provided (see Appendix A).

Materials
For comparison with previous results, we assessed the performance of AMPLE against the set of 25 matching (100% sequence-identical) NMR search models and target crystal structures, recently solved, previously used with the FindCore method of search-model preparation (Mao et al., 2011).
Additionally, we used a set of ten NMR ensembles of thioredoxin-fold proteins to try to solve the crystal structure of Streptomyces coelicolor thioredoxin (PDB entry 1t00; 112 residues; diffraction data to 1.51 Å resolution; Stefankova et al., 2005) both with AMPLE and FindCore. Thioredoxin proteins were chosen as providing a broad range of sequence identities versus the target from 16 to 52% (Table 1), allowing a better definition of the limits of success. NMR ensembles were assessed for structural quality using the NRG-CING server (Doreleijers, Sousa da Silva et al., 2012), resulting in an ROG overall molecular classification  of red (lower quality), orange (intermediate) or green (higher).

Methods
NMR ensembles were processed into search models by AMPLE in two ways. The first treats the NMR ensemble in the same way as described previously for processing a set of ab initio models (Bibby et al., 2012). Briefly, AMPLE determines the conformational diversity of C atoms in the NMR ensemble along the protein chain using THESEUS (Theobald & Wuttke, 2006). This guides the truncation of the NMR ensemble in 5% steps starting with the most variable regions and with application to whole residues. The set of truncated ensembles are subclustered at different radii and subjected to three modes of side-chain treatment: retention of all side chains, elimination of all side chains beyond C or retention of only a subset. The subset are those that the side-chain prediction program SCWRL (Canutescu et al., 2003;Krivov et al., 2009) places most accurately, a consideration that is not relevant to the processing of NMR models but is related indirectly to side-chain conformation variability in a way that might help to preferentially eliminate ill-defined surface residues. Alternative side-chain treatments oriented specifically towards NMR ensembles, e.g elimination according to conformational variability, will be explored in the future. Processing an NMR ensemble into a set of search models typically takes around 15 min. The resulting set is then passed to MrBUMP  for MR with both Phaser  and MOLREP (Vagin & Teplyakov, 2010). The resulting top placements are then treated to rapid phase modification and C tracing in SHELXE (Usó n et al., 2007;Sheldrick, 2010): resulting CC scores of >25 are reliably indicative of correct placement and often result from nearcomplete automatic tracing of the structure (Rodríguez et al., 2012). Thus, a CC score of >25 was our stringent measure of the success of a given search model.
Where simple truncation as above failed to give a correct solution, additional processing of the NMR ensemble with Rosetta was tried. This is based on previous observations that the phasing power of NMR ensemble-derived search models can be improved by Rosetta (Qian et al., 2007). Our refinement consisted of an initial idealization of each model of the NMR ensemble using the idealize application of Rosetta (Leaver-Fay et al. Results of attempts to solve the crystal structure of S. coelicolor thioredoxin (PDB code 1t00) with various NMR structures using FindCore and AMPLE for search-model preparation. Squares indicate success with both programs, diamonds indicate AMPLE-only successes with a truncation-only protocol, triangles indicate AMPLE-only successes with a Rosetta remodelling protocol and the circle indicates a case that was not solved with either program. present work, no electron density is provided to this application. The comparative modelling protocol was applied using the sequence of the NMR structure separately to each member of the ensemble. 1000 models were generated, sampling each member of the NMR ensemble equally. Since the number of conformers in each deposited NMR ensemble varies, the number of times that each conformer is used as the basis for remodelling will vary. A typical run time for generating the 1000-model set is 13 h, making it comparable in overall timing to similarly sized ab initio modelling cases. This set of models was then treated in the same way as the sets of 1000 decoys generated ab initio previously (Bibby et al., 2012).
For comparison, FindCore was also applied to the thioredoxin test set. For each NMR ensemble, FindCore indicated a list of core residues. Non-core residues were removed and the result was used for MR and rebuilding. This was performed with AMPLE invoking the -ensembles flag without any further modification. Structural superpositions were performed with TM-align (Zhang & Skolnick, 2005).

Sequence-identical NMR ensembles and target crystal structures
Previous work had shown a good success rate (22 from 25) using the FindCore program to prepare MR search models from NMR ensembles sharing 100% sequence identity (Mao et al., 2011). In the same work, Rosetta refinement of the NMR ensemble prior to MR solved a further two cases. As Supplementary Table S1  AMPLE processes NMR structures into successful search models of various sizes. The structure of E. coli thioredoxin (PDB entry 1xoa; Jeng et al., 1994) yields successful search models to solve the crystal structure of S. coelicolor thioredoxin (PDB entry 1t00) containing, for example, (a) 108 residues (untruncated) retaining all side chains, (b) 60 residues with side chains trimmed to C and (c) 15 residues with only selected side chains retained. A stereo comparison of the 1xoa ensemble (green) and the target crystal structure, S. coelicolor thioredoxin (PDB entry 1t00; magenta), is shown in (d). The figure was produced using PyMOL (http://www.pymol.org).
Rosetta remodelling, performs similarly well with this test set. Truncation alone in AMPLE solves 19 of 22 successes of FindCore to the point of automatic tracing in SHELXE, while Rosetta remodelling leads to success for the same additional two cases. In three cases previously successfully solved by FindCore the AMPLE pipeline failed: for these, diffraction data to only 2.4-2.5 Å resolution were available, which is at the limit of the range in which SHELXE is reliable.

Thioredoxin-fold test cases of non-sequence-identical NMR models
The successes of FindCore and AMPLE on a set of thioredoxin-fold NMR structures with various levels of sequence and structural similarity to a selected crystallographic target are shown in Table 1 and Fig. 1. Four cases (the easiest, with sequence identities of >49% and r.m.s.d. values of <1.6 Å ) were solved with both programs. Straightforward truncation of the NMR ensembles was successful with AMPLE alone for two cases with very low sequence identity (<20%) but moderate structural conservation (1.7-2.1 Å r.m.s.d.). An additional remodelling step prior to the clustering and truncation protocol of AMPLE allowed the solution of three cases that were around 25% sequence-identical to the target but ranged widely in their structural difference from it. These include a case in which the r.m.s.d. was very high at 3.30 Å . One case, PDB entry 2b5x (Zhang et al., 2006), with the lowest sequence identity (16%) was not soluble, even with the remodelling. Interestingly, this ensemble contained only 11 conformers, the fewest among the NMR structures used: it remains to be seen whether this relative lack of sampling of structural space contributed to its failure.
As described previously (Bibby et al., 2012), successful AMPLE-derived ensembles ranged broadly in size (Table 1; Figs. 2 and 3). The smallest, derived from Escherichia coli thioredoxin, contained 15 residues, which was only 14% of the NMR structure. Below this, presumably, even extremely accurate search models contain too little phasing information for success. The largest was 133 residues from human protein disulfide-isomerase A6 domain 2. Also as described previously, there is a correlation between the r.m.s.d. of the search model versus the target and successful search-model size (Fig. 3): for both the simple truncations and the remodelling cases a larger r.m.s.d. is tolerated for larger search models, whereas smaller search models must be more accurate for success.

Discussion
We tested the cluster-and-truncate methods of AMPLE on NMR structures even though they were specifically developed and optimized to process a very different type of structure: ab initio protein models. The comparison with recent work using the FindCore program to process NMR ensembles is illustrative (Mao et al., 2011). With a set of sequence-identical test cases performance is very similar, but a current limitation of AMPLE leads to failure in three cases that were solved with FindCore. We ascribe this to the resolution of the data available in these cases of 2.4-2.5 Å , which is at the limit of the capabilities of SHELXE. Thus, although SHELXE is a very powerful and convenient tool, particularly for its ability to distinguish correct MR solutions using a reliable statistic, it can constrain the success of AMPLE as a whole in some cases. As well as its resolution limits, its much better performance with -helical proteins compared with all-proteins must also be borne in mind. Future development of AMPLE will allow a case-dependent choice of rebuilding tool. The performance of AMPLE in the thioredoxin test set was very encouraging, solving cases with low sequence identity (18%) and/or high structural divergence from the target (3.3 Å r.m.s.d.). Although based on a single fold and calling for further confirmation, these results compare very well with FindCore, which only solved the thioredoxin structure with NMR structures of >49% sequence identity. This is in line with previous FindCore results, in which structures with >40% sequence identity were solved routinely but those with <30% sequence identity were solved only rarely (Mao et al., 2011). The broad positive correlation seen in Fig. 3   Success and failure of structure solution as a function of search-model size and r.m.s.d. difference from the target. Structure comparisons were performed with TM-align (Zhang & Skolnick, 2005), aligning a majority of the structures with a mean ten residues of the search model left unaligned. Search models relating to cases solved by simple truncation are shown as red diamonds and those relating to cases solved by rebuilding are shown as blue triangles. Filled symbols indicate successful search models. Grey circles indicate search models for PDB entry 2b5x, the single unsolved case. the available NMR structure and the target crystal structure and hence those that are the most advantageous to remove. It is likely that more extensive sampling also contributes to the additional success of AMPLE. Applied to ab initio models, AMPLE can generate up to 500 or so search models per case. The numbers were smaller here since AMPLE processes the three largest clusters of ab initio models while the NMR ensemble was treated here as a single cluster: the number of search models per case here ranged from 183 to 213. The benefits of sampling a range of sizes and side-chain treatments are graphically illustrated by the single success obtained using the tryparedoxin NMR structure (PDB entry 1okd; Krumme et al., 2003; Table 1). The unique successful search model was 105 residues long and had all side chains cropped back to C .
Taken as a whole, the results from the application of AMPLE to NMR structures are already very promising and suggest that it is a useful alternative to FindCore or manual processing. In particular, there are clear suggestions that it can extend success to harder cases of lower sequence identity and structural similarity between NMR structure and target crystal structure (Table 1, Fig. 1). Encouragingly, there are obvious possibilities to improve the performance further. At present, the side-chain methods in AMPLE are tailored to the ab initio model scenario: an explicit consideration of side-chain variability in the NMR ensemble would allow a better treatment in the resultant search models. For example, only those side chains that are experimentally poorly defined could be eliminated. Such protocols will be implemented in future versions of AMPLE.
We used NMR ensemble validation (Doreleijers, Sousa da Silva et al., 2012) to assess whether structural quality could be limiting MR performance in some cases. The validation, based on residue-level stereochemical analyses, results in moleculelevel quality ROG ratings of red (lower), orange (intermediate) or green (better). Although the numbers are too small to draw firm conclusions, there are hints that red-rated or orange-rated ensembles are less prone to solve crystal structures straightforwardly. In the comparison with FindCore (Supplementary Table S1), most NMR ensembles are of high structural quality (green) and can typically be solved, without Rosetta refinement, using either FindCore or AMPLE. Of the two 'red' ensembles, one solves straightforwardly and the other requires Rosetta rebuilding for success with both Find-Core and AMPLE. The single case that does not solve with either program, even with Rosetta rebuilding, is 'orange'. The thioredoxin cases (Table 1) are harder to interpret since the percentage sequence identity between NMR ensembles and crystal structure varies, but within the nine AMPLE successes two of the three in which Rosetta rebuilding was required are 'orange', whereas only one 'orange' ensemble solved the target without rebuilding. If confirmed, this suggests that future improvements in NMR methodology and consequently ensemble quality would feed through into improved performance in MR. Also interestingly, the single structure that failed to solve (PDB entry 2b5x), although 'green', contains a minimized average structure in its ensemble, a practice that is now deprecated, and is the only structure in the set to do so.
In conclusion, we have previously shown that the clusterand-truncate methodology is an effective tool for processing ab initio models, and in the current article we have shown that it is also powerful for processing NMR ensembles. This central idea can also be applied to other scenarios, and we are currently investigating its use in completing partial MR solutions and its application to specific structural classes such as transmembrane domains and coiled-coil proteins.
APPENDIX A Using the AMPLE software AMPLE has several modes of operation. As well as allowing the user to have some control over its procedures, these different options allow different approaches to solving a particular molecular-replacement problem. Some of this functionality is exposed in the CCP4i interface to AMPLE, but the full range of options is only available from the command line. Here, we give an overview of the protocols in AMPLE along with a brief description of how to interpret the output of the program. For detailed user documentation on AMPLE, the reader is directed to the CCP4 wiki site at http:// ccp4wiki.org.

A1. Basic procedure
The primary function of the program is to create or receive as input ab initio models ('decoys') and to prepare them for use as MR search models. Currently, the program can use the Rosetta package to produce these decoys given the target sequence. The decoys are assembled from fragments that can be derived locally if additional programs and databases are installed (Gront et al., 2011). Alternatively, fragments can be obtained from the Robetta web server (Kim et al., 2004; http:// www.robetta.org) and supplied to AMPLE. Decoy models can also be obtained by the user from other programs such as QUARK (Xu & Zhang, 2012). AMPLE accepts these from the user and subjects them to the MR search-model preparation procedures. In all cases, the user must provide the structurefactor amplitudes in the form of an MTZ file for use in the MR step.

A2. Accepting and remodelling externally provided structures
In order to allow maximum flexibility, AMPLE can be directed to retrieve user-provided models from a given directory. This allows AMPLE to work with NMR ensembles, as described above, but also to accept models that the user may have obtained in other ways. Prior to the generation of search models, AMPLE can be directed to carry out Rosetta remodelling. Using a fragment-based technique, Rosetta can repeatedly remodel and refine an input structure to produce an ensemble of structures that can be dealt with by clustering and truncation.

A3. Missing domains
As mentioned above, AMPLE can be used to create search models for locating missing domains in cases where one or more components have already been found. This procedure can take advantage of the information provided by the existing model regarding the separation of the termini of the missing domain. Rosetta uses this information to restrain the distance between the termini of the decoy models. The user specifies the value for this distance along with the model for the already known part of the target structure.

A4. Interpreting the AMPLE output
The MR preparation procedures in AMPLE can create hundreds of search models. In many of our test cases, such a broad sampling of parameter space was necessary since only a small number of combinations of clustering, truncation and side-chain treatment produced successful search models. As a result, running AMPLE can take several days of CPU time on a single machine. It is possible to speed up the processing by taking advantage of multiple cores or by submitting the underlying decoy-generation and MR steps to a compute cluster (currently, Sun Grid Engine is supported). An early indication of the likelihood of success is given by the results of the decoy-clustering step and this is reported by AMPLE. A large top cluster consisting of many more decoys than the subsequent clusters is indicative of potentially more accurate ab initio modelling, which can be expected to result in more successful search models. Once the program enters the molecular-replacement stage, the user is presented with a summarized table of the results for each of the search models that have completed their processing. Refinement statistics along with CC scores from SHELXE (where used) are presented and ordered according to the final R free value after 30 cycles of restrained refinement of the MR solution in REFMAC (Murshudov et al., 2011).
The field of ab initio modelling is developing rapidly: improvements in accuracy (Xu & Zhang, 2012) and, excitingly, in the size of the target that can potentially be addressed (see, for example, Karakaş et al., 2012) have recently been made. In addition, developments in the area of molecular-replacement, phase-improvement and model-building software continue apace. We anticipate that these developments will help AMPLE to become an increasingly valuable tool in structure solution and make it applicable to an ever-broadening range of target structures. However, users of AMPLE should be aware of the current limitations of the program. To date, we have only been able to determine structures with experimental data to 2.5 Å resolution or better and the ab initio protocol is highly unlikely to work on targets that are longer than around 130 residues since this is the upper size limit accessible to the modelling.
Special thanks go to Randy Read and Airlie McCoy for fruitful discussions on Phaser, to Andrey Lebedev for his expert advice on molecular replacement, and to George Sheldrick, Andrea Thorn and Tim Gruene for their help with