Automated Structure Solution With autoSHARP
Clemens Vonrhein, Eric Blanc, Pietro Roversi, and Grard Bricogne

Summary

We present here the automated structure solution pipeline autoSHARP. It is built around the heavy-atom renement and phasing program SHARP, the density modication program SOLOMON, and the ARP/wARP package for automated model building and renement (using REFMAC). It allows fully automated structure solution, from merged reection data to an initial model, without any user intervention. We describe and discuss the preparation of the user input, the data ow through the pipeline, and the various results obtained throughout the procedure. Key Words: SHARP; autoSHARP; automation; structure solution.
1. Introduction Automation of crystal structure determination, from processed data to an initial macromolecular model, is desirable in two areas: in the high-throughput structural genomics efforts, and in making the power of todays most sophisticated methods accessible to the structural biologist without much crystallographic experience. This requires user-friendliness both in dening the experimental phasing protocol and in presenting and commenting the results, and also in the tracking of work ow. Here we present a system (autoSHARP) that tries to accommodate the novice user as well as the expert. 2. Materials The automatic structure solution pipeline autoSHARP is a set of computer programs and scripts designed to run on a variety of modern, UNIX-like computers. The main autoSHARP tasks are written in Bourne shell (1) and Perl (2) with some helper applications used as system-specic binaries. Further details on requirements and installation instructions can be found at our website http://www.globalphasing.com/sharp/. The development of autoSHARP started
From: Methods in Molecular Biology, vol. 364: Macromolecular Crystallography Protocols: Volume 2: Structure Determination Edited by: S. Doubli Humana Press Inc., Totowa, NJ

Vonrhein et al.

Fig. 1. Flowchart of autoSHARP procedure.
in 2000, with the rst public release in January 2002 and an active user base with more than 1500 installations at the date of writing. 3. Methods The methods described next outline the various steps for successfully running the automatic structure solution pipeline in autoSHARP. An on-line manual is also available at: http://www.globalphasing.com/sharp/. The steps include (1) preparation of the input data, (2) steps performed by the program, (3) analysis of the program output, and (4) presentation of results. The logic follows closely the structure solution process as shown in the owchart of Fig. 1.
Fig. 2. User input form for SIR(AS) experiment.
3.1. Input In order to start autoSHARP some mandatory and some optional information has to be provided for each wavelength in a SAD/MAD experiment, or for each derivative in a SIR(AS)/MIR(AS) experiment (see Fig. 2). Although autoSHARP tries to apply sensible defaults whenever it encounters a missing input value, the automatic structure solution process (and the decisions that
need to be taken within this process) will be only as good as the starting information given to the program. 3.1.1. Mandatory Information This information needs to be supplied or given as part of the le header for reection data.

3.1.1.1. HEAVY-ATOM TYPE

Usually, only one type of heavy atom is present in the crystal (e.g., selenium [Se] or mercury [Hg]). In some cases, however, a metal ion might be bound to the macromolecule (e.g., calcium [Ca2+]) or a cofactor contains a heavy atom (e.g., a heme group with an iron [Fe] atom). Although autoSHARP can handle a list of different heavy atoms it is usually enough to dene the heavy atom that will be most visible during the specic experiment (when collecting the peak wavelength of a Se-MAD experiment this will be the Se atomseven if in nal maps the sulfur atoms of cysteine residues might be visible through their weak contribution to anomalous scattering) (see Note 1).

3.1.1.2. NUMBER

EXPECTED HEAVY ATOMS
The exact number of heavy atoms bound to the macromolecule is usually not known a prioriunless the number of molecules in the asymmetric unit is known and the heavy atom is an intrinsic part of the macromolecule (selenomethionine [Se-Met], sulfurs of Met and/or Cys residues, metals of cofactors, and so on). In general, autoSHARP will dynamically adjust the number of heavy atoms actually used for phasingup to a maximum limit given by the user. Therefore, overestimating the number of heavy atoms usually has no negative effect on the performance (see Note 2).

3.1.1.3. SPACE GROUP

This will be picked up automatically from the header of the reection le. However, quite often data are processed in a space group without any of the possible screw axes (e.g., P222 instead of P212121), in order to check for systematic absences along the various axes. The data les going into autoSHARP should have the correct space group in their header. If no data were collected along a potential screw axis, autoSHARP should probably be run with and without the screw-axis component. The ambiguity in handedness of a screw axis (41 vs 43) will be taken into account automatically and does not need to be checked.

3.1.1.4. WAVELENGTH

In order to dene the scattering properties of the heavy atom for a given experiment, the various wavelengths at which data were collected need to be
known. They will be used for calculating f and f values for all heavy atoms declared (see Note 3).

3.1.1.5. REFLECTION DATA

The MTZ (3) and SCALEPACK (4) reection le formats are supported. Other formats can easily be converted into one of these widely used ones (see Note 4). 3.1.2. Optional Input Additionally, the following information is usually available, and giving it to autoSHARP will improve its efciency.

3.1.2.1. CONTENT

ASYMMETRIC UNIT
This can be given either as molecular weight, number of residues, or in the form of a sequence le. If the asymmetric unit can accommodate more than one copy of the macromolecule, but the exact number is not yet known, it is recommended to specify this sequence for a single molecule only. The use of a sequence le is preferred, because all other parameters can be calculated from this (see Note 5).

3.1.2.2. F AND F VALUES

Although these can be calculated from the knowledge of the heavy atoms chemical type and the wavelength (5), the actual value observed during the experiment can vary substantially from the theoretical value when collecting data close to the absorption edge of an element. This may be caused by the specic environment of the anomalous scatterer, and also by the spread of wavelengths around its mean value in the X-ray beam. Therefore, it is always advisable to perform a uorescence scan on the same (or very similar) crystal as was used for data collection (see Note 6). 3.2. Checks The rst step within the autoSHARP procedure is to check the user input for syntax errors. Any mandatory input parameter that is missing will be reported and will result in the program halting after all other checks have been performed. 3.3. Data Analysis During the next few steps, the program performs as much data analysis as possible. 3.3.1. Crystal-Independent Analysis The autoSHARP scripts carry out the following tasks:
1. Calculation of molecular weight and the number of residues based on a sequence le (if given).

2. Calculation of f and f values (in case only the wavelength was given) using the program CROSSEC (3). 3. Analysis of f and f values for a MAD experiment to automatically determine which of the datasets belongs to peak, inection, high-energy remote, or low-energy remote. 4. In case of sulfur or Se-Met phasing, a comparison between the specied number of heavy atoms and the number of Cys/Met residues in the sequence le is done.
3.3.2. Crystal- and Dataset-Dependent Analysis For each crystal and each dataset pertaining to that crystal, the following tasks are performed:
1. Extraction of cell parameters, space group name (and number) from the data les, and a test for consistency of these values between datasets. 2. Determination of resolution limits of each dataset, as well as overall and common resolution limits. 3. Estimation of overall temperature factor by using the least-squares line t for a Wilson plot (6) (see Note 7). 4. Estimation of the most likely number of monomers per asymmetric unit, as well as the possible range for this number using the Matthews coefcient (7).
3.3.3. Dataset-Dependent Analysis For each dataset, the autoSHARP scripts carry out the following analysis:
1. Determination of overall and anomalous completeness per dataset (see Note 8). 2. Detection and removal of anomalous difference outliers. By analyzing the value of |ano|/F, those reections with unusually large values are reported. Reection where this value is larger than 1.9 are also removed from the dataset. Unless a very strong anomalous signal is expected these nearly always point back to problems during data processing (see Note 9). 3. Calculation of self-rotation functions (8), assuming a globular shape of the macromolecule and a radius based on the size of the monomer. These are plotted with the program POLARRFN (3) to help dene possible noncrystallographic symmetry (NCS). This NCS information is at the moment not actively used within autoSHARP, but it will be used in future versions to impose constraints on the heavy-atom substructure or to perform real-space averaging during density modication. 4. Calculation and analysis of the native Patterson function (9,10). This can bring to light the presence of purely translational NCS, an extreme case of which could lead to a mistake in the assignment of the correct space group where a centering operation has been missed.
3.4. Scaling Apart from SAD experiments, other types of structure solution protocols (SIR[AS], MIR[AS], MAD) require more than one dataset. These need to be scaled relative to each other, which will give additional information such as:
1. Outlier analysis based on normalized structure factors. Reections for which the normalized structure factor (E-value) in different datasets differs by more than ve are reported. Furthermore, reections for which two values differ by more than 10 are deleted from the dataset (see Note 10). 2. Outlier analysis of isomorphous/dispersive and anomalous differences. During the scaling of two datasets using either Krauts method (11) as implemented in FHSCAL (3) or a simple anisotropic scaling model as implemented in SCALEIT (3), reections with unlikely large differences are detected and an appropriate cutoff value determined. 3. Cross-table of R-factor values between all dataset pairs within a common overall resolution range. 4. Table of correlation coefcients between ano values for all dataset pairs in case of a MAD experiment.

All of the previously listed statistics are used for determining a reasonable high-resolution limit for the heavy-atom detection step. Furthermore, in cases of signicantly larger R-factors in the lowest resolution shell the low-resolution limit will also be restricted. 3.5. Heavy-Atom Detection Several powerful programs for heavy-atom substructure detection are availablelike SHELXD (12), Shake-and-Bake (SnB) (13), HySS (14), SOLVE (15), or CRUNCH2 (16)the procedures for heavy-atom substructure solution currently implemented in autoSHARP use either SHELXD or RANTAN (17). However, solutions from other programs can be input through a simple le, in either Protein Data Bank (PDB) (18) or in our internal fractional coordinates format. The heavy-atom substructure for a given dataset (or a collection of datasets in the case of MAD) can be determined from a variety of data:
1. Anomalous differences measured for each dataset. 2. Isomorphous differences between a derivative and a native dataset. 3. Dispersive differences between different wavelengths of a MAD experiment, especially between inection and remote wavelengths.
The strategy for nding a substructure solution is:
1. When using RANTAN: each type of difference is converted into normalized differences using the program ECALC (3) and is given to RANTAN, which then generates three sets of phases that are most consistent with the data. Using these phases and the input E-values, a real-space map can be calculated which should show peaks at the heavy-atom positions. For further checking, the list of peaks is analyzed for consistency with an origin-removed Patterson function (calculated using [E2-1] coefcients based on the same E-values as are input into RANTAN). Furthermore, any peaks on special positions or in unusual relation to each other (e.g., several peaks differing only in the coordinate along a polar axis) are removed. The resulting list is sorted by peak height and used for further analysis.
For MAD experiments, the optimized value of normalized anomalous scattering as calculated by the program REVISE (19) can be used in the same manner for the determination of the heavy-atom substructure. Because any solution from the direct methods program RANTAN is automatically checked against the corresponding difference Patterson map (anomalous, isomorphous, or dispersive), the space group-specic Harker sections of this Patterson map are plotted as well. With experience it is very easy to judge the quality of the signal visually from these plots. 2. When using SHELXD: different types of protocols (MAD, SIR[AS], SAD) are used to generate a set of FA-values using SHELXC (20). To not only assess the correctness of a substructure solution, but also to allow for possible errors in the initial estimation of the number of heavy atoms, the top N peaks of this list are used in a stepwise manner (where the cut-off value on peak height is slowly lowered to include more and more lower scoring peaks in the analysis). These peaks are converted into atoms of the given heavy atom type, with a temperature factor based on the Wilson plot (determined in Subheading 3.3.2.) and an occupancy dependent on both the type of experiment and the peak height (in case of a Se- or S-MAD/SAD experiments all atoms will have the same occupancy). A correlation coefcient between the (E21) values (for RANTAN) or the FA-values (for SHELXD) of the observed data and the heavy-atom structure factor amplitudes calculated from this set of atoms is used as a score to judge the quality of the substructure solution. This correlation coefcient can either use all reections or only a test set of reections, which were not used for detecting the substructure (similar to the free R-value as introduced by ref. 21). In case of a MIR(AS) experiment, the various lists of heavy-atom sites for each derivative cannot be used together directly, because each substructure may be dened with a different origin and enantiomorph. Therefore, the solution with the best correlation coefcient is used as a starting point (effectively starting as a SIR[AS] experiment).

3.6. Heavy-Atom Renement Once an initial set of heavy atoms is available, it is given to SHARP (22) for renement. SHARP will rst estimate the absolute scale (based on the analysis of asymmetric unit content done previously) so that the initial heavy-atom occupancies are on a meaningful scale. Also, an overall anisotropic temperature factor is estimated to accommodate any anisotropy that might be present in the data. If more than a single dataset is used, the scale and temperature factors relative to the rst dataset (the rst wavelength in a MAD experiment, or the native dataset in a SIR[AS] or MIR[AS] experiment) are estimated. SHARP will then rene coordinates, occupancy, and temperature factors for the heavy-atom sites, together with scaling parameters and nonisomorphism parameters for each dataset. This renement is done in a stepwise manner, rst rening the most critical parameters, then adding more parameters until all are rened
together to convergence. Coefcients for a map of the nal log-likelihood gradient (LLG) are calculated for further analysis of the current model (see ref. 22). 3.7. Heavy-Atom Phasing Once the heavy-atom renement in SHARP has converged, a set of phases and map coefcients is calculated. This also includes HendricksonLattman coefcients (23) for subsequent phase combination in density modication programs. Furthermore, additional information about the two-dimensional phase probability for each reection is output (24). The main statistical descriptors regarding heavy-atom phasing are:
1. Mean phasing power in resolution shells for each derivative (separate for isomorphous/dispersive and anomalous differences). 2. Cullis R-factor (25) in resolution shells for each derivative (separate for isomorphous/dispersive and anomalous differences). 3. Mean gure-of-merit values within resolution shells (separate for centric and acentric reections).
3.8. Analysis of LLG Maps Once convergence has been reached during the renement of a given heavyatom model, SHARP produces maps for the calculated heavy-atom density and for the LLG corresponding to each dataset. The nal heavy-atom parameters and the associated maps are then analyzed in order to:
1. Detect heavy-atom sites for which the rened parameters indicate a nonexistent site (i.e., very low occupancy, very large temperature factor, or no signicant density in the calculated model map): these are deleted from the current model. 2. Detect new and additional heavy-atom sites, where the LLG maps have signicant positive density within a minimum distance of already existing sites; these will be automatically added. 3. Detect cases where the anomalous LLG map shows signicant positive or negative features directly at most of the current heavy atom positions: for these datasets the renement of f-values will be switched on.

The current heavy-atom model is then updated by deletion and/or addition of sites as well as update or renement protocol, and another heavy-atom parameter renement with SHARP is run (see Note 1). This is done automatically, but a detailed description of the decisions made is presented to the user, who has the option of overriding them through the graphical user interface. 3.9. Determination of Hand Once no further update of the heavy-atom model seems necessary, a nal set of phases is calculated in both hands. For this purpose, the original heavy-atom conguration is inverted (mostly around the origin accompanied by a switch to
the enantimorph space groupbut for I41, I4122, and F4132 the inversion is around a different point without the need to change the space group). The two phase sets are then analyzed to determine which heavy-atom conguration (and space group) is correct. The criterion for a correct set of phases is that the resulting electron density map has features that are consistent with a macromolecule. Because exactly the same criterion is used in density modication methods (26), the latter can be used to decide on the correct hand. For that purpose, a single cycle of solvent ipping with SOLOMON (27) is performed for each phase set, and the resulting statistics are used to decide which is the correct hand. Two criteria can be used as a basis for this decision:
1. The value of the correlation coefcient between observed E2 values and E2 values based on the modied map (which should be higher for the correct hand). 2. The value of the ratio between the standard deviation in the solvent region (which should have a low value because this region contains mainly disordered solvent) and the standard deviation in the protein region (which should be high because of well-dened, positive features).
3.10. Density Modication In order to have the best electron density map for building and analyzing the crystal structure, density modication procedures are used for phase improvement. Here, the solvent ipping method as implemented in SOLOMON (27) is appliedtaking into account some specic features of the two-dimensional probability information coming from SHARP (24). Because even at this stage the exact number of molecules per asymmetric unit might not be known precisely, a simple optimization of the solvent content (rst stepwise and nally through a parabolic t) is performed. The resulting map should show all molecules within the asymmetric unit, without the danger of attening density for unexpected molecules or articially including density of the disordered solvent in the protein region. 3.11. Automatic Building The best density modied map is nally handed over to the ARP/wARP suite of programs for automatic model building (28). The protocol used within this program suite takes into account the available experimental phase information from SHARP (in form of HendricksonLattman coefcients) as well as the possible presence of heavy atoms in the macromolecule (in case of SAD or MAD experiments). 3.12. Results From autoSHARP Results are produced by autoSHARP at various stages within the automated pipeline (see Fig. 3): if the data resolution is sufcient, a PDB le of a partially built model will often be produced.

Automated Structure Solution With autoSHARP 225
Fig. 3. Main autoSHARP output le for SIR(AS) experiments.
226 The main categories of results are:
1. A rened heavy-atom model, comprising: (1) position, occupancy, and temperature factor of the heavy-atom sites, (2) scaling parameters (including anisotropic temperature factors), and (3) nonisomorphism parameters. 2. LLG or residual maps, which can be further analyzed to ne-tune the renement parameters, e.g., typically to introduce anisotropic thermal parameters for certain sites. 3. Electron density maps from heavy-atom phasing, density modication, and automatic building. 4. PDB le of a partially built model.
There are separate graphical user interfaces available to: (1) change the parameterization of the heavy-atom renement model, (2) analyze and view LLG maps, (3) view electron density maps, (4) change parameters for density modication, (5) perform automatic model building with ARP/wARP, and (6) detect NCS with GETAX (29). 4. Notes
1. It can be very helpful to check the features of heavy-atom sites in the nal LLG maps, especially the ones based on anomalous differences. Different heavy atoms should have different scattering properties (f), which should be visible there. However, a wrongly assigned heavy atom type for a specic site might be accommodated for through an occupancy value that renes to very low or very high values. Furthermore, the heavy-atom positions in sulfur or Se-Met-phasing experiments can be used to help assign the correct sequence to the macromolecular chain during manual model building. Finally, during the nal stages of structure renement and model completion, when the solvent structure is being completed (either automatically or by hand), the difcult differentiation between a water atom or a bound metal ion can be helped by checking the anomalous residual (LLG) or Fourier maps at this position. 2. In cases where the actual content of the asymmetric unit is unknown and could vary substantially (within a sensible value for the solvent content resulting from a given number of monomers), it is recommended to give the asymmetric unit content for a single copy of the macromolecule, as well as the number of heavy atoms expected to be present for a monomer. autoSHARP will adjust the actual number of molecules during structure solution at two stages: initially during data analysis (to have a reasonable solvent content of around 50%) and during the density modication step (where the optimization of the solvent content should give a better estimate of the actual number of molecules in the asymmetric unit). 3. Accurate values for f and f are needed for a MAD experiment because only a single occupancy and temperature factor for each site (which is shared between the different datasets) is being rened. In other types of experiments (SAD, SIR[AS], or MIR[AS]) any error in these values will probably be compensated by the renement of occupancy and temperature factor. However, if these parameters are far

from their true values, the starting occupancy at the rst renement cycle can be quite wrong (which is additionally inuenced by the data not being on absolute scale, when the correct number of molecules is not known). The reection data used within autoSHARP should have been processed as thoughtfully as possible. In particular, the treatment of resolution limits has to be done carefully: including high-resolution data with hardly any signal can give difculties during the heavy-atom detection because the direct methods program RANTAN uses normalized structure factors (E-values) so that no automatic down-weighting of weak and noisy data in the outer resolution shells occurs. Furthermore, errors in the handling of the area behind the beam stop during the integration of diffraction images can lead to large errors in low-resolution strong reections, hence, to problems in the detection of outliers during internal scaling or merging of data. A substantial part of the statistics and analyses tries to deal with these problematic reections (or resolution shells). Any warning message from autoSHARP about data quality usually points to problems during data integration and scaling/merging effects that autoSHARP can only try to diminish but is unable to correct. In cases where the asymmetric unit can accommodate a large range of numbers of molecules, some additional information is usually available to help decide on an appropriate value. Crystals that diffract only to low resolution quite often have a large solvent content or a loose packing of few molecules. On the other hand, welldiffracting crystals can have very tight packing of several molecules. Moreover, some biochemical data might be available to suggest an oligomerization state in vivoand the asymmetric unit (maybe in conjunction with special symmetry elements) should probably take this into account. If no uorescence scan was performed, then some standard values for the typical wavelengths at which to perform a MAD experiment (e.g., Se-Met) can be input. These can usually be obtained from the beam line scientist. Using a more uncommon heavy atom with more complicated absorption edges will require the measurement of a uorescence scan which can be analyzed, e.g., with the program CHOOCH (30). The overall temperature factor obtained from a Wilson plot is used as a starting value for the temperature factor of newly found heavy-atom sites (additional sites that are found through analysis of the LLG maps will have the mean temperature of all already existing heavy-atom sites within this dataset). Although heavy atoms introduced through soaking of the crystal (e.g., Hg or Pt soaks) usually have a much larger temperature factor, it seems a good starting point. During data collection, the exposure time, crystal-to-detector distance, oscillation angle, and strategy can easily be adapted to give a nearly complete dataset. If there are time restraints during data collection it might be better to collect a complete, lower resolution dataset (by increasing detector-to-crystal distance, which will allows larger oscillation ranges and/or shorter exposure) than a high resolution but incomplete dataset. This is especially true when the missing data are not randomly distributed, but rather a wedge of data is missing. This not only can lead to streaks in the resulting electron density maps, but also makes the analysis of LLG

maps more difcult. The estimation of overall anisotropic temperature factors can give wrong results because this determination can be ill determined. 9. As with many of these statistics, any outlier or unlikely value for low-resolution reections should be seen as suspicious. Usually, low-resolution reections are strongest and should have been integrated most accurately (whereas at the outer diffraction limit reections are weak and show greater noise). However, if some systematic errors were introduced during data processing, these strong reections not only can have a wrong value, but their corresponding variance might be wrongly estimatedleading to a wrong weighting of these reections during maximumlikelihood calculations as those performed within SHARP. 10. If reections within a very narrow resolution range are listed here, it might be because an ice-ring hasnt been processed appropriately for one of the datasets. Also, a larger number of rejections at very low resolution can again show problems with the data processing and the treatment of the beamstop.
Acknowledgments The authors wish to acknowledge partial nancial support for this work from European Commission Grant no. QLRI-CT-2000-00398 within the AUTOSTRUCT project.

References

1. Bourne, S. R. (1978) Unix time-sharing system: The Unix shell. Bell Sys. Tech. J. 57, 19711990. 2. Wall, L., Christiansen, T., and Orwant, J. (2000) Programming Perl, 3rd ed., OReilly and Associates, Inc., Sebastopol, CA. 3. Collaborative Computational Project, Number 4 (1994) The CCP4 suite: programs for protein crystallography. Acta Cryst. D50, 760763. 4. Otwinowski, Z. and Minor, W. (1997) Methods in Enzymology, Vol. 276, Macromolecular Crystallography, Part A, (Carter, C. W., Jr and Sweet, R. M., eds.), Academic Press, New York, NY, pp. 307326. 5. Cromer, D. T. (1983) Calculation of anomalous scattering factors at arbitrary wavelengths. J. Appl. Cryst. 16, 437438. 6. Wilson, A. J. C. (1942) Determination of absolute from relative X-ray intensity. Nature 150, 151152. 7. Matthews, B. W. (1968) The solvent content of protein crystals. J. Mol. Biol. 33, 491497. 8. Rossmann, M. G. and Blow, D. M. (1962) The detection of sub-units within the crystallographic asymmetric unit. Acta Cryst. 15, 2431. 9. Patterson, A. L. (1934) A Fourier series method for the determination of the components of interatomic distances in crystals. Phys. Rev. 46, 372376. 10. Eagles, P. A. M., Johnson, L. N., Joynson, M. A., McMurray, C. H., and Gutfreund, H. (1969) Subunit structure of aldolase: chemical and crystallographic evidence. J. Mol. Biol. 45, 533544.

11. Kraut, J., Sieker, L. C., High, D. F., and Freer, S. T. (1962) Chymotrypsinogen: a three-dimensional Fourier syntheis at 5AA resolution. Proc. Nat. Acad. Sci.USA 48, 14171424. 12. Schneider, T. R. and Sheldrick, G. M. (2002) Substructure solution with SHELXD. Acta Cryst. D58, 17721779. 13. Weeks, C. M. and Miller, R. (1999) The design and implementation of SnB v2.0. J. Appl. Cryst. 32, 120124. 14. Grosse-Kunstleve, R. W. and Adams, P. D. (2003). Substructure search procedures for macromolecular structures. Acta Cryst. D59, 19661973. 15. Terwilliger, T. C. (2003) Automated structure solution, density modication and model building. Acta Cryst. D58, 19371940. 16. Ness, S. R., de Graaff, R. A. G., Abrahams, J. P., and Pannu, N. S. (2004). Crank: new methods for automated macromolecular crystal structure solution. Structure 12, 17531761. 17. Yao, J. (1981) On the application of phase relationships to complex structures. XVIII. RANTAN-random MULTAN. Acta. Cryst. A37, 642644. 18. Westbrook, J. and Fitzgerald, P. M. (2003) The PDB format, mmCIF formats and other data formats. In: Structural Bioinformatics (Bourne, P. E. and Weissig, H., eds.), John Wiley and Sons, Inc., Hoboken, NJ, pp. 161179. 19. Fan, H.-F., Woolfson, M. M., and Yao, J.-X. (1993) New techniques of applying multi-wavelength anomalous scattering data. Proc. R. Soc. Lond. A 442, 1332. 20. Sheldrick, G. M., Hauptman, H. A., Weeks, C. M., Miller, R., and Uson, I. (2001) Direct methods: ab initio phasing. In: International Tables for Macromolecular Crystallography, Vol. F (Rossmann, M. G., and Arnold, E., eds.), Kluwer Academic Publishers, Dordrecht, Germany, pp. 333345. 21. Brnger, A. T. (1997) Free R value: cross-validation in crystallography. Meth. Enzym. 277, 366396. 22. De La Fortelle, E. and Bricogne, G. (1997) Maximum-likelihood heavy-atom parameter renement for the multiple isomorphous replacement and multiwavelength anomalous diffraction methods, Meth. Enzym. 276, 472494. 23. Hendrickson, W. A. and Lattman, E. E. (1970) Representation of phase probability distributions for simplied combination of independent phase informaton. Acta Cryst. B26, 136143. 24. Bricogne, G., Vonrhein, C., Flensburg, C., Schiltz, M., and Paciorek, W. (2003) Generation, representation and ow of phase information in structure determination: recent developments in and around SHARP 2.0. Acta Cryst. D59, 20232030. 25. Cullis, A. F., Muirhead, H., Perutz, M. F., Rossmann, M. G., and North, A. C. T. (1961) The structure of haemoglobin. VIII. A three-dimensional Fourier synthesis at 5.5 resolution: Determination of the phase angles. Proc. Roy. Soc. A 265, 1538. 26. Cowtan, K. D. and Zhang, K. Y. (1999) Density modication for macromolecular phase improvement. Prog. Biophys. Mol. Biol. 72, 245270. 27. Abrahams, J. P. and Leslie, A. G. W. (1996) Methods used in the structure determination of bovine mitochondrial F1 ATPase. Acta Cryst. D52, 3042.

28. Perrakis, A., Morris, R. J., and Lamzin, V. S. (1999) Automated protein model building combined with iterative structure renement. Nature Struct. Biol. 6, 458463. 29. Vonrhein, C. and Schulz, G. E. (1999). Locating proper non-crystallographic symmetry in low-resolution electron-density maps with the program GETAX. Acta Cryst. D55, 225229. 30. Evans, G. and Pettifer, R. F. (2001) CHOOCH: a program for deriving anomalousscattering factors from X-ray uorescence spectra. J. Appl. Cryst. 34, 8286.