research papers
REFMAC5 for the of macromolecular crystal structures
^{a}Structural Biology Laboratory, Department of Chemistry, University of York, Heslington, York YO10 5YW, England,^{b}Biophysical Structural Chemistry, Leiden University, PO Box 9502, 2300 RA Leiden, The Netherlands,^{c}Randall Division of Cell and Molecular Biophysics, New Hunt's House, King's College London, London, England, and ^{d}STFC Daresbury Laboratory, Warrington WA4 4AD, England
^{*}Correspondence email: garib@ysbl.york.ac.uk
This paper describes various components of the macromolecular crystallographic REFMAC5, which is distributed as part of the CCP4 suite. REFMAC5 utilizes different likelihood functions depending on the diffraction data employed (amplitudes or intensities), the presence of and the availability of SAD/SIRAS experimental diffraction data. To ensure chemical and structural integrity of the refined model, REFMAC5 offers several classes of restraints and choices of model parameterization. Reliable models at resolutions at least as low as 4 Å can be achieved thanks to lowresolution tools such as secondarystructure restraints, restraints to known homologous structures, automatic global and local NCS restraints, `jellybody' restraints and the use of novel longrange restraints on atomic displacement parameters (ADPs) based on the Kullback–Leibler divergence. REFMAC5 additionally offers TLS parameterization and, when highresolution data are available, fast of anisotropic ADPs. in the presence of is performed in a fully automated fashion. REFMAC5 is a flexible and highly optimized package that is ideally suited for across the entire resolution spectrum encountered in macromolecular crystallography.
programKeywords: REFMAC5; refinement.
1. Introduction
As a final step in the process of solving a macromolecular crystal (MX) structure, ARP/wARP (Perrakis et al., 1999), SOLVE/RESOLVE (Terwilliger, 2003) and Buccaneer (Cowtan, 2006)] and are of paramount importance in guiding manual model updates using moleculargraphics software [Coot (Emsley & Cowtan, 2004), O (Jones et al., 1991) and XtalView (McRee & Israel, 2008)].
is carried out to maximize the agreement between the model and the Xray data. Model parameters that are optimized in the process include atomic coordinates, atomic displacement parameters (ADPs), scale factors and, in the presence of twin fraction(s). Although procedures are typically designed for the final stages of MX analysis, they are also often used to improve partial models and to calculate the `best' electrondensity maps for further model (re)building. protocols are therefore an essential component of modelbuilding pipelines [The first software tools for MX ). This was followed a few years later by reciprocalspace algorithms for the of individual atomic parameters with added energy (Jack & Levitt, 1978) and restraints (Konnert, 1976) in order to deliver chemically reasonable models. The energy and restraints approaches differ only in terminology as they use similar information and both can be unified using a Bayesian formalism (Murshudov et al., 1997). Early programs used the well established statistical technique of leastsquares residuals with equal weights on all reflections (Press et al., 1992), with gradients and second derivatives (if needed) calculated directly. This changed when Fourier methods, which were developed for smallmolecule structure (Booth, 1946; Cochran, 1948; Cruickshank, 1952, 1956), were formalized for macromolecules (Ten Eyck, 1977; Agarwal, 1978). The use of the FFT for structurefactor and gradient evaluation (Agarwal, 1978) sped up calculations dramatically and the of large molecules using relatively modest computers became realistic. Later, the introduction of (Brünger, 1991), the generalization of the FFT approach for all space groups (Brünger, 1989) and the development of a modular approach to programs (Tronrud et al., 1987) dramatically changed MX solution procedures. Also, the introduction of the very robust and popular smallmolecular program SHELXL (Sheldrick, 2008) to the macromolecular community allowed routine analysis of highresolution MX data, including the of and nonmerohedral twins.
appeared in the 1970s. Realspace using torsionangle parameterization was introduced by Diamond (1971More sophisticated statistical approaches to MX structure ; Ramachandran et al., 1963; Srinivasan & Ramachandran, 1965; see also Srinivasan & Parthasarathy, 1976, and references therein), their implementation for MX started in the middle of the 1990s (Pannu & Read, 1996; Bricogne & Irwin, 1996; Murshudov et al., 1997). It should be emphasized that prior to the application of (ML) techniques in MX the importance of advanced statistical approaches to all stages of MX analysis had been advocated by Bricogne (1997) for two decades. Nowadays, most MX programs offer likelihood targets as an option. Although ML can be very well approximated using the weighted leastsquares approach in the very simple case of against structurefactor amplitudes (Murshudov et al., 1997), ML has the attractive advantage that it is relatively easy (at least theoretically) to generalize for the joint utilization of a variety of sources of observations. For example, it was immediately extended to use experimental phase information (Bricogne, 1997; Murshudov et al., 1997; Pannu et al., 1998). In the last two decades, there have been many developments of likelihood functions towards the exploitation of all available experimental data for thus increasing the reliability of the refined model in the final stages of and improving the electron density used in model building in the early stages of MX analysis (Bricogne, 1997; Skubák et al., 2004, 2009).
started to emerge in the 1990s. Although the basic formulations and most of the necessary probability distributions used in crystallography were developed in the 1950s and 1960s (Luzzati, 1951MX crystallography can now take advantage of highly optimized software packages dealing with all of the various stages of structure solution, including BUSTER/TNT (Blanc et al., 2004), CNS (Brünger et al., 1998), MAIN (Turk, 2008), MOPRO (Guillot et al., 2001), phenix.refine (Adams et al., 2010), REFMAC5 (Murshudov et al., 1997), SHELXL (Sheldrick, 2008) and TNT (Tronrud et al., 1987). While MOPRO was specifically designed for niche ultrahighresolution and is able to model deformation density, all of the other programs can deal with a multitude of MX problems and produce highquality electrondensity maps, although with different emphases and strengths.
There are several programs available that either are designed to perform or offer as an option. These includeThis contribution describes the various components of the macromolecular crystallographic REFMAC5, which is distributed as part of the CCP4 suite (Collaborative Computational Project, Number 4, 1994). REFMAC5 is a flexible and highly optimized package that is ideally suited for across the entire resolution spectrum that is encountered in macromolecular crystallography.
program2. Target functions in REFMAC5
As in all other REFMAC5 has two components: a component utilizing geometry (or prior knowledge) and a component utilizing experimental Xray knowledge,
programs, the target function minimized inwhere f_{total} is the total target function to be minimized, consisting of functions controlling the geometry of the model and the fit of the model parameters to the experimental data, and w is a weight between the relative contributions of these two components. In macromolecular crystallography, the weight is traditionally selected by trial and error. REFMAC5 offers automatic weighting, which is based on the fact that both components are the natural logarithm of a probability distribution. However, this `automatic' weight may lead to unreasonable deviations from ideal geometry (either too tight or too relaxed) in some cases, as the ideal geometry is difficult to describe statistically. For these cases, the weight parameter may need to be selected manually to produce more reasonable geometry, e.g. such that the rootmeansquare deviation of the bond lengths from the ideal values is 0.02 Å and at resolutions lower than 3 Å perhaps even smaller.
From a Bayesian viewpoint (O'Hagan, 1994), these functions have the following probabilistic interpretation (ignoring constants which are irrelevant for minimization purposes):
From this point of view, MX via f_{xray}, i.e. likelihood is a mechanism that controls information flow from the experimental data to the derived model. Consequently, it is important to design a likelihood function that allows optimal information transfer from the data to the derived model. f_{geom} ensures that the derived model is consistent with the presumed chemical and structural knowledge. This function plays the role of regularization, reduction of the effective number of parameters and transfer of known information to the new model. If care is not taken, then wrong information may be transferred to the model; removing the effect of such errors may be difficult if possible at all. The design of such functions should be performed using verifiable invariant information and it should be testable and revisable during the and modelbuilding procedures.
is similar to a well known technique in statistical analysis: maximum posterior (MAP) estimation. The model parameters are linked with the experimental dataFunctions dealing with geometry usually depend only on atomic parameters. We are not aware of any function used in crystallography that deals with the prior geometry probability distributions of overall parameters. A possible reason for the lack of interest in (and necessity of) this type of function may be that, despite popular belief, the statistical problem in crystallography is sufficiently well defined and that the main problems are those of model parameterization and completion.
The existing R and R_{free}).
programs differ in the target functions and optimization techniques used to derive model parameters. Most MX programs use likelihood target functions. However, their form, implementations and parameterizations are different. Therefore, it should not come as a surprise if different programs give (slightly) different results in terms of model parameters, electrondensity maps and reliability factors (such as2.1. Xray component
The Xray likelihood target functions used in REFMAC5 are based on a general multivariate probability distribution of E observations given M model structure factors. This function is derived from a multivariate complex Gaussian distribution of N = E + M structure factors for acentric reflections and from a multivariate real Gaussian distribution for centric reflections and has the following form:
where P = P(F_{1}, …, F_{E}; F_{E+1}, …, F_{N}), F_{i} = F_{i}exp(ια_{i}}, F_{1}, …, F_{E} denote the observed amplitudes, F_{E+1}, …, F_{N} are the model structure factors, C_{N} is the covariance matrix with the elements of its inverse denoted by a_{ij}, C_{M} is the bottom right square submatrix of C_{N} of dimension M with the elements of its inverse denoted by c_{ij}. We define c_{ij} = 0 for i ≤ 0 or j ≤ 0. C_{N} and C_{M} are the determinants of matrices C_{N} and C_{M}, = (α_{1}, …, α_{E}) is the vector of the unknown phases of the observations that need to be integrated and is a probability distribution expressing any prior knowledge about the phases.
In the simplest case of one observation, one model and no prior knowledge about phases, the integral in (3) can be evaluated analytically. In this case, the function follows a Rice distribution (Bricogne & Irwin, 1996), which is a noncentral χ^{2} distribution of F_{o}^{2}/Σ and F_{o}^{2}/2Σ with noncentrality parameters D^{2}F_{c}^{2}/Σ and D^{2}F_{o}^{2}/2Σ with one and two for centric and acentric reflections, respectively (Stuart & Ord, 2009),
where D in its simplest interpretation is 〈cos(Δxs)〉, a Luzzati error parameter (Luzzati, 1952) expressing errors in the positional parameters of the model, F_{c} is the model F_{o} is the observed amplitude of the and Σ is the uncertainty or the second central moment of the distribution. Both Σ and D enter the equation as part of the covariance matrices C_{N} and C_{M} from (3). Σ is a function of the multiplicity of the (∊ factor), experimental uncertainties (σ_{o}), model completeness and model errors. For simplicity, the following parameterization is used:
The current version of REFMAC5 estimates D and Σ_{mod} in resolution bins. Working reflections are used for estimation of D and free reflections are used for Σ_{mod} estimation. Although this simple parameterization works in many cases, it may give misleading results for data from crystals with pseudo translation, OD disorder or modulated crystals in general. Currently, there is no satisfactory implementation of the error model to account for these cases.
2.2. Incorporation of experimental phase information in model refinement
2.2.1. MLHL likelihood
MLHL likelihood (Bricogne, 1997; Murshudov et al., 1997; Pannu et al., 1998) is based on a special case of the probability distribution (3) where we have one observation, one model and phase information derived from an experiment available as a prior distribution P_{pr}(α),
where F_{o} = F_{o}exp(ια), F_{c} = F_{c}exp(ια_{c}), α is the unknown phase of the and α_{1} and α_{2} are its possible values for a centric reflection. The prior phase probability distribution P_{pr}(α) is usually represented as a generalized von Mises distribution (Mardia & Jupp, 1999) and is better known in crystallography as a Hendrickson–Lattman distribution (Hendrickson & Lattman, 1970),
where A, B, C and D are coefficients of the Fourier transformation of the logarithm of the phase probability distribution and N is the normalization coefficient. The distribution is unimodal when C and D are zero; otherwise, it is a that reflects the possible phase uncertainty in experimental phasing. For centric reflections C and D are zero.
2.2.2. SAD/SIRAS likelihood
The MLHL likelihood is dependent on the reliability and accuracy of the prior distribution P_{pr}(α). However, the phase distributions after density modification (or even after phasing), which are usually used as P_{pr}(α), often suffer from inaccurate estimation of the phase errors. Furthermore, MLHL [as well as any other special case of (3) with a nonuniform P_{pr}(α)] assumes independence of the prior phases from the model phases. These shortcomings can be addressed by using experimental information directly from the experimental data, instead of from the P_{pr}(α) distributions obtained in previous steps of the structuresolution process. Currently, SAD and SIRAS likelihood functions are implemented in REFMAC5.
The SAD probability distribution (Skubák et al., 2004) is obtained from (3) by setting E = 2, M = 2, P_{pr}(α) = constant and F_{1} = F_{o}^{+}, F_{2} = (F_{o}^{−})*, F_{3} = F_{c}^{+}, F_{4} = (F_{c}^{−})*, where F^{+} and F^{−} are the structure factors of the Friedel pairs. The model structure factors are constructed using the current parameters of the protein, the heavyatom substructure and the inputted parameters. Similarly, the SIRAS function (Skubák et al., 2009) is a special case of (3) with E = 3, M = 3, P_{pr}(α) = constant and F_{1} = F_{o}^{N}, F_{2} = F_{o}^{+}, F_{3} = (F_{o}^{−})*, F_{4} = F_{c}^{N}, F_{5} = F_{c}^{+}, F_{6} = (F_{c}^{−})*, where F_{1} and F_{4} correspond to the observation and the model of the native crystal, respectively, and F_{2}, F_{3}, F_{5} and F_{6} refer to the observations and models of the derivative crystal. If any of the E observations are symmetrically equivalent, for instance centric intensities, the equation is reduced appropriately so as to only include nonequivalent observations and models.
The incorporation of prior phase information by the
function is especially useful in the early and middle stages of model building and at all stages of structure solution at lower resolutions, owing to the improvement in the observationtoparameter ratio. The of a well resolved highresolution structure is often best achieved using the simple Rice function.Fig. 1 shows the effect of various likelihood functions on automatic model building using ARP/wARP (Perrakis et al., 1999).
2.3. Twin refinement
The function used for twin
is a generalization of the Rice distribution in the presence of a linear relationship between the observed intensities. This function has the formwhere N_{o} and N_{model} are normalization coefficients. In the first equation, the first term inside the integral, P(I_{o}; F), represents the probability distribution of observations if `ideal' structure factors are known. Here, all reflections that are twinned and that can be grouped together are included. Models representing the datacollection instrument, if available, could be added to this term. The second term, P(F; model), represents a probability distribution of the `ideal' structure factors should an atomic model be known for a single crystal. Here, all reflections from the that contribute to the observed `twinned' intensities are included. If the data were to come from more than one crystal or if, for example, SAD should be used simultaneously with then this term would need to be modified appropriately. F_{c} is a function of atomic and overall parameter D. Overall parameters also include Σ and twinfraction parameters. f represents the way structure factors from the contribute to the particular `twinned' intensity. The above formula is more symbolic rather than precise; further details of twin will be published elsewhere.
REFMAC5 performs the following preparations before starting against twinned data.
All integrals necessary for evaluation of the minus loglikelihood function and its derivatives with respect to the structure factors are evaluated using the Laplace approximation (McKay, 2003).
2.4. Modelling bulksolvent contribution
Typically, a significant part of a macromolecular crystal is occupied by disordered solvent. Accurate modelling of this part of the crystal is still an unsolved problem of MX. The contribution of bulk solvent to structure factors is strongest at low resolution, although its effect at high resolution is still nonnegligible.
The absence of good models for disordered solvent may be one of the reasons why R factors in MX are significantly higher than those in smallmolecular crystallography. For small molecules R factors can be around 1%, whereas for MX they are rarely less than 10% and more often around 20% or even higher.
REFMAC5 uses two types of bulk (disordered) solvent models. One of them is the socalled Babinet's bulksolvent model, which is based on the assumption that the only difference between solvent and protein at low resolution is their scale factor (Tronrud, 1997). Here, we use a slight modification of the formulation described by Tronrud (1997) and assume that if protein electron density is convoluted using the Gaussian kernel and multiplied by an appropriate scale factor, then protein and solvent electron densities are equal,
where * denotes convolution, denotes the Fourier transform and k_{babinet} = k_{babinet0}exp(−B_{babinet}s^{2}/4). Here, we used the convolution theorem, which states that the Fourier transform of the convolution of two functions is the product of their Fourier transforms.
The second bulksolvent model is derived similarly to that described by Jiang & Brünger (1994). The basic assumption is that disordered solvent atoms are uniformly distributed over the region of the that is not occupied by the atoms of the modelled part of the The region of the occupied by the atomic model is masked out. Any holes inside this mask are removed using a cavitydetection algorithm. A constant value is assigned outside this region and the structure factors F_{mask} are calculated using an FFT algorithm. These structure factors, multiplied by appropriate scale factors (estimated during the scaling procedure), are added to those calculated from the atomic model. Additionally, various mask parameters may optionally be optimized.
One should be careful with bulksolvent corrections, especially when the atomic model is incomplete. This type of bulksolvent model may result in smearedout electron density that may reduce the height of electron density in lessordered and unmodelled parts of the crystal.
The final total structure factors with scale and solvent contributions included take the following form:
where the ks are scale factors, s is the reciprocalspace vector, s is the length of this vector, U_{aniso} is the crystallographic anisotropic tensor that obeys crystal symmetry, F_{mask} is the contribution from the mask bulk solvent and F_{protein} is the contribution from the protein part of the crystal. Usually, either mask or Babinet bulksolvent correction is used. However, sometimes their combination may provide better statistics (lower R factors) than either individually.
The overall parameters of the solvent models, the overall anisotropy and the scale factors are estimated using a leastsquares fit of the amplitude of the total structure factors to the observed amplitudes,
In the case of twin
the following function is used to estimate overall parameters including twin fractions (details of twin will be published elsewhere),where f(α, F) is as defined in (8).
Both (11) and (12) are minimized using the Gauss–Newton method with eigenvalue filtering to solve linear equations, which ensures that even very highly correlated parameters can be estimated simultaneously. However, one should be careful in interpretating these parameters as the system is highly correlated.
Once overall parameters such as the scale factors and twin fractions have been estimated, REFMAC5 estimates the overall parameters of one of the abovementioned likelihood functions and evaluates the function and its derivatives with respect to the atomic parameters. A general description of this procedure can be found in Steiner et al. (2003).
2.5. Geometry component
The function controlling the geometry has several components.

Standard restraints on the covalent structure have the general form
where b_{m} represents a geometric parameter (e.g. bonds, angles, chiralities) calculated from the model and b_{i} is the ideal value of this particular geometric parameter as tabulated in the dictionary.
Apart from ω (the angle of the peptide bond) and χ (the angles of aminoacid side chains), torsion angles in general are not restrained by default. However, the user can request to restrain a particular torsion angle defined in the dictionary or can define general torsion angles and use them as restraints. In general, it is not clear how to handle the restraint on torsion angles automatically, as these angles may depend on the covalent structure as well as the chemical environment of a particular ligand.
2.6. restraints
2.6.1. Automatic NCS definition
Automatic NCS identification in REFMAC5 is performed using the following procedure.
Steps (i)–(v) are performed once during each session of
Step (vi) is performed during every cycle of in order to allow conformational changes to occur.2.6.2. Global NCS
For global NCS restraints, transformation operators (R_{ij} and t_{ij}) that optimally superpose all NCSrelated molecules are estimated and the following residual is added to the total target function,
where the weight w is a usercontrollable parameter. Note that the transformation matrices are estimated using x_{i} and x_{j} and thus they are dependent on these parameters. Therefore, in principle the gradient and secondderivative calculations should take this dependence into account, although this dependence is ignored in the current version of REFMAC5. Ignoring the contribution of these terms may reduce the rate of convergence, although in practice it does not seem to pose a problem.
2.6.3. Local NCS
The following function (similar to the implementation in BUSTER) is used for local NCS restraints,
where GM is the Geman–McClure robust estimator function (Geman & McClure, 1987), which can be written
Fig. 2 shows that for small values of r this function is similar to the usual leastsquares function. However, it behaves differently for large r: leastsquare residuals do not allow conformational changes to occur, whereas this type of function is more tolerant to such changes.
2.6.4. External structure restraints
The interatomic distances within the structure being analysed may be similar to a known (deposited) structure, particularly in localized regions. In cases where it makes sense, this information can be exploited in order to aid the et al. (2009) is used for external structure restraints.
of the target structure. In doing so, the target structure is pulled towards the conformation adopted by the known structure. The mechanism for generic external restraints described by MooijIn our implementation, structural information from external known structures is utilized by applying restraints to the distances between atom pairs based on a presumed atomic correspondence between the two structures. The following function is used for external structure restraints,
where the atoms a_{i} belong to the set A of atoms for which a correspondence is known, d_{ij} is the distance between the positions of atoms a_{i} and a_{j}, d_{ij}^{*} is the corresponding distance in the known structure, σ_{ij} is the estimated standard deviation of d_{ij} about d_{ij}^{*} and d_{max} ensures that atom pairs are only restrained within localized regions, allowing insensitivity to global conformational changes. External structure restraints should be weighted differently to the other geometry components in order to allow the restraint strength to be separately specified. Consequently, a weight w_{ext} is applied, which should be appropriately chosen depending on the data quality and resolution, the structural similarity between the external known structure and the target, and the choice of d_{max}. The Geman–McClure function with sensitivity parameter σ_{GM} is used to increase robustness to outliers, as with the local NCS restraints.
Prior information from the external known structure(s) is generated using the software tool PROSMART. Specifically, this includes the atomic correspondence A, distances d_{ij}^{*}, standard deviations σ_{ij} and the distance cutoff d_{max}.
Potential sources of prior structural information include different conformations of the target chain (such as those that may result from using different crystallization conditions or in a different binding state) as well as those from homologous or structurally similar proteins. It is possible to use multiple known structures as prior information. The combination of this information results in modified values of d_{ij}^{*} and σ_{ij} as appropriate. This allows a structure to be refined utilizing information from a whole class of similar structures, rather than just a single source. Furthermore, it opens up the future possibility for multicrystal corefinement.
The employed formalism also allows the application of atomic distance restraints to secondarystructure elements (and, in principle, other motifs). Consequently, external restraints may be applied without requiring the prior identification of known structures similar to the target. This is intended to help to refine such motifs towards the expected/presumed local conformation.
This technique has been found to be particularly useful for lowresolution crystals and in cases where the target structure is unable to be refined to a satisfactory level. When used appropriately, external structure restraints should increase R and R_{free} values is expected to decrease in successful cases.
reliability. Consequently, the difference between theFig. 3 shows the resulting from using external restraints to refine a lowresolution bluetongue virus VP4 enzyme (Sutton et al., 2007). A sequenceidentical structure solved at a higher resolution is used as prior information. are compared after ten cycles with and without using external restraints. Using the external restraints results in a 2.8% improvement in R_{free}. Furthermore, the difference between the R and R_{free} values is reduced from 11.5 to 4.3%, suggesting greatly increased reliability.
2.6.5. `Jellybody' restraints
The ratio of the number of observations to the number of adjustable parameters is very small at low resolution. Even after accounting for chemical restraints, this ratio stays very small and R and R_{free} values. External structure restraints and the use of experimental phase information (described above) provide ways of dealing with this problem. Unfortunately, it is not always possible to find similar structures refined at high resolution (or at least ones that result in a sufficiently successful improvement in statistics) and experimental phase information is not always available or sufficient. Fortunately, statistical techniques exist to deal with this type of problem. Such techniques include ridge regression (Stuart et al., 2009), the lasso estimation procedure (Tibshirani, 1997) and Bayesian estimation with prior knowledge of parameters (O'Hagan, 1994).
in such cases is usually unstable. The danger of overfitting is very high; this is reflected in large differences between theREFMAC5 has a regularization function in interatomic distance space that has the form
for pairs of atoms i, j from the same chain, with maximum radius d_{max}, which can be controlled (default 4.25 Å). Note that this term does not contribute to the value of the function or its gradient; it only changes the second derivative, thus changing the search direction. It should be noted that a similar technique has been implemented in CNS (Schröder et al., 2010).
Note that if all interatomic distances were constrained, then individual atomic ). This simple formula has been found to work surprisingly well.
would become rigidbody The effect of `jellybody' restraints is the implicit parameterization between the rigid body and individual atoms. This technique has strong similarity to elastic network model calculations (Trion, 19962.6.6. Atomic displacement parameter restraints
Unlike positional parameters, where prior knowledge can be designed using basic knowledge of the chemistry of the building blocks of macromolecules and analysis of highresolution structures, it is not obvious how to design restraints for atomic displacement parameters (ADPs). Ideally, restraints should reflect the geometry of the molecules as well as their overall mobility. Various programs use various restraints (Sheldrick, 2008; Adams et al., 2010; Konnert & Hendrickson, 1980; Murshudov et al., 1997). In the new version of REFMAC5, restraints on ADPs are based on the distances between distributions. If we assume that atoms are represented as Gaussian distributions, then we are able to design restraints based on the distance between such distributions.
For a given two distributions in threedimensional space P(x) and Q(x), the symmetrized Kullback–Liebler (KL) divergence (McKay, 2003) is defined as follows:
It can be verified that the symmetrized KL divergence satisfies the conditions of a metric distance in the space of distributions. The KL divergence can also be represented as follows:
This distance changes more smoothly than the L_{2} distance between functions and seems to be a useful criterion for the design of approximate probability distributions (McKay, 2003; O'Hagan, 1994).
When both distributions are Gaussian with mean zero, this distance has an elegant form. Assume that both atoms have Gaussian distribution:
In this case, the KL divergence becomes
In the case of isotropic ADPs, KL has an even simpler form:
REFMAC5 uses restraints based on the KL divergence:
The summation is over all atom pairs with distance less than r_{max}. The weights depend on the nature of the bonds as well as on the distance between the atoms. If atoms are bonded or anglerelated then the weight is larger. However, the weight is smaller if the atoms are not related by covalent bonds. Moreover, if the distance between the atoms is more than 3 Å then the weight decreases as follows:
where w_{0,ij} is the weight for nonbonded atoms that are closer than 3 Å to each other.
2.6.7. Rigidbond restraints
For anisotropic atoms there are socalled rigidbond restraints, based on the idea of rigidbond tests of anisotropic atoms (Hirshfeld, 1976). The idea is that projections of U values on the bond vector joining two atoms should be similar. In other words, if two atoms are bonded then an oscillation across the bond is more likely than an independent oscillation along the bond. Atoms oscillate along the bond in a concerted fashion.
Rigidbond restraints are designed as follows. Let us assume that two atoms have positions x_{1} and x_{2} and their corresponding ADPs are U_{1} and U_{2}; the unit vector joining these atoms is then calculated,
The projections of corresponding U values on this vector are then calculated as
Now, using these projections, the KL divergence is formed for all pairs and added to the target function:
Again, the weights depend on the nature of the bonds between the atoms and the distances between them. Note that if the ADPs of both bonded atoms are isotropic then the rigidbond restraint is equivalent to the abovedescribed KL restraint.
2.6.8. Sphericity restraints
To avoid atoms exploding and becoming too elliptical or, even worse, nonelliptical, REFMAC5 uses restraints on sphericity. It is a simple restraint: an isotropic equivalent of the anisotropic tensor,
where k indexes the anisotropic atoms, i, j are components of the anisotropic tensor and w_{k} are weights for this particular type of restraint. The weights depend on the number of other restraints (KL, rigid bond) on this atom. Atoms that have fewer restraints have stronger weights on sphericity, since these atoms are more likely to be unstable.
It should be noted that similar restraints on ADPs are used in several other ; Adams et al., 2010).
programs (Sheldrick, 20083. Parameterization
3.1. General parameters
REFMAC5 uses the standard parameterization of molecules in terms of atomic coordinates and isotropic/anisotropic atomic displacement parameters. The of these parameters is performed using an FFT formulation for gradients and approximations for second derivatives. Details of these formulations have been published elsewhere (Murshudov et al., 1997, 1999; Steiner et al., 2003). Once the gradients and approximate second derivatives have been calculated for these parameters, they are used to calculate the derivatives of derived parameters. Derived parameters include those for rigidbody and TLS refinement.
3.2. Rigid body
Rigidbody parameterization is achieved as follows. For each rigid group, transformation operators are defined and new positions are calculated from the starting positions using the formula
where R_{j} is the rotation matrix, t_{origin} is the centre of mass of the rigid group and t_{j} is the translational component of the transformation. The x_{old} are the starting coordinates of the atoms and x_{new} are their positions after application of the transformation operators. There are six parameters per rigid group, defining the rotation matrix and the translational component. At each cycle of an eigenvaluefiltering technique is used to avoid potential singularities arising from the shape of the rigid groups. It should be noted that no terms between rigid groups are calculated for the approximate secondderivative matrix. For large rigid groups this does not pose much of a problem. However, for many small rigid groups it may slow down convergence substantially. In any case, it is not recommended to divide molecules into very small rigid groups. For these cases, `jellybody' should produce better results.
Once derivatives with respect to the positional parameters have been calculated, those for rigidbody parameters are calculated using the chain rule. The current version of REFMAC5 uses an Euler angle parameterization.
3.3. TLS
Atomic displacement parameters describe the spread of atomic positions and can be derived from the Fourier transform of a Gaussian probability distribution function for the atomic centre. The atomic displacement parameters are an important part of the model. Traditionally, a single parameter describing isotropic displacements has been used, namely the B factor. However, it is well known that atomic displacements are likely to be anisotropic owing to directional bonding and at high resolutions the six parameters per atom of a fully anisotropic model can be refined. TLS is a way of modelling anisotropic displacements using only a few parameters, so that the method can be used at medium and low resolutions. The TLS model was originally proposed for smallmolecule crystallography (Schomaker & Trueblood, 1968) and was incorporated into REFMAC5 almost ten years ago (Winn et al., 2001).
The idea behind TLS is to suppose that groups of atoms move as rigid bodies and to constrain the anisotropic displacement parameters of these atoms accordingly. The rigidbody motion is described by translation (T), libration (L) and screw (S) tensors, using a total of 20 parameters for each rigid body. Given values for these 20 parameters, anisotropic displacement parameters can be derived for each atom in the group (and this relationship also allows one to calculate derivatives via the chain rule). Usually, an extra isotropic displacement parameter (the residual B factor) is refined for each atom in addition to the TLS contribution. The sum of these two contributions can be output using the supplementary program TLSANL (Howlin et al., 1993) or optionally directly from REFMAC5.
TLS groups need to be chosen before REFMAC5. More detailed choices can be made using methods such as TLSMD (Painter & Merritt, 2006). By default, REFMAC5 also includes waters in the first hydration shell, which it seems reasonable to assume move in concert with the protein chain.
and constitute part of the definition of the model for the macromolecule. Groups of atoms should conform to the idea that they move as a quasirigid body. Often the choice of one group per chain suffices (or at least serves as a reference calculation) and this is the default inFig. 4 shows the effect of TLS and orientation of libration tensors. In this case, TLS improves R/R_{free} and the derived libration tensors make biological sense.
4. Optimization
REFMAC5 uses the Gauss–Newton method for optimization. For an elegant and comprehensive review on optimization techniques, see Nocedal & Wright (1999). In this method, the exact second derivative is not calculated, but rather approximated to make sure it is always nonnegative. Once derivatives or approximations have been calculated, the following linear equation is built,
where H is the approximate second derivative and G is the gradient vector. The contribution of most of the geometrical terms are calculated using algorithms designed for quadratic optimization or leastsquares fitting (Press et al., 1992). To calculate the contribution from the Geman–McClure terms, the following approximation is used (Huber & Ronchetti, 2009),
This approximation ensures that H stays nonnegative and consequently directions calculated as a result of the solution of (32) point towards a reduction of the total function.
The contribution of the Xray term to the gradient is calculated using FFT algorithms (Murshudov et al., 1997). The Fisher information matrix, as described by Steiner et al. (2003), is used to calculate the contribution of the likelihood functions to the matrix H. Tests have demonstrated that using the diagonal elements of the Fisher information matrix and both diagonal and nondiagonal elements of the geometry terms results in a more stable refinement.
Once all of the terms contributing to H and G have been calculated, the linear equation (32) is solved using preconditioned conjugategradient methods (Nocedal & Wright, 1999; Tronrud, 1992). A diagonal matrix formed by the diagonal elements of H is used as a preconditioner. This brings parameters with different overall scales (positional and B values) onto the same scale and controlling convergence becomes easier.
If the conjugategradient procedure does not converge in N_{maxiter} cycles (the default is 1000), then the diagonal terms of the H matrix are increased. Thus, if the matrix is not positive then ridge regression is activated. In the presence of a potential (near) singularity, REFMAC5 uses the following procedure to solve the linear equation.
5. Conclusions
is an important step in macromolecular elucidation. It is used as a final step in structure solution, as well as as an intermediate step to improve models and obtain improved electron density to facilitate further model rebuilding.
REFMAC5 is one of the programs that incorporates various tools to deal with some crystal peculiarities, lowresolution MX structure and highresolution There are also tabulated dictionaries of the constituent blocks of macromolecules, cofactors and ligands. The number of dictionary elements now exceeds 9000. There are also tools to deal with new ligands and covalent modifications of ligands and/or proteins.
Lowresolution MX structure analysis is still a challenging task. There are several outstanding problems that need to be dealt with before we can claim that lowresolution MX analysis is complete. Statistics, image processing and computer science provide general methods for these and related problems. Unfortunately, these techniques cannot be directly applied to MX structure analysis, either because of the huge computer resources needed or because the assumptions used are not applicable to MX.
In our opinion, the problems of stateoftheart MX analysis that need urgent attention include the following.
Further improvement may consist of a combination of various experimental techniques. For example, the simultaneous treatment of electronmicroscopy (EM) and MX data could increase the reliability of EM models and put MX models in the context of larger biological systems.
The direct use of unmerged data is another direction in which
procedures could be developed. If this were achieved, then several longstanding problems could be easier to deal with. Two such problems are the following. (i) In general, the of a crystal should be considered as an adjustable parameter. If unmerged data are used, then spacegroup assumptions could be tested after every few sessions of and model building. (ii) Dealing with the processes in the crystal during data collection requires unmerged data. One of the bestknown such problems is radiation damage.Acknowledgements
We thank the CCP4 staff and CCP4 Bulletin Board participants for continuous stimulating discussions. We also thank the wider user community. Without their continuous feedback and bug reports, development of programs such as REFMAC5 would be impossible. This work was supported by Wellcome Trust Grant No. 064405/Z/01/A to GNM and FL. AAV was supported by a CCP4 grant, PS and NSP are supported by the Netherlandse Organisatie voor Wetenschappelijk Onderzoek (NWO), and AAL and RAN were supported by a BBSRC grant. CCP4 is supported by BBSRC grant BB/F0202281.
References
Adams, P. D. et al. (2010). Acta Cryst. D66, 213–221. Web of Science CrossRef CAS IUCr Journals Google Scholar
Agarwal, R. C. (1978). Acta Cryst. A34, 791–809. CrossRef CAS IUCr Journals Web of Science Google Scholar
Altomare, A., Cuocci, C., Giacovazzo, C., Kamel, G. S., Moliterni, A. & Rizzi, R. (2008). Acta Cryst. A64, 326–336. Web of Science CrossRef CAS IUCr Journals Google Scholar
Blanc, E., Roversi, P., Vonrhein, C., Flensburg, C., Lea, S. M. & Bricogne, G. (2004). Acta Cryst. D60, 2210–2221. Web of Science CrossRef CAS IUCr Journals Google Scholar
Booth, A. (1946). Proc. R. Soc. Lond. A Math. Phys. Sci. 188, 77–92. CrossRef CAS Google Scholar
Bricogne, G. (1997). Methods Enzymol. 276, 361–423. CrossRef CAS Web of Science Google Scholar
Bricogne, G. & Irwin, J. (1996). Proceedings of the CCP4 Study Weekend. Macromolecular Refinement, edited by E. J. Dodson, M. Moore, A. Ralph & S. Bailey, pp. 85–92. Warrington: Daresbury Laboratory. Google Scholar
Brünger, A. (1991). Annu. Rev. Phys. Chem. 42, 197–223. Google Scholar
Brünger, A. T. (1989). Acta Cryst. A45, 42–50. CrossRef Web of Science IUCr Journals Google Scholar
Brünger, A. T., Adams, P. D., Clore, G. M., DeLano, W. L., Gros, P., GrosseKunstleve, R. W., Jiang, J.S., Kuszewski, J., Nilges, M., Pannu, N. S., Read, R. J., Rice, L. M., Simonson, T. & Warren, G. L. (1998). Acta Cryst. D54, 905–921. Web of Science CrossRef IUCr Journals Google Scholar
Cochran, W. (1948). Acta Cryst. 1, 138–142. CrossRef CAS IUCr Journals Google Scholar
Collaborative Computational Project, Number 4 (1994). Acta Cryst. D50, 760–763. CrossRef IUCr Journals Google Scholar
Cowtan, K. (2006). Acta Cryst. D62, 1002–1011. Web of Science CrossRef CAS IUCr Journals Google Scholar
Cruickshank, D. W. J. (1952). Acta Cryst. 5, 511–518. CrossRef IUCr Journals Web of Science Google Scholar
Cruickshank, D. W. J. (1956). Acta Cryst. 9, 747–753. CrossRef CAS IUCr Journals Web of Science Google Scholar
DeLano, W. L. (2002). PyMOL. http://www.pymol.org . Google Scholar
Diamond, R. (1971). Acta Cryst. A27, 436–452. CrossRef CAS IUCr Journals Web of Science Google Scholar
Emsley, P. & Cowtan, K. (2004). Acta Cryst. D60, 2126–2132. Web of Science CrossRef CAS IUCr Journals Google Scholar
Geman, S. & McClure, D. (1987). Bull. Int. Stat. Inst. 52, 5–21. Google Scholar
Guillot, B., Viry, L., Guillot, R., Lecomte, C. & Jelsch, C. (2001). J. Appl. Cryst. 34, 214–223. Web of Science CrossRef CAS IUCr Journals Google Scholar
Hendrickson, W. A. & Lattman, E. E. (1970). Acta Cryst. B26, 136–143. CrossRef CAS IUCr Journals Google Scholar
Hirshfeld, F. L. (1976). Acta Cryst. A32, 239–244. CrossRef IUCr Journals Web of Science Google Scholar
Howlin, B., Butler, S. A., Moss, D. S., Harris, G. W. & Driessen, H. P. C. (1993). J. Appl. Cryst. 26, 622–624. CrossRef Web of Science IUCr Journals Google Scholar
Huber, P. J. & Ronchetti, E. M. (2009). Robust Statistics. Hoboken: John Wiley & Sons. Google Scholar
Hue, H. & Held, L. (2005). Gaussian Markov Random Field Models. Boca Raton: Chapman & Hall/CRC. Google Scholar
Jack, A. & Levitt, M. (1978). Acta Cryst. A34, 931–935. CrossRef CAS IUCr Journals Web of Science Google Scholar
Jiang, J.A. & Brünger, A. (1994). J. Mol. Biol. 243, 100–115. CrossRef CAS PubMed Web of Science Google Scholar
Jones, T. A., Zou, J.Y., Cowan, S. W. & Kjeldgaard, M. (1991). Acta Cryst. A47, 110–119. CrossRef CAS Web of Science IUCr Journals Google Scholar
Konnert, J. H. (1976). Acta Cryst. A32, 614–617. CrossRef CAS IUCr Journals Web of Science Google Scholar
Konnert, J. H. & Hendrickson, W. A. (1980). Acta Cryst. A36, 344–350. CrossRef CAS IUCr Journals Google Scholar
Lebedev, A. A., Vagin, A. A. & Murshudov, G. N. (2006). Acta Cryst. D62, 83–95. Web of Science CrossRef CAS IUCr Journals Google Scholar
Luzzati, V. (1951). Acta Cryst. 4, 367–369. CrossRef IUCr Journals Web of Science Google Scholar
Luzzati, V. (1952). Acta Cryst. 5, 802–810. CrossRef IUCr Journals Web of Science Google Scholar
Mardia, K. V. & Bibby, J. E. (1979). Multivariate Analysis. London/San Diego: Academic Press. Google Scholar
Mardia, K. V. & Jupp, P. E. (1999). Directional Statistics. Chichester: John Wiley & Sons. Google Scholar
McKay, D. J. C. (2003). Information Theory, Inference and Learning Algorithms. Cambridge University Press. Google Scholar
McRee, D. E. & Israel, M. (2008). J. Struct. Biol. 163, 208–213. Web of Science CrossRef PubMed CAS Google Scholar
Mooij, W., Cohen, S., Joosten, K., Murshudov, G. & Perrakis, A. (2009). Structure, 17, 183–189. Web of Science CrossRef PubMed CAS Google Scholar
Mouilleron, S. & GolinelliPimpaneau, B. (2007). Protein Sci. 16, 485–493. Web of Science CrossRef PubMed CAS Google Scholar
Murshudov, G. N., Vagin, A. A. & Dodson, E. J. (1997). Acta Cryst. D53, 240–255. CrossRef CAS Web of Science IUCr Journals Google Scholar
Murshudov, G. N., Vagin, A. A., Lebedev, A., Wilson, K. S. & Dodson, E. J. (1999). Acta Cryst. D55, 247–255. Web of Science CrossRef CAS IUCr Journals Google Scholar
Needleman, S. B. & Wunsch, C. D. (1970). J. Mol. Biol. 48, 443–453. CrossRef CAS PubMed Web of Science Google Scholar
Ness, S. R., de Graaff, R. A. G., Abrahams, J. P. & Pannu, N. S. (2004). Structure, 12, 1753–1761. Web of Science CrossRef PubMed CAS Google Scholar
Nocedal, J. & Wright, S. J. (1999). Numerical Optimization. New York: Springer. Google Scholar
O'Hagan, A. (1994). Kendal's Advanced Theory of Statistics, Vol. 2B, Bayesian Inference. London: Hodder Arnold. Google Scholar
Painter, J. & Merritt, E. A. (2006). Acta Cryst. D62, 439–450. Web of Science CrossRef CAS IUCr Journals Google Scholar
Pannu, N. S., Murshudov, G. N., Dodson, E. J. & Read, R. J. (1998). Acta Cryst. D54, 1285–1294. Web of Science CrossRef CAS IUCr Journals Google Scholar
Pannu, N. S. & Read, R. J. (1996). Acta Cryst. A52, 659–668. CrossRef CAS Web of Science IUCr Journals Google Scholar
Perrakis, A., Morris, S. & Lamzin, V. S. (1999). Nature Struct. Biol. 6, 458–463. Web of Science CrossRef PubMed CAS Google Scholar
Potterton, L., McNicholas, S., Krissinel, E., Gruber, J., Cowtan, K., Emsley, P., Murshudov, G. N., Cohen, S., Perrakis, A. & Noble, M. (2004). Acta Cryst. D60, 2288–2294. Web of Science CrossRef CAS IUCr Journals Google Scholar
Press, W. H., Flannery, B. P., Teukolsky, S. A. & Vetterling, W. T. (1992). Numerical Recipes in FORTRAN. Cambridge University Press. Google Scholar
R Development Core Team (2007). R: A Language and Environment for Statistical Computing. Vienna: R Foundation for Statistical Computing. http://www.Rproject.org . Google Scholar
Ramachandran, G. N., Srinivasan, R. & Sarma, V. R. (1963). Acta Cryst. 16, 662–666. CrossRef CAS IUCr Journals Web of Science Google Scholar
Schomaker, V. & Trueblood, K. N. (1968). Acta Cryst. B24, 63–76. CrossRef CAS IUCr Journals Web of Science Google Scholar
Schröder, G. F., Brünger, A. T. & Levitt, M. (2010). Nature (London), 464, 1218–1222. Web of Science PubMed Google Scholar
Schüttelkopf, A. W. & van Aalten, D. M. F. (2004). Acta Cryst. D60, 1355–1363. Web of Science CrossRef IUCr Journals Google Scholar
Sheldrick, G. M. (2008). Acta Cryst. A64, 112–122. Web of Science CrossRef CAS IUCr Journals Google Scholar
Skubák, P., Murshudov, G. N. & Pannu, N. S. (2004). Acta Cryst. D60, 2196–2201. Web of Science CrossRef IUCr Journals Google Scholar
Skubák, P., Murshudov, G. & Pannu, N. S. (2009). Acta Cryst. D65, 1051–1061. Web of Science CrossRef IUCr Journals Google Scholar
Skubák, P., Waterreus, W.J. & Pannu, N. S. (2010). Acta Cryst. D66, 783–788. Web of Science CrossRef IUCr Journals Google Scholar
Srinivasan, R. & Parthasarathy, S. (1976). Some Statistical Applications in Xray Crystallography. Oxford: Pergamon Press. Google Scholar
Srinivasan, R. & Ramachandran, G. N. (1965). Acta Cryst. 19, 1008–1014. CrossRef CAS IUCr Journals Web of Science Google Scholar
Steiner, R. A., Lebedev, A. A. & Murshudov, G. N. (2003). Acta Cryst. D59, 2114–2124. CrossRef CAS IUCr Journals Google Scholar
Stuart, A. & Ord, K. (2009). Kendall's Advanced Theory of Statistics, Vol. 1, Distribution Theory. Hoboken: John Wiley & Sons. Google Scholar
Stuart, A., Ord, K. & Arnold, S. (2009). Kendall's Advanced Theory of Statistics, Vol. 2A, Classical Inference. Hoboken: John Wiley & Sons. Google Scholar
Sutton, G., Grimes, J., Stuart, D. & Roy, P. (2007). Nature Struct. Mol. Biol. 14, 449–451. CrossRef CAS Google Scholar
Ten Eyck, L. F. (1977). Acta Cryst. A33, 486–492. CrossRef CAS IUCr Journals Web of Science Google Scholar
Terwilliger, T. C. (2003). Acta Cryst. D59, 1174–1182. Web of Science CrossRef CAS IUCr Journals Google Scholar
Tibshirani, R. J. (1997). Stat. Med. 16, 385–395. CrossRef CAS PubMed Web of Science Google Scholar
Trion, M. M. (1996). Phys. Rev. Lett. 77, 1906–1908. Google Scholar
Tronrud, D. E. (1992). Acta Cryst. A48, 912–916. CrossRef CAS Web of Science IUCr Journals Google Scholar
Tronrud, D. E., Ten Eyck, L. F. & Matthews, B. W. (1987). Acta Cryst. A43, 489–501. CrossRef CAS Web of Science IUCr Journals Google Scholar
Tronrud, G. (1997). Methods Enzymol. 277, 306–319. CrossRef CAS PubMed Web of Science Google Scholar
Turk, D. (2008). Acta Cryst. A64, C23. CrossRef IUCr Journals Google Scholar
Vagin, A. A., Steiner, R. A., Lebedev, A. A., Potterton, L., McNicholas, S., Long, F. & Murshudov, G. N. (2004). Acta Cryst. D60, 2184–2195. Web of Science CrossRef CAS IUCr Journals Google Scholar
Winn, M. D., Isupov, M. N. & Murshudov, G. N. (2001). Acta Cryst. D57, 122–133. Web of Science CrossRef CAS IUCr Journals Google Scholar
This is an openaccess article distributed under the terms of the Creative Commons Attribution (CCBY) Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original authors and source are cited.