Restraints, Constraints and Group Fitting, and Disorder

RESTRAINTS, CONSTRAINTS AND GROUP FITTING, AND DISORDER


In crystal structure refinement, there is an important distinction between a 'constraint' and a 'restraint'. A constraint is an exact mathematical condition which enables one or more least-squares variables to be expressed exactly in terms of other variables or constants, and hence eliminated. An example is the fixing of the x, y and z coordinates of an atom on an inversion center. A restraint takes the form of additional information which is not exact but is subject to a probability distribution; for example we could restrain two chemically but not crystallographically equivalent bonds to be approximately equal, with an effective standard deviation of (say) 0.01 Angstroms.

A restraint is incorporated in the least-squares refinement as if it were an additional experimental observation; w(yt-y)2 is added to the quantity Sigma[w(Fo2-Fc2)2] to be minimized, where a quantity y (which is a function of the least-squares parameters) is to be restrained to a target value yt, and the weight w (for either a restraint or a reflection) is 1/sigma2. In the case of a reflection sigma2 is estimated using a weighting scheme; for a restraint sigma is simply the effective standard deviation. In SHELXL-93 the restraint weights are multiplied by the square of the Goodness of Fit for the reflection data, which allows for the possibility that the reflection weights may be relative rather than absolute, and also gives the restraints more influence at the early stages of refinement (when the Goodness of Fit is invariably much greater than unity), which improves convergence.

Most of the constraints and restraints available in SHELXL-93 have already been widely used in other programs, especially for macromolecular refinement. In SHELXL-93 an effort has been made to make them simple to understand and use, while at the same time avoiding the bias which is introduced when specific target values etc. have to be assumed. For example it is more realistic to assume that a phenyl group is planar and has mm (C2v) symmetry (in both cases within a reasonable tolerance) rather than that it is an exactly regular hexagon with a bond length of 1.39 Angstroms; however both approaches may conveniently be applied using SHELXL-93. The following general categories of constraints and restraints are available using SHELXL-93:

1. Constraints for the coordinates and anisotropic displacement parameters for atoms on special positions: these are generated automatically by the program for ALL special positions in ALL space groups, in conventional settings or otherwise. If the user applies (correct or incorrect) special position constraints using free variables etc., the program assumes this has been done with intent and reports but does not apply the correct constraints. Thus the accidental application of a free variable to a Uij term of an atom on a special position can lead to the refinement 'blowing up'!

2. Two or more atoms sharing the same site: the xyz and Uij parameters may be equated using the EXYZ and EADP constraints respectively (or by using 'free variables'). The occupation factors may be expressed in terms of a 'free variable' so that their sum is constrained to be constant (e.g. 1.0). If more than two different chemical species share a site, a linear free variable restraint (SUMP) is required to restrain the sum of occupation factors. EADP is also useful for equating the Uij of 'opposite' fluorines of disordered -CF3 groups.

3. Floating origin restraints: these are generated automatically by the program as and when required by the method of H.D. Flack and D. Schwarzenbach, Acta Cryst., A44 (1988) 499-506, so the user should not attempt to fix the origin in such cases by fixing the coordinates of a heavy atom.

4. Geometrical constraints: these include rigid-group refinements (AFIX 6), variable-metric rigid-group refinements (AFIX 9) and various riding models (AFIX/HFIX) for hydrogen atom refinement, for example torsional refinement of a methyl group about the local threefold axis.

5. Fragments of known geometry may be fitted to target atoms (e.g. from a previous Fourier peak search), and the coordinates generated for any missing atoms. Four standard groups are available: regular pentagon, regular hexagon, naphthalene and pentamethylcyclopentadienyl; any other group may be used simply by specifying orthogonal or fractional coordinates in a given cell (AFIX mn with m > 16 and FRAG...FEND). This is usually, but not always, a preliminary to rigid group refinement.

6. Geometrical restraints: a particularly useful restraint is to make chemically but not crystallographically equivalent distances equal (subject to a given or assumed esd) without having to invent a value for this distance (SADI). The SAME instruction can be used to generate such restraints automatically, e.g. when chemically identical molecules or residues are present. This has the same effect as making equivalent bond lengths and angles but not torsion angles equal. The FLAT instruction restrains a group of atoms to lie in a plane (but the plane is free to move and rotate). DFIX and CHIV restrain distances and chiral volumes respectively to target values. When 'free variables' are used for the target values, it is possible to restrain different distances etc. to be equal and to refine their mean value (for which an esd is thus obtained). ALL types of geometrical restraints may involve ANY atom, even if it is part of a rigid group or a symmetry equivalent generated using EQIV $n ... and referenced by _$n, except for hydrogen atoms which ride on rigid group atoms (see preceding section).

7. 'Anti-bumping' restraints may be applied individually, by means of DFIX distance restraints with the distance given as a negative number, or generated automatically by means of the BUMP instruction, which operates on all atoms which have been designated by CONN 0 instructions (and so are excluded from the connectivity array). DFIX restraints with negative distance d are ignored if the two atoms are further from one another than |d| in the current refinement cycle; if they are closer than |d|, a restraint is applied to increase the distance to |d| with the given (or assumed) esd. The automatic generation of anti-bumping restraints takes all possible symmetry equivalents into account, and allows a safety margin of 0.5 A so that atoms which move towards one another during the refinement are also covered. In combination with the SWAT instruction for diffuse solvent, BUMP provides a very effective way of handling solvent water in macromolecules.

8. Restraints on anisotropic displacement parameters: three different types of restraint may be applied to Uij values. DELU applies a 'rigid-bond' restraint to Uij of two bonded (or 1,3) atoms; the anisotropic displacement components of the two atoms along the line joining them are restrained to be equal. This restraint was suggested by J.S. Rollett (in Crystallographic Computing, Ed. F.R. Ahmed, S.R. Hall and C.P. Huber, Munksgaard, Copenhagen, (1970) pp. 167-181), and corresponds to the rigid-bond criterion for testing whether anisotropic displacement parameters are physically reasonable (F.L. Hirshfeld, Acta Cryst., A32 (1976) 239-244; K.N. Trueblood and J.D. Dunitz, Acta Cryst., B39 (1983) 120-133). J.J. Didisheim and D. Schwarzenbach (Acta Cryst., A43 (1987) 226-232) have shown that in many but not all cases, rigid-bond restraints are equivalent to the TLS description of rigid body motion in the limit of zero esd's; however this requires that (almost) all atom pairs are restrained in this way, which for molecules with conformational flexibility is unlikely to be appropriate. An extensive study (E. Irmer, Ph.D. Thesis, University of Goettingen, 1990) has shown that this condition is fulfilled within the experimental error for routine X-ray studies of bonds and 1,3-distances between two first-row elements (B to F inclusive), and so may be applied as a 'hard' restraint (low esd). A rigid bond restraint is not suitable for systems with unresolved disorder, e.g. AsF6- anions and dynamic Jahn-Teller effects, although it may be useful in detecting such effects.

Isolated (e.g. solvent water) atoms may be restrained to be approximately isotropic, e.g. to prevent them going 'non-positive-definite'; this is a rough approximation and so should be applied as a 'soft' restraint with a large esd (ISOR). Similarly the assumption of 'similar' Uij values for spatially adjacent atoms (SIMU) is useful so that (for example) the thermal ellipsoids increase and change direction gradually going along a side-chain in a polypeptide, but this treatment is approximate and thus also appropriate only for a soft restraint; it is also useful for partially overlapping atoms of disordered groups. A simple way to apply SIMU to all such overlapping atoms is to give a SIMU instruction with no atoms (i.e. all atoms implied) and the third number set to a distance less than the shortest bond, i.e.

 SIMU 0.02 0.04 0.8
which applies the restraint to all pairs of atoms separated by less than 0.8 Angstroms. Additional SIMU restraints may be included in the same job.

SHELXL-93 does not permit DELU, SIMU and ISOR restraints to reference symmetry generated atoms, although this is allowed for all geometrical restraints. To permit such references for displacement parameter restraints as well would considerably complicate the program, and is rarely required in practice.

9. 'Shift limiting restraints' may be applied in SHELXL-93 by the Marquardt algorithm (J. Soc. Ind. Appl. Math., 11 (1963) 431-441). Terms proportional to a 'damping factor' (the first parameter on the DAMP instruction) are added to the least-squares matrix before inversion. Shift limiting restraints are particularly useful in the refinement of structures with a poor data to parameter ratio, and for pseudosymmetric problems. The 'damping factor' should be reduced towards the end of the refinement, otherwise the least-squares estimates of the esd's in the less well determined parameters will be too low (the program does however make a first order correction to the esds for this effect). The shifts are also scaled down if the maximum shift/esd exceeds the second DAMP parameter. In addition, if the actual and target values for a particular restraint differ by more than 100 times the given esd, the program will temporarily increase the esd to limit the influence of this restraint in any one cycle to that produced by a discrepancy of 100 times the esd. This helps to prevent a bad initial model and tight restraints from causing dangerously large shifts in the first cycle.

10. Further constraints may be applied to atom coordinates, occupation and displacement parameters, and to restrained distances (DFIX) and chiral volumes (CHIV), by the use of 'free variables'. Linear combinations of free variables may in turn be restrained (SUMP). Free variables were required for special position constraints and for refining more than one atom on the same site in SHELX-76; their use in this way is allowed (for upwards compatibility) in SHELXL-93, but it is more convenient to use the fully automatic handling of special positions in SHELXL-93, and atoms on multiply occupied sites may be constrained using EXYZ and EADP. For further details see the description of the FVAR instruction.

A major advantage of applying chemically reasonable restraints is that a subsequent difference electron density synthesis is often more revealing, because the parameters were not allowed to 'mop up' any residual effects. The refinement of pseudosymmetric structures, where the X-ray data may not be able to determine all of the parameters, is also considerably facilitated, at the cost of making it much easier to refine a structure in a space group of unnecessarily low symmetry!

By way of example, assume that the structure contains a cyclopentadienyl (Cp) ring pi-bonded to a metal atom, and that as a result of the high thermal motion of the ring only three of the atoms could be located in a difference electron density map. We wish to fit a regular pentagon (default C-C 1.42 A) in order to place the remaining two atoms, which are input as dummy atoms with zero coordinates. Since the C-C distance is uncertain (there may well be an appreciable librational shortening in such a case) we refine the C5-ring as a 'variable metric' rigid group, i.e. it remains a regular pentagon but the C-C distance is free to vary. In SHELXL-93 this may all be achieved by inserting one instruction (AFIX 59) before the five carbons and one (AFIX 0) after them:

 AFIX 59                 ! AFIX mn with m = 5 to fit pentagon (default C-C
 C1 1 .6755 .2289 .0763  ! 1.42 A) and n = 9 for v-m rigid-group refinement
 C2 1 .7004 .2544 .0161
 C3 1 0 0 0              ! the coordinates for C3 and C4 are obtained by the
 C4 1 0 0 0              ! fit of the other 3 atoms to a regular pentagon
 C5 1 .6788 .1610 .0766
 AFIX 0                  ! terminates rigid group
Since Uij values were not specified, the atoms would refine isotropically starting from U = 0.05. To refine with anisotropic displacement parameters in the same or a subsequent job, the instruction:

 ANIS C1 > C5
should be inserted anywhere before C1 in the '.ins' file. The SIMU and ISOR restraints on the Uij would be inappropriate for such a group, but:

 DELU C1 > C5
could be applied if the anisotropic refinement proved unstable. The five hydrogen atoms could be added and refined with the 'riding model' by means of:

 HFIX 43 C1 > C5
anywhere before C1 in the input file. For good data, in view of possible librational effects, a possible alternative would be:

 HFIX 44 C1 > C5
 SADI 0.02 C1 H1 C2 H2 C3 H3 C4 H4 C5 H5
(which retains a riding model but allows the C-H bond lengths to refine, subject to the restraint that they should be equal within about 0.02 A).

In analogous manner it is possible to generate missing atoms and perform rigid group refinements for phenyl rings (AFIX 66) and Cp* groups (AFIX 109). Very often it is possible and desirable to remove the rigid group constraints (by simply deleting the AFIX instructions) in the final stages of refinement; there is good experimental evidence that the ipso-angles of phenyl rings differ systematically from 120 degrees [P.G. Jones, J. Organomet. Chem., 345 (1988) 405; T. Maetzke and D. Seebach, Helv. Chim. Acta, 72 (1989) 624-630; A. Domenicano, Accurate Molecular Structures, eds. Domenicano and Hargittai, Chapter 18, OUP 1992].

As a second example, assume that the structure contains two molecules of poorly defined THF solvent, and that we have managed to identify the oxygen atoms. A rigid pentagon would clearly be inappropriate here, except possibly for placing missing atoms, since THF molecules are not planar. However we can RESTRAIN the 1,2- and the 1,3-distances in the two molecules to be similar by means of a 'similarity restraint' (SAME). Assume that the molecules are numbered O11 C12 ... C15 and O21 C22 ... C25, and that the atoms are given in this order in the atom list. Then we can either insert the instruction:

 SAME O21 > C25
before the first molecule, or:

 SAME O11 > C15
before the second. These SAME instructions define a group of five atoms which are considered to be the same as the five (non-hydrogen) atoms which immediately follow the SAME instruction. The entries in the connectivity table for the latter are used to define the 1,2- and 1,3-distances, so the SAME instruction should be inserted before the group with the best geometry. This one SAME instruction restrains five pairs of 1,2- and five pairs of 1,3- distances to be nearly equal, i.e.

 d(O11-C12) = d(O21-C22),  d(C12-C13) = d(C22-C23),  d(C13-C14) = d(C23-C24),
 d(C14-C15) = d(C24-C25),  d(C15-O11) = d(C25-O21),  d(O11-C13) = d(O21-C23),
 d(C12-C14) = d(C22-C24),  d(C13-C15) = d(C23-C25),  d(C14-O11) = d(C24-O21),
 and  d(C15-C12) = d(C25-C22).
In addition, it would also be reasonable to restrain the distances on opposite sides of the same ring to be equal. This can be achieved with one further SAME instruction in which we count the other way around the ring. For example we could insert:

 SAME O11 C15 < C12
before the first ring. The symbol '<' indicates that one must count up the atom list instead of down. The above instruction is exactly equivalent to:

 SAME O11 C15 C14 C13 C12
This generates 10 further restraints, but two of them [d(C13-C14) = d(C14-C13) and d(C12-C15) = d(C15-C12)] are identities, and each of the others appears twice, so only four are independent and the rest are ignored. It is not necessary to add a similar instruction before the second ring, because the program also automatically generates all 'implied' restraints, i.e. restraints which can be derived by combining two existing distance restraints which refer to the same atom pair.

In contrast to other restraint instructions, the SAME instructions must be inserted at the correct positions in the atom list. These similarity restraints provide a very general and powerful way of exploiting non-crystallographic symmetry; in this example two instructions suffice to restrain the THF molecules so that they have (within an assumed standard deviation) twofold symmetry and are the same as each other. However we have not imposed planarity on the rings nor restricted any of the torsion angles.

To complicate matters, let us assume that the two molecules are two alternative conformations of a THF molecule disordered on a single site. We must then ensure that the site occupation factors of the two molecules add to unity, and that no spurious bonds linking them are added to the connectivity table. The former is achieved by employing site occupation factors of 21 (i.e. 1 times free-variable 2) for the first molecule and -21 [ 1*(1-fv(2)) ] for the five atoms of the second molecule. Free variable 2 is then the occupation factor of the first molecule; its starting value must be specified on the FVAR instruction. The possibility of spurious bonds is eliminated by inserting PART 1 before the first molecule, PART 2 before the second, and PART 0 after it. Hydrogen atoms can be inserted in the usual way using the HFIX instruction since the connectivity table is 'correct'; they will automatically be assigned the site occupation factors of the atoms to which they are bonded.

Finally we would like to refine with anisotropic displacement parameters because the thermal motion of such solvent molecules is certainly not isotropic, but the refinement will be unstable unless we restrain the anisotropic displacement parameters to behave 'reasonably' by means of rigid bond restraints (DELU) and 'similar Uij' restraints (SIMU); fortunately the program can set up these restraints automatically. The DELU restraints restrain the differences in the components of the displacement parameters of two atoms to zero along the 1,2- and 1,3-vector directions, and are derived with the help of the connectivity table. Since the SIMU restraints are much more approximate, we restrict them to atoms which, because of the disorder, are almost overlapping (i.e. are within 0.7 A of each other). Note that the SIMU restraints ignore the connectivity table and are based directly on a distance criterion specifically because this is a sensible way of handling disorder. In order to specify a non-standard distance cutoff which is the third SIMU parameter, we must also give the first two parameters which are the restraint esd's for distances involving non-terminal atoms (0.02) and at least one terminal atom (0.04) respectively. The '.ins' file now contains:

 HFIX 23 C12 > C15 C22 > C25
 ANIS O11 > C25
 DELU O11 > C25
 SIMU O11 > C25 0.02 0.04 0.7
 FVAR ..... 0.75
 ....
 PART 1
 SAME O21 > C25
 SAME O11 C15 < C12
 O11 4 ..... ..... ..... 21
 C12 1 ..... ..... ..... 21
 C13 1 ..... ..... ..... 21
 C14 1 ..... ..... ..... 21
 C15 1 ..... ..... ..... 21
 PART 2
 O21 4 ..... ..... ..... -21
 C22 1 ..... ..... ..... -21
 C23 1 ..... ..... ..... -21
 C24 1 ..... ..... ..... -21
 C25 1 ..... ..... ..... -21
 PART 0
An alternative type of disorder common for THF molecules and proline residues in proteins is when one atom (say C14) can flip between two positions (i.e. it is the flap of an envelope conformation). If we assign C14 to PART 1, C14' to PART 2, and the remaining ring atoms to PART 0 then the program will be able to generate the correct connectivity, and so we can also generate hydrogen atoms for both disordered components (with AFIX, not HFIX):

 SIMU C14 C14'
 ANIS O11 > C14'
 FVAR ..... 0.7
 ....
 SAME O11 C12 C13 C14' C15
 O11 4 ..... ..... .....
 C12 1 ..... ..... .....
 AFIX 23
 H12A 2 ..... ..... .....
 H12B 2 ..... ..... .....
 AFIX 0
 C13 1 ..... ..... .....
 PART 1
 AFIX 23
 H13A 2 ..... ..... ..... 21
 H13B 2 ..... ..... ..... 21
 PART 2
 AFIX 23
 H13C 2 ..... ..... ..... -21
 H13D 2 ..... ..... ..... -21
 AFIX 0
 PART 1
 C14 1 ..... ..... ..... 21
 AFIX 23
 H14A 2 ..... ..... ..... 21
 H14B 2 ..... ..... ..... 21
 AFIX 0
 PART 0
 C15 1 ..... ..... .....
 PART 1
 AFIX 23
 H15A 2 ..... ..... ..... 21
 H15B 2 ..... ..... ..... 21
 PART 2
 AFIX 23
 H15C 2 ..... ..... ..... -21
 H15D 2 ..... ..... ..... -21
 AFIX 0
 C14' 1 ..... ..... ..... -21
 AFIX 23
 H14C 2 ..... ..... ..... -21
 H14D 2 ..... ..... ..... -21
 AFIX 0
 PART 0
It will be seen that six hydrogens belong to one conformation, six to the other, and two are common. The generation of the idealized hydrogen positions is based on the connectivity table but also takes the PART numbers into account. These procedures should be able to set up the correct hydrogen atoms for all cases of two overlapping disordered groups. In cases of more than two overlapping groups the program will usually still be able to generate the hydrogen atoms correctly by making reasonable assumptions when it finds that an atom is 'bonded' to atoms with different PART numbers, but it is possible that there are examples of very complex disorder which can only be handled by using dummy atoms constrained (EXYZ and EADP) to have the same positional and displacement parameters as atoms with different PART numbers (in practice it may be easier - and quite adequate - to ignore hydrogens except on the two components with the highest occupancies!).

When the site symmetry is high, it may be simpler to apply similarity restraints using SADI or DFIX rather than SAME. For example the following three instruction sets would all restrain a perchlorate ion (CL,O1,O2,O3,O4) to be a regular tetrahedron:

 SAME CL O2 O3 O4 O1
 SADI  O1 O2  O1 O3
followed immediately by the atoms CL, O1... O4; the SAME restraint makes all the Cl-O bonds equal but introduces only FOUR independent restraints involving the O..O distances, which allows the tetrahedron to distort retaining only one -4 axis, so one further restraint must be added using SADI.

or:

 SADI  CL O1  CL O2  CL O3  CL O4
 SADI  O1 O2  O1 O3  O1 O4  O2 O3  O2 O4  O3 O4
or:

 DFIX 31  CL O1  CL O2  CL O3  CL O4
 DFIX 31.6330  O1 O2  O1 O3  O1 O4  O2 O3  O2 O4  O3 O4
in the case of DFIX, one extra least-squares variable (free variable 3) is needed, but it is the mean Cl-O bond length and refining it directly means that its esd is also obtained directly. If the perchlorate ion lies on a three-fold axis through CL and O1, the SADI method would require the use of symmetry equivalent atoms (EQIV $1 y, z, x and O2_$1 etc. for R3 on rhombohedral axes) so DFIX would be simpler (same DFIX instructions as above with distances involving O3 and O4 deleted) [the number 1.6330 in the above example is of course twice the sine of half the tetrahedral angle].

If you wish to test whether you have understood the full implications of these restraints, try the following problems:

(a) A C-O-H group is being refined with AFIX 87 so that the torsion angle about the C-O bond is free. How can we restrain it to make the 'best' hydrogen-bond to a specific Cl- ion, so that the H..Cl distance is minimized and the O-H..Cl angle maximized, using only one restraint instruction (it may be assumed that the initial geometry is reasonably good)?

(b) Restrain a C6 ring to an ideal chair conformation using one SAME and one SADI instruction. Hint: all 1-2, 1-3 and 1-4 distances are respectively equal for a chair conformation, which also includes a regular planar hexagon as a special case. A non-planar boat conformation does not have equal 1-4 distances. To force the ring to be non-planar, the ratio of the 1-2 and 1-3 distances would have to be restrained using DFIX and a free variable.


Ahead to Macromolecules and Other Structures with a Poor Data/Parameter Ratio

Back To Treatment of Hydrogen Atoms

Back to Table of Contents