Structure refinement
Completing the structure and evaluating how good
your data and model agree
Why you should refine a structure
We have considered how atoms are located by Patterson, direct
methods or specialized methods, as well as from Fourier
difference maps
The atomic positions extracted from these methods are close to
the correct values, but very rarely in exactly the right place
During a process called refinement, the starting atomic
positions are optimized
- the goal is to get an as good fit between your model and your data as
possible
Possible refinement methods
You can modify your model in order to obtain better
agreement between observed and calculated Fourier maps
Alternatively, you can modify your model to get better
agreement between observed and calculated structure
factor magnitudes (|Fo| and |Fc|)
For most small molecule structure solutions, the latter
process is used
Letting the computer do the work
The refinement process can be described as a minimization process
- minimization of the difference between |Fo| and |Fc|
- easily automated
The best model is obtained by minimizing the following expression:
R' = whkl ( Fo k Fc ) 2
hkl
- called a least squares minimization
- wkhl are weighing factors related to the quality/reliability of the data
- k is a scale factor (inverse of scale factor that would need to be applied to |Fo|
Least squares minimization
Imagine a function F that is linearly dependent on a set of
parameters, xj
- F(x1,x2,xn) = p1x1 + p2x2 + pnxn
Suppose we make m independent measurements of F for different
values of xj
- we want to get the parameters pj
- if m = n, we just have to solve a set of simultaneous equations
- if m > n, the system is overdetermined
Crystallography and least squares
The crystallographic function F(hkl) (with the variables: atomic
coordinates and thermal parameters of each atom) is not a linear
function of the model parameters
N
F(hkl) = f j exp[2i (hx j + ky j + lz j ) (hkl )]
j =1
The equation cannot be solved for the correct parameters in one go
Use an iterative process instead
Refinement iterations
2
R
'
=
w
(
F
k
F
)
Our goal is to minimize
hkl o
c
hkl
We can use an iterative least squares approach to give parameter
shifts that will lead to an improved agreement between |Fo| and |Fc|
The process is repeated until the suggested shifts are insignificantly
small
- usually considered to be achieved when parameter shift << standard deviation
- at this minimum, the derivative of R with respect to each parameter (xj, yj, zj
Bj) should be zero
Convergence
A refinement process is usually considered as finished when
convergence is reached
- when all parameter shifts << standard deviation
However, the process only works well if the starting model is
sufficiently good
- many local minima in which the refinement could get stuck
Other methods offer better chances to find the absolute minimum
from a bad starting model
- simulated annealing, random walk
False minima
Errors
The errors of all observations are included in the refinement process
via the weight factors
- weak or uncertain reflections will have less weight than strong reflections
The least squares output will proved esds (estimated standard
deviations) for all refined parameters
Watch out for high correlation coefficients
- correlation coefficients tell you whether two model parameters really are
independent
Judging the refinement
Statistical values are used to judge the goodness of a refinement
The level of agreement between observed and calculated structure
factors is often indicated by R factors and Goodness of Fit (GooF)
values
RF = ( Fo (hkl ) k Fc (hkl ) ) / | Fo (hkl ) |
hkl
hkl
2 2
wRF 2 = { [ w( Fo (hkl ) Fc (hkl ) ) ] / [ w | Fo (hkl ) |2 ]}1/ 2
2
hkl
hkl
GooF = S = { [ w( Fo (hkl ) Fc (hkl ) ) 2 ] /( n p )}1/ 2
2
hkl
- for n observations and p parameters
R factors
For a good small molecule refinement, the final RF values
are expected to be ~0.02-0.08
Placing random atoms in a unit cell is expected to give R
factors of 0.83 and 0.59 for centric and acentric space
groups, respectively
Obtaining RF < 0.20 usually means that the structural model
has no major errors in it
Refinement strategy
Unit cell constants need to be refined
- original indexing just gives approximate values
Atomic positions obtained by Patterson searches, direct methods or
other approaches should be refined
- look at interatomic distances to judge whether refined positions make sense
- atoms in special positions cannot move in all directions
Atomic displacement parameters should be varied
- can indicate wrong atomic weight: Zmodel > (<) Zreal leads to large (small) ADPs
- isotropic overall temperature factors can be obtained from Wilson plots as a
first approximation
Estimating the overall temperature factor
We know that observed structure factor magnitudes are smaller than
real values because of thermal motion and scaling issues
K Fo (hkl ) F (hkl ) = f e
2
2 2 B sin 2 / 2
ln[ Fo (hkl ) / f 2 ] = ln K 2 B sin 2 / 2
2
- for random atom placement in the unit cell
Both K and B can be obtained from a Wilson plot
Wilson plots
Restraints and constraints
It is possible to impose certain restrictions on the refinement due to
some knowledge that is not inherent in the diffraction data
- restraints and constraints
Chemical knowledge, such as bond distances and angles, can be used
as a restraint in a minimization procedure
- a restraint will make certain moves unfavorable, but will not prohibit them
Knowledge about molecular connectivity can be used
- e.g., a rigid aromatic ring can be described by three positional and three
rotational parameters instead of 4n parameters (n = number of atoms)
- this would be a rigid body constraint
When to use restraints and constraints
Restraints and constraints can improve your data to parameter ratio
by reducing the number of parameters or adding observations
- very useful if your data is limited or of low quality
They improve the convergence properties of your refinement and
may allow a refinement to converge to the correct answer even if the
starting model is poor
- they make potentially disastrous parameter changes (e.g., one carbon atom
moving so far that the aromatic ring will no longer be connected) unfavorable
or prohibit them altogether
