Back to the tutorial main page

REFMAC5 Tutorial - Part 3

Twin Refinement

In this tutorial, we will use REFMAC5 to refine the model of heme oxygenase with PDB code 1irm, which corresponds to a twinned crystal in P32, collected to 2.55Å.

A basic working knowledge of the CCP4i2 GUI is assumed.

Before We Start

Required files:

• 1irm.pdb - the current model;
• 1irm.mtz - the reflection data.

Please click these links to access the files and save them to your computer.

Contents

• Attempt Standard Refinement
• Perform Twin Refinement
• Inspecting Twin Refinement Results

Attempt Standard Refinement

We will begin by attempting to refine the structure without specifying that it is twinned:

• Click on the "Task menu" at the top-left of the CCP4i2 interface, and navigate to "Refinement - REFMAC5". Show Me This
• A new job will be created. The main inputs may already be populated with data from a previous job, in which case select "No" from the "Use data from job" selection box. Show Me This
• Give the job a meaningful title, e.g. "Tutorial Example 3 - Twin Refinement - No twin". Show Me This
• Provide the files 1irm.pdb and 1irm.mtz as inputs in the "Atomic model", "Reflections", and "Free R set" fields. Show Me This
• Click "Run" to start the job, and continue with the tutorial while this job runs.... Show Me This

Perform Twin Refinement

Since the crystal is twinned, we should perform twin refinement:

• Click on the "Clone job" button at the top of the interface. This will create a new job, pre-populated with the inputs from the previous job.
• Give the job a meaningful title, e.g. "Tutorial Example 3 - Twin enabled". Show Me This
• The "Advanced Options" tab includes a text box where additional keywords can be written - any REFMAC5 keywords placed here will be used. Since "twin" is the keyword to enable twin refinement in REFMAC5, replace the contents of the keywords box with the word "twin". This will tell refmac to test whether the crystal is twinned, determine the number of twin domains, and refine the twin fractions. Show Me This
• Click "Run" to start the job. Show Me This

Intensity based vs amplitude based twin refinement

In REFMAC5, you can choose to perform either "intensity based" or "amplitude based" twin refinement. The "intensity based" option uses reflection intensities directly, whilst the "amplitude based" option will convert structure factor amplitudes into (approximate) intensities for subsequent use. Theoretically, using intensities directly is better. In practice, this may not always be the case. When using "amplitude based" twin refinement, the R-factors are usually lower than when you use "intensity based" twin refinement. However, this does not mean that models derived using "amplitude based" refinement are better.

Inspecting Twin Refinement Results

Now wait for the job to finish, and inspect the results:

Show Me This

• The Results page now contains an addition section labelled "Refmac analysis of twinning". Expand this section.
• Refmac identified that there are three unique symmetry operators (thus four potential twin domains) possible for the lattice, given the assumed space group. The Results page lists the Rmerge calculation for each of the identified symmetry operators. This scores the agreement between potentially twin-related reflections. If Rmerge is very small, then the crystal symmetry may be higher than was assumed during data processing. Any symmetry operators with corresponding Rmerge greater than 0.44 are automatically removed. Since all three symmetry operators have Rmerge less than 0.44, all four twin domains are kept in this case.
• Refmac then refined the twin fractions for the four domains as 0.385, 0.116, 0.385 and 0.116, respectively. These are considered sufficiently large, so all four twin domains are kept in this case.
• Now compare the refinement statistics for the two refinement runs - i.e. with and without assuming that the crystal is twinned. Clearly, the Rfactors are greatly improved when enabling the twinning option.
  Show Me The Results - Without Assuming Twinning
  Show Me The Results - Assuming Twinning

Notes on twin refinement

If a domain has a very small twin fraction, then the twin fraction may not be reliable, especially during earlier stages of refinement. In such cases, it could be that the data are not twinned. REFMAC5 automatically removes domains with a small twin fraction (less than 0.07).

It is important to remember that reduction in R-factors is not sufficient evidence to suggest twinning. Even when data are not twinned, models refined under the assumption of twinning can systematically have lower R-factors, despite not resulting in a model of improved quality. For more information, see Murshudov (2011).

Although R-factors are substantially smaller with twin refinement than without twin refinement, it does not mean that model also is substantially better. Most parts of the electron density will be very similar with/without using twin refinement. However, for weak parts of the electron density, the signal may be more visible in correctly treated twin refinement cases. Again, care should be exercised - usually, maps after twin refinement are more biased towards errors in the model. The reason for this is that when crystal is twinned then the amount of information in the data about the model is decreased. This is an intrinsic property of data from twinned crystals.

Notes on free R flag selection and twinning

Note that in this case the R-factor and R-free diverge during refinement, and R-free increases. One reason for this is that the data were taken directly from the PDB, and the model was not originally refined with the twin flag. Furthermore, twinning relates two (or more) sets of reflections. Usually, selection of the free R set does not take this into account. REFMAC5 puts twin-related reflections into the same set (i.e. either all related reflections belong to the "free" or "working" sets).

In general, free R set selection should be done before starting refinement, accounting for potential twin operators. One way of achieving this is to select free reflections in a higher space group and expand them, which would ensure that all related reflections belong to the same set throughout refinement and all other treatment of data.

You can use Coot to view the output coordinates and density map. Compare the models that were refined with and without using twin refinement. Can you see any difference in the quality of the maps?

Back to the tutorial main page