TLS Refinement in Crystallography
TLS stands for Translation Libration Screw-motion (the dash makes it acronym-ically fine) which is a method of refinement in the program REFMAC5 within the CCP4 suite or in phenix.refine. According to developer, Martyn Winn, TLS refinement can be at almost any resolution.
Why should I use it?
The benefits of using TLS refinement is that it can reduce your Rfree and Rwork values. The implication being that the produced structural model will be a better representation of the collected data.
How Does it Work?
TLS refines ’sequence groups’ that are described using 20 parameters per each group.
How Do I Determine the Groups?
The TLS Motion Determination (TLSMD) is a server that allows for the submission of your amino acid sequence and recommends how to segment your sequence (ref). A number of different TLS groups are possible for the same sequence (ref).
—
How do I actually do this?
1) Do a rigid body refinement followed by ~10 rounds of restrained refinement
2) Take this output and submit it to the TLSMD
3) Take the segments that are produced by the TLSMD and fix the B factors to 40 (ref)
Note: the B factor was set to 20 in the literature reference
4) REFMAC needs the following inputs: REFI TLSC 20, TLSIN, BFAC SET 40 (more details and here)
5) Perform TLS refinement
6) Perform restrained refinement followed by the addition of ligands, ions and solvent
–
How to do you know if TLS helped?
A decrease in the Rfree value as well as an improvement in the electron density maps.
I have done my best to condense about 100 pages of websites, presentations and literature into 250 words. Please let me know what I need to change/add/remove to make this post more helpful, thanks!
Sili
December 5th, 2009 at 9:11 AM #
I’m greatly in favour of Transient Libations Screws.
Oh … nevermiiiind …
Sean
December 6th, 2009 at 9:25 PM #
I think you are getting more out of this than myself.
Matt k
December 7th, 2009 at 12:37 PM #
I don’t understand the logic of picking a fixed arbitrary number for fixing the B-factor for all possible datasets. I tend to use the mean B from the Wilson plot – it should be close to the best global B-factor for the dataset in question. Also, I find that the best way to optimize TLS groups is simply to iteratively divide the structure into smaller and smaller groups while monitoring Rfree. For high resolution structures, the optimal group size can be as small as 3 residues.
Sean
December 7th, 2009 at 8:01 PM #
Hi Matt k,
I didn’t come across an explanation of why selecting 20 or 40 for the B-factor was done. I imagine that the Wilson is giving a value around 40 or maybe a little higher (although it is hard to generalize). In a high resolution structure the data to parameter ratio should be quite good and therefore allow for more groups. If the structure is very high then anisotropic refinement should be possible.
I would really like to see a clear explanation of the data to parameter ratios vs. resolution that have deposited in the PDB. I fear that this gets into the hand waiving of what the final resolution of a particular structure should be.
Ethan A Merritt
December 11th, 2009 at 6:55 PM #
Sean: The constant value picked as a starting point for B factors in TLS refinement is purely arbitrary, as is the origin for the TLS group. In both cases the refined set of 20 TLS parameters will factor out these arbitrary constants, yielding the same net description of anisotropy at the end of the day. In other words, it doesn’t matter what number you choose as a starting point. I myself favor choosing the Wilson B, just because if I have to pick an arbitrary number it might as well mean _something_.
Matt: There is a combinatorial explosion of possible ways to subdivide a chain into smaller segments. The chance of happening across an “optimal” choice of segment boundaries by the procedure you describe is very low. That is the rationale behind using TLSMD. It defines a measure of optimality and then uses a clever but computationally intensive trick to find the choice of segment boundaries that satisfies that criterion for being optimal. In principle one can argue about the definition of “optimal” that the TLSMD server should use, but so far no one has proposed a specific alternative definition that we could try instead.
Sean
December 20th, 2009 at 4:03 PM #
Hello Ethan,
Thanks so much for this insightful reply and creating this invaluable resource.
Dima
December 21st, 2009 at 11:22 AM #
Someone please enlighten me on how to rationally pick a number of TLS groups. So far – without a single exception – the more groups I chose (typically up to 20 as suggested by Ethan’s server), the less Rfree becomes. I have a feeling that after a certain threshold the effect is simply increasing a number of parameters in the model (2 groups vs 20 groups = 40 vs 400 parameters).
Thanks!
Sean
December 22nd, 2009 at 11:18 PM #
Hi Dima,
I don’t know if there is one (or at least not one that I have found). TLSMD describes three scenarios:
1) increasing TLS groups decreases the Rfree
2) increasing TLS groups can eventually increase the Rfree
3) TLS alone can produce lower Rfree than with refining individual B factors
The algorithm in TLSMD is what is capable of generating these ‘ideal sub-models.’ The optimal divisions do not appear to precisely match hinge or domains therefore a rational approach seems difficult.
Hope that helps.
Pavel
November 8th, 2011 at 9:39 PM #
phenix.find_tls_groups will find TLS groups automatically, and, give you unique choice for TLS groups selection. Using it from PHENIX GUI will allow seeing (and editing, if desired) TLS groups graphically. Also, it works from 10 to 90 times faster than TLSMD, and can benefit from multiple CPUs.
Pavel
Pavel
November 8th, 2011 at 9:43 PM #
The most recent and comprehensive review on the subject of TLS:
see “TLS for dummies” here (~50 pages):
http://www.phenix-online.org/newsletter/
and may be also relevant “On atomic displacement parameters (…)”.
Hope that answers most of the questions above.
Pavel