Hi Tristan,

  what you observe makes a lot of sense - and I would advocate the use of TLS for sure. Any B-factor model is going to have problems when the underlying coordinates are incorrect. 

  Cheers,
	Paul

> On Feb 12, 2016, at 2:08 AM, Tristan Croll <[log in to unmask]> wrote:
> 
> Hi Paul,
> 
> The structures that I've been primarily focusing on are mostly historical ones like the one that belongs to the attached validation slider, that were largely or entirely hand-built into 3.5-4 Angstrom density using mid-2000s tools. It's no secret that large errors are common in such structures - not just poor backbone conformation or incorrect rotamers, but substantial shifts in register. In such situations, if you give the refinement an inch it will take a mile. When you have register shifts that leave a Ser in place of a Met, or loops threaded through density that turns out to be fictional, *any* B-factor you apply will be wrong, and if it has the freedom to do so the refinement will do its best to pick a B-factor that makes the model seem less wrong. Nearest-neighbour restraints certainly help, but when the neighbour (and its neighbour, and *its* neighbour) is also wrong, then you're still left with scope for a lot of over-fitting - and every individual instance of over-fitting has a ripple effect throughout the structure.
> 
> In these situations, I've found time and again that the simpler TLS-only model leads to maps that are *far* more interpretable, where I can clearly and easily identify the source and solution of a problem that was previously obscure. The 2mFo-DFc maps become sharper and stronger - but more importantly, so do the mFo-DFc maps. A great deal of noise seems to go away, and weak regions in particular tend to become more connected (not model bias as far as I can tell, because the new connections tend to point to improvements in the structure).
> 
> All that being said, once the model is closer to a finished state where the bulk (>>90%) is well-fitted and all the quality indicators are in good shape (by "good shape" here, I mean that I currently have this structure at a clashscore of 2.6, Rama favoured/outliers at 93.5/0.48%, and rotamer outliers at 1.17%, with Rfree almost 3% lower than the original), then I find that adding restrained individual B-factor refinement can help - *if* the TLS-only model has converged first. If I start from an entirely reset model, it's very variable - sometimes it converges well, sometimes it runs away entirely... and sometimes, like in my original example here, it seems to converge well but starts losing previously-resolved details.
> 
> Best regards,
> 
> Tristan
> 
> ________________________________________
> From: CCP4 bulletin board <[log in to unmask]> on behalf of Paul Adams <[log in to unmask]>
> Sent: Friday, 12 February 2016 4:23 AM
> To: [log in to unmask]
> Subject: Re: [ccp4bb] Individual B-factors at low resolution: a cautionary tale?
> 
> We observe the same thing with the ADP restraints used in Phenix. The local distance-based restraints on displacements are very effective even at low resolution. There is scope for improvement no doubt, but users are encouraged to use individual restrained ADPs even at low resolution rather than grouped displacements.
> 
>> On Feb 11, 2016, at 9:45 AM, Zbyszek Otwinowski <[log in to unmask]> wrote:
>> 
>> Refmac has made a lot of progress in refining individual B-factors at very
>> low resolution. At such resolution, the procedure is not always stable;
>> however, with some effort, we have found a considerable improvement of
>> R-free in all our 4-5 A resolution limit data.
>> 
>> The same applies to the refinement of anisotropic B-factors, it may work
>> at much lower resolution (e.g. 2A) than typically assumed.
>> 
>> To achieve improvement in the above situations, we perform many
>> alternative strategies of the refinement. These strategies depend on the
>> version of Refmac and type of problem and quality of the model. The better
>> the model, the more improvement we observe.
>> 
>> Our experience may, or may not, apply to Phenix and other refinement
>> programs as the differences between programs are likely to be significant.
>> Our observations are consistent with past negative experience with regard
>> to B-factor refinement; however, in most situations we could find a
>> work-around.
>> 
>> Zbyszek Otwinowski
>> 
>> 
>>> There must be some middle ground here, in the form of a restraint scheme
>>> that allows one to gracefully reduce the effective parameter count as the
>>> resolution decreases, without "step changes" between different schemes.
>>> Perhaps by applying some variation on a LOWESS smooth to the B-factors,
>>> with the strength of the weighting term determined by the Wilson B factor?
>>> 
>>> 
>>> ________________________________
>>> From: Bjørn Pedersen <[log in to unmask]>
>>> Sent: Thursday, 11 February 2016 6:05 PM
>>> To: Tristan Croll
>>> Cc: [log in to unmask]
>>> Subject: Re: [ccp4bb] Individual B-factors at low resolution: a cautionary
>>> tale?
>>> 
>>> Hi Tristan,
>>> I think your reasoning makes a lot of sense. I would be skeptical of the
>>> use of individual B-factors at this resolution of 3.6A in general, working
>>> with 'low-quality' anisotropic data etc. Regions with extremely high
>>> b-factors should help to warn you that you could be 'falling' into a local
>>> minimum hole.
>>> With think kind of data I often find it useful to 'reset' all B-factors in
>>> my current model to the estimated Wilson B-factor once in a while. That
>>> helps me to escape the local minimum of the model (along with multiple
>>> parallel rounds of simulated annealing tracking the R-factors of the
>>> low-resolution shells).
>>> 
>>> The followup question I have is more along the lines of whether it would
>>> ever be justifiable to use individual B-factors at this resolution? I have
>>> always been a proponent of the one/two B-factors per residue in this
>>> resolution range, but I was wondering what the community thinks of this
>>> these days, with current improvements in target functions, parametizations
>>> etc etc?
>>> 
>>> All the best
>>> -Bjørn
>>> 
>>> 
>>> 
>>> 
>>> 
>>> 
>>> On Mon, Feb 8, 2016 at 7:46 PM, Tristan Croll
>>> <[log in to unmask]<mailto:[log in to unmask]>> wrote:
>>> 
>>> For the most part they haven't moved far (1-2 tenths of an Angstrom in the
>>> backbone, further for sidechains) - and secondary structure remains
>>> essentially unchanged, which is a plus. I think it will come good with a
>>> little more tinkering.
>>> 
>>> 
>>> ________________________________
>>> From: Eleanor Dodson
>>> <[log in to unmask]<mailto:[log in to unmask]>>
>>> Sent: Tuesday, 9 February 2016 4:28 AM
>>> 
>>> To: Tristan Croll
>>> Cc: [log in to unmask]<mailto:[log in to unmask]>
>>> Subject: Re: [ccp4bb] Individual B-factors at low resolution: a cautionary
>>> tale?
>>> 
>>> Yes - the really isnt enough information at that resolution to support the
>>> number of parameters with xy,z,b per atom. Do your atoms move away from
>>> their correct positions too?
>>> 
>>> On 8 February 2016 at 18:22, Tristan Croll
>>> <[log in to unmask]<mailto:[log in to unmask]>> wrote:
>>> 
>>> 
>>> That's OK - I tend to be very hands-on with my corrections. :)
>>> 
>>> 
>>> But what has me interested is not so much this, but that refining with
>>> individual B-factors actually seems to end up obscuring the information
>>> that says where the wrong atoms should go! In this particular case I went
>>> through multiple rounds of rebuilding/refinement of this domain, where
>>> successive adjustments simultaneously improved fit to the map, resolved
>>> clashes and improved the secondary structure, and each refinement with a
>>> TLS-only model led to sharper and stronger density. Then with a few rounds
>>> of rebuilding elsewhere combined with individual B-factor refinement, it's
>>> all but gone. I think it really argues for the idea that at these
>>> resolutions the B-factor model should be kept as simple as possible while
>>> rebuilding, and only extended to individual B-factors (if at all) in a
>>> final round for deposition.
>>> 
>>> 
>>> Cheers,
>>> 
>>> 
>>> Tristan
>>> 
>>> ________________________________
>>> From: Eleanor Dodson
>>> <[log in to unmask]<mailto:[log in to unmask]>>
>>> Sent: Tuesday, 9 February 2016 4:02 AM
>>> To: Tristan Croll
>>> Cc: [log in to unmask]<mailto:[log in to unmask]>
>>> Subject: Re: [ccp4bb] Individual B-factors at low resolution: a cautionary
>>> tale?
>>> 
>>> Yes - I think you are right. We use "B factors" as mop-up-error factors.
>>> If the atoms are in the wrong place a very high B factor is a useful
>>> indicator that the atom should be deleted or moved! But you will probably
>>> need to do some hands-on correction to use the information
>>> Eleanor
>>> 
>>> 
>>> 
>>> On 8 February 2016 at 10:18, Tristan Croll
>>> <[log in to unmask]<mailto:[log in to unmask]>> wrote:
>>> 
>>> Hi all,
>>> 
>>> 
>>> The attached image depicts the weakest region of the 3.6 Angstrom
>>> structure I've been working on. The three maps shown are 2mFo-DFc at 1
>>> sigma, from three different refinements. The purple one is the first,
>>> after extensive rebuilding and refinement using strictly a TLS-only
>>> B-factor model. Not strong, but after sharpening and cross-checking with
>>> its slightly better resolved NCS partner, enough to be happy with it. The
>>> green map is the result of taking the refined TLS-only model and further
>>> refining with individual B-factors. So far so good - the maps are more or
>>> less the same.
>>> 
>>> 
>>> The blue surface is the current map, after multiple rounds of rebuilding
>>> in the (much) more strongly resolved regions, with TLS plus restrained
>>> individual B-factor refinement from a blank slate in between each round.
>>> It's looking... not so great.
>>> 
>>> 
>>> This result make a lot of sense when I think about it further - but just
>>> to check if my reasoning is correct:
>>> 
>>> 
>>> One way to look at refinement with a single overall B-factor is that
>>> you're implicitly "flattening" your model - increasing the contribution of
>>> the weakly resolving regions, and decreasing the contribution of the
>>> stronger regions - akin to adjusting the contrast in a photograph. That's
>>> reflected (no pun intended) in the maps becoming stronger in these areas
>>> and a general sharpening throughout, even if the R factors are 1-2% higher
>>> than with individual B-factors. Most importantly, though, I think it
>>> forces the refinement algorithms to pay more attention to the coordinates
>>> in these regions. Once these are refined to convergence in the TLS-only
>>> B-factor model, then it seems safe to introduce individual B-factors since
>>> the refinement will simply fall further into the current local minimum.
>>> But if the model is refined from scratch with individual B-factors, then
>>> it's much easier for the refinement to over-fit the strongly resolving
>>> regions, balanced by smearing out the weak ones - significantly reducing
>>> the interpretability of weaker regions and resulting in an overall
>>> poorer-quality model.
>>> 
>>> 
>>> Does this make sense?
>>> 
>>> 
>>> Best regards,
>>> 
>>> 
>>> Tristan
>>> 
>>> 
>>> 
>>> 
>> 
>> 
>> Zbyszek Otwinowski
>> UT Southwestern Medical Center at Dallas
>> 5323 Harry Hines Blvd.
>> Dallas, TX 75390-8816
>> Tel. 214-645-6385
>> Fax. 214-645-6353
> 
> --
> Paul Adams
> Interim Division Director, Molecular Biophysics & Integrated Bioimaging, Lawrence Berkeley Lab
> Division Deputy for Biosciences, Advanced Light Source, Lawrence Berkeley Lab
> Adjunct Professor, Department of Bioengineering, U.C. Berkeley
> Vice President for Technology, the Joint BioEnergy Institute
> Laboratory Research Manager, ENIGMA Science Focus Area
> 
> Building 33, Room 347
> Building 80, Room 247
> Building 978, Room 4126
> Tel: 1-510-486-4225, Fax: 1-510-486-5909
> http://cci.lbl.gov/paul
> 
> Lawrence Berkeley Laboratory
> 1 Cyclotron Road
> BLDG 33R0345
> Berkeley, CA 94720, USA.
> 
> Executive Assistant: Louise Benvenue [ [log in to unmask] ][ 1-510-495-2506 ]
> --
> <multipercentile_validation.png>

-- 
Paul Adams
Interim Division Director, Molecular Biophysics & Integrated Bioimaging, Lawrence Berkeley Lab
Division Deputy for Biosciences, Advanced Light Source, Lawrence Berkeley Lab
Adjunct Professor, Department of Bioengineering, U.C. Berkeley
Vice President for Technology, the Joint BioEnergy Institute
Laboratory Research Manager, ENIGMA Science Focus Area

Building 33, Room 347
Building 80, Room 247
Building 978, Room 4126
Tel: 1-510-486-4225, Fax: 1-510-486-5909
http://cci.lbl.gov/paul

Lawrence Berkeley Laboratory
1 Cyclotron Road
BLDG 33R0345
Berkeley, CA 94720, USA.

Executive Assistant: Louise Benvenue [ [log in to unmask] ][ 1-510-495-2506 ]
--