Re: [ccp4bb] How high a B factor is too high to assume a loop is in place, in the AlphaFold era?

James Holton Thu, 01 Aug 2024 17:14:38 -0700

I submit that modern B factor restraints make them much less trashy thanthey were in the early days. As Pavel points out the exact strategiesdiffer from program to program, but I don't think anybody doesunrestrained B factor refinement. Not by default.

Besides, all we are really doing is fitting Gaussian-shaped peaks to the"curve" of the data. These peaks have a width and a height. Forexample, a carbon atom with B=20 has a peak density of 1.6 e-/A^3 and afull-width-at-half-max (FWHM) of 1.4 A. That is it! That is the modeldensity being fit. If you increase to B=80 the peak drops to 0.3 e-/A^3and the FWHM increases to 2.6 A. At the largest B you can stuff into aPDB file (999.99), the peak height is 0.008 e-/A^3 and the "peak" is8.45A wide. Your disordered loop, however, is probably not sampling froma symmetric Gaussian distribution like that. This is the real problemwith large B factors. They can fit better than sharper B atoms, but thatdoesn't mean they fit well.

Occupancy is easy because all it does is scale the height withoutaffecting the width. So, an 0.5 occupancy atom model is half the heightof a full-occupancy one. The width is unchanged. B factors impact bothwidth and height because they must preserve the number of electrons inthe peak. This is perhaps why they are often confusing and mysterious. We should also never forget that bulk solvent gets excluded with exactlythe same radii rules from every modeled atom, regardless of B factor andoccupancy. So, the "change in density" from adding or deleting an atomis a little more complicated than adding or subtracting a Gaussian peak.

Nevertheless, if you want to fit peak height and width independently(like we do in pretty much every other kind of curve fitting), then youshould refine occupancy and B factors at the same time.

Over-fitting you say? Hardly. Polynomials are easy to over-fit, but notGaussians. Observations/parameters is a useful guide for polynomialfits, but in general the hallmark of over-fitting is that the predictionpasses exactly through all the observed points (and not thecross-validation or "Rfree" points). I have never seen a macromolecularrefinement end up with Rwork = 0. Have you?

At the end of the day, what we do with our models is look at theirparameters and try to extract the physically meaningful reality they aretrying to capture. Restraints are very helpful in preventing many typesof unrealistic situations, but ultimately it is up to you to decide ifthe fitted model makes sense.


-James Holton
MAD Scientist

On 7/30/2024 11:30 AM, Ian Tickle wrote:

Obviously no refined parameters can ever be completely error-free,it's just that for the co-ordinates we have very accurate geometricrestraints so that the relative uncertainty in the refinedco-ordinates is small (but try refining co-ordinates withoutrestraints!). For the B factors we don't have accurate estimates (ifany) for their restraints so their relative uncertainty afterrefinement is much greater.


-- Ian

On Tue, Jul 30, 2024 at 6:57 PM Oganesyan, Vaheh<vaheh.oganes...@astrazeneca.com> wrote:


    Yes, it is and I like the definition of shared “trash bin”. It
    will have more physical meaning if we can separate those
    contributions into separate bins.

    Vaheh

    *From:* Pavel Afonine <pafon...@gmail.com>
    *Sent:* Tuesday, July 30, 2024 1:51 PM
    *To:* Oganesyan, Vaheh <vaheh.oganes...@astrazeneca.com>
    *Cc:* CCP4BB@jiscmail.ac.uk
    *Subject:* Re: [ccp4bb] How high a B factor is too high to assume
    a loop is in place, in the AlphaFold era?

    Vaheh,

    I think coordinates are no different from B factors, occupancies,
    f', or f'' in this respect. Coordinates can play their "trash bin"
    role by adjusting to the noise at the expense of violated geometry
    (bonds, angles, planes, torsions, etc.). As I mentioned in my
    previous email, their trash bin capacity is much smaller (but
    definitely not zero!) because the number and strength (confidence)
    of geometry restraints are much greater than those of ADP restraints.

    I agree that all refined parameters share this trash bin capacity,
    but to varying degrees. Isn't this essentially what we call the
    error on the refined parameter? All refined parameters have their
    error bars, which we have referred to as the "trash bin" in this
    thread.

    Pavel

    On Tue, Jul 30, 2024 at 10:09 AM Oganesyan, Vaheh
    <vaheh.oganes...@astrazeneca.com> wrote:

        Your point is taken, Pavel. However, despite resolution, you
        define coordinate of the atom as a geometric point with no
        width. Although coordinates are “refineable”, they have no
        capacity for “trash”. Their “trash” still goes into B-factor
        “trash bin”. At least this is how I see it.

        Thank you.

        *Vaheh Oganesyan, Ph.D.*

        *R&D **| Biologics Engineering*

        One Medimmune Way, Gaithersburg, MD 20878

        T:  301-398-5851

        _vaheh.oganes...@astrazeneca.com
        <mailto:oganesy...@medimmune.com>_

        *From:* Pavel Afonine <pafon...@gmail.com>
        *Sent:* Tuesday, July 30, 2024 11:45 AM
        *To:* Oganesyan, Vaheh <vaheh.oganes...@astrazeneca.com>
        *Cc:* CCP4BB@jiscmail.ac.uk
        *Subject:* Re: [ccp4bb] How high a B factor is too high to
        assume a loop is in place, in the AlphaFold era?

        From this perspective, all refinable atomic model parameters
        can be viewed as trash bins, with the size of these bins being
        proportional to the amount of prior information (restraints)
        imposed on these parameters. For example, coordinates have the
        most restraints and thus are the smallest trash bins, while B
        factors have the least restraints and thus are one of the
        largest bins.

        Pavel

        On Tue, Jul 30, 2024 at 8:25 AM Oganesyan, Vaheh
        <vaheh.oganes...@astrazeneca.com> wrote:

            Early in my Crystallography life I was postdoc with Robert
            Huber in Munich. We had those gatherings once a week when
            in very informal way we can ask and answer questions. I
            remember my question about B factors: how is it possible
            to have high resolution structure and average B-factor of
            100A^2 . I think it was Robert or Albrecht Messerschmidt
            who told that B-factor is a “trash can” that describes not
            only loosely positioned atoms but also all other problems
            that either you created during processing, harvesting or
            crystal had from the beginning.

            *Vaheh Oganesyan, Ph.D.*

            *R&D **| Biologics Engineering*

            One Medimmune Way, Gaithersburg, MD 20878

            T:  301-398-5851

            _vaheh.oganes...@astrazeneca.com
            <mailto:oganesy...@medimmune.com>_

            *From:* CCP4 bulletin board <CCP4BB@JISCMAIL.AC.UK> *On
            Behalf Of *James Holton
            *Sent:* Tuesday, July 30, 2024 10:35 AM
            *To:* CCP4BB@JISCMAIL.AC.UK
            *Subject:* Re: [ccp4bb] How high a B factor is too high to
            assume a loop is in place, in the AlphaFold era?

            How high B factors can go depends on the refinement
            program you are using.

            In fact, my impression is that the division between the
            "let the B factors blow up" and "delete the unseen" camps
            is correlated to their preferred refinement program. You
            see, phenix.refine is relatively aggressive with B factor
            refinement, and will allow "missing" atoms to attain very
            high B factors. Refmac, on the other hand, has restraints
            that try to make B factor distributions look like those
            found in the PDB, and so tends to keep nearby B factors
            similar. As a result, you may get "red density" for
            disordered regions from refmac, inviting you to delete the
            offending atoms, but not from phenix, which will raise the
            B factor until the density fits.

            Then there are programs like VagaBond that don't formally
            have B factors, but rather let an ensemble of chains
            spread out in the loopy regions you are concerned about. 
            This might be the way to go?

            You can also do ensemble refinement in the latest Amber.
            That is, you run an MD simulation of a unit cell (or more)
            and gradually increase structure factor restraints. This
            would probably result in the "fan" of loops you have in mind?

            -James Holton
            MAD Scientist

            On 7/28/2024 8:13 AM, Javier Gonzalez wrote:


                        

                Dear CCP4bb,

                I'm refining the ~3A crystal structure of a big
                protein, largely composed of alpha helices connected
                by poorly-resolved loops.

                In the old pre-AlphaFold (AF) days I used to simply
                remove those loops/regions with too high B factors,
                because there was little to none density at 1 sigma in
                a 2Fo-Fc map.

                However, considering that the quality of a
                readily-computable AF model is comparable to a 3A
                experimental structure, and that the UniProt database
                is flooded with noodle-like AF models, I was
                considering depositing a combined model in the PDB.

                Once R/Rfree reach a minimum for the model truncated
                in poorly resolved loops, I would calculate an
                augmented model with AF calculated missing regions
                (provided they have an acceptable pLDDT value), assign
                them zero occupancy, and run only one cycle of
                refinement to calculate the formal refinement statistics.

                Would that be acceptable? Has anyone tried a similar
                approach?

                I'd rather do that instead of depositing a
                counterintuitive model with truncated regions that few
                people would find useful!!

                Thank you for your comments,

                Javier

--

                Dr. Javier M. González
                Instituto de Bionanotecnología del NOA
                (INBIONATEC-CONICET)
                Universidad Nacional de Santiago del Estero (UNSE)
                RN9, Km 1125. Villa El Zanjón. (G4206XCP)
                Santiago del Estero. Argentina

                Tel: +54-(0385)-4238352

                Email <mailto:bio...@gmail.com> Twitter
                <https://twitter.com/_biojmg>

                
------------------------------------------------------------------------

                To unsubscribe from the CCP4BB list, click the
                following link:
                https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1
                
<https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1>


            
------------------------------------------------------------------------

            To unsubscribe from the CCP4BB list, click the following link:
            https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1
            <https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1>


            
------------------------------------------------------------------------

            To unsubscribe from the CCP4BB list, click the following link:
            https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1
            <https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1>



    ------------------------------------------------------------------------

    To unsubscribe from the CCP4BB list, click the following link:
    https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1
    <https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1>


------------------------------------------------------------------------

To unsubscribe from the CCP4BB list, click the following link:

https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1<https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1>


########################################################################

To unsubscribe from the CCP4BB list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCP4BB&A=1

This message was issued to members of www.jiscmail.ac.uk/CCP4BB, a mailing list 
hosted by www.jiscmail.ac.uk, terms & conditions are available at 
https://www.jiscmail.ac.uk/policyandsecurity/

Re: [ccp4bb] How high a B factor is too high to assume a loop is in place, in the AlphaFold era?

Reply via email to