[phenixbb] geometry_minimization makes molprobity score worse

James Holton jmholton at lbl.gov
Mon Jul 12 17:55:58 PDT 2021

Thank you Dale for the thoughtful explanation.

I understand and agree that the "ideal" geometry is just an 
approximation.  But is our approximation so bad that that it is 
impossible to construct a model that doesn't have "problems"?  As in: do 
we always have to decide between strained bonds and angles vs bad 
clashes and torsional outliers?  Even in the absence of an x-ray term?

I don't believe that is so.  My evidence to support this is that I can 
get pretty close to perfect molprobity scores using amber (imin=1), 
refmac (refi type ideal), and pdb2ins/shelxl (WGHT 0.1 0 0 0 0 0), but 
not with phenix.geometry_minimization. I thought perhaps the phenix 
developers might find that interesting?  I certainly do.

I do get a better clash score if I provide a model with phenix.reduce 
hydrogens, but I always get Rama outliers. CDL or no CDL. And, of 
course, if I turn on Rama restraints I always get clashes, no matter 
what nonbonded_weight I use.

1aho is a small, 1.0 A structure so it is not hard to see what is 
specifically happening.  The one Rama outlier is Gly17. After amber 
minimization it is on the edge of the "favored" region, but after 
geometry_minimization it is pushed into outlier territory.  The other 
thing bumping up the Molprobity score is a steric clash between OG:Ser40 
and O:Gly43.  Yes, there is a hydrogen between them, but the O--O 
distance goes from 2.61 to 2.33 A, which is pretty tight for an H-bond. 
I realize that in common practice a single outlier is considered "OK", 
but in reality they are physically unreasonable.

I bring up my "whirlygig hypothesis" because the contour length of the 
main chain is the only aspect of macromolecular models that is sensitive 
to small "errors" in the geometry. Or, at least, errors that are not 
counter-balanced by other errors.  Yes, the difference in the ideal (aka 
"goal") CA-CA distance in phenix is only 2% shorter than it is in the 
other three packages, but a 2% stretch of a 100 aa chain requires ~4 
kcal/mol.  This is significantly more than any conventional dihedral 
and/or anti-bumping restraints, which is where I think the energy is going.

Or do you think something else is going on?

Thank you for taking the time to consider this,


On 7/9/2021 2:12 PM, Dale Tronrud wrote:
> Hi,
>    I thought I'd toss in an observation that might clarify some of 
> your problems.  As Pavel noted earlier, the uncertainty in the "ideal" 
> values in libraries based on high resolution crystal structures are 
> about 0.02 A for most bonds and one to two degrees for bond angles 
> (with a higher number for the NCaC angle).  The origin of these 
> uncertainties is not "error" but the actual variability of atomic 
> arrangements caused by their environment. The variability these 
> "uncertainties" represent are, in fact, real.
>    Not only are the angles and bonds in individual molecules perturbed 
> by their surroundings, but the distortions of one angle will, 
> systematically, be connected to distortions of other angles.  For 
> example, when the NCaC angle is larger than average both the NCaCb and 
> CbCaC angles will tend to be smaller.  This particular relationship is 
> ignored in most libraries of standard geometry but is implicit in the 
> CDLs that Andy Karplus has produced.
>    What this fact indicates to me is that the standard libraries 
> describe an ensemble of molecular models that all are consistent with 
> the library, and long as their deviations from the mean are within the 
> uncertainty.  This means that there is nothing at all special about 
> the conformation that has average bond length and angles and makes me 
> question to utility of creating an "idealized" model.  There really is 
> no evidence that a significant number of molecules actually exist with 
> that set of lengths and angles.  In fact, since we don't really 
> understand all the interrelationships between the lengths and angles, 
> there is a good chance that a molecule with all "perfect" angles and 
> lengths is impossible! There may be a conflict which we identify as a 
> Molprobaty clash which forbids that arrangement of atoms despite them 
> being within the simple-minded ensemble limits of a library w/o 
> interrelationships.
>    In the absence of any other information you may decide that an 
> "idealized" set of coordinates is the safest statement you can make 
> about your molecule's structure but this limitation of traditional 
> geometry libraries will mean that you have little to no confidence 
> that any molecule will actually have that structure.  It is kind of 
> like the expectation value of a structure factor whose magnitude you 
> know but you have no idea of the phase.  There is a center to that 
> distribution, but that doesn't mean that the probability at that 
> center is non-zero.
> Dale Tronrud
> On 7/8/2021 10:24 AM, James Holton wrote:
>> Thank you Pavel for your prompt response!
>> I agree with everything you wrote below, and that is a good point 
>> about 2nd derivatives.
>> However, what I'm seeing is the opposite of what you might predict. 
>> See below.
>> On 7/7/2021 11:27 PM, Pavel Afonine wrote:
>>> Hi James,
>>> thanks for email and sharing your observations!
>>>> Greetings all, and I hope this little observation helps improve 
>>>> things somehow.
>>>> I did not expect this result, but there it is. My MolProbity score 
>>>> goes from 0.7 to 1.9 after a run of phenix.geometry_minimization
>>>> I started with an AMBER-minimized model (based on 1aho), and that 
>>>> got me my best MolProbity score so far (0.7). But, even with 
>>>> hydrogens and waters removed the geometry_minimization run 
>>>> increases the clashscore from 0 to 3.1 and Ramachandran favored 
>>>> drops from 98% to 88% with one residue reaching the outlier level.
>>> It is not a secret that 'standard geometry restraints' used in 
>>> Phenix and alike (read Refmac, etc) are very simplistic. They are 
>>> not aware of main chain preferential conformations (Ramachandran 
>>> plot), favorable side chain rotamer conformations. They don't even 
>>> have any electrostatic/attraction terms -- only anti-bumping 
>>> repulsion! Standard geometry restraints won't like any NCI 
>>> (non-covalent interaction) and likely will make interacting atoms 
>>> break apart rather than stay close together interacting.
>> Yes, there's the rub: I'm not seeing "interacting atoms break apart", 
>> but rather they are being smashed together.  Torsion angles are also 
>> being twisted out of allowed regions of the Ramachandran plot.
>> All this with the x-ray term turned off!
>>> With this in mind any high quality (high-resolution) atomic model or 
>>> the one optimized using sufficiently high-level QM is going to have 
>>> a more realistic geometry than the result of geometry regularization 
>>> against very simplistic restraints target. An example:
>>> https://journals.iucr.org/d/issues/2020/12/00/lp5048/lp5048.pdf
>>> and previous papers on the topic.
>> I agree, but what doesn't make sense to me is how the "simplistic 
>> restraints" of phenix.geometry_minimization would be so inconsistent 
>> with the "simplistic restraints" in phenix.molprobity ?
>> What I am doing here is starting with an energy-minimized model of a 
>> 1.0 A structure (1aho). It's not a fancy QM, just the ff14SB 
>> potential in AMBER.  I get my best molprobity scores this way, but I 
>> need an x-ray refinement program like phenix.refine to compare these 
>> models with reality.  It troubles me that the "geometry" in the x-ray 
>> refinement program all by itself messes up my molprobity score.
>>>> Just for comparison, with refmac5 in "refi type ideal" mode I see 
>>>> the MolProbity rise to 1.13, but Clashscore remains zero, some 
>>>> Ramas go from favored to allowed, but none rise to the level of 
>>>> outliers.
>>> I believe this is because of the nature of minimizer used. Refmac 
>>> uses 2nd derivative based one, which in a nutshell means it can move 
>>> the model much less (just a bit in vicinity of a local minimum) than 
>>> any program that uses gradients only (like Phenix).
>> good point.
>> So, what should I do to stabilize phenix.geometry_minimization? Crank 
>> up the non-bonded weight?  Restrain to starting coordinates?
>>>> Files and logs here:
>>>> https://bl831.als.lbl.gov/~jamesh/bugreports/phenixmin_070721.tgz
>>>> I suspect this might have something to do with library values for 
>>>> main-chain bonds and angles?  They do seem to vary between 
>>>> programs. Phenix having the shortest CA-CA distance by up to 0.08 
>>>> A. After running thorough minimization on a poly-A peptide I get:
>>>> bond   amber   refmac  phenix  shelxl Stryer
>>>>  C-N   1.330   1.339   1.331   1.325     1.32
>>>>  N-CA  1.462   1.482   1.455   1.454     1.47
>>>> CA-C   1.542   1.534   1.521   1.546     1.53
>>>> CA-CA  3.862   3.874 *3.794* 3.854
>>>> So, which one is "right" ?
>>> I'd say they are all the same, within their 'sigmas' which are from 
>>> memory about 0.02A:
>> I note that 3.874 - 3.794 = 0.08 > 0.02
>> This brings me to my pet theory.  I think what is going on is small 
>> errors like this build up a considerable amount of tension in the 
>> long main chain. For this 64-mer, the contour length of the main 
>> chain after idealization is ~5 A shorter after 
>> phenix.geometry_minimization than it is after shelxl or amber. That 5 
>> A has to come from somewhere.  Without stretching bonds or bending 
>> angles the only thing left to do is twisting torsions. A kind of 
>> "whirlygig" effect.
>> The question is: is the phenix CA-CA distance too short?  Or is the 
>> amber CA-CA distance too long?
>> Shall we vote?
>> -James
>> _______________________________________________
>> phenixbb mailing list
>> phenixbb at phenix-online.org
>> http://phenix-online.org/mailman/listinfo/phenixbb
>> Unsubscribe: phenixbb-leave at phenix-online.org

More information about the phenixbb mailing list