[phenixbb] convergence in Phenix real space refinement

[Pavel Afonine]

Hi Rob,

rmsd between cycles is a global metric that isn't sensitive to local 
changes. One can also look at maximal shift but then you'll face problem 
of oscillations -- for example, a side chain may flip back and forth 
between rotameric conformations for whatever reason (poor map, or 
equally good density for a few alternative conformations,...): in this 
case global rmsd won't show anything but max shift will be always large. 
All in all, it is not easy to find a reliable solution that will always 
work. This is why choosing number of cycles is left on the user to 
decide on (with the default 5 being fine enough in most cases). Also, 
part of the reason why it is not easy is because 
phenix.real_space_refine combined gradient-based minimization with local 
grid searches (side chain rotamer fitting).

Pavel

On 10/20/23 00:29, Robert Oeffner wrote:
> Thanks Pavel, In one test example I noted that "moved from start" 
> didn't change much after about 12 macro cycles. This would then be a 
> necessary but not sufficient condition for convergence If I understand 
> you correctly. Would there be any other statistics indicating 
> convergence? In my test example CCmask hardly changed at all.
>
> Rob
>
> On 20/10/2023 04:03, Pavel Afonine wrote:
>> Hi Rob,
>>
>>> I'm curious to know when convergence has been obtained in 
>>> phenix_real_space_refine. I see the number "moved from start" in the 
>>> log file and assume that once it doesn't change much after several 
>>> macro cycles then convergence has been obtained when refining a 
>>> structure. True? What is that value based on, the RMSD between the 
>>> initial structure and the final structure? I didn't find this 
>>> documented in the 2018 publication of the program.
>>
>> phenix.real_space_refine runs a set number of cycles that you can 
>> adjust. In the log file it reports the rmsd between the initial and 
>> current model. While this number can give you some idea about the 
>> amount of model movement between cycles, it also may hide some useful 
>> information: for example, if you refine a large structure and only a 
>> few residues flipped their side chain conformations during 
>> refinement, then this overall number probably won't tell you this.
>>
>> Pavel
>>
>>