Last week I asked a question about geometry restraints at low resolution. I would like to thank you all that took the time to answer, specially Pavel for looking at the data. First, here is the original question:

__________________________________________

Is there a known procedure to deal with this? I feel that my geometry weighing term is too loose but I don't know exactly how to tighten it other than decreasing the wxc_scale parameter. I have gone and manually corrected bad positions before but given the low resolution nature of the structure the clashes tend to be reestablished during refinement. 

Thank you in advance for your help,

__________________________________________

It turns out that the reason to my nightmares was a highly anisotropic dataset. When I tried to fix geometry related problems in coot, phenix.refine was twisting the geometry back again, probably because it was trying to fit a lot of noise due to the anisotropy. Everything was getting worst, geometry, clashscore, and Ramachandran. 

Pavel finally corrected it by using this server: http://services.mbi.ucla.edu/anisoscale/  I should have been more diligent regarding that, since I could clearly see that my data was anisotropic from the diffraction images. I just didn't know how important this parameter could be. Anyways, phenix.xtriage will tell you if your data is anisotropic. The lesson is to pay attention to it. 

In my case, the quality of the electron density improved considerably, the geometry improved massively (manual intervention in coot was absolutely necessary, but now it stays correct after I manually fix the problems), clashscore lowered, etc. 

As a side note, it worked better without hydrogens in my case than with. It is certainly a parameter to try if you are having problems. 

Thank you all again for the tips. 

Mario Sanches