How many macro-cycles are you doing?

  I've also found that fixing things can drive Rfree up for a while, sometimes hundreds of cycles, but if you give the refinement program a chance to converge, Rfree usually settles back down again to a new, lower, value.  This is especially true with REFMAC.  I define "convergence" as the point when the model stops changing in xyz, B and occupancy, not when the R factors appear to stabilize.  In fact, the value of the Rfree and Rwork at "coordinate convergence" can be helped by using a less "greedy" minimizer, like jelly body or DEN refinement.  Yes, the model can start to oscillate, but the atoms doing the oscillation are usually a good indicator of the next thing to rebuild.

That said, it is hard to apply my convergence criterion with phenix.refine since it always "jiggles" the atoms a bit.  Coordinate convergence is then the point when the atoms settle about a central value of x,y,z,B, and occ.  The phenix BB will perhaps contain more detailed advice on convergence.

At the risk of giving you the "try a different program" answer, I suspect that if you want to say something general about refinement at different resolutions you should be using at least two different refinement programs.

-James Holton
MAD Scientist

On 7/31/2015 4:07 PM, Tristan Croll wrote:

At the risk of setting the cat amongst the pigeons...

In the course of developing and demonstrating some new tools I've been working on, I've run quite a large number of re-refinements of various structures from the PDB, focusing on those that currently sit in the low percentiles according to the validation metrics. At resolutions better than about 3 Angstroms, things work pretty much as one would intuitively expect: I fix a set of outliers, improve the fit to the map, and get a 2-5% reduction in Rfree. Past 3 Angstroms, however, I find almost as a rule that the opposite is the case: I can fix large numbers of clear, large-scale errors (loops sitting entirely out of density, bulky sidechains with rotamers flipped 180 degrees from where they should be, strands out of register) as well as much larger numbers of smaller-scale problems, improving the fit to the map and the geometry at every step, and be rewarded with significantly improved map interpretability... and typically a 0.5-2% increase in Rfree.

Take the structure I've spent the last few weeks on. Almost 17,000 residues at 3.8 Angstroms, with >6% Ramachandran outliers (75% favoured), >15% rotamer outliers, clashscore above 30, and Rfree of 0.285. I have this down to 0.4% Ramachandran outliers (and almost 96% favoured), 2% rotamer outliers, clashscore of 2, and have both cleaned up a host of issues (e.g. 6 residues stretched into the density of 10 at the N-terminal tails) and picked out some really interesting new details. I have been through the entire structure, viewing every residue against the map, and see vastly improved correspondence compared to the original, with very few remaining issues apparent in the Fo-Fc map. Yet my Rfree is now 0.291.

Now, by watching closely the weight optimisation during refinement (I'm using Phenix for this step) I can see that by reducing my restraints I could easily shave perhaps 1-2% off Rfree - but at the expense of significant increases (doubling, tripling) in outliers and clashes. While that would probably make it easier to get past reviewers, I'm not keen to do so. Instead, I've been trying to work out the reasons behind this pattern. Potential explanations I've come up with:

- My approach involves very substantial adjustments in real space. As such, I've been absolutely strict about excluding the free set from map generation (and filling missing F_obs with F_calc). It's currently common for people to include free reflections in maps for this purpose, and I wonder if this is having a big effect at these resolutions?

- A low resolution map is primarily low resolution because the protein retains some mobility within the crystal (static or dynamic disorder) - it's a superposition of many slightly different conformations. In places where mobility becomes significant, the average of all the different conformations is going to be something decidedly non-physical - to which we're trying to fit a single structure. The fit to the data could, in general, be improved by allowing the structure to distort into unrealistic arrangements, but it doesn't really make sense to do so.

- Unlike at higher resolutions, at low resolution it's considered quite standard to accept quite high clashscores (>20) and other unrealistic conformations (Ramachandran and rotamer outliers) as long as the Rfree is decent. But I'm beginning to think (and, with a little more thought, suspect I will begin to contend) that this is back to front. Allowing such issues gives the structure a huge (if difficult to quantify) number of extra degrees of freedom, when perhaps the biggest problem at low resolution is a low observations:DOF ratio. After all, this is why we restrain bond lengths, angles and planarity to known favourable values - if these were relaxed, it would be easily to obtain wondrous but entirely spurious reductions in Rfree. Surely the same reasoning holds at the larger scale? In other words, as resolution reduces we need to start relying more strongly on a realistic forcefield for guidance, while of course making sure that the structure still remains consistent with the data.

Hopefully an interesting conversation-starter for the weekend, and I'd be very interested to hear others' input. The moral, in my opinion, is that one shouldn't discard a structure with better stereochemistry simply because the Rfree is higher - at least visually check the changes against the map first, and compare the structures to see which is the more physically sensible.

Best regards,

Tristan