going to high res

Submitted by sstagg on Fri, 2010-04-23 20:36

Forums

Frealign

I am writing to get some advice on what frealign parameters I should vary for trying to go to high (better than 7A) resolution. I have a data set that will go to 6A with EMAN right now, and it only goes to 8A with frealign. Some of that resolution in EMAN is noise bias, I'm sure, but the details with the EMAN reconstruction are better than what I'm currently getting with frealign, even after tweaking the B factor. Should I try different RREC, RMAX1, and RMAX2? If so, do you have recommendations for those values? Any other suggestions?

Thanks,
Scott

niko Fri, 2010-04-23 22:45

Hi Scott,

One parameter to tweak is the weighting factor (PBC). This will alter the weighting applied to the particles in a reconstruction. A small value means that the weight difference between a particle with low phase residual and a high phase residual will be large while a large value will reduce this difference. You should recalculate your reconstruction with different values for PBC (e.g. 5, 10, 20, 50, 100) and see if this makes a difference.

Sometimes it is also important to scale the reconstruction appropriately. EMAN might have a different way of scaling that might attenuate the low-resolution terms compared to Frealign and thus emphasize high-resolution details. To see if scaling is the issue you could use the program diffmap to scale the Frealign reconstruction against the EMAN reconstruction (both need to have the same pixel size and overall box dimensions). This will ensure that both maps are scaled the same way (in resolution zones). Remaining differences will mainly be due to phase differences.

You can also set the Frealign flag FCREF to T (for true). This will switch on figure-of-merit weighting of the reconstruction (Rosenthal & Henderson, 2003, JMB 333, 721–745). You should still apply a negative B-factor to the output map.

If you think the particle angles and shifts are not optimally aligned you can also play with the RBFACT parameter. By setting this parameter to a negative number (e.g. -500) you can amplify the high-resolution terms. This will probably lead to some over-fitting of the high-resolution noise but may also get you out of a local minimum in the refinement. If you run 2 or 3 refinement cycles with RBFACT set to a negative value and then another 2 or 3 cycles with RBFACT = 0 the effects of over-fitting should mostly disappear again.

To test how reliable your determined particle angles are you can run a few cycles of randomized search (IFLAG = 2) and see how many particles change their angles significantly. If you do not have symmetry in your particle the angles of correctly aligned particles should not change much. With symmetry in your particle you will have to check if changes in the angles correspond the actual changes in the views or if they simply correspond to a different symmetry-related views.

Finally, instead of the weighted correlation coefficient (subroutine cc3m.f) you can use the linear correlation coefficient (cc3m.f.cclin). Just copy cc3m.f.cclin to cc3m.f and recompile (to go back to the weighted correlation coefficient, copy cc3m.f.cc3m to cc3m.f). The linear correlation coefficient sometimes performs better in the alignment.