Refining initial model using negative-stain data

Submitted by Vu on Sun, 2011-07-24 13:56

Forums

Frealign

Hi,

I am trying to refine several initial models obtained from negative-stain data using ~20,000 low-tilt (15 deg) negatively stained particles. My previous attempts with Spider and EMAN2 could not improve the resolution (0.5 FSC at 50-45 angstroms). I tried to run a few FREALIGN sessions (Mode 1) yesterday with particle parameters obtained from previous projection matching done with a Spider script and CTFTILT. The resolution seemed to stabilize at 49 A after 3 iterations. I would like to try FREALIGN again and have a few questions,

1. Would you recommend using the program with my data?
2. Do I have to invert the contrast of my negatively stained particles to use as input? If yes, would NEG (or NEG A) in Spider do the job?
3. For the last session, I used 4-fold decimated, CTF-corrected particles. I set WGH to -1.0 to turn off CTF correction. I also set RBFACT to -1000 to boost the high-resolution terms for 3 cycles. This seemed to improve the resolution a bit; however, the resolution did not improve after subsequent iterations with RBFACT=0. Here are the rest of my input parameters,
thresh_reconst 69.0
thresh_refine 15.0
pixel_size 6.6
dstep 15.0
WGH -1.0
CS 6.2
kV1 120.0
radius 200.0
PBC 100.0
BOFF 35.0
DANG 10.0
ITMAX 10
MODE 1
XSTD 0.0
RBFACT -1000
FMAG F
FDEF F
FASTIG F
FPART F
dfsig 150.0
IEWALD 0
res_reconstruction 15.0
res_low_refinement 100.0
res_refinement 25.0
start_process 1
end_process 3
first_particle 1
last_particle 19387

I am hoping to get some feedback/suggestions from you on how to improve the refinement results.

Thank you,

Vu

niko Sun, 2011-07-24 18:10

I suspect that your data will not go to higher resolution than what you have already achieved. I do not think that at 40 or 50 A one piece of software will perform significantly better than another. Your results seem to agree with this. Frealign might give you a reconstruction that is somewhat better corrected for the CTF than some other packages but in terms of resolution it sounds like you may be limited by sample preparation issues or heterogeneity in the sample. Maybe classification would help (using EMAN or Spider).

Setting RBFACT to a negative number will boost the high-resolution terms. This will most likely give you a better-looking FSC but it is also likely that you will over-fit your data (read more here). If you run a few cycles with RBFACT = 0 over-fitting might be reduced again but you should be careful since a badly over-fitted structure often cannot be rescued by more refinement due to reference bias.

If you turn off CTF correction your protein density on input should be positive with respect to the background. With CTF correction switched on, it should be negative with respect to the background. You can use Spider to do the inversion. I always use pixel arithmetic to do this where I multiply by -1.0.

A final word about FSC: If you do not over-fit your data, a more appropriate resolution cut-off is usually FSC = 0.143. When you over-fit your data and get an artificially high FSC, it may be necessary to use a higher threshold to measure the resolution, such as FSC = 0.5. For more discussion on this, please read our recent review, page 268.

Vu Tue, 2011-07-26 23:13

Thank you. Those are very helpful references. I do want to give FREALIGN another try with CTF correction and also work around the parameters a bit more. I'll let you know how it goes.