Hi Mark,
One guess I would have is increasing the box size to help with delocalization is making it much harder for the bayesian particle polishing to fit reliably. My reasoning for this is that the polishing (I think) is done against the model, not the local region, so including additional background area will only make things worse (whereas fitting against each local patch, you’d expect the fitting to become more robust, but less accurate, with a larger area).
The particle polishing executable implements a —new_box feature, I wonder what would happen if you were to extract, train and polish on the unbinned dataset using a very tight box, but extract with the larger box size. I believe the extracted box size uses the pixel size as the input particle set (so will resample if needed to match the input particles).
Cheers,
-Craig
> On Aug 29, 2018, at 10:06 AM, Mark Herzik <[log in to unmask]> wrote:
>
> Hi All,
>
> First, apologies for the long email.
>
> We have a dataset of a small protein (<150 kDa) collected at a fairly high magnification (0.56 Å/pixel) that we have been trying to process to its fullest using the new Bayesian particle polishing implementation in RELION 3.0. We have hit a proverbial wall in our processing and are hoping some of our fellow EMists could provide some insights as to how best to proceed.
>
> Implementing Bayesian PP with data that have been binned 2x2 (1.12 Å/pixel) using a 384 pixel box size yields robust particle trajectory alignments, nice dose-weighting diagnostics, and yields a 3D reconstruction that is Nyquist limited (2.3 Å resolution) without CTF refinement. Great.
>
> However, we are now trying to extract shiny particles with a smaller downsampling (to lower the Nyquist-limiting frequency) and larger box size (to decrease the effects of CTF delocalization, which is quite severe at 0.56 Å/pixel and 200 kV) without much success.
>
> What we have tried:
> 1) Polishing the unbinned data or data binned 1.5x1.5 using parameters trained against those data results in very noisy particle trajectories, a poor B-factor and dose-weighting scheme, and a worse resolving reconstruction as a result of these efforts (~2.6 Å resolution). Quite surprising given that these steps utilize the same particle stack that resulted in the 2.3 Å resolution reconstruction mentioned above.
>
> 2) Polishing the unbinned data or data binned 1.5x1.5 using parameters trained against the binned 2x2 data yields better results than those obtained in scenario 1 but worse than the binned 2x2 data processing in their entirety (~2.5 Å resolution versus 2.3 Å resolution).
>
> 3) Training the binned 1.5x1.5 data using a small box size (256 pixels, which aids the training step) and then trying to extract the shiny particles with a larger box size (512 pixels) -> RELION crashes with the following error:
>
> ERROR:
> Box size cannot be changed without re-estimating motion - reference pixel size (0.84 A) is not an integral multiple of movie pixel size (0.56 A)!
>
> SO we wonder, is there a protocol currently implemented within RELION 3.0 that can use the particle trajectories from downsampled data (the binned 2x2 data in our case) but then extract the shiny particles using a user-inputted pixel size and box size?
>
> Any and all help will be greatly appreciated.
>
> Thanks,
> -Mark
>
> Laboratory of Dr. Gabriel Lander
> Helen Hay Whitney Foundation Postdoctoral Fellow
> The Scripps Research Institute
> Integrative Structural and Computational Biology
> 10550 N. Torrey Pines Rd. HZ 175L
> La Jolla, CA 92037
> Office: (858) 784-9499
>
>
> ########################################################################
>
> To unsubscribe from the CCPEM list, click the following link:
> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
########################################################################
To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
|