Print

Print


I confess that the jet lag remaining from my China trip impacted the readability of my message. 

When I said 'All 3 of these should have the same "resolution" which should extend to 3 Å' I was trying to say that the PDB -> density conversion should have a low-pass filter at the lower sampling Nyquist (3 Å) for all 3 cases. My entire point was exactly what you are saying about sampling, but the problem doesn't only exist AT Nyquist. Partial ambiguity extends to 1/2 Nyquist. My PDB suggestion was targeted at people whose brains aren't deeply embedded in Fourier space, to give them a visual example to better understand the issue.

--------------------------------------------------------------------------------------
Steven Ludtke, Ph.D. <[log in to unmask]>                      Baylor College of Medicine 
Charles C. Bell Jr., Professor of Structural Biology
Dept. of Biochemistry and Molecular Biology                      (www.bcm.edu/biochem)
Academic Director, CryoEM Core                                        (cryoem.bcm.edu)
Co-Director CIBR Center                                    (www.bcm.edu/research/cibr)



On Sep 3, 2018, at 10:06 AM, Penczek, Pawel A <[log in to unmask]> wrote:

Steve: 

As far as I can tell your postulated numerical experiment is not applicable to the problem discussed. Sampling theorem states that only band-limited signals can be properly sampled while X-ray model atomic coordinates have “infinite” resolution. Properly, in such a case one would have to apply low-pass filtration before sampling and then the three structures you suggest would be identical.


A minor remark, sampling cannot be done at exact Nyquist frequency as one would get ambiguities you describe and this is what sampling theory states. 

Regards,
Pawel Penczek
UT Houston

PS I resisted for hours and then caved. 

On 02/09/2018 23:38, Ludtke, Steven J wrote:
[log in to unmask]" class=""> Ok, can't resist chiming in any more. 

The basis of this general consensus, IS mathematical. The specific threshold (2/3 Nyquist) which pretty much everyone agreed on back in the 90's, is empirical. The reason (at least the one I've always used) is that you do not have a complete representation of data at Nyquist. Nyquist corresponds to a +1/-1/+1/-1 sequence in adjacent pixels, if you phase shift this pattern by 90 degrees, it becomes 0,0,0,0. At 1/2 Nqyuist, you can still represent arbitrary sinusoidal waveforms with arbitrary phase. Between 1/2 Nyquist and Nyquist, you get patterns which are dependent on position within the "box". 

When you do X-ray crystallography, you are measuring the Fourier intensities experimentally, and provided that you can get the correct phase, you can then oversample the real-space representation of the crystal pattern (with specified phases), such that it is fully sampled. In CryoEM, we image in real-space, and between 1/2 Nyquist and Nyquist the phases and amplitudes are convolved in spatially dependent ways, such that information is actually lost, and over-sampling cannot recover the information fully.

This is NOT saying you cannot achieve FSC curves that remain close to 1 all the way to Nyquist. You can, of course, but the resulting reconstruction will not be properly sampled and features will be distorted from what you would see if you had the same structure measured with 2x finer sampling.

Try this little experiment. Take a PDB model and convert to electron density with 1.5 Å/voxel sampling, repeat the same process, but translate the PDB by 0.75 Å in x/y/z before doing the conversion. Finally, generate a PDB with 0.75 Å/pixel sampling. All 3 of these should have the same "resolution" which should extend to 3 Å. Look at the 3-maps. Overlay the PDB. Take a look at the sidechains. Ostensibly, these 3 maps are all "identical", but you will see that they are definitely not...

PS - please note that I am NOT saying that the FFT is a lossy process. It is not, of course. Information is exactly preserved by the FFT. The point, is that an arbitrary real-space periodicity requires 4 pixels, not 2 pixels, to unambiguously represent. 

--------------------------------------------------------------------------------------
Steven Ludtke, Ph.D. <[log in to unmask]>                      Baylor College of Medicine 
Charles C. Bell Jr., Professor of Structural Biology
Dept. of Biochemistry and Molecular Biology                      (www.bcm.edu/biochem)
Academic Director, CryoEM Core                                        (cryoem.bcm.edu)
Co-Director CIBR Center                                    (www.bcm.edu/research/cibr)



On Sep 2, 2018, at 7:27 PM, Dimitry Tegunov <[log in to unmask]> wrote:

Dear Marin,

quoting from the 1997 paper:

"The crossing point between the two curves here lies at about 78% of the Nyquist frequency. The theoretical limit for all processing lies at the Nyquist frequency; yet, considerably before that limit, practical limitations due to the 2D or 3D interpolation procedures used limit, or at least interfere with, the information transfer through the processing chain. Moreover, since both 3D maps are processed by the same programs, with the same interpolation routines, the same systematic round-off errors may be introduced in both reconstructions, which the FSC program may see as common “information”. It is thus good practice not to interpret resolution curves at this high end of the resolution range. The sampling of the data at a sampling interval of 5 Å, in our experience, effectively limits the attainable resolution to ∼15 Å rather than to the theoretical Nyquist limit of 10 Å."

You seem to agree that getting close to Nyquist is possible if sloppy real-space interpolation is avoided. As the latter has been the case for at least the past decade, perhaps it is time to let go of an arbitrary rule derived from personal experience rather than signal processing fundamentals?

Cheers,
Dimitry

########################################################################

To unsubscribe from the CCPEM list, click the following link:
https://urldefense.proofpoint.com/v2/url?u=https-3A__www.jiscmail.ac.uk_cgi-2Dbin_webadmin-3FSUBED1-3DCCPEM-26A-3D1&d=DwIFaQ&c=ZQs-KZ8oxEw0p81sqgiaRA&r=GWA2IF6nkq8sZMXHpp1Xpg&m=F5Rudew3hA-EqzY67760waCGQCsGcLHKXLRjv420UKw&s=oBlVi2P88nOAHy2CQuhxXeHTFSwH9xmKLYJ5Xhl7ViI&e=



To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1


-- 
==============================================================

    Prof Dr Ir Marin van Heel

    Laboratório Nacional de Nanotecnologia - LNNano
    CNPEM/LNNano, Campinas, Brazil

    tel:    +55-19-3518-2316
    mobile  +55-19-983455450 (current)
    mobile  +55-19-981809332  
                 (041-19-981809332 TIM)
    Skype:  Marin.van.Heel
    email:  marin.vanheel(A_T)gmail.com
            marin.vanheel(A_T)lnnano.cnpem.br
    and:    mvh.office(A_T)gmail.com  

--------------------------------------------------
    Emeritus Professor of Cryo-EM Data Processing
    Leiden University
    Mobile NL: +31(0)652736618 (ALWAYS ACTIVE SMS)
--------------------------------------------------
    Emeritus Professor of Structural Biology
    Imperial College London
    Faculty of Natural Sciences
    email: m.vanheel(A_T)imperial.ac.uk
--------------------------------------------------

I receive many emails per day and, although I try, 
there is no guarantee that I will actually read each incoming email. 



To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1