Hi,

I realized Randy's reply did not go to the mailing list, so I'm forwarding it now.

Pavel


-------- Original Message --------
Subject: Re: [phenixbb] anisotropic induced noise amplification
Date: Mon, 7 May 2012 09:49:57 +0100
From: Randy Read <[email protected]>
To: Pavel Afonine <[email protected]>
CC: [email protected] <[email protected]>, Peter Zwart <[email protected]>


In the absence of an answer from Peter (who implemented this and therefore ought to know best), I'm pretty sure that the idea is to measure how much difference there is in the amount of signal in the data that have received the biggest anisotropy corrections, compared to the ones that have received the smallest anisotropy corrections.  If there's very little signal in the data that have been scaled up the most, then the noise is being amplified (nearly-random amplitudes have gone from being relatively small to the same size on average as the well-measured amplitudes).

But  what the Z-scores mean is not at all clear from that output, and I'm not aware that Peter has written anything describing this.

This amplified noise is what I suspect is creating problems in phenix.refine refinement with highly anisotropic data, because if the experimental sigmas are ignored then the program doesn't know (after anisotropic scaling) which amplitudes of a comparable size contain signal and which are essentially noise.  Anisotropic truncation gets rid of it, which would explain why refinement works better with anisotropically-truncated data.

Randy

On 6 May 2012, at 20:21, Pavel Afonine wrote:

Hi,

did anyone address this post? I'm interested to know the answer myself: I have absolutely no idea what is this.

Pavel

-------- Original Message --------
Subject: [phenixbb] anisotropic induced noise amplification
Date: Wed, 2 May 2012 16:06:55 -0500
From: James Thompson <[email protected]>
Reply-To: PHENIX user mailing list <[email protected]>
To: [email protected]


Anyone willing to summarize Xtriage's algorithm that determines whether "anisotropic induced noise amplification" is present in diffraction data?  What is this noise amplification?  Is it anisotropic correction factors that over-fit the data thereby inducing noise?

Many thanks,
Jim T

----------------    Anisotropy analyses     ----------------

Anisotropy    ( [MaxAnisoB-MinAnisoB]/[MaxAnisoB] ) :  3.596e-01
                          Anisotropic ratio p-value :  0.000e+00

     The p-value is a measure of the severity of anisotropy as observed in the PDB.
     The p-value of 0.000e+00 indicates that roughly 100.0 % of datasets available in the PDB have
     an anisotropy equal to or worse than this dataset.


For the resolution shell spanning between 2.36 - 2.20 Angstrom,
the mean I/sigI is equal to  3.37. 42.2 % of these intensities have
an I/sigI > 3. When sorting these intensities by their anisotropic
correction factor and analysing the I/sigI behavior for this ordered
list, we can gauge the presence of 'anisotropy induced noise amplification'
in reciprocal space.

  The quarter of Intensities *least* affected by the anisotropy correction show
    <I/sigI>                 :   3.82e+00
    Fraction of I/sigI > 3   :   5.01e-01     ( Z =     3.59 )

  The quarter of Intensities *most* affected by the anisotropy correction show
    <I/sigI>                 :   2.02e+00
    Fraction of I/sigI > 3   :   1.84e-01     ( Z =    10.78 )

The combined Z-score of    11.36 indicates that there probably is significant
systematic noise amplification that could possibly lead to artefacts in the
maps or difficulties in refinement

<Attached Message Part.txt>

------
Randy J. Read
Department of Haematology, University of Cambridge
Cambridge Institute for Medical Research      Tel: + 44 1223 336500
Wellcome Trust/MRC Building                   Fax: + 44 1223 336827
Hills Road                                    E-mail: [email protected]
Cambridge CB2 0XY, U.K.                       www-structmed.cimr.cam.ac.uk