[cctbxbb] use_internal_variance in iotbx.merging_statistics

Tue Nov 1 04:10:42 PDT 2016

As a side comment on what Aimless does, I’ve been looking at SD estimates and the SAMPLESD option in particular, in my development version (I think there is a bug in the current distributed version). Using the internal variance does provide an alternative estimate of var(Imean), but doesn’t answer the question of how to weight the individual observations in determining the mean, which is not straightforward. The error model used in Aimless and XDS is purely dependent on intensity, so it takes no account of radiation damage, which is hard to do.

My development version of Aimless will by default “optimise” the error model (and I’m trying out alternative methods of doing this), even if SAMPLESD is switched on, and there is an analysis of the difference between the internal and external error estimates. I’m wondering about automatically defaulting to switching SAMPLESD on for high multiplicity (but how high?). Or doing some weighted mean between the two estimates.

Error models are hard :-(

Phil

> On 1 Nov 2016, at 09:21, <richard.gildea at diamond.ac.uk> <richard.gildea at diamond.ac.uk> wrote:
> 
> Dear Keitaro,
> 
> iotbx.merging_statistics does have the option to change the parameter use_internal_variance. In xia2 we use the defaults use_internal_variance=False, eliminate_sys_absent=False, n_bins=20, when calculating merging statistics which give comparable results to those calculate by Aimless:
> 
> $ iotbx.merging_statistics 
> Usage: 
> phenix.merging_statistics [data_file] [options...]
> 
> Calculate merging statistics for non-unique data, including R-merge, R-meas,
> R-pim, and redundancy.  Any format supported by Phenix is allowed, including
> MTZ, unmerged Scalepack, or XDS/XSCALE (and possibly others).  Data should
> already be on a common scale, but with individual observations unmerged.
>  Diederichs K & Karplus PA (1997) Nature Structural Biology 4:269-275
>    (with erratum in: Nat Struct Biol 1997 Jul;4(7):592)
>  Weiss MS (2001) J Appl Cryst 34:130-135.
>  Karplus PA & Diederichs K (2012) Science 336:1030-3.
> 
> 
> Full parameters:
> 
>  file_name = None
>  labels = None
>  space_group = None
>  unit_cell = None
>  symmetry_file = None
>  high_resolution = None
>  low_resolution = None
>  n_bins = 10
>  extend_d_max_min = False
>  anomalous = False
>  sigma_filtering = *auto xds scala scalepack
>    .help = "Determines how data are filtered by SigmaI and I/SigmaI. XDS"
>            "discards reflections whose intensity after merging is less than"
>            "-3*sigma, Scalepack uses the same cutoff before merging, and"
>            "SCALA does not do any filtering. Reflections with negative SigmaI"
>            "will always be discarded."
>  use_internal_variance = True
>  eliminate_sys_absent = True
>  debug = False
>  loggraph = False
>  estimate_cutoffs = False
>  job_title = None
>    .help = "Job title in PHENIX GUI, not used on command line"
> 
> 
> Below is my email to Pavel and Billy when we discussed this issue by email a while back:
> 
> The difference between use_internal_variance=True/False is explained in Luc's document here:
> 
> libtbx.pdflatex $(libtbx.find_in_repositories cctbx/miller)/equivalent_reflection_merging.tex
> 
> Essentially use_internal_variance=False uses only the unmerged sigmas to compute the merged sigmas, whereas use_internal_variance=True uses instead the spread of the unmerged intensities to compute the merged sigmas. Furthermore, use_internal_variance=True uses the largest of the variance coming from the spread of the intensities and that computed from the unmerged sigmas. As a result, use_internal_variance=True can only ever give lower I/sigI than use_internal_variance=False. The relevant code in the cctbx is here:
> 
> https://sourceforge.net/p/cctbx/code/HEAD/tree/trunk/cctbx/miller/merge_equivalents.h#l379
> 
> Aimless has a similar option for the SDCORRECTION keyword, if you set the option SAMPLESD, which I think is equivalent to use_internal_variance=True. The default behaviour of Aimless is equivalent to use_internal_variance=False:
> 
> http://www.mrc-lmb.cam.ac.uk/harry/pre/aimless.html#sdcorrection
> 
> "SAMPLESD is intended for very high multiplicity data such as XFEL serial data. The final SDs are estimated from the weighted population variance, assuming that the input sigma(I)^2 values are proportional to the true errors. This probably gives a more realistic estimate of the error in <I>. In this case refinement of the corrections is switched off unless explicitly requested."
> 
> I think that the "external" variance is probably better if the sigmas from the scaling program are reliable, or for low multiplicity data. For high multiplicity data or if the sigmas from the scaling program are not reliable, then "internal" variance is probably better.
> 
> Cheers,
> 
> Richard
> 
> Dr Richard Gildea
> Data Analysis Scientist
> Tel: +441235 77 8078
> 
> Diamond Light Source Ltd.
> Diamond House
> Harwell Science & Innovation Campus
> Didcot
> Oxfordshire
> OX11 0DE
> 
> ________________________________________
> From: cctbxbb-bounces at phenix-online.org [cctbxbb-bounces at phenix-online.org] on behalf of Keitaro Yamashita [k.yamashita at spring8.or.jp]
> Sent: 01 November 2016 07:23
> To: cctbx mailing list
> Subject: [cctbxbb] use_internal_variance in iotbx.merging_statistics
> 
> Dear Phenix/CCTBX developers,
> 
> iotbx/merging_statistics.py is used by phenix.merging_statistics,
> phenix.table_one, and so on. By upgrading phenix from 1.10.1 to 1.11,
> merging statistics-related codes were significantly changed.
> 
> Previously, miller.array.merge_equivalents() was always called with
> argument use_internal_variance=False, which is consistent with XDS,
> Aimless and so on. Currently, use_internal_variance=True is default,
> and cannot be changed by users (see below).
> 
> These changes were made by @afonine and @rjgildea in rev. 22973 (Sep
> 26, 2015) and 23961 (Mar 8, 2016). Could anyone explain why these
> changes were introduced?
> 
> https://sourceforge.net/p/cctbx/code/22973
> https://sourceforge.net/p/cctbx/code/23961
> 
> 
> My points are:
> 
> - We actually cannot control use_internal_variance= parameter because
> it is not passed to merge_equivalents() in class
> filter_intensities_by_sigma.
> 
> - In previous versions, if I gave XDS output to
> phenix.merging_statistics, <I/sigma> values calculated in the same way
> (as XDS does) were shown; but not in the current version.
> 
> - For (for example) phenix.table_one users who expect this behavior,
> it can give inconsistency. The statistics would not be consistent with
> the data used in refinement.
> 
> 
> cf. the related discussion in cctbxbb:
> http://phenix-online.org/pipermail/cctbxbb/2012-October/000611.html
> 
> 
> Best regards,
> Keitaro
> _______________________________________________
> cctbxbb mailing list
> cctbxbb at phenix-online.org
> http://phenix-online.org/mailman/listinfo/cctbxbb
> 
> -- 
> This e-mail and any attachments may contain confidential, copyright and or privileged material, and are for the use of the intended addressee only. If you are not the intended addressee or an authorised recipient of the addressee please notify us of receipt by returning the e-mail and do not use, copy, retain, distribute or disclose the information in or attached to the e-mail.
> Any opinions expressed within this e-mail are those of the individual and not necessarily of Diamond Light Source Ltd. 
> Diamond Light Source Ltd. cannot guarantee that this e-mail or any attachments are free from viruses and we cannot accept liability for any damage which you may sustain as a result of software viruses which may be transmitted in or with the message.
> Diamond Light Source Limited (company no. 4375679). Registered in England and Wales with its registered office at Diamond House, Harwell Science and Innovation Campus, Didcot, Oxfordshire, OX11 0DE, United Kingdom
> 
> 
> _______________________________________________
> cctbxbb mailing list
> cctbxbb at phenix-online.org
> http://phenix-online.org/mailman/listinfo/cctbxbb