Yes, 0.8 would actually be very conservative, so I was somewhat surprised.
Useful threshold values are often around 0.2-0.7. 0.08 indeed sounds very
low.
S


> Sorry there was a typo - 0.08 was what I used, not 0.8. That should make
> more sense!
>
> On Mon, Oct 27, 2014 at 9:05 AM, Liz Montabana <[log in to unmask]>
> wrote:
>
>> Hi Sjors
>>
>> Thanks for the advice. I had originally gone through the described
>> FOM/particle threshold picking with a few micrographs and ended up
>> picking
>> a threshold of 0.8, which looked like it was getting mostly what I would
>> have picked out as particles by eye. I guess I would be a pretty greedy
>> manual picker too! I'll go back and pick something more conservative and
>> see how that turns out. For the auto-picking I was low-pass filtering to
>> 20A, but I'll change that to 30 for this round and see what happens.
>>
>> Do you think that it would be better to not mask your original mini-set
>> of
>> manual picked particles when doing 2D class averages, so when using them
>> for templates there is no chance of that affecting the auto-picking?
>>
>> thanks
>> Liz
>>
>>
>>
>>
>> On Sat, Oct 25, 2014 at 4:52 AM, Sjors Scheres
>> <[log in to unmask]>
>> wrote:
>>
>>> Dear Liz,
>>>
>>> Don't apologize for this excellent and well-documented question!
>>>
>>> The rings you observe are a typical sign of Einstein-from-noise
>>> artifacts
>>> that can occur with (any) template-based particle picking. I attach a
>>> PDF
>>> of the manuscript I wrote on the auto-picking procedure in RELION. Note
>>> this hasn't been back from peer-review yet, so for the moment just
>>> consider it as my own opinion! :-) We can observe the same effects with
>>> many more data sets. The cause is a too low threshold for the
>>> auto-picking
>>> algorithm, which then reproduces the templates from noise areas in the
>>> micrographs. I suppose you used the smaller diameter to calculate the
>>> 2D
>>> classes that you used as templates for auto-picking? The normalization
>>> in
>>> relion (and many other packages) makes that the average background is
>>> often slightly negative (i.e. black); this is caused by the presence of
>>> particles (with white signal) in the background area. This results in
>>> 2D
>>> class averages that have a white ribosome and a slightly negative
>>> (black)
>>> background. When you use those for auto-picking and you are too
>>> "greedy",
>>> i.e. you set the threshold too low, then you start up picking noise
>>> that
>>> also describes your templates. 2D class averaging with those large,
>>> noisy
>>> data sets then gives rise to the typical black rings in the class
>>> averages
>>> (see also Fig 5 in the attached paper). The good class averages (e.g.
>>> first row, number 5 and second row number 1 &2 in your figure) do not
>>> show
>>> these template-induced artefacts.
>>> You can now do 2 things: 1) hope 2D classification deals with this, and
>>> just throw away all Einstein-from-noise classes; or 2) (probably
>>> better):
>>> re-run auto-picking with a higher threshold and then re-run 2D
>>> classification.
>>> Many of your good classes seem to represent a similar view, so you may
>>> have a strong preferred orientation, but this is better assessed once
>>> you've cleaned up your data from the false positives of the
>>> auto-picking
>>> approach.
>>>
>>> The Einstein-from-noise artifacts are an important reason why you
>>> should
>>> ALWAYS use low-pass filtered templates for auto-picking. Probably using
>>> beyond 30A information is not useful anyway, and information that did
>>> not
>>> go into the process can also not come out as an Einstein-from-noise
>>> artifact. Using high-resolution templates for auto-picking is probably
>>> what went wrong in the HIV trimer structure that caused such a stir-up
>>> last year. Both Richard Henderson and Marin van Heel wrote excellent
>>> commentaries in PNAS on that issue. See
>>> http://www.ncbi.nlm.nih.gov/pubmed/24106306 and
>>> http://www.ncbi.nlm.nih.gov/pubmed/24106301.
>>>
>>> HTH,
>>> Sjors
>>>
>>>
>>> > Hi All
>>> >
>>> > I hope this is not too bone-headed of a question. I am new to using
>>> Relion
>>> > (1.3), and I am doing 2D classification and I have been getting some
>>> > results that look suspicious. I am using my own DD data set of yeast
>>> 80S
>>> > particles. I followed the steps outlined in "Tutorial (The quickest
>>> way
>>> to
>>> > learning RELION-1.3) ", for ctf-correction, autopicking, etc. - Now I
>>> am
>>> > trying to do some particle pruning with the 2D classification. There
>>> are
>>> > two things going on:
>>> > 1) There are two highly populated classes (40% of ~60k particles)
>>> that
>>> > look the same - they seem to be neither junk nor good classes.
>>> >
>>> > 2) There are two visible rings for each class - one is the soft mask,
>>> as I
>>> > would expect, but the other seems to be related to the rotation of
>>> the
>>> > particle during alignment - I've attached an image - it seems to
>>> exist
>>> on
>>> > both good and bad classes - these are sorted by class distribution. I
>>> > can't figure out what causes this, but I feel like its a good
>>> indicator
>>> > I've picked some bad parameter. When I did 2D Classification on a
>>> very
>>> > small manual picked set with a very slightly smaller mask (350 A vs
>>> 360
>>> A)
>>> > I did not see the same thing. My box size is 300 pix = 396 Ang.
>>> >
>>> > Thanks for any help -  please excuse such a possible inane question!
>>> > Liz Montabana
>>> >
>>> > Here is the command:
>>> > mpirun -np 240 --bynode `which relion_refine_mpi` --o
>>> > Class2D/autopick_sort2 --i particles_autopick_rd2.star
>>> --particle_diameter
>>> > 360 --angpix 1.32 --ctf  --iter 25 --tau2_fudge 2 --K 300
>>> > --flatten_solvent  --oversampling 1 --psi_step 10 --offset_
>>> > range 5 --offset_step 2 --norm --scale  --j 1 --memory_per_thread 8
>>> >
>>> > Class2D/autopick_sort2.out
>>> >
>>> > and settings:
>>> > Class2D/autopick_sort2.gui_class2d.settings
>>> > is_continue == false
>>> > Output rootname: == Class2D/autopick_sort2
>>> > Continue from here:  ==
>>> > Input images STAR file: == particles_autopick_rd2.star
>>> > Number of classes: == 300
>>> > Do CTF-correction? == Yes
>>> > Have data been phase-flipped? == No
>>> > Ignore CTFs until first peak? == No
>>> > Number of iterations: == 25
>>> > Regularisation parameter T: == 2
>>> > Mask individual particles with zeros? == No
>>> > Limit resolution E-step to (A):  == -1
>>> > Perform image alignment? == Yes
>>> > In-plane angular sampling: == 5
>>> > Offset search range (pix): == 5
>>> > Offset search step (pix): == 1
>>> > Number of MPI procs: == 240
>>> > Number of threads: == 1
>>> > Available RAM (in Gb) per thread: == 8
>>> > Submit to queue? == Yes
>>> >
>>> >
>>> >
>>>
>>>
>>> --
>>> Sjors Scheres
>>> MRC Laboratory of Molecular Biology
>>> Francis Crick Avenue, Cambridge Biomedical Campus
>>> Cambridge CB2 0QH, U.K.
>>> tel: +44 (0)1223 267061
>>> http://www2.mrc-lmb.cam.ac.uk/groups/scheres
>>
>>
>>
>


-- 
Sjors Scheres
MRC Laboratory of Molecular Biology
Francis Crick Avenue, Cambridge Biomedical Campus
Cambridge CB2 0QH, U.K.
tel: +44 (0)1223 267061
http://www2.mrc-lmb.cam.ac.uk/groups/scheres