3. Particle picking
This next set of steps will produce individual particle images from the micrographs. We will pick some particles from a small sample of the micrographs using a reference-free method. This particle set will then be cleaned up by 2D classification and used to train a model for the more advanced Topaz picker.
3.1. Get a subset of the micrographs
Under the Subset Selection category select the Random Sample from Starfile job and run it with the following parameters:
Sample from micrographs file:: CtfFind/job003/micrographs_ctf.star
Number of samples:: 5
Then press RUN. The job should take less than 10 secs to execute. The RESULTS tab will show a montage of all your randomly selected pre-processed micrographs.
3.2. LoG Picking
We will now use a template-free auto-picking procedure based on a Laplacian-of-Gaussian (LoG) filter to select an initial set of particles. These particles will then be used in a 2D classification job to generate 2D class averages.
In Automated Particle Picking select a RELION Autopick LoG (single particle) job.
Input micrographs for autopick:: Select/job004/sampled_micrographs.star
Pixel size in micrographs (A):: -1
NOTE: The pixel size will be set automatically from the information in the input STAR file
Min. diameter for LoG filter (A):: 150
Max. diameter for LoG filter (A):: 180
NOTE: The min and max diameter should correspond to the smallest and largest size of your particles’s projections in Ångstroms.
Are the particles white?:: No
NOTE: They are black
Maximum resolution to consider (A):: 20
Adjust default threshold (stddev):: 0
Upper threshold (stddev):: 5
Write FOM maps?:: No
Read FOM maps?:: No
HEADER: Running options
Number of MPI procs:: 12
Then press RUN. The job should take less than 10 secs to execute. The RESULTS tab will show a carousel of all your micrographs along with bounding boxes for the selected particles. It should have picked ~250 particles per micrograph for ~1200 in total.
3.3. Particle Extraction
Next, we must extract the individual particle images from the micrographs. They will be downsampled to speed up later processing steps.
In Extract Particles select a RELION Extract Particles (single particle) job.
micrograph STAR file:: CtfFind/job003/micrographs_ctf.star
Input coordinates:: AutoPick/job005/autopick.star
HEADER: Extraction Options
Particle box size (pix):: 256
NOTE: This should always be an even number
Invert contrast?:: Yes
NOTE: Makes white instead of black particles
Normalize particles?:: Yes
NOTE: We always normalise
Diameter background circle (pix):: 200
NOTE: Particles will be normalized to a mean value of zero and a standard-deviation of one for all pixels in the background area.The background area is defined as all pixels outside a circle with this given diameter in pixels (before rescaling). When specifying a negative value, a default value of 75% of the Particle box size will be used.
Stddev for white dust removal:: -1
Stddev for black dust removal:: -1
NOTE: We only remove very white or black outlier pixels if we actually see them in the data. In such cases we would use stddev values of 5 or so. In this data set there are no outlier pixels, so we don’t correct for them, and leave the default values at -1 (i.e. don’t do anything)
Rescale particles?:: Yes
NOTE: Down-scaling particles will speed up computations. Therefore, we often down-scale particles in the initial stages of processing, in order to speed up the initial classifications of suitable particles. Once our reconstructions get close to the Nyquist frequency, we then re-extract the particles without down-scaling
Re-scaled size (pixels):: 64
Use autopick FOM threshold: No
NOTE: This option allows to only extract those particles with the highest figure-of-merits from the autopicking procedure. We will use this later on to extract particles picked by topaz.
Write output in float16?:: Yes
HEADER: Running options
Number of MPI procs:: 10
Then press RUN. The job should take less than 10 secs to execute. The RESULTS tab will show a montage of 30 random particles extracted from the micrographs.
3.4. 2D Classification
We almost always use reference-free 2D class averaging to throw away bad particles. Because bad particles do not average well together, they often go to relatively small classes that yield ugly (i.e. blurred images with no identifiable protein-like features) 2D class averages. Throwing those away then becomes an efficient way of cleaning up your data.
Under 2D Classification select a new RELION Class2D (EM, Single Particle) job.
Input images STAR file:: Extract/job006/particles.star
Number of iterations:: 25
Do CTF-correction?:: Yes
NOTE: We will perform full phase+amplitude correction inside the Bayesian framework
Ignore CTFs until first peak?:: No
NOTE: This option is occasionally useful, when amplitude correction gives spuriously strong low-resolution components, and all particles get classified together in very few, fuzzy classes.
Number of classes:: 50
NOTE: For cryoEM data we like to use on average at least approximately 100 particles per class. However, with this small number of particles, we have observed a better separation into different classes by relaxing these numbers. Possibly, always having a minimum of 50 classes is not a bad idea.
Regularisation parameter T:: 2
NOTE: For the exact definition of T, please refer to Scheres, JMB, 2012. For cryoEM 2D classification we typically use values of T=2-3, and for 3D classification values of 3-4. For negative stain, sometimes slightly lower values are better. In general, if your class averages appear very noisy, then lower T; if your class averages remain too-low resolution, then increase T. The main thing is to be aware of overfitting high-resolution noise.
Mask diameter (A):: 200
NOTE: This mask will be applied to all 2D class averages. It will also be used to remove solvent noise and neighbouring particles in the corner of the particle images. On one hand, you want to keep the diameter small, as too much noisy solvent and neighbouring particles may interfere with alignment. On the other hand, you want to make sure the diameter is larger than the longest dimension of your particles, as you do not want to clip off any signal from the class averages.
Mask individual particles with zeros?:: Yes
Limit resolution E-step to (A):: -1
NOTE: If a positive value is given, then no frequencies beyond this value will be included in the alignment. This can also be useful to prevent overfitting. Here we don’t really need it, but it could have been set to 10-15A anyway. Difficult classifications, i.e. with very noisy data, often benefit from limiting the resolution.
Center class averages?:: Yes
NOTE: This will re-center all class average images every iteration based on their center of mass. This is useful for their subsequent use in template-based auto-picking, but also for the automated 2D class average image selection in the next section
Perform image alignment?:: Yes
In-plane angular sampling:: 6
Offset search range (pix):: 5
Offset search step (pix):: 1
Allow coarser sampling?:: No
Number of pooled particles:: 30
Use GPU acceleration?:: Yes
Which GPUs to use:: <leave blank>
HEADER: Running options
Number of MPI procs:: 2
NOTE: Set this to one plus the number of GPUs you want to use, for example use 3 here if you have 2 GPU cards on the system, or 5 if you have 4 GPUs.
Number of threads:: 10
NOTE: The total number of CPUs used is the number of MPI procs multiplied by the number of threads. Set this so that (MPIs x threads) is equal to or slightly less than the number of CPU cores you want to use.
Then press RUN. The job should take less than 10 mins to execute, it took 2 mins on the tested system. The RESULTS tab will show a montage of 30 random particles extracted from the micrographs. Although not used in this tutorial, for future reference take notice of the SELECT MODE button. This allows for manual selection of classes you’d consider good, which will create a new Select job containing the particles only from the manually selected classes.
3.5. Automated 2D class selection
Next, we will select the good 2D classes and their particles using an automated procedure based on a neural network that was trained on thousands of 2D class averages. This requires Relion’s 2D class auto selection tool.
Under Automated Class Selection select a new RELION Automatic 2D Class Selection job
Select classes from optimiser.star:: Class2D/job007/run_it025_optimiser.star
Minimum threshold for auto-selection:: 0.21
NOTE: The score ranges from 0 for terrible class average images to 1 for excellent ones. We are using a relatively low value here, because we only have a few particles, so the 2D class averages will probably not look very good. When working on your own data sets, you will probably want to run the program once, sort your class averages on their predicted score and decide what a good value is for those class averages; also see below.
Then press RUN. The job should take less than 10 secs to execute. Once the job has finished, examine the results. You can hover the mouse over the rejected classes to see their threshold scores (the rejected particles are at the bottom of the RESULTS tab). If you are not satisfied with the results, you can run the job again with a modified threshold.
3.6. Fine-tuning the TOPAZ neural network
In the next step, we will select good particles from the rest of the micrographs using an automated procedure based on a neural network that will be fine-tuned to our data using the good set we just extracted in the earlier steps. In this step, we will fine-tune the network to our data.
Under Automated Particle Picking select a new RELION Topaz Training job.
Input micrographs for autopick:: Select/job004/sampled_micrographs.star
Pixel size in micrographs (A):: -1
OR train on a set of particles?:: Yes
Particles STAR file for training:: Select/job008/particles.star
NOTE: This will be the same job name regardless of which class selection method you chose above.
Particle diameter (A):: 180
Nr of particles per micrograph:: 300
Which GPUs to use:: <leave blank>
NOTE: This is set automatically if left blank
Then press RUN. The job should take less than 20 mins to execute, it took 12 mins on the tested system. We do not need to worry about the RESULTS tab in this case, as it mostly contains some training statistics.
3.7. Particle picking with Topaz
Now we will pick particles from all the micrographs with the Topaz picker that was fine-tuned on our good set.
From Automated Particle Picking select a new RELION Autopick Topaz (Single Particle) job.
Input micrographs for autopick:: CtfFind/job003/micrographs_ctf.star
NOTE: This is the CTF info for all micrographs from the precalculated results
Pixel size in micrographs (A):: -1
NOTE: A negative value here means the pixel size will be read from the file
Particle diameter (A):: 180
Nr of particles per micrograph:: 300
Trained topaz model:: AutoPick/job009/model_epoch10.sav
NOTE: This is the trained model from your previous job.
Use GPU acceleration?:: Yes
Which GPUs to use:: <leave blank>
HEADER: Running Options
Number of MPI procs:: 1
Then press RUN. The job should take less than 10 mins to execute, it took 2 mins on the tested system. Examine the RESULTS tab for this job. Compare them to the results of AutoPick/job005 (the LoG picking). Look how many more particles have been found!
3.8. Extracting the Full Particle Set
At this point you should have several thousand particles (typically 3,000 - 10,000). If you don’t, stop and check that you have run all the previous jobs correctly!
We will now extract the full particle set from all the micrographs. The particles will again be down sampled for faster processing. In Extract Particles select a Relion Extract Particles (Single Particle) job.
micrograph STAR file:: CtfFind/job003/micrographs_ctf.star
Input coordinates:: AutoPick/job010/autopick.star
HEADER: Extraction Options
Particle box size (pix):: 256
Invert contrast?:: Yes
Normalize particles?:: Yes
Diameter background circle (pix):: 200
Stddev for white dust removal:: -1
Stddev for black dust removal:: -1
Rescale particles?:: Yes
Re-scaled size (pixels):: 64
Write output in float16?:: Yes
HEADER: Running options
Number of MPI procs:: 12
Then press RUN. The job should take less than 10 secs to execute. Once the job has finished, the RESULTS tab will once again show a montage of extracted particles, but this time, they should be of good quality, having been extracted by an automated neural network fine-tuned on our selected set.
3.9. Cleaning up the particle set
We will next run another round of 2D classification and 2D class selection to improve the particle set. This time we will use the gradient based 2D classification algorithm VDAM.
From 2D Classification select a new RELION Class2D (VDAM, Single Particle) job.
Input images STAR file:: Extract/job011/particles.star
Number of VDAM mini-batches:: 100
Do CTF-correction?:: Yes
Ignore CTFs until first peak?:: No
Number of classes:: 100
Regularisation parameter T:: 2
Mask diameter (A):: 200
Mask individual particles with zeros?:: Yes
Limit resolution E-step to (A):: -1
Center class averages?: Yes
Perform image alignment?: Yes
In-plane angular sampling:: 6
Offset search range (pix):: 5
Offset search step (pix):: 1
Allow coarser sampling?:: No
Number of pooled particles:: 30
Use parallel disc I/O?:: Yes
Use GPU acceleration?:: Yes
Which GPUs to use:: <leave blank>
HEADER: Running options
Number of threads:: 24
Then press RUN. The job should take about 20-30 mins to execute, it took 20 mins on the tested system. Once the job has finished, the RESULTS tab will once again show a montage of selected 2d averages.
3.10. Select the particles from the good classes
This uses Relion’s automated 2D class selection on the new full data set. Alternatively you can do a manual class selection using the SELECT MODE button as described in 3.4 section of this tutorial.
Under Automated Class Selection select a new RELION Automatic 2D Class Selection job.
Select classes from optimiser.star:: Class2D/job012/run_it0100_optimiser.star
Minimum threshold for auto-selection:: 0.20
NOTE: The correct threshold varies, see below.
Then press RUN. The job should take less than 10 secs to execute. Once the job has finished, examine the RESULTS tab. Scroll down in the results page to find the images of the rejected classes. You can hover the mouse over the rejected classes to see their scores. If you are not satisfied with the results (e.g. if there are rejected classes that look good and you want to include them), you can run the job again with a modified threshold.
