EMAN Refinement

Files needed

• threed.0a.mrc -- the strating model, perhaps produced by common lines
• start.hed, start.img -- the particle stack
• optional: structure factor file, if doing full CTF correction
• optional: start.filt.hed, start.filt.img: if using filtered particles for projection matching

Description of command line options

• EMAN uses a single command, refine, to do all refinement
• A typical command is refine 8 hard=25 pad=120 sym=C1 mask=40 ctfcw=structure.sf dfilt classkeep=1 classiter=8 refine
• There are a large number of switches, some notes on what the commands mean are as follows

Size options

• ang=[Angular increment in degrees]
  • Take arctan (desired resolution / particle radius) as a first estimate
  • A fairly critical parameter. Highly resolution dependant, dependant on size of model and level of detail
  • Do not be too aggressive at first
  • If you entered all the info in the first eman window, the program will give you an appropriate value to use
  • Time varies with (1/n) squared
• mask=[Mask radius in A]
  • should be bigger than the particle size, in most cases substantially bigger
  • most of the time, use 1/2 box size
• pad=[padded box size in pixels]
  • used to prevent artifacts caused by direct Fourier inversion
  • spreads particles out so that ripples are pushed to edge of padded box
  • why not automatically do this? Memory considerations
  • Should be 25% larger than box size and have small prime factors

CTF correction parameters (one of the following three options)

• ctfc=[RES in Angstroms]
  • No structure factor file needed
  • Good for low-res work only (within 1st 0 of CTF)
  • Advantage of using it is that it avoids the HP filtration effect of CTF
  • Do not use in most instances, as ususally you want full CTF correction
• ctfcw=[SF filename]
  • Need structure factor file, which must be the same as was used in determining the ctf value
• median
  • Math of this is poorly defined (don't use often)
• Note that ctfc and ctfcw require that ctf params have been added to the particle stack, otherwise the command will likely fail or give wrong results

Phase error options

• dfilt, fscls, phasecls
  • if no CTF correction is being done, use fscls or phasecls (particle dependant)
  • fscls uses Fourier shell correlation to determine similarity
  • phasecls uses phase residual
  • dfilt is an optimized variance with a phase filter
  • if doing CTF correction, use dfilt
  • dfilt : works slightly better than the others, expect a 5 - 10% improvement in final resolution over fscls
• hard=[phase error in Angstroms]
  • looks at mean phase error between class average and reconstruction projection
  • decides whether or not a given class average should be used
  • discards class averages which do not match
  • calculates mean phase error over predifined frequency range
  • Need to watch output in order to see which class averages it is throwing away
  • In most cases, 25 is a good value
  • However, if starting with a Gaussian blob, may need to start bigger (45-55)
  • Don't make much smaller than 25 -- throw away too much

• • amost always specified
  • refers to 2d alignment process
  • for alignment, EMAN uses an autocorrelation function, the translates to polar coordinates aligns, then back to real space
  • refine then does a final refinement in real space, to sub-pixel resolution (0.02 pixel, 0.05 degrees)
  • Why not use? Takes twice as long
• sym=[cn, dn, oct, or icos]
  • symmetry enforcement
  • for asymmetric particles, use sym=C1 or omit entirely
  • cn describes single n-fold rotational symmetry about the z-axis
  • dn describes n-fold dihedral symmetry (cn with n 2-folds in xy plane)
  • oct is octahedral (2-3-4 symmetry of cube)
  • icos is icosahedral (2-3-5)

Iteration options

• [N]
  • Number of refinement iterations to do
  • It will start with the most previously completed ieteration, so refine 10 after refine 8 will start with iteration 9
• classiter=[number]
  • Actually uses (N-2) iterations (why???)
  • so classiter=3 will do 1 iteration (1 and 2 are undefined)
  • 0 can be used, but risky -- use at end, or alternate with 3 to get more resolution
  • The number refers to the number of reference-free alignments of each class average done at each overall iteration
• classkeep=[number]
  • happens same time as classiter
  • Decides how many particles will be kept for making class average
  • goes "classkeep" number of std deviations away from mean
  • Classkeep=0: throw away 50% of particles
  • classkeeep=1: throw away 10-20%
  • Usually use 1-2

Additional options

• sep=[n]
  • useful -- "a poor-man's maximum liklihood"
  • each particle is put into best n classes, rather than just the best class
  • Use with oversampled ang values at end of refinement
  • no weighting, just puts into n bins
  • combine with a low class-keep value (<1, maybe <0) to throw away bad particles
• xfiles=[a/pix,mass(kD),ali-to]
  • for each threed.1a.mrc, makes a x.1.mrc file scaled so that surface threshold of 1 gives mass specified
  • ali-to means to align to a specific iteration number, to prevent moleculae from drifting.
    • Using 99 is common, to prevent it from doing the alignment
  • a/pix -- need to specify pixel size for proper scaling
• amask=<radius,threshold,iteration>
  • "Very important option"
  • does a 3d floodfill outside of particle (solvent flattening)
  • radius -- in pixels -- just big enough to touch inside density of particle
  • isosurface -- the line that is floodfilled
  • iter -- add a bunch of trailing shells outside of tight floodfilled mask
  • must use with xfiles option since threshold selection requires volume normalization
  • radius -- radius of particle
  • threshold --slightly lower than surface threshold needed to see particle (if 1 is edge, use 0.7)
  • iteration: 10% of box size
  • If too tight, get artificial increases in resolution
• shrink=[n]
  • reduces image size n-fold
  • use if you don't want to bother shrinking data at start
  • Don't use on final model
  • shrink=2: gives an 8-fold speed improvement
• usefilt
  • No longer experimental, is now in frequent use
  • most helpful in full CTF correction
  • do any filtration you like (lowpass, Weiner, ...) and save particle stack as start.filt.hed
• refmaskali
  • Use mask that was projected from 3d volume to do alignment (2d), but do nor apply mask to aligned particles
  • prevent mask from throwing away extra chunks
  • good idea in early stages, but will give you lower resolution
  • A good safe option at first, until you are sure

Options not recommended

• tree=[2,3]
  • uses a 2-step classification, to speed things up
  • thinks it is working ok, but may not work on cluster
  • shrinking particle beforehand is a better way to speed things up
• filt3d
  • "If you need it, then something else is wrong"
• euler2
  • Not suggested -- often only makes things worse
• perturb
  • "probably not useful"
• imask=[radius]
  • Apparently still works, but not tested in a long time
  • only useful in specific circumstances
• collapse
  • Don't use
• any other options not mentioned in this list
  • "Lots of options have been added to EMAN over the years, but never removed"
  • "Documentation has not been updated in several years"

Resolution determination

• use eotest with same options as refine
• Delete any options that give you an error

• Set ALLOWTOPICVIEW =

-- BillRice - 18 Aug 2008