This is an old revision of the document!


Gold Standard reconstruction

Starting with the version 4.05 of Auto3DEM, it is possible to follow what has been recently defined as the 'gold standard procedure' (called 'unbiased model reconstruction method' in the original paper on Auto3DEM by Yan et. al, 2007), in a way that is almost transparent to the user. The procedure consists of splitting the data into two half sets of particles, at the beginning of the iterative alignment: at each iteration each set generates an independent reconstruction, which is used as a reference, against the same set, for the next cycle of origin and orientation refinement. In this approach two independent maps are generated, and any possible bias introduced from the data by overfitting, will not contribute to overestimate the resolution of the final structure.

Generation of two independent stacks of data

The input particles need to be organized into two independent sets. This is achieved by generating two stack parameter files, each pointing to a different subset of particles. The actual files containing the images, i.e. the .pif files, don't need to be manipulated. The two stack parameter files can be generated by using the general purpose program autopp (option 6). If each stack of particles has a separate parameter file, it is recommended to consolidate them in one single file, still using autopp (option 5). The approach used to generate two independent sets of particles is to split them according to their sequential numbering in each stack: even-numbered particles are assigned to one set and odd-numbered to the other. The convention adopted for the stack parameter files is that their extension reflects the set their particles belong:

.dat_XXX   -> all particles from the stack
.date_XXX  -> only even-numbered particles from the stack
.dato_XXX  -> only odd-numbered particles from the stack

where XXX is the iteration number the parameters refer to. It is to remark that the strategy to split the data into two independent sets can actually be arbitrary, as long as the parameter files follow this convention. However, for different strategies the user needs to generate the .date and .dato files on his/her own, since autopp only splits by sequential numbering.

Reference maps

Each half set of particles needs to be aligned against a different reference. In the parameter file controlling the workflow of Auto3DEM (usually ending with _master or _continue) the reference maps for the two independent sets are specified in the fields

auto  start_map_even
auto  start_map_odd

When starting the iterative procedure, the same map can be used as initial reference for both the two sets, as long as only the low resolution features are used determining the alignment parameters. This is achieved by setting either the parameter ppft resolution_high or po2r res_max, depending if auto mode is set to search or refine, respectively, to 30 Å or higher.

Aligning the particles

In the parameter file controlling the workflow of Auto3DEM a flag, auto gold_standard, controls the use of the gold standard procedure. If set to 1, the fields data can only contain a list of stack parameter files properly split, that is only .dato and .date files are accepted. If the flag is set to 0, no gold standard procedure is followed, all the stack parameter files are treated as one single data set, and their particles are aligned against the map listed in the parameter auto start_map. In the gold standard procedure, after each iteration three maps are actually generated: one from each set and one from all the particles. The last one is just for visualization purposes, and only its radial amplitude profile is utilized to re-calibrate some masking parameters.

If you use the Random Model method implemented in Auto3DEM to generate an ab-initio reference map, the program takes care to generate the two independent stack parameter files, along with a _master file properly set for the gold standard procedure.

Limitations

Currently there are two caveats in the procedure:

Handedness
Currently, in case the two reconstructions converge to maps with opposite handedness, the program is not able to detect it, and therefore it will not be able to improve the reconstruction beyond the resolution where the handedness becomes evident. At the moment the simplest solution is to compare visually the two independent maps, and if the handedness is different then:

  1. modify the stack parameter files from the last iteration, for only one set, using the command handflip;
  2. change the handedness of the corresponding reconstruction. This can be done either recomputing the reconstruction from the modified parameter files (in the last _continue file, just set 'auto generate_map' to 'only' and launch it), or by modifying the map using an external program (e.g. using the program bimg from bsoft, with the option '-reslice xy-z').

Background masking
Currently the two independent reference maps generated at each iteration are not low-pass filtered, and therefore the level of noise from the high frequency regions is very high. For this reason the option to perform a solvent flattening of these maps, when used as references, is not successful, since the algorithm can not discriminate between density and noise. Therefore, when using the gold standard procedure the flattening option must be avoided by setting

auto flatten_map 0