Debian Codegraph @

docs/source/saxs/saxs_bead_models.rst

raw
1 Bead model reconstructions
2 ----------------------------------------------------------
3 .. _saxs_bead_models:
4
5 This tutorial covers basic principles and best practices for creating
6 bead (dummy atom) model reconstructions of macromolecule shape from
7 SAXS data. This is not a tutorial on how to use RAW for this type of analysis.
8 For that, please see the RAW tutorial for :ref:`DAMMIF/N <dammif>`.
9
10
11 Overview
12 ^^^^^^^^^^^^^^^^^
13
14 A natural desired outcome for many SAXS experiments is determining the
15 'solution structure' of the sample, i.e. the structure of the macromolecule
16 as it exists in solution. Unfortunately, unlike crystallography, cryoEM,
17 and NMR, SAXS data cannot be used to generate a high resolution 3D shape
18 (though it can be used to constrain other methods of structure determination).
19 What SAXS can often provide, and what is a common end point of SAXS analysis,
20 is a low resolution shape reconstruction of the sample.
21
22 For many years, bead modeling was the state of the art approach
23 for generating these low resolution shape reconstructions. Recently,
24 other techniques have been developed, such as direct reconstruction of
25 the electron density at low resolutions. Despite this, bead modeling
26 remains the de facto standard for shape reconstruction.
27
28
29 Why do we do bead model reconstructions?
30 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
31
32 Bead models, despite their low resolution, can be powerful tools for understanding
33 the system in solution. High resolution structures can be docked into the bead
34 model, allowing visual analysis of how well the high resolution structure
35 agrees with the solution structure. In the absence of other structure
36 information, bead models can be used to provide important
37 clues to the overall shape and size of the system, which can often be
38 enough to draw conclusions about important aspects of the macromolecular
39 function or interactions with other systems.
40
41 While bead modeling can be quite useful, it is important to always keep in
42 mind two things. First, it is very easy to get poor reconstructions, even
43 with good quality data, and you must carefully evaluate the results of your
44 reconstructions before using them. Second, SAXS is much more accurate at
45 hypothesis testing than it is at generating bead models. Put another way
46 SAXS is very good at telling you what something isn't, but not so good
47 at telling you what it is. For example, if you want to compare a high
48 resolution structure to SAXS data and see if they agree, you are always
49 better off testing the calculated scattering profile of the high resolution
50 structure against the measured scattering profile than docking the structure
51 into the bead model. In this case, the bead model might be useful in visualizing
52 the differences between the solution shape and your high resolution structure.
53
54
55 How do we do bead model reconstructions?
56 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
57
58 There are a number of different programs available to do bead modeling, some
59 suited for general systems and some tuned for very specific applications
60 like reconstructions of membrane proteins with detergent halos. Regardless,
61 all of these methods share a similar approach.
62
63 #. Generate a volume of beads (aka 'dummy atoms') and randomly assign the
64 beads to be one of the allowed phases. In general, beads are either
65 solvent or macromolecule, but some programs allow more than 2 phases,
66 such as differentiating between protein and RNA or macromolecule and lipid.
67 #. Calculate a scattering profile from the bead model and fit that against the data.
68 #. Flip a randomly chosen set of beads between phases (e.g. from solvent to
69 macromolecule or vice versa).
70 #. Recalculate the scattering profile from the model and the to the data.
71 #. If the fit is better, accept the changes to the beads. If the fit is
72 worse, accept the changes to the beads with some probability (avoids
73 local minima).
74 #. Repeat steps 3-5 until the convergence criteria is met.
75
76 Additional, programs usually impose physical constraints on the bead models
77 to improve the model. Common constraints are requiring connectivity of the model,
78 imposing penalties for extended models, and constraining the size of the model
79 based on the |Rg| and/or |Dmax|.
80
81 It turns out that a scattering profile does not generate a unique reconstruction,
82 so to account for this a Monte Carlo like approach is taken where a number
83 (usually 10-20) of models are generated, and then averaged to give a consensus
84 reconstruction.
85
86 The most common program for creating bead model reconstructions (though far from
87 the only such program) is DAMMIF (or DAMMIN) [1-2] from the ATSAS package.
88 For the remainder of this tutorial we will exclusively discuss using
89 DAMMIF/N, though much of the discussion should apply to other programs as well.
90
91
92 Using DAMMIF/N for bead model reconstructions
93 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
94
95 DAMMIF is the most commonly used program for bead model reconstructions. Here
96 we discuss some of the practical aspects. A detailed tutorial of
97 :ref:`how to use DAMMIF in RAW <dammif>` is also available.
98
99 Input data
100 *************
101
102 DAMMIF requires a P(r) function generated by GNOM (a .out file) as input. Note
103 that, as mentioned in the :ref:`IFT tutorial <saxs_ift>`, the scattering profile
104 for the IFT should be truncated to a maximum *q* value of 8/R\ :sub:`g` or
105 ~0.25-0.30 1/Angstrom, whichever is smaller. This is because the hydration layer
106 and internal structure are not modeled by DAMMIF, which leads to errors at
107 higher *q*. The truncation removes the potentially problematic higher *q* data.
108
109 Generating models
110 **********************
111
112 As bead modeling does not generate a unique solution. In order to generate a
113 reasonable solution, we create 10-20 bead model reconstructions and then average them.
114 I recommend 15 reconstructions. This means that we need to run DAMMIF 15 different
115 times.
116
117 The most accessible settings for DAMMIF are the Mode, Symmetry, and Anisometry.
118
119 **Mode:** For mode, options are Fast or Slow. Fast is quick, but less detailed, Slow is
120 the opposite. For a final reconstruction, use Slow mode.
121
122 **Symmetry:** Adding in symmetry constraints can improve the reconstruction.
123 If you know the symmetry of the particle, you can specify this. However,
124 it is always recommend that you do an additional set of reconstructions
125 with P1 symmetry, to verify that the symmetry did not overly constrain the
126 reconstructions.
127
128 **Anisometry:** Adding in anisometry constraints can improve the reconstruction.
129 If you know the anisometry of the particle, you can specify this. However,
130 it is always recommend that you do an additional set of reconstructions
131 with no anisometry, to verify that the symmetry did not overly constrain the
132 reconstructions.
133
134 Additional advanced options are available, and are described in the
135 `DAMMIF manual <https://biosaxs-com.github.io/atsas/4.0.0/manuals/dammif.html>`_.
136
137 If you only want a quick look at the shape (such as when collecting data
138 at a beamline) 3 reconstructions in Fast mode will work for that purpose.
139
140 Averaging and clustering models
141 **********************************
142
143 After models are generated the next step is to average and cluster the models.
144 Averaging generates a consensus shape from the individual models, and provides
145 statistics on how stable the reconstruction is. This is done with DAMAVER [3].
146 The average outputs both damaver.pdb and damfilt.pdb model files. These correspond
147 to two different consensus shapes of the model, loosely and tightly defined
148 respectively. However, neither of these models actually fits the data, and so
149 generally should not be used to display your reconstructions. DAMAVER will
150 also specify the most probably individual model. If you do not refine the
151 results of DAMAVER (below) you should use the most probable model as your
152 final result.
153
154 Clustering is done with DAMCLUST [4] and clusters models that are more
155 similar to each other than they are to the rest of the models. This is a
156 way of assessing the ambiguity of the reconstruction, and we will discuss
157 it further in the section on evaluating reconstructions below.
158
159 Creating a final refined model
160 ********************************
161
162 The output of DAMAVER, specifically the damstart.pdb file, can be used
163 as input for DAMMIN to create a final refined model. Essentially, the
164 damstart.pdb represents a conservative core of the most probably
165 occupied volume as determined by averaging all the reconstructions using
166 DAMAVER. DAMMIN keeps this core fixed, and refines the outside of the model
167 to match the scattering profile. I've seen mixed recommendations (even from
168 the makers of the software) on whether you should do a refinement. I typically
169 do, but it seems you can often do just as well with the most probable model
170 determined by DAMAVER.
171
172 Evaluating DAMMIF/N reconstructions
173 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
174
175 SAXS data contains very limited information, both because it is measured at
176 relatively low *q*, and because it is measured from a large number of particles
177 in solution oriented at random angles. The SAXS scattering profile represents the
178 scattering from a single particle, averaged over all possible orientations.
179 The practical consequence of this is that there are often several possible
180 shapes that could generate the same (or so similar as to be indistinguishable
181 within experimental noise) scattering profiles. As such, it may simply not
182 be possible to generate a bead model reconstruction from a dataset that
183 accurately represents the solution shape, regardless of the overall data
184 quality. If the sample is flexible or otherwise exists in multiple conformational
185 or oligomeric states in solution the reconstruction is also challenging or
186 impossible. **In summary, high quality SAXS data is not a guarantee of a good
187 bead model reconstruction. This makes it very important to critically evaluate
188 every reconstruction done, regardless of the underlying data quality.**
189
190 The information needed to evaluate the reconstructions is generated when
191 running DAMMIF, DAMAVER, DAMCLUST, SASRES [5] (run as part of DAMAVER) and
192 AMBIMETER [6]. While it can all be accessed through the files these programs
193 generate, RAW gathers and presents it for you when you run DAMMIF in RAW.
194
195 |dammif_results_png|
196
197 Criteria for a good DAMMIF/N reconstruction
198 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
199
200 * Ambiguity score < 2.5 (preferably < 1.5)
201 * NSD < 1.0
202 * Few (0-2) models rejected from the average
203 * Only one cluster of models
204 * Model :math:`\chi^2` near 1.0 for all models
205 * Model |Rg| and |Dmax| close to values from P(r) function for all models
206 * M.W. estimated from model volume close to expected M.W.
207
208 More about these criteria can be found below.
209
210 Ambiguity
211 *************
212
213 It is possible to evaluate the potential ambiguity of your bead model reconstructions
214 before doing the reconstructions. The AMBIMETER program in the ATSAS package
215 can be run on P(r) functions from GNOM to assess how likely you are to get
216 a good reconstruction. The program has a database of scattering profiles
217 representing all possible shapes made out of up to 7 beads. Your scattering profile
218 is compared against these shapes, and AMBIMETER reports how many match your
219 profile. The more profiles from AMBIMETER that match yours, the more possible
220 shapes could have generated your profile.
221
222 AMBIMETER reports both the number of shapes and the log (base 10) of the number
223 shapes, which is the Ambiguity score. They provide the following interpretations:
224
225 * Ambiguity score < 1.5 - Reconstruction is likely unique
226 * Ambiguity score of 1.5-2.5 - Take care when doing the reconstruction
227 * Ambiguity score > 2.5 - Reconstruction is most likely ambiguous.
228
229 This provides a quick initial assessment of whether you should even attempt
230 a shape reconstruction for your dataset. :ref:`You can run AMBIMETER from RAW <raw_ambimeter>`.
231
232 Normalized spatial discrepancy
233 **********************************
234
235 DAMAVER reports a number of different results. The most useful is the normalized
236 spatial discrepancy (NSD). This is essentially a size normalized metric for comparing
237 how similar two different models are. When DAMAVER is run, it reports the
238 average and standard deviation of the NSD between all the reconstructions. It
239 also reports the average NSD for each model.
240
241 The average NSD is commonly used to evaluate the stability of the reconstruction.
242 Roughly speaking we evaluate reconstruction stability as:
243
244 * NSD < 0.6 - Good stability of reconstructions
245 * NSD between 0.6 and 1.0 - Fair stability of reconstructions
246 * NSD > 1.0 - Poor stability of reconstructions
247
248 Generally speaking, if your average NSD is less than 1.0, the reconstruction can
249 probably be trusted (if all of the other validation metrics also check out),
250 while if it is greater than 1.0 you should proceed with caution, or not use
251 the reconstructions at all.
252
253 The NSD is also used to determine which models to include in the average.
254 If the average NSD of a given model is more than two standard deviations above
255 the overall average NSD, that model is not included in the average. If more
256 than ~2 models are rejected (out of 15), that may be a sign of an unstable
257 reconstruction.
258
259 Clusters
260 ***********
261
262 DAMCLUST creates clusters of models that are more similar to each other
263 than they are to the rest of the models. This is a way of assessing the
264 ambiguity of the reconstruction. If you have more than one cluster of models
265 in your reconstructions, you may have several distinct shapes that are being
266 reconstructed by the DAMMIF algorithm. This typically indicates that there
267 are several distinct shapes in solution that could generate the measured
268 scattering profile, and so is another indication of a highly ambiguous
269 reconstruction.
270
271 The caveat to this is that with good quality data that is very low ambiguity
272 (ambiguity score from AMBIMETER < 0.5) and yields a set of reconstructions with
273 a very small average NSD (<0.5, typically) and NSD standard deviation (~0.01),
274 I have seen several (often >5) clusters identified with DAMCLUST. I believe
275 that in this case there are not actually multiple clusters, but the extremely
276 low deviation between the models is fooling the DAMCLUST algorithm.
277
278 Note that the different clusters should not be taken as representatives of different
279 distinct shapes in solution. Even if there are a finite number of distinct shapes
280 scattering in the solution (such as an open and closed state of a protein),
281 the measured scattering profile is an average of the scattering from each component, and
282 each individual reconstruction fits that measured scattering profile. As such,
283 there is no way for an individual reconstruction to fit just the scattering from
284 one of the components and so the different clusters cannot be representative
285 of the different shapes in the solution.
286
287 Model fit and parameters
288 *****************************
289
290 Each model has the following parameters that can be used to evaluate the success
291 of an individual reconstruction: :math:`\chi^2`, |Rg|, |Dmax|, volume, molecular
292 weight estimated from volume, and the normalized residual of the model fit to
293 the data. For a good fit to the data, the model :math:`\chi^2` should be close
294 to 1 and the normalized residual between the model fit and the data should
295 be flat and randomly distributed about zero. However, in my experience the normalized
296 residual often shows some small systematic deviations, and so this should not
297 be too concerning. A :math:`\chi^2` value significantly larger than 1 (1.5-2 or
298 larger) indicates either a poor fit to the data or that the uncertainty for
299 the data is underestimated. To differentiate between these two cases, look
300 at the normalized residual. If it is flat and randomly distributed, then
301 the uncertainty is most likely underestimated. If it shows significant
302 systematic deviations then the fit quality is poor.
303
304 The |Rg| and |Dmax| obtained from the model should be close to those calculated
305 from the P(r) function. If that is not the case, you should reevaluate your P(r)
306 function and redo the reconstruction if necessary. If the discrepancy persists,
307 it is an indication that your reconstruction isn't a good representation of what
308 is in solution, and shouldn't be trusted. While there's no hard and fast rule
309 here on how closely |Rg| and |Dmax| should agree, my experience is generally
310 that for high quality data |Rg| agrees to better than ~5% and |Dmax| to ~10%.
311
312 The volume is reported for each bead model, but it is usually easier to
313 compare the molecular weight calculated from that volume with the expected
314 molecular weight. In this case, M.W. is calculated by dividing the volume
315 (nominally representing the sample's excluded volume) by an empirically
316 determined constant [4] of 1.66 (used in RAW, other programs may use different
317 values). This value is approximate, and varies between roughly 1.5 and 2.0
318 depending on the shape of the macromolecule. This M.W. is less well determined
319 than :ref:`other SAXS methods <saxs_mw>`, given the variation in the coefficient.
320 As such, it is mostly useful for indicating general agreement between the overall
321 size of the reconstruction and the expected size. If the M.W. is different from
322 the expected M.W. by more than 20-25% you should consider the reconstructions
323 to be suspect.
324
325
326 Limitations of bead models
327 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
328
329 While bead models can be quite useful, they have a number of limitations,
330 many of which are mentioned in previous parts of the tutorial. In summary:
331
332 * Bead models can be ambiguous, even if the data quality is very high.
333 This is because multiple different shapes in solution can produce the
334 same scattering profile, so there is no guaranteed unique solution to
335 a reconstruction, and the success of the reconstruction depends not just
336 on the input data quality but also the inherent shape of the particle
337 and how ambiguous that shape is for SAXS. Because of this, all models
338 should be thoroughly evaluated as described above.
339
340 * Ignoring ambiguity, bead models still only work best with particular
341 particle shapes. An excellent discussion of how well bead models work for
342 different types of shapes is found in [3]. The summary is that bead
343 models tend to be less reliable for high aspect ratio objects, such as
344 long rods or thin discs, objects with voids (such as a spherical shell),
345 and rings. They are most reliable for things that are generally globular.
346
347 * Bead models are low resolution. Small variations of the surface of the
348 model are likely insignificant. I rarely see estimated model resolutions
349 less than ~20 Angstroms, often they are much larger.
350
351 * Bead models do not (typically) model the hydration layer or internal
352 structure of the particle. This requires that you use only data out to
353 a maximum *q* of 8/R\ :sub:`g` or ~0.25-0.30 1/Angstrom, whichever is less.
354
355 * The most common bead modeling programs cannot model multiple electron
356 densities within a sample, such a protein-nucleic acid complex or
357 a membrane protein with a detergent halo. There are specialized
358 programs (such as `MONSA <https://biosaxs-com.github.io/atsas/4.0.0/manuals/monsa.html>`_
359 or `Memprot <https://pubmed.ncbi.nlm.nih.gov/25615863/>`_) that can handle
360 these cases, but these require the input of additional information to provide
361 extra constraints.
362
363 * The bead model is only as good as the input data. In particular,
364 bead models are quite sensitive to the presence of larger particles
365 in solution, either oligomers or non-specific aggregate. In one simple
366 simulation I've seen, as little as 0.7% aggregate caused a significant
367 change in the bead model. Non-specific aggregation usually manifests
368 as an extended protrusion from the main model.
369
370 As you can see, while bead models can certainly be useful for your research,
371 you should proceed with caution and ensure that you have a trustworthy
372 reconstruction before proceeding with your bead models.
373
374
375 Visualizing DAMMIF/N reconstructions
376 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
377 Visualizing DAMMIF/N bead model reconstruction is slightly different from
378 displaying a typical macromolecular structure. There are two main ways that
379 these are visualized, either as individual beads or, more commonly, as an
380 envelope that defines the edges of the model. Both representations are usually
381 made semi-transparent so that a high resolution structure docked with the
382 bead model is simultaneously visible.
383
384 The main detail to remember is that to get a correct visualization you have to
385 set the correct bead size for the model, which is given in the header of the
386 DAMMIF/N .cif file.
387
388 Below are two quick tutorials for visualizing models in Chimera (or ChimeraX)
389 and PyMOL.
390
391 Visualizing bead models with Chimera
392 **************************************
393
394 Note: There are some differences between Chimera and the newer ChimeraX.
395 Differences for ChimeraX are noted in **bold**.
396
397 #. Open Chimera.
398
399 #. Load in the DAMMIF/N .cif file of interest.
400
401 #. If necessary, open the Model Panel and the Command Line from the Tools->
402 General Controls menu.
403
404 #. In the Select->Chain menu choose the bead model ("no ID" or **the filename**).
405
406 #. In the Actions->Ribbons (**Actions->Cartoon**) menu, choose "Hide".
407
408 #. In the Actions->Atoms/Bonds menu, choose "Show".
409
410 #. In the Actions->Atoms/Bonds (**Actions->Atoms/Bonds->Atom Style**) menu
411 choose "Sphere".
412
413 #. **In the Camera section of the ChimeraX "Graphics" ribbon, click "View
414 selected".**
415
416 #. Open the .cif file in a text editor.
417
418 #. Find the "Dummy atom radius" value. That is the bead size you need to set.
419
420 #. In the command line, enter the command ``vdwdefine x #y`` where ``x``
421 is the bead size from the PDB header and ``y`` is the ID number of the
422 bead model shown in the model panel.
423
424 * **The command is "size atomradius x" in ChimeraX.**
425
426 #. Your beads are now the right size. If you want to make an envelope
427 proceed with the following optional steps. If you'd rather use the
428 individual bead display, you can stop here.
429
430 #. To make an envelope, in the command line enter the command ``molmap #y z``
431 where ``y`` is the ID number of the bead model shown in the model panel
432 and ``z`` is 3x the bead size that you found in the previous steps.
433
434 * *Tip:* The last number controls the smoothness of the envelope. You
435 may need to vary it from 3*(bead size), depending on the size of
436 your beads and how smooth you want your envelope. I recommend leaving
437 at least a hint of the beads visible (not overly smoothing the envelope) to
438 help whoever sees the graphic to remember that an envelope is not
439 an electron density contour.
440
441 #. Hide the bead model using the "Hide" button in the model panel.
442
443 #. In the Volume Viewer window that appeared when you entered the molmap
444 command, in the Features menu select the "Surface and Mesh options'.
445
446 * *Note:* This menu doesn't exist in ChimeraX.
447
448 #. Check the box for Surface smoothing and set the iterations to 2 and the
449 factor to 1.
450
451 * *Note:* This option doesn't exist in ChimeraX.
452
453 #. Check the box for Subdivide surface and set it to 2 times.
454
455 * *Note:* This option doesn't exist in ChimeraX.
456
457 #. Click on the color box to set color and opacity. I find that 0.4 (**40%**)
458 is a good opacity for overlaying with high resolution models.
459
460 #. Load in your aligned (such as with :ref:`SUPCOMB <raw_supcomb>`) high
461 resolution structure if available.
462
463 Visualizing bead models with PyMOL
464 **************************************
465
466 #. Open PyMOL
467
468 #. Load in the DAMMIF/N .cif file of interest.
469
470 #. Using the model Hide menu ('H'), hide 'everything'.
471
472 #. Using the model Show menu ('S'), show 'spheres'.
473
474 #. Open the .cif file in a text editor and find the "Dummy atom radius"
475 value in the header.
476
477 #. In the PyMOL command line, enter the command ``alter <model_name>, vdw=x``
478 where ``x`` is the size you found in the previous step. This sets the
479 spheres to be the correct size of the beads in the model.
480
481 #. Click the 'Rebuild' button to refresh the view of the model.
482
483 #. Your model is now displayed correctly with beads. If you want to make an
484 envelope, proceed with the following optional steps. If you'd rather use
485 the individual bead display, you can stop here.
486
487 #. Using the model Show menu, show 'surface'.
488
489 #. Using the model Hide menu, hide 'spheres'.
490
491 #. You can smooth the surface by increasing the probe radius using
492 the command ``set solvent_radius, 3.0`` (where you can vary the size
493 from 3.0).
494
495 #. You can improve the quality of the surface using the command
496 ``set surface_quality, 1``.
497
498 * *Note:* Values larger than 1 may take a long time to render.
499
500 #. Set your surface transparency to 50% for overlaying with high resolution
501 models using the command ``set transparency, 0.5``
502
503 #. Load in your aligned (such as with :ref:`SUPCOMB <raw_supcomb>`) high
504 resolution structure if available.
505
506 FAQ
507 ^^^^^^^^
508
509 Do I have to make a bead model?
510 **********************************
511
512 No. It all depends on what you're trying to say about the data. However,
513 particularly if your system shows signs of flexibility or AMBIMETER reports
514 a high ambiguity score you probably shouldn't bother making a bead model
515 even if you want to.
516
517 How do I fit my high resolution structure into my bead model?
518 *****************************************************************
519
520 If your high resolution structure is relatively complete (contains all
521 residues in solution, and ideally post-translational modifications), you
522 can use a program like SUPCOMB [7] to automatically
523 :ref:`fit the structure into the bead model <raw_supcomb>`. If you are missing
524 significant amounts of the structure (such as a large flexible loop) or have
525 only one subunit of a multi-subunit complex you may have to manually dock
526 the structure in the envelope.
527
528 My bead model and my high resolution structure disagree. Which one is right?
529 ******************************************************************************
530
531 Maybe both, maybe neither! It really depends on your inputs. If you've validated
532 the bead model as above and it seems good, then it likely represents the low
533 resolution shape in solution. You should also verify that your high resolution
534 shape contains all of the residues in your system, often high resolution
535 structures are missing things like flexible loops or N and C terminus regions.
536
537 If both models are good, then depending on how you obtained your high resolution
538 shape it might also be correct, but represent the shape under different conditions.
539 For example, it is common in crystallography to see structural artifacts induced
540 by the packing of the macromolecule into the crystal.
541
542 Of course, the best way to compare your high resolution structure to SAXS data
543 isn't by docking it in a bead model, but by fitting it against the data using
544 a program like `CRYSOL <https://biosaxs-com.github.io/atsas/4.0.0/manuals/crysol.html>`_ or
545 `FoXS <https://modbase.compbio.ucsf.edu/foxs/>`_. If these fits are bad, then
546 your high resolution structure doesn't match the data, regardless of what the
547 bead model shows. If these fits are good, and the bead model doesn't agree with
548 the high resolution structure, then the bead model is wrong.
549
550 My bead model isn't good, what should I do instead?
551 *********************************************************
552
553 There are many more approaches available than I can list here, but a couple
554 of the more common ones are:
555
556 * If your data is flexible, you can try some kind of ensemble based approach,
557 such as `EOM <hhttps://biosaxs-com.github.io/atsas/4.0.0/manuals/eom.html>`_,
558 `SASSIE <https://sassie-web.chem.utk.edu/sassie2/>`_, or
559 `BilboMD <https://bl1231.als.lbl.gov/bilbomd>`_.
560
561 * If your data is more rigid and consists of several subunits you can consider
562 rigid body modeling such as
563 `SASREF <https://biosaxs-com.github.io/atsas/4.0.0/manuals/sasref.html>`_.
564
565
566
567 References
568 ^^^^^^^^^^^^
569
570 1. Franke, D. and Svergun, D.I. (2009) DAMMIF, a program for rapid ab-initio
571 shape determination in small-angle scattering. J. Appl. Cryst., 42, 342-346.
572
573 2. D. I. Svergun (1999) Restoring low resolution structure of biological
574 macromolecules from solution scattering using simulated annealing. Biophys J. 2879-2886.
575
576 3. V. V. Volkov and D. I. Svergun (2003). Uniqueness of ab-initio shape
577 determination in small-angle scattering. J. Appl. Cryst. 36, 860-864.
578
579 4. Petoukhov, M.V., Franke, D., Shkumatov, A.V., Tria, G., Kikhney, A.G.,
580 Gajda, M., Gorba, C., Mertens, H.D.T., Konarev, P.V. and Svergun, D.I.
581 (2012) New developments in the ATSAS program package for small-angle
582 scattering data analysis. J. Appl. Cryst. 45, 342-350
583
584 5. Anne T. Tuukkanen, Gerard J. Kleywegt and Dmitri I. Svergun(2016) Resolution
585 of ab initio shapes determined from small-angle scattering IUCrJ. 3, 440-447.
586
587 6. M.V. Petoukhov and D.I. Svergun (2015) Ambiguity assessment of small-angle
588 scattering curves from monodisperse systems Acta Cryst. D71, 1051-1058.
589
590 7. M.Kozin & D.Svergun (2001) Automated matching of high- and low-resolution
591 structural models J Appl Cryst. 34, 33-41.
592
593 .. |Rg| replace:: R\ :sub:`g`
594
595 .. |Dmax| replace:: D\ :sub:`max`
596
597 .. |dammif_results_png| image:: ../tutorial/images/dammif_results.png
598 :target: ../_images/dammif_results.png