| 1 |
Molecular weight calculation |
| 2 |
------------------------------ |
| 3 |
.. _saxs_mw: |
| 4 |
|
| 5 |
This tutorial covers basic principles and best practices for calculating |
| 6 |
molecular weight from SAXS data. This is not a tutorial on how to use RAW |
| 7 |
for this type of analysis. For that, please see the :ref:`RAW tutorial <raw_mw>`. |
| 8 |
|
| 9 |
|
| 10 |
Overview |
| 11 |
^^^^^^^^^^^^^^^^^ |
| 12 |
|
| 13 |
There are a number of ways to calculate molecular weight from SAXS data. RAW |
| 14 |
supports four of the most common methods natively: |
| 15 |
|
| 16 |
#. Molecular weight from absolute scaled I(0) [1]. |
| 17 |
#. Molecular weight by comparing to a reference standard [1]. |
| 18 |
#. Molecular weight from the Porod volume [2]. |
| 19 |
#. Molecular weight from the volume of correlation [3]. |
| 20 |
|
| 21 |
There are two additional methods supported in the ATSAS software, which RAW |
| 22 |
will show if ATSAS is installed: |
| 23 |
|
| 24 |
#. Molecular weight estimation by comparison of scattering to known structures [4]. |
| 25 |
#. Molecular weight calculation by Bayesian inference from the other molecular |
| 26 |
weight methods [5]. |
| 27 |
|
| 28 |
All of these methods have advantages and disadvantages, and which to use and |
| 29 |
trust depends to a certain extent on the details of your data. Regardless, |
| 30 |
calculating molecular weight is important to verify the oligomeric state |
| 31 |
of your sample. |
| 32 |
|
| 33 |
|
| 34 |
Why do we do calculate molecular weight? |
| 35 |
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
| 36 |
|
| 37 |
SAXS molecular weight calculations are not terribly accurate, a usual rule of thumb |
| 38 |
is ~10% uncertainty (or more). For this reasons, SAXS should not be used to |
| 39 |
determine the molecular weight of your sample, something like multi-angle |
| 40 |
light scattering (MALS) is going to be much more accurate. The main reason |
| 41 |
to calculate molecular weight from SAXS data is to determine the oligomeric |
| 42 |
state of the protein in solution. Particularly if you're studying a system |
| 43 |
that can form homodimers or higher order oligomers, the SAXS molecular |
| 44 |
weight calculations should be accurate enough to tell what oligomeric state |
| 45 |
you're measuring. For this reason, the molecular weight is an important diagnostic |
| 46 |
for verifying that what you think is in solution is actually what's in solution. |
| 47 |
|
| 48 |
|
| 49 |
How do we do calculate molecular weight? |
| 50 |
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
| 51 |
|
| 52 |
Because there are so many methods for calculating molecular weight from SAXS data, |
| 53 |
we will only present a short summary of the methods listed above here. |
| 54 |
Generally speaking the methods can be broken up into two categories: |
| 55 |
concentration dependent and concentration independent. The concentration |
| 56 |
dependent methods require knowing the concentration of the sample in the SAXS cell, |
| 57 |
which often means they are incompatible with SEC-SAXS measurements. The concentration |
| 58 |
independent methods do not require knowledge of the concentration, and so are |
| 59 |
useful for SEC-SAXS measurements. |
| 60 |
|
| 61 |
Molecular weight from absolute scaled I(0) |
| 62 |
******************************************** |
| 63 |
|
| 64 |
SAXS data is often put on an absolute scale, so that the intensity values have |
| 65 |
units (by convention of cm\ :sup:`-1` ). This means that the I(0) value can be |
| 66 |
directly related to the number of electrons in the sample and the concentration |
| 67 |
and contrast of the sample [6]. If the concentration and contrast of the sample |
| 68 |
is known, the total number of electrons can be calculated, and then that can be |
| 69 |
used to determine the molecular weight of the sample. The accuracy of this |
| 70 |
approach can be improved by measuring or calculating (for example, using `MULCh |
| 71 |
<http://smb-research.smb.usyd.edu.au/NCVWeb/>`_) the protein partial specific |
| 72 |
volume. If necessary, adjusting the electron densities of the sample and buffer |
| 73 |
will also improve the accuracy. Molecular weight is calculated as: |
| 74 |
|
| 75 |
.. math:: |
| 76 |
|
| 77 |
MW = \left(\frac{N_A I(0)}{c \Delta \rho^2_M}\right) |
| 78 |
|
| 79 |
where *MW* is the molecular weight, *c* is the concentration, :math:`N_A` |
| 80 |
is Avagadro's constant, and :math:`\Delta \rho_M` is the scattering contrast |
| 81 |
per mass. |
| 82 |
|
| 83 |
The accuracy of this method was assessed in [1] where they found it to be <~10% |
| 84 |
for most proteins. |
| 85 |
|
| 86 |
Molecular weight by comparing to a reference standard |
| 87 |
******************************************************** |
| 88 |
|
| 89 |
The scattering at zero angle, I(0) is proportional to the molecular weight of |
| 90 |
the macromolecule, and the concentration and contrast of the macromolecule in |
| 91 |
solution. If a reference sample of known molecular weight and concentration is |
| 92 |
measured, it can be used to calibrate the molecular weight of any other |
| 93 |
scattering profile with known concentration (assuming constant contrast between |
| 94 |
reference and sample, and a monodisperse sample) [1]. Molecular weight is |
| 95 |
calculated as: |
| 96 |
|
| 97 |
.. math:: |
| 98 |
|
| 99 |
MW_m = \left(\frac{I(0)_m}{c_m}\right)\left(\frac{MM_{st}}{\left(\tfrac{I(0)_{st}}{c_{st}}\right)}\right) |
| 100 |
|
| 101 |
where *MW* is the molecular weight, *c* is the concentration, and the subscripts |
| 102 |
*m* and *st* designate quantities from the macromolecule of interest and the standard |
| 103 |
respectively. |
| 104 |
|
| 105 |
Molecular weight from Porod volume |
| 106 |
**************************************** |
| 107 |
|
| 108 |
The Porod volume is nominally the excluded volume of the macromolecule in |
| 109 |
solution. It can be calculated directly from the scattering profile, first |
| 110 |
by calculating the Porod invariant: |
| 111 |
|
| 112 |
.. math:: |
| 113 |
|
| 114 |
Q_p = \int^\infty_0 q^2 I(q) dq |
| 115 |
|
| 116 |
where :math:`Q_p` is the Porod invariant. The Porod volume is then calculated as: |
| 117 |
|
| 118 |
.. math:: |
| 119 |
|
| 120 |
V_p = \frac{2\pi^2 I(0)}{Q_p} |
| 121 |
|
| 122 |
where *I(0)* is the scattering at zero angle. Once this volume is determined, |
| 123 |
the molecular weight is determined by simply multiplying by the macromolecule's |
| 124 |
density. |
| 125 |
|
| 126 |
Because :math:`Q_p` is determined by an integral from 0 to :math:`\infty`, the actual |
| 127 |
integration must be approximated in some way. There are various approaches, including |
| 128 |
the the 'Porod' and 'Qp' methods implemented in the ATSAS datmw program. RAW |
| 129 |
uses the SAXSMoW 2 approach described in [2], which applies a correction factor |
| 130 |
to the Porod volume based on the available range of data in the scattering profile. |
| 131 |
Then an average protein density is used to calculate the molecular weight. |
| 132 |
The accuracy of this method can be improved if a more accurate density |
| 133 |
for the macromolecule is known. |
| 134 |
|
| 135 |
In [2] they found a median uncertain of 12% for calculated molecular weight |
| 136 |
from globular proteins. By contrast, in [5] they found using a large test set |
| 137 |
of simulated curves that the median molecular weight of this method was 3% high, |
| 138 |
and the median absolute deviation from this method was 5%. The uncertainty was higher |
| 139 |
for data with simulated noise. It seems reasonable to say the uncertainty in |
| 140 |
molecular weight is ~10% for most systems, though there are outliers. |
| 141 |
|
| 142 |
Molecular weight from volume of correlation |
| 143 |
************************************************** |
| 144 |
|
| 145 |
In [3] they defined the volume of correlation as |
| 146 |
|
| 147 |
.. math:: |
| 148 |
|
| 149 |
V_c = \frac{I(0)}{\int_0^{\infty}qI(q)dq} |
| 150 |
|
| 151 |
They empirically found the ratio :math:`V_c^2/R_g` is logarithmically |
| 152 |
proportional to molecular weight, with the following formula: |
| 153 |
|
| 154 |
.. math:: |
| 155 |
|
| 156 |
MW = \left(\frac{V_c^2/R_g}{c}\right)^k |
| 157 |
|
| 158 |
where *c* and *k* are empirically determined constants via fitting results from |
| 159 |
theoretical scattering profiles. They found different constants for proteins |
| 160 |
and RNA. For proteins, :math:`c=0.1231` and :math:`k=1` while for RNA |
| 161 |
:math:`c=0.00934` and :math:`k=0.808` (note: *c* and *k* are defined slightly |
| 162 |
differently in the original paper). |
| 163 |
|
| 164 |
In [3] they found a molecular weight uncertainty of ~5% from theoretical |
| 165 |
profiles and ~10% from experimental profiles. By contrast, in [5] they found |
| 166 |
using a large test set of simulated curves that the median molecular weight |
| 167 |
of this method was 2% low, and the median absolute deviation from this method was 7%. |
| 168 |
The uncertainty was higher for data with simulated noise. It seems reasonable to say |
| 169 |
the uncertainty in molecular weight from this is ~10% for most systems, though there are outliers. |
| 170 |
|
| 171 |
Molecular weight by comparison to known structures |
| 172 |
***************************************************** |
| 173 |
|
| 174 |
In [4] they describe a machine learning method that categorizes SAXS data into |
| 175 |
shape categories based on comparison with a catalog of known structures from the PDB. |
| 176 |
By finding the nearest structures in shape and size (also the name of the method: |
| 177 |
Shape&Size), they can obtain estimates for the molecular weight of the sample. |
| 178 |
|
| 179 |
In [4] they found that, for the theoretical scattering profiles used for testing, |
| 180 |
the method calculated molecular weights within 10% of the expected value for 90% |
| 181 |
test data. In [5] they found that for the test dataset the median molecular |
| 182 |
weight was correct and the median absolute deviation was 4%. Again, it seems |
| 183 |
reasonable to say that the uncertainty in molecular weight from this method is |
| 184 |
~10% for most systems, though there are outliers. |
| 185 |
|
| 186 |
Molecular weight from Bayesian inference |
| 187 |
******************************************* |
| 188 |
In [5] they describe a method for calculating a molecular weight using Bayesian |
| 189 |
inference with the molecular weight calculations from the Porod volume, volume of |
| 190 |
correlation, and comparison to known structures methods as the evidence. |
| 191 |
Essentially, it takes a large test dataset of theoretical scattering profiles, |
| 192 |
calculates the molecular weight for each using each method, then creates a |
| 193 |
probability distribution for each method that describes the probability of obtaining |
| 194 |
a particular calculated molecular weight given the true molecular weight. These |
| 195 |
probabilities are combined across all the methods, and the most likely molecular |
| 196 |
weight is thus estimated. |
| 197 |
|
| 198 |
They found that for the theoretical scattering profiles used, the median molecular |
| 199 |
weight from this method was accurate and the median absolute deviation was 4%. |
| 200 |
Overall, they reported that it was more accurate than any individual method. |
| 201 |
It may be that the uncertainty in this method is usually closer to ~5% than 10% |
| 202 |
for the other methods. |
| 203 |
|
| 204 |
What are the strengths and weaknesses of different MW methods? |
| 205 |
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ |
| 206 |
|
| 207 |
Each method has distinct strengths and weaknesses, and tend to be better with |
| 208 |
certain types of data. Every method requires a good determination of I(0), |
| 209 |
and all of the concentration independent methods require R\ :sub:`g`, which generally |
| 210 |
means a good Guinier fit is required in all cases. Also, in [5] it is reported that |
| 211 |
all of the concentration independent methods struggle with flat and ring-shaped |
| 212 |
proteins. |
| 213 |
|
| 214 |
Molecular weight from absolute scaled I(0) |
| 215 |
******************************************** |
| 216 |
|
| 217 |
**Advantages:** |
| 218 |
|
| 219 |
* Can be highly accurate when all parameters are well known. |
| 220 |
* With correct parameters can be used for proteins or RNA/DNA. |
| 221 |
|
| 222 |
**Disadvantages:** |
| 223 |
|
| 224 |
* Requires accurate sample concentration. |
| 225 |
* Requires accurate absolute calibration. |
| 226 |
* Best when the scattering contrast of the macromolecule is well known. |
| 227 |
* Best when partial specific volume is well known. |
| 228 |
|
| 229 |
Molecular weight by comparing to a reference standard |
| 230 |
******************************************************** |
| 231 |
|
| 232 |
**Advantages:** |
| 233 |
|
| 234 |
* Can be highly accurate for similar standards and samples under the same conditions. |
| 235 |
* With correct standards can be used for proteins or RNA/DNA. |
| 236 |
|
| 237 |
**Disadvantages:** |
| 238 |
|
| 239 |
* Requires accurate sample concentration. |
| 240 |
* The reference standard should have the same scattering contrast as the sample |
| 241 |
(i.e.. is in a similar buffer). |
| 242 |
* The standard and sample should be similar shapes (i.e. the same partial specific |
| 243 |
volume). |
| 244 |
|
| 245 |
|
| 246 |
Molecular weight from Porod volume |
| 247 |
**************************************** |
| 248 |
|
| 249 |
Particular to the MoW method described in [2]. |
| 250 |
|
| 251 |
**Advantages:** |
| 252 |
|
| 253 |
* Accurate for most molecule shapes [5]. |
| 254 |
* More accurate than the volume of correlation method when the signal to noise |
| 255 |
level of the data is reasonable [5]. |
| 256 |
|
| 257 |
**Disadvantages:** |
| 258 |
|
| 259 |
* Should struggle when the macromolecule is flexible or extended in solution |
| 260 |
(though [5] found this was not always the case). |
| 261 |
* May need to have the protein density adjusted in some cases (default: 0.83 |
| 262 |
kDa/\ :sup:`3`) |
| 263 |
* Will fail if the macromolecule is not a protein. |
| 264 |
* Sensitive to subtraction errors. |
| 265 |
|
| 266 |
Molecular weight from volume of correlation |
| 267 |
************************************************** |
| 268 |
|
| 269 |
**Advantages:** |
| 270 |
|
| 271 |
* More accurate than other methods when signal to noise is low [5]. |
| 272 |
* More accurate than other methods when there are subtraction errors [5]. |
| 273 |
* Should be accurate for flexible or extended extended macromolecules [3] |
| 274 |
(though [5] found that was not always the case). |
| 275 |
* Works for both proteins and RNA/DNA. |
| 276 |
|
| 277 |
**Disadvantages:** |
| 278 |
|
| 279 |
* Less accurate than other methods for high signal to noise data [5]. |
| 280 |
* Less accurate than the Porod volume MoW method for extended macromolecules [5]. |
| 281 |
* Large uncertainty for macromolecules less than ~15-20 kDa (based on experience, |
| 282 |
and the fact that the empirical coefficients were generated from size ranges |
| 283 |
20 kDa and larger). |
| 284 |
* Doesn't work for protein nucleic acid complexes. |
| 285 |
* The integral of *qI(q)* needs to have converged (see figure below). |
| 286 |
|
| 287 |
.. figure:: images/vc_convergence.png |
| 288 |
:target: ../_images/vc_convergence.png |
| 289 |
:width: 600 px |
| 290 |
|
| 291 |
Both plots show the integral of *qI(q)* as a function of *q*. The plot on |
| 292 |
the left shows data where the integrated value has converged, i.e. it is essentially |
| 293 |
unchanging at high q as q increases. The plot on the right shows data where |
| 294 |
the integrated value has not converged, i.e. it is increasing at high q |
| 295 |
as q increases. The data on the right will not give an accurate molecular weight |
| 296 |
by the volume of correlation method. |
| 297 |
|
| 298 |
Molecular weight by comparison to known structures |
| 299 |
***************************************************** |
| 300 |
|
| 301 |
**Advantages:** |
| 302 |
|
| 303 |
* Most accurate individual concentration independent method except |
| 304 |
for low signal to noise data [5]. |
| 305 |
* Relatively accurate when there are subtraction errors. |
| 306 |
|
| 307 |
**Disadvantages:** |
| 308 |
|
| 309 |
* Provides no result for flexible systems. |
| 310 |
* Only works for proteins. |
| 311 |
|
| 312 |
Molecular weight from Bayesian inference |
| 313 |
******************************************* |
| 314 |
|
| 315 |
**Advantages:** |
| 316 |
|
| 317 |
* More accurate than individual concentration independent methods in most |
| 318 |
cases [5]. |
| 319 |
|
| 320 |
**Disadvantages:** |
| 321 |
|
| 322 |
* Struggles with significant subtraction errors. |
| 323 |
* Only works for proteins. |
| 324 |
|
| 325 |
FAQ |
| 326 |
^^^^^^^^ |
| 327 |
|
| 328 |
I don't get the expected molecular weight from my SAXS data, what do I do? |
| 329 |
***************************************************************************** |
| 330 |
|
| 331 |
Molecular weight from good SAXS data has relatively large uncertainties (often |
| 332 |
~10%), and for low signal to noise data can be significantly worse. What you |
| 333 |
need to do in the case where it's wrong depends on what you're trying to determine. |
| 334 |
|
| 335 |
If you know your sample is stable in solution (not prone to aggregation/degradation), |
| 336 |
or you have evidence it was all in one state (such as elution in a single sharp |
| 337 |
peak in a SEC-SAXS experiment), if your MW is a bit off that's okay. In this |
| 338 |
case you're just trying to determine the oligomeric state of the sample. If you |
| 339 |
can clearly make the distinction, then you're fine. If not, you need to measure |
| 340 |
the molecular weight with a different method. |
| 341 |
|
| 342 |
If your sample is unstable in solution (prone to aggregation/degradation), |
| 343 |
you need to measure the molecular weight of your sample with another method. |
| 344 |
Good methods include multi-angle light scattering (MALS) or analytical |
| 345 |
ultra centrifugation (AUC). The best approach is to do a SEC-MALS-SAXS experiment, |
| 346 |
where MALS data is collected on the same elution as the SAXS data, removing any |
| 347 |
question about changes in the sample between the MALS and SAXS measurements. |
| 348 |
|
| 349 |
|
| 350 |
I need to determine if *<small molecule>* is bound to *<big molecule>*. Or I want to determine the binding stoichiometry. Can I do that with SAXS? |
| 351 |
******************************************************************************************************************************************************* |
| 352 |
|
| 353 |
It depends to a certain extent on the relative sizes of your molecules. However, |
| 354 |
if you have something that's small, say ~20 kDa, and something much larger, say |
| 355 |
~250 kDa, SAXS data is unlikely to be reliable enough to accurately determine the difference |
| 356 |
between bound and unbound (250 kDa or 270 kDa), or between 1:1 and 2:1 binding |
| 357 |
(270 kDa or 290 kDa). In this case you should pursue additional characterization |
| 358 |
of the molecular weight. The best approach is to do a SEC-MALS-SAXS experiment, |
| 359 |
where MALS data is collected on the same elution as the SAXS data. |
| 360 |
|
| 361 |
References |
| 362 |
^^^^^^^^^^^^ |
| 363 |
|
| 364 |
1. Mylonas, E. & Svergun, D. I. (2007). J. Appl. Crystallogr. 40, s245–s249. |
| 365 |
DOI: `10.1107/S002188980700252X <https://doi.org/10.1107/S002188980700252X>`_ |
| 366 |
|
| 367 |
2. Piiadov, V., de Araújo, E. A., Oliveira Neto, M., Craievich, A. F. & |
| 368 |
Polikarpov, I. (2018). Protein Sci. 2–22. DOI: `10.1002/pro.3528 |
| 369 |
<https://doi.org/10.1002/pro.3528>`_ |
| 370 |
|
| 371 |
3. Rambo, R. P. & Tainer, J. A. (2013). Nature. 496, 477–481. DOI: |
| 372 |
`10.1038/nature12070 <https://doi.org/10.1038/nature12070>`_ |
| 373 |
|
| 374 |
4. Franke, D., Jeffries, C. M. & Svergun, D. I. (2018). Biophys. J. 114, |
| 375 |
2485–2492. DOI: `10.1016/j.bpj.2018.04.018 |
| 376 |
<https://doi.org/10.1016/j.bpj.2018.04.018>`_ |
| 377 |
|
| 378 |
5. Hajizadeh, N. R., Franke, D., Jeffries, C. M. & Svergun, D. I. (2018). |
| 379 |
Sci. Rep. 8, 7204. DOI: `10.1038/s41598-018-25355-2 |
| 380 |
<https://doi.org/10.1038/s41598-018-25355-2>`_ |
| 381 |
|
| 382 |
6. Orthaber, D., Bergmann, A. & Glatter, O. (2000). J. Appl. Crystallogr. |
| 383 |
33, 218–225. DOI: `10.1107/S0021889899015216 |
| 384 |
<https://doi.org/10.1107/S0021889899015216>`_ |
| 385 |
|
| 386 |
.. |Rg| replace:: R\ :sub:`g` |
| 387 |
|
| 388 |
.. |qRg| replace:: qR\ :sub:`g` |
| 389 |
|