Help with Analyzing my data with EOM

[Nicholas-Kl]

Hello everyone. I am using ATSAS 4.1.1 locally and try to fit my SAXS data on a protein in 3 different nucleotide states, batch mode SAXS, different concentrations. So far I tried to be very thorough with reading in this forum and analyzing my data. I am new to the field and have the feeling I need guidance as I am reaching my ends regarding oligomer fit and EOM fit.
On the system: Flexible. Domain-linker (17 AS)-domain (2 subdomains). Regarding the literature, it’s likely that depending on nucleotide state the protein builds Monomers, Dimers, Trimers. At the same time the monomers may be docked or undocked (linker and 2nd domain detach from domain one and are independent). Regarding that, EOM may be the best choice followed by oligomer as a close 2nd.
Regarding sample quality an example:

autorg quality score in most cases >0.8: But St.dev. is often large (e.g. 5+- 0.65 nm^-1 ,which may mean that the different by autorg fitted regions give different Rg?). Aggr. Scores are distributed but in most cases 0 to 0.05,seldomly clearly aggregated as around 0.08.

I ran Ranch with default settings but different assignments + structures (Monomer, Dimer, pool of both) many times on chosen nucleotide states and concentrations and also zero extrapolated curves. FFmaker and Gajoe also always with default settings. I tried s <=0.5 and s <= 0.2 (Angstrom^-1).
The most likely best set of monomers (30000 structures, 3 ranch runs) gives the green curve in the picture. This set plus 10000 homodimers gives the pink one. Shown is a Zero extrapolated data set. Both chi squares are around 35. The best I get with single concentrations is chi square of 11, even with s<=0.2 Angstrom^-1.

In summary I am at my limit and want to exclude that my data is too bad/ I overlook something obvious.
I apologize for the long thread and hope you can help me.

Best,

Nicho

[franke-biosaxs]

Hi again. More information is better than too little, so please, be as expressive as you can.

Having said that: personally I would try to avoid zero-extrapolated data. That is not exactly witchcraft, but certainly vague. In my book at least.
What do the fits look like with the actual data? Any better? Instead of gajoe, you may try nnlsjoe using the sam input data from ffmaker; nnlsjoe does not do a random selection, but attempts to get the analytical best subset of models that fit the data. Never tried with 30k inputs, so it might take either very long, or not work at all.

Otherwise: it might be something else completely in solution? Does the calculated MW match the expectation?

[franke-biosaxs]

I can not comment on any of the biological stuff you are looking at, but I can comment on chi-square values.

First and foremost, your assumption is that the error estimates are as correct as can be. If not, the chi-square values are useless. Assuming that this is recently collected data, the errors are, likely, ok (not so likely for extrapolated data). In such a case, the only chi-square value that indicates a fit is 1.028, everything else is, statistically, significantly "not similar" (aka "different").

Anyone saying that a chi-square of, say, 2.0 indicates a better fit than a chi-square of 5.0 is correct in the sense that the differences are less significant, but they are still very significant and not "a better fit". Neither fits. Note that my statistical interpretation is very literal here, I am a mathematician, not a biologist.

Short summary how to check your errors:

• take five consecutive frames collected within a short period of time (seconds, not minutes; assumption, no systematic differences in this short a time frame, only random, do not use samples as they might be affected by the beam)
• calculate chi-square for all pair-wise comparisons (in primus)
• if all chi-square values are, roughly, in the range 0.9 to 1.1, you are likely good (repeat with five other frames to be more sure)
• if the starting errors are ok, make sure that in your processing you never re-use any data, e.g. subtract the same buffer from multiple samples, then average the subtractions that use the same buffer, as this introduces correlations, which in turn bias error estimates

If you have specific questions on how to set up ranch, I may be able to help with that, though.

[Nicholas-Kl]

Hello, thank you very much for considering my post and for your answer. I tried to follow your instructions on loading frames. I wanted to compare them using Primus Compare. I am confused about which chi-square value I should use for the comparison. In the statistics tab, where the cormap, red. Chi^2, and A-D test are depicted, is it the value in the row "chi^2" and the column "tval"? It seems to me it is rather the tval and not the chi^2, right? So does Compare not give the chi^2 value? All the statistics are green, though. I am sorry for the beginner question, but of all the software in the ATSAS package, Compare made the most sense to me, and I couldn’t find a description of the statistics tab, although it’s the same for all the Primus options. Is Compare the way to go?

Regarding the help with Ranch, I would gladly make use of it! From the chi^2 comparison of ATP, ADP, and APO, it seems that my pool is relatively well suited to describe the scattering curves of the ATP state but not the other states. My idea is that the dimers are not diverse enough or that I maybe need some trimers and so on. I made the pool with Ranch by using three variants of assignment file 1 (monomer) and one variant of assignment 2 (homodimer).

assignment.txt
assignment_homo_oligo_2_domains2.txt

I used an AlphaFold structure for both. Unfortunately, I could not use the “style” of assignment file 1 to build a homodimer with it. This would be two monomers, each with 3 domains. No matter what I fix or unfix, Ranch either can’t compute it (visible by it running but no files showing up), or chain A is not connected to another chain. Therefore, I built the dimer by leaving the subdomains of domain 2 together. Even then, only two arrangements work (assignment file 2 as well as reversing the fixation, which only gives new information if the dimers are asymmetric). For myself, I don’t see a way to make the dimer pool more diverse or even compute trimers. With your experience, do you have any ideas or workarounds? Maybe a parameter I can change? As always, thank you very much in advance.

Best,

Nicho

[franke-biosaxs]

The "t-val" is the test value, aka the chi-square value. In this screenshot it is 0.208 (old data, errors vastly overestimated).

For completeness: the cormap is a measure of similarity of the data that does not take the error estimates into account. And A-D is the Anderson-Darling test for standard-normal distribution of the residuals (experimental data - fit); also see the "Residual Distribution" tab.

On ranch: the first that comes to mind: I assume that all your domains are individual files? If so, put everything you want "fixed" into a single file (cat or copy-paste), and remove all "fixed" keywords from the assignment.txt - it will keep the combined model together as stated and attach any free additional domains. This has reportedly helped others.

[Nicholas-Kl]

Hello again :),

First of all, I checked the frames, and overall, I think the error estimates are reliable.

I now have tried to build trimers using various approaches, such as assembling three monomer structures or a combination of monomer and dimer structures. However, I still cannot get Ranch to work with a trimer. I also remain limited by the assignments for the dimer, and as a result, the chi^2 of the apo state fit is still above 30. No matter what I do, I cannot seem to get EOM to run for a certain part of my data. This is unfortunate, as I believe EOM would be the best way to analyze the data. I will recheck oligomer.

Until now, I have never truncated the scattering data in the low angle region. Do you think this is a viable option in my case? For example, would it make sense to cut the data (only two parts of the whole curve shown) below 0.008 or 0.01 (i.e., angles near the beamstop)?

I have noticed that the more I cut, the lower the chi^2 value gets. For example, if I cut the curve at 0.02, chi^2 drops from 33 to 11. While I realize this is consistent with underfitting—since trimming low-angle data also simplifies the data—I’m wondering if there is a commonly accepted range of s/q values used for such analyses. I came across a 2007 forum post mentioning omission of points before the Guinier range in certain cases.

Thank you very much for considering my problem.

Nicholas

[franke-biosaxs]

This is armchair analysis, but looking at your data, the first few (10?) data points seem indicative of a "larger species" being present. Whether those are meaningful trimers, tetramers, or larger oligomers, or simply aggregation I can not say, of course. It is unlikely that anybody would tell you off for dropping those points "due to aggregation". Careful Guinier analysis would be required to determine a reasonable range, though.

If doing so solves your problem, great. If you still want to build trimers (tetramers?), and have trouble with ranch, you can send me your sequence, atomic models and data, together with any configuration files, and I have a quick look (franke AT biosaxs DOT com). Might be an issue in ranch, after all.