There are a number of reasons why your Rfree might not be not dropping. Though the missing 50 residues could indeed introduce instability and variability into your protein molecules, the fact that you got diffracting crystals makes me think this is probably not the main problem. More often than not, high Rfree values stem from wrong indexing due to pseudosymmetry and translational elements or because your crystals are twinned. Resolution of the data also affects the Rfree, if your resolution is close to or worse than 3 Ang, then your Rfree can certainly stay close to 25-30%.  Could you give us a bit more information? What are your current Rfactor, Rmerge, I/sig, completeness, multiplicity and resolution? What are your Ramachandran plot statistics (maybe the structure still needs a lot of refinement)? Have you already build any visible solvent molecules?

Thanks Antonio! Here is a list of the parameters:

R-factor=0.3020

R-free=0.3320

R-merge=0.207

I/sig=4.1

Completeness=99.6

multiplicity=5.0

resolution=1.58 A

Ramachandran outlliers: 0.3% (goal 0.2%)

Ramachandran favored: 95.5% (98%)

Phenix refinement has automatically added water molecules to the models.

difraction data was collected to 720 frames 360 degree. mosflm indexing gives solutions in p222 p2, the final model was searched in p21 with 500 frames.  Reducing the resolution to 2.2 A doesn't improve the Rfree much. 

So you guess the miss 50 amino acids shouldn't affect the R values too much? The Rmerge also appears a bit high, I wonder what could the be problem.

an additional question, can I see the electron density for missing 50 amino acids in the map? my Phaser MR search model doesn't contain the structure for the extra amino acids. Thanks!

Your resolution is very high, which means the crystal you shot was very stable and this generally indicates that there's little variation between the individual protein molecules in the crystal. If they were slightly disordered because of the extra 50 amino acids, the resolution would have been worse. At the resolution you quote, I would expect an Rfactor around 0.13 - 0.19 and an Rfree around 0.18 - 0.24 for a finished and well refined structure.

Your Rmerge is either very low or very high depending on which resolution shell you're quoting. Please provide the overall Rmerge (this should usually be lower than 0.1 for an acceptable data set) and the outer shell Rmerge (this should fall between  0.6 and 0.4, if it is much higher or lower than these numbers you should reprocess your data at a lower or higher resolution, respectively). The exact resolution you should eventually use depends on a number of different factors, but the Rmerge is a good starting point to decide where to cut your data at the start of refinement.

Your Ramachandran stats indicate fairly good refinement and overall build, which means the problem isn't a badly build structure. Regarding the 50 extra residues, if you have build everything there is to see and there is no obvious extra positive density at the terminus where the extra residues lie, then they are disordered. If they were ordered, you would have seen at least some partial positive density for them right after molecular replacement or at the latest towards the end of the build.

I think it looks like there is a problem with the data set, it was either indexed wrongly because of unusual symmetry elements or your crystal is twinned. If you use Phenix, check your data with Xtriage. The output will tell you if the data should be indexed in a different space group, or a different origin and if there are pseudosymmetry/translational elements or twinning.

Thanks Antonio!

The Rmerge I quoted is the overall value. While the InnerShell and OuterShell values are at 0.164 (47.15A - 8.66A) and 0.488 (1.61A - 1.58A) respectively.

Therefore, according to your judgement, there may be a problem with the dataset. Actually my supervisor suggests the poor R values may be due to anisotropy, in our case, the protein tends to crystallize very quickly into a thin sheet form, but we managed to get some good form crystals, the difraction was performed in cryo condition.

I try to re-indexing using mosflm, there are always a few warnings after integration, I have listed these warnings below, you can take a quick look at them. What we did is just ignoring these warnings and proceed to the scaling step. While these warning always occur with all the solutions we have, do you think there may be something we can do to modify mosflm settings to eliminate these warnings?

1. Negative mosaicity or beam divergence

Details:

the bombined mosaic spread and beam divergence has refined to a negative value

This can result from inaccurate cell parameters

it can also arise when images have split spots or the spots are not resolved in phi. In this case, estimate a value of mosaicity by checking the predictions and fix the mosaicity at this value

Hints:

Check for split spots or inaccurate cell parameters

Notes: refined mosaic spread was reset to 0.05.

2. Refined mosaic spread too small

Details:

Refined mosaic spread is less than 0.25*input value

This can result from inaccurate cell parameters

It can also arise when images have split spots or the spot are not resolved in phi.

In this case, estimate a value of mosaicity by checking the predictions and fix the mosaicity at this value.

Hints:

Fix mosaic spread at visually estimated value.

3. Crystal slippage excessive

Details:

There is significant slippage between successive images.

Maximum slippage is 0.01 degrees for image 490 which is more than a quarter of the sum of the beam divergence and mosaic spread. There may be serious errors in the intensities of partially recoreded reflections.

Check the images, indexing solution and cell parameters.

4. Large weighted residual

Details:

The weighted residual is rather large (it should be unity , actual maximum value 4.27).

This is usually due to poor cell parameters.

Checking indexing solution, cell and pssibly try post refinement

Hints,

Check or refine cell parameters

BTW, I will use Xtriage to check my dataset and post the results later so that other people could also learn from our case. Thanks!

I went back to check the log file of ccp4/aimless and found this:

TWINNING SUMMARY

Twinning fraction from H-test: 0.35

L-statistic from L-Test: 0.48

Relation between L statistics and twinning fraction:

Twinning fraction = 0.000 L statistics = 0.500:

Twinning fraction = 0.100 L statistics = 0.440:

Twinning fraction = 0.500 L statistics = 0.375:

It is highly probable that your crystal is TWINNED.

Please use twin refinement after your model is almost completed and R-free is below 40%.

When I check my dataset with Phenix/Xtriage:

Possible pathologies

Twinning:

0 possible merohedral twin operators and 1 pseudo-merohedral twin operator found.

The results of the L-test indicate that the intensity statistics behave as expected. No twinning is suspected. The symmetry of the lattice and intensity however suggests that the input space group is too low. See the relevant sections of the log file for more details on your choice of space groups. As the symmetry is suspected to be incorrect, it is advisable to reconsider data processing.

The 1 pseudo-merohedral twin operator found is:

Type  Axis      R-metric (%)  delta (lc-page)   delta (Lcbedev)  Twin law

PM     -fold     0.136                 0.120                 0.002                     h,-k,-l 

Statistics depending on twin laws

Operator    Type      R-abs obs     R-abs calc       Britton-alpha   H-alpha    ML-alpha

h,-k,-l           PM           0.132            0.382                0.380                 0.377        0.392

The estimated twin fraction is around 0.380

Now I am doing phenix refine again by applying the twin law to check the R values.

In regards to the potential higher space group (p222) we suspect it is not true. May be the high R value is due to twining in the crystal? If this is the case, how do I reduce the Rmerge, Rfactor and Rfree? Thanks!

BTW, I look online, some guys say high Rmerge is acceptable, because it is dependant on a lot factors while Rpim may be more reliable. In our case Rpim overall stands at 0.156. Thanks!

A high Rmerge is indeed acceptable but, as I said, the values I quoted are a good starting point for choosing a resolution shell to work at. As you mention, the Rpim is sometimes a better value as it is not influenced by excessively high multiplicity, but again, I don't think your Rmerge values are the problem.

You had slippage during your collection, so I would try to process your first set of images (1-489) and then your second set of images (490 or 491 to 720) separately. As your space group is either orthorhombic or monoclinic, you will have enough data for a complete data set from the first set of images and it might give you a clearer result of the actual space group if it isn't affected by the deviation in values caused by the slippage. Also, you might get better statistics and maps from this smaller set, particularly if your crystal suffered radiation damage during the collection. You can also merge both sets of images after you've indexed them separately and then scale them together, which would side step the slippage problem. You can then compare the result of:

If you still get the same twinning message from 3. in 1. and 2., then it might be worthwhile refining your structure with the appropriate twinning operators to see if you get a marked improvement of your R factors.

Finally, if the processing of 1. and 2. still has problems with finding the correct space group, then you should re-process 2. in P1 (use 2. and not 1. for this as you will require 360o  for this space group) and feed it into ZANUDA from ccp4, which will try to find the correct space group for you.

Hi Antonio, we managed to improve the R-factor and R-free value significantly by applying the twin operator: h,-k,-l in Phenix refinement.

Firstly, with the resolution at 1.58A, the addition of the twin law improves the R-factor and R-free to 0.2619/0.2996 respectively, compared to 0.3020/0.3320 without the twin law.

Secondly, we did the same test with a 2.2A cuttoff, in this case the application of the twin law reduced the R-factor and R-free to 0.2504/0.2642, compared to 0.2975/0.3239 without twin law.

All these tests are based on the first 500 frames by indexing at P21. So according to your suggestion, there might be a slippage after frame490, right? Judging from the mosflm statement, how serious you reckon the slippage problem is? I will try to process 1-489 and 490-720 separately to see what happens.

The thing is that we have also tested the possibility with P222, and we only use the first 400 frames for this lattice group. aimless and Xtriage twining analyses showed that twinning is not possible for this solution. So we reckon P222 may not be a choice for our crystal (the statistic for P222 is even worse, R-free generally above 0.4).

At the moment, I guess the next step would be to determine the optimal  frame number and resolution cuttoff?

BTW, it seems that twinning is a problem with our case, do you think it could also be the reason we have high R-merge value?

one point I am worried about is that, during the Phaser MR step, I used a chain from another refined model as the search model, I wonder this is acceptable or not. Thanks!

It looks like I was right and your crystal is indeed twinned. At least that's one problem solved. LOL These days twinning is not such a huge deal anymore and many programmes can circumvent the problem to the extent that the resulting data are good enough for publication and deposition.

I would definitely try to find the optimal range of frames for the best statistics. Your space group (P21) requires a minimum of 180o for good completeness, since you collected with 0.5o oscillation you will need a minimum of 360 images for an acceptable dataset. So, I would compare the result of processing 360, 425 and 489 images to see which one gives the best statistic AND best maps. If 489 images is better than 360, it's quite possible that your crystal did not suffer radiation damage and therefore adding even more images might improve your data even further. As I mentioned last time, in this case you will need to index the images after the slippage separately and then merge them with the data set containing images 1-489 before you scale them together as a single data set.

You can use whatever you like as a MR search model, anything that gives you a solution is ok. The immediate result of MR will be fairly biased by the search model (but even then it's only ever noticeable when the search model is very different from the actual result), but after a fair number of refinement cycles this bias will disappear.

Yes, the result looks very promising. Thanks Antonio!

I guess I still have to bring up more questions to you, there are still some weird observations.  As you might have noticed. The phenix refine with the resolution at 1.58A is still far from satisfactory, now the R-factor and R-free values are like this:

1.58 A =0.3020/0.3320

1.58 A detwin=0.2619/0.2996.

2.2 A = 0.2975/0.3239

2.2 A detwin=0.2504/0.2642

does it make sense to you that the R values varied so dramatically with different resolutions of the same data set? And also with twining, shouldn't there be a true spacegroup and an apparent spacegroup? Since mosflm indexing solution and also Xtriage always suggest a higher solution at p222, and MR at p222 gives a few solutions as well, do you have any idea how this could be true? 

I have checked the ccp4 webpage lookup table, and it seems that twinning is not even possible for p222, and no twin operator could operate between p222 and p21.

One last point is about the Rmerge value which is still pretty high for publication, although it could be affected by multiplicity and resolution cuttoff etc, but as you mentioned, more frames with a good dataset shouldn't affect this statistic too much, right? Only in except of crystal degradation, which in our case may not be that likely since the diffraction was performed in cryo condition with cryo protectant?

I will play around with frame number and resolution cuttoff to see how much we can improve the Rmerge, and will post later about the statistics just in case other people may also be interested. Thanks a lot for your instruction.

Refining data using twin laws generally leads to a small improvement in the stats even when the data set does not suffer from twinning. If the crystal suffers from a high percentage of twinning, then the improvement should be significant, as is the case in your 2.2 A trial. However, I think you really need to use the data set from the images collected before the slippage instead of all the images to compare the R factors at different resolutions before making a decision on what's best.

I'm pretty sure that the distortion in the data created by the slip is what is producing all the suboptimal stats (though PHENIX should not detect twinning just because there was some slippage, so I'm pretty sure that you also have twinning).

Regarding radiation damage, using a cryo protectant and a liquid nitrogen cryo stream significantly reduces radiation damage, but they do not stop it. Generally, if you collect around 360o while trying to get the highest resolution possible, you will damage your crystal. Unless you already cut back the maximum resolution from 1.0 A to 1.6 A using high attenuation and the fastest frame rate possible, I would expect that the last part of your 360o dataset is suffering from radiation damage. You can easily test this by seeing if the inclusion of the data makes the statistics worse. When comparing the maps you can see that certain amino acids like Glu and Asp will quickly lose one of their oxygen atoms because of the damage and they will have a bit of negative density towards the end of the side chain, or at least it looks like there isn't any density for both the oxygen atoms. In general there will be more noise (small bits of random positive and negative density that don't correspond to actual density) in a damaged crystal. The easiest way to detect radiation damage is to look at the resolution of the images, as the resolution decreases as radiation damage increases. For example: If you compare how far the spots go in your first image (0o) and in your last image (360o = 0o), they should look the same if there is no radiation damage. If the spots in the last image don't go as far as in the first image (e.g. 1.8 A or worse in the last image and 1.6 A in the first image), then the crystal has suffered radiation damage.

Thanks Antonio,

I checked the diffraction images for potential radiation, the later images indeed showed smaller resolution range, but still good.

We found something interesting. It seems there is something tricky when indexing using mosflm. I keept repeating the indexing refinement and integration steps in mosflm with the same solution, and the results varied greatly from time to time. As posted before, most often, mosflm gives a few warnings (red and yellow), and it just happen one time when I repeat the data processing, the previous warnings:

1. Negative mosaicity or beam divergence

2. Refined mosaic spread too small

3. Crystal slippage excessive

4. Large weighted residual

all disappeared, two significant differences in this case:

there is less much less full reflections detected for each image, and the mosaicity refined to 0.88 instead of the negative value. And most importantly the Rmerge drops to 0.091 with 360 frames and 99.5% completeness, 3.5 multiplicity, 5.6 I/sig(I). Subsequent refinement with and without the twin law resulted in Rfactor/Rfree at 0.2371/0.2605 and 0.3242/0.3490 respectively with the highest resolution at 1.58A. This means twining is indeed true in our case, but why mosflm tends to be so "unstable", the indexing solutions were all based on frames 1 and 180, but the cell parameter refinement results varied subtlly by just repeating it. Thanks!

I tend to use at least 4 (sometimes as many as 8) images for indexing and I play around with this, choosing images from just 0o to 90o, or 0o to 180o or even 360o. During the cell refinement step I do the same, and this will often give me a clue to where the data changes (i.e. if there is a slip in the crystal), as the cell parameter refinement goes a bit wonky at those points. After you have integrated the images, the display on iMOSFLM will show you if there is a slip in the data. Also. looking at the scaling graphs that AIMLESS (or its older sister programme SCALA) spit out will tell you where there is data you should probably cut out and not include in you finished integrated and scaled data set (look at the Rmerge graph for each image and delete individual images or blocks of images that have a much higher Rmerge than the other images). I like mosflm, but I have to say that often XDS does a much better job when processing less-than-ideal data.

Thanks Antonio! mosflm automatically picks two images 90 degree apart for indexing, I am playing around with mosflm with more indexing images. I got another dataset which looks a bit worse than the previous one, the diffraction images looks a bit dirty but still with some good frames, with this dataset, mosflm integration always gets disrupted giving a warning "Mosflm has encountered a fatal error in the subrountine 'MASKIT' ". So far I haven't even been able to integrate the images, actually this same error all occurs with my last dataset, but not so often, I wonder what's going on here, it seems to do with the diffraction quality. Anyway, I will see what I can do. BTW, I will check XDS. Thanks!

Hi Yong,

I did not find the structure in PDB deposited by you, which would have this resolution. So, it looks like you did not deposit this structure.

Sorry, I found your question too late, but anyway...

You mentioned in your messages that you could merge your data in P222.

But you think that real space group is P21. That would mean that your angles are close to 90. It this case the indexing may choose the wrong beta angle. And you are actually merging your data with two fold pseudo symmetry. That would explain your overall high R merge. The best way to find this would be to process the data in P1, and check what kind of R merge you will get. If your R merge in P1 still will be high, that would mean bad data. You had multiple crystal patterns, in some orientation spots matched in other orientation they split.

But if your R merge in P1 will be low (for this resolution you can expect below 10%). Than you can try to merge your data in different setting for beta angle. You should get three data sets with three R merges. Two of them should be high (pseudo two fold) one should be low. And the data with low R merge is your P21 set.

Your crystals could be P1. If you have multiple domains in your protein and several chains, it may happened that one of the domains has slightly different conformation, but the rest very well matches P21 pattern.

You should try P1.

George.

Slippage of crystals during data collection, mild twinning, flexible residues which are not defined by the electron density and radiation damage will contribute to poor refinement statistics including a large difference between R-factor and Rfree factors.

Well, you can try these

1) Perform, a L-test and see if your crystal is twinned. Often it happens that after using the twin law (if there is twining) your stats improve.

2) See the Phenix.refine table and it should show you the Rfree at different res bins. You can cut out the resolution at that res. Alternatively try paired refinement.

3) last one, if nothing works try to reprocess the data, it might be in the wrong spacegroup. The DIALS reciprocal lattice viewer is one such which will tell you how your diffraction pattern looks like. Its a great tool to see and remove outliers.

This happened with me as well. The problematic crystal diffracted to 2.6A and had twining, nothing of the above worked out and R/Rfree was 0.30/0.33. Fortunately I got better untwined crystal in a different condition and processed it at 1.9A.

Twinning can also be get ridden during data collection. Use microfocus and hit the corner of untwinned region (sometimes you are lucky, if you are able to visualize twin crystals)

good luck!!

Additionally, use higher restraints to control the rising values of Rfree to prevent the over-fitting of the model.