I am little bothered that my structure refined at 3.2 Angstrom, have a difference of 10 % without and with solvent contents. The difference is not changing. completeness is 97.6 % at 3.2 Angstrom. This completeness I got when I merged the two data set one with 360 images and other with 56 images (as 56 images have nice reflection spots but with 360 little overlapped at some places). Redundancy is 2.36 and mosaicity is 0.3. and overall Figure of Merit is 0.777. only 2.34 % outliers in Ramachandran plot. At 2.8 Angstrom almost same statistics has been shown but have to compromise with completeness (i.e = 92%). Has my structure solved in a correct way ?
Over-fitting of poor data is the most obvious explanation, as has already been mentioned.
I don't know how you collected your data or how much experience you have, but here are a few points that might help you if you're new to X-ray crystallography.
If you can't improve your crystals, then maybe you can improve your data quality with a better collection strategy. Try to run a strategy prediction programme with a few test images BEFORE you randomly collect the minimum number of degrees need for a particular space group (many synchrotron beamlines will do this automatically nowadays but you can also use iMOSFLM and HKL2000 to do this). These programmes will tell you what wedge of data to collect to obtain the highest completeness with the lowest number of images. This is important if your crystals decay quickly in the X-ray beam.
You mention you have merged two data sets: one with 360 images and one with 56. They should obviously be as isomorphous as possible, if they deviate too much, this will affected your data quality.
Also, this seems a fairly low number of images (unless you are working with very high symmetry). How many degrees did you collect and what's your space group? If you can, try to collect more images by reducing the oscillation angle and either lowering the transmission or the collection time for each image. This way most spots will be measured more than once and this should improve your scaling and your overall data quality. Alternatively collecting more degrees than you need will also improve most of your stats (as long as the crystal isn't decaying of course), but "fine slicing" is better if you have many closely set spots (where one or more of your cell edges are very large or your data is highly mosaic).
Another thing to keep in mind is the need to work with the right resolution. Try not to be too greedy and collect data to a resolution that is beyond the point to which your crystal diffracts. If your diffraction is poor (smeary, closely set spots, etc...) and you collect a data set to 2.0A but the crystal only diffracts to 2.8A, then your low resolution data will be of poor quality as the spots may not resolve well and merge into each other close to the centre of the diffraction pattern. All these reflections will then be discarded by the integration/scaling programmes and drastically reduce your number of parameters. It would be better to collect data to a radius of 2.7 and even 3.2A (you will still collect data to at least 2.8A in the corners of the image) and make sure your low resolution spots near the centre are well resolved.
If your crystals diffract to lets say 2.5A but they start to die early on, with a quick drop in resolution to 3.5A or worse, then it might be better to reduce the transmission or collection time to the point where the crystal only diffracts to 3.0A and collect the data set that way. The crystal should decay a bit more slowly and give you more complete data in the higher resolution shells.
If your crystals are very isomorphous, then you might consider frying a few crystals with very high transmission to get the highest possible resolution out of them and then merge to data to get the necessary completeness in the higher resolution shells. I would also collect good low-resolution data at 3.5 to 4.0A from a crystal to merge with the high resolution data for the reasons outlined earlier. Remember that when you "fry" a crystal, you still need to tweak the transmission and time settings to make sure you don't get any overloads. Spots that are overloaded will be discarded during scaling and lead to an unnecessary loss of reflections/parameters that will your data worse.
I hope this will help you.
How many reflections do you have in your Rfree test set? According to the paper of Brunger (Methods in Enzymology, 277:366-396.) you need at least 500 reflections in the test set for it to have any meaning
It is clearly indicated the over-fitting of the building model as far as the data itself and process were fair.
Please use higher geometrical weight in the refinement. Recent default weight seemingly adopts for very good diffraction data set like 1.5 Å resolution or better.
The situation is dependent on the how to phase and how good the diffraction data. There are many pitfall if you phased MR using poor search model.
If the search model is excellent, you may retry the the exhaustive rigid body refinement of the search model structure, may decrease the much R values.
If the data set is questionable, it should be tried to collect better diffraction data set including the improving crystal quality again. One criteria is the Wilson's plot shape, and the plotted points bending horizontally at high resolution, such region is already no meaningful signals. The I/sigma(I) value is more than 2 at the highest shells.
The protein is stable, to obtain better crystals is better way to solve the situation. No protein structure better then the diffraction data quality.
good luck
Dear Prem,
in addition to what said by Masashi, if you used MR for phasing, please check your model, since, at that resolution, model bias might be an issue. Before running the search, please use a pruned model without ligands, solvent, mutated side chains (chainsaw in ccp4i does this for you). This should make you start from a structure with R/Rfree
Hello Prem,
As Masashi says, a gap of 10% between R and Rfree clearly indicates over-fitting. The suggestion to use tighter geometric restraints is a very good one, as is Adriana's suggestion to be very careful about model bias. In addition, I would look carefully at anything else you can do to restrain the numbers of parameters in your refinement. For instance, it may well be worth refining group temperature factors (eg. per residue rather than per atom) and/or TLS refinement. If you have a good MR search model and are using Refmac then using jelly-body retraints can be very helpful in the early stages of refinement.
You also mention that your statistics are almost as good when including data out to 2.8A. I would need to see the processing statistics to say for certain, but I would be inclined to use thos data in refinement - that's really quite a large number of extra reflections and will greatly improve your observation : parameter ratio, which may in itself help to improve your over-fitting problems.
Good luck!
Over-fitting of poor data is the most obvious explanation, as has already been mentioned.
I don't know how you collected your data or how much experience you have, but here are a few points that might help you if you're new to X-ray crystallography.
If you can't improve your crystals, then maybe you can improve your data quality with a better collection strategy. Try to run a strategy prediction programme with a few test images BEFORE you randomly collect the minimum number of degrees need for a particular space group (many synchrotron beamlines will do this automatically nowadays but you can also use iMOSFLM and HKL2000 to do this). These programmes will tell you what wedge of data to collect to obtain the highest completeness with the lowest number of images. This is important if your crystals decay quickly in the X-ray beam.
You mention you have merged two data sets: one with 360 images and one with 56. They should obviously be as isomorphous as possible, if they deviate too much, this will affected your data quality.
Also, this seems a fairly low number of images (unless you are working with very high symmetry). How many degrees did you collect and what's your space group? If you can, try to collect more images by reducing the oscillation angle and either lowering the transmission or the collection time for each image. This way most spots will be measured more than once and this should improve your scaling and your overall data quality. Alternatively collecting more degrees than you need will also improve most of your stats (as long as the crystal isn't decaying of course), but "fine slicing" is better if you have many closely set spots (where one or more of your cell edges are very large or your data is highly mosaic).
Another thing to keep in mind is the need to work with the right resolution. Try not to be too greedy and collect data to a resolution that is beyond the point to which your crystal diffracts. If your diffraction is poor (smeary, closely set spots, etc...) and you collect a data set to 2.0A but the crystal only diffracts to 2.8A, then your low resolution data will be of poor quality as the spots may not resolve well and merge into each other close to the centre of the diffraction pattern. All these reflections will then be discarded by the integration/scaling programmes and drastically reduce your number of parameters. It would be better to collect data to a radius of 2.7 and even 3.2A (you will still collect data to at least 2.8A in the corners of the image) and make sure your low resolution spots near the centre are well resolved.
If your crystals diffract to lets say 2.5A but they start to die early on, with a quick drop in resolution to 3.5A or worse, then it might be better to reduce the transmission or collection time to the point where the crystal only diffracts to 3.0A and collect the data set that way. The crystal should decay a bit more slowly and give you more complete data in the higher resolution shells.
If your crystals are very isomorphous, then you might consider frying a few crystals with very high transmission to get the highest possible resolution out of them and then merge to data to get the necessary completeness in the higher resolution shells. I would also collect good low-resolution data at 3.5 to 4.0A from a crystal to merge with the high resolution data for the reasons outlined earlier. Remember that when you "fry" a crystal, you still need to tweak the transmission and time settings to make sure you don't get any overloads. Spots that are overloaded will be discarded during scaling and lead to an unnecessary loss of reflections/parameters that will your data worse.
I hope this will help you.
Our suggestion is to not use TEST and Rfree.
See explanation on the CCP4 forum and in our paper (attachment).
https://www.mail-archive.com/[email protected]/msg39381.html
Highly recommend this paper that answers the question about Rwork and Rfree ratio http://scripts.iucr.org/cgi-bin/paper?AD0082.
Our recent paper Refining the macromolecular model - achieving the best agreement with the data from X-ray diffraction experiment also touched this aspect
Article Refining the macromolecular model – achieving the best agree...
- Badges
- Science method
- Similar topics
- Methodology
- Crystallography
- X-ray Diffraction