Paul Adams

B-factors for the model refer to mean square displacements for the position of the atoms. The phrase B-factor is used in many contexts (unfortunately), which can be confusing. I prefer atomic displacement parameters instead.

Claudia Lucía Millán Nebot

Yes, atomic displacement variation gets called also bfactor but it is not an overall, is a value for a given atom in this case. So mixed up terms yes!

Paul Adams

Atomic B-factor (ADP) refinement is usually performed when you have resolutions of 4-5A or better. Currently this is group B-factors - Pavel will probably talk about what that is. At lower resolution you still want to refine the displacements for atoms, but that might be limited to a single B-factor for the whole molecule, or domains.

Paul Adams

@Valeria there is no technical limit (i.e. you can run them at any resolution). However, in practice they are likely to be useful at resolutions better than ~4-5A.

Lisandro Otero

How does Phenix real_space optimize the refinement into a map with different local resolutions? Does it detect the different resolutions along the map modifying the parameters (i.e. Constraints/Restraints)?

Krishan Pandey

Do map-map FSC and Model-map FSC provide different information? Why generally model-map FSC is higher value than map-map FSC (from published data).

Paul Adams

@Lisandro The restraints for refinement are quite tight, so it isn’t clear that making the restraint weight change locally in the map will improve anything, but we are thinking about looking at this.

Injae Chung

In phenix.real_space_refine, rotamer restraints (in terms of sigma values), seem to be titrated down from 5 in the first macro cycle to 1 in the final macro cycle. Is it possible to manually change this initial sigma value?

Mutum YAikhomba

Paul, Can I clarify something '' At lower resolution you still want to refine the displacements for atoms, but that might be limited to a single B-factor for the whole molecule, or domains......''. so can we say that ADP refinement is the same as Rigid body fit in resolutions of ~7A-8A (where we see helices as tubes)? Do we have to supply what atom/ molecules move togetherDP refinement at low resolutions (7A-8A)?

Fengwei Zheng

The structure table seems have several cc values, e.g. CC mask, CC box, CC peaks, which one should we include in the paper table? Many thanks!

Mutum YAikhomba

Also, Paul, what makes ADP refinement better at 4-5 compared to say ''Reference Mode lbased global minimisation''?

Paul Adams

@Mutum ADPs are different to rigid body refinement. The latter is about the position of the atoms in space. ADPs describe how smeared out the atom is around their position in space (it is a Gaussian probability density function).

Claudia Lucía Millán Nebot

@fengwei in this paper there is a description and references for each of the coefficients https://journals.iucr.org/d/issues/2018/09/00/kw5139/index.htmlv

Paul Adams

@Mutum Reference model is a way to apply restraints to the coordinates of the atoms. This is unrelated to atomic B-factors (ADPs)

Claudia Lucía Millán Nebot

They differ in the data used in their calculation and point to different aspects

Dorothee

@Fengwei: I think there is no clear consensus in the community for the parameters to be shown in the paper table. If in doubt, several could be cited, along with the definition.

Injae Chung

I’ve found that the default Ramachandran restraints often generate a false backbone twist, which even after manual correction in Coot, returns following another phenix real-space refinement run. How would one eliminate such false twists given that the map is good enough to be 100% sure of backbone positions?

Mutum YAikhomba

Paul, regarding the ADP, I get the idea that B-factor describes/measures the mobility of the atoms. But why is this recommended at 4-5 A when there is not much side chain information?

Dorothee

@Injae: You can switch off Ramachandran restraints or you can change the settings. See the documentation for parameter names etc (https://www.phenix-online.org/documentation/reference/real_space_refine.html#full-set-of-parameters)

Paul Adams

@Mutum You are trying to generate a model that fits (describes) the experimental data as well as possible. The mean square displacement of the atoms are important is describing the variability of the map locally. If you have a model where the ADPs are all constant (and maybe too large or too small) this model doesn’t fit the data very well. You should be careful making any detailed interpretation of the ADPs at low resolution - which is maybe what you are worried about.

Oleg Sobolev

@Injae, It is important to also check the surroundings since something else could push the backbone out of the desired position. If it is already in favored region, Ramachandran restraints are not that strong to twist it out.

André Graça

Basic question but for what I do not have an answer right now: ordered solvent, water molecules, should be at a maximum distance from other solvent molecules (or atoms of a protein side chain) which suppose is the maximum allowed distance of an H-bond. Between what range are H-bond expected?

Tristan

Also map magnification - I've seen quite recent cryo-EM maps that have magnification out by up to 7%!

Paul Adams

@Valeria Water placement typically requires map resolutions of 2.8A or better - and the number of waters that are visible will increase with the resolution (you might only see a handful of very well ordered waters, or ions, at low resolution).

Injae Chung

Thanks @Oleg. that was my first reaction and discarded any surrounding waters for example, but to no improvement

André Graça

@Dorothee, yes, 2.8Å is what I typically expect, but how much longer could it be? Could 3.2Å still be H-bond?

Fenghui Fan

The stastics of the current cryo-em models (Ramachandran, molprobity, et al) are often betten the old crystal structures. Then is any possibility the current cryo-em model has been over idalized than it is real structure?

Oleg Sobolev

@Injae, not new scope “ramachandran_plot_restraints” in parameters (Dorothee shared the link). It gives a great flexibility in defining what and how should be restrained. E.g. you can skip restraints for residues in favored region completely or apply different type of Rama restraints.

Oleg Sobolev

@Hugo, the current standard is “No unexplained Ramachandran outliers” and typical Cryo-EM resolution usually does not allow to justify an outlier.

Paul Adams

@Andre 3.2A can still be a H-bond. However, as the distance gets longer, the energy gets lower. In all cases you need to have some good experimental data to support the interpretation of details like this. I.e. a 3.5A cryo-EM map is going to make it hard to make strong statements about the lengths of H-bonds.

Dorothee

@André: There is no clear cut-off in the literature. It depends on the types of atoms involved for example. See this review: https://onlinelibrary.wiley.com/doi/10.1002/1521-3773(20020104)41:1%3C48::AID-ANIE48%3E3.0.CO;2-U

Dave Lawson

What about adding hydrogens in refinement. Should give better validation statistics eg. Clash score

Anna Loveland

Is the ramachandra Z score reported in the new versions of Phenix? Is there a specific command we can use to get this score for older structures?

Vincent Chen

addition of hydrogens is important to do before calculating clashes. I'm sure Jane will cover that in her talk

Oleg Sobolev

@Anna, it is shown in GUI (validation tabs), log-file and there is a command-line tool (phenix.rama_z)

Gayathri

I am try to understand why one gets weird plots aka outliers when the atomic model is built with good statistics. Does this have something to do with the cryo-EM map itself? due to uncertainties in the calculation?

Oleg Sobolev

@Injae, we made it possible for the flexibility and we still experiment ourselves. You are welcome to try as well.

André Graça

Thank you @Dorothee and @Paul At the moment I am working with a cryo-EM map with average resolution below (0.143) 2.8Å and I am considering to analyze ordered solvent molecules. Seems to be feasible, but still I want to be careful and therefore have distances in consideration.

Injae Chung

What is the maximum value (practically) you can specify for ‘nonbonded weight’ in phenix.real_space_refine?

Dorothee

@Andre: Maximum distance of 3.2 Å seems reasonable. I think that’s the default for phenix.douse. Going further might require looking at each case to make sure they are OK.

Paul Adams

@Gayathri The cryo-EM maps are typically low resolution and don’t have enough information to well define the conformation of the model, that’s why restraints are required. The lower the resolution the more restraints or constraints are required.

Joseph Wang

In the real space refinement, we are using density guided method to refine the model. Say, at ~3A of the cryo-EM density map, in stead of the "wrong structure density", what would also cause high number of the outlier in the Ramachandran Plot? How much of outliers can be tolerated? 2% or 10%?

Ibrahim Moustafa

In refining icosahedral structure, cutting out the map around monomer is never perfect and because of that sidechains invade neighboring density during refinement and requires manual correction after every cycle. Any tips to help with this issue? Also, refining the whole structure always report a lot of clashes (talking about 2.4 A resolution map)? Any suggestion on refinement strategy in that case?

Paul Adams

@Injae The nonbonded weight isn’t something that should be manipulated to change the refinement results. This is modeling physics and we know what the value should be for the kind of target we use. If you increase the value it will reduce clashes, but lead to very strange features in other distributions like Ramachandran.

Xiaolong

I got a symmetric map from Relion imposing C4 symmetry and I refined my model in Phenix with NCS checked. But the model out is still not exactly 4 fold symmetry. I wonder what could be the reasons?

Dorothee

@Joseph: 10% would be considered quite poor. Typically, there should not be more than 0.5 % outliers in crystal structures. Cryo-EM models often have slightly poorer geometry. Jane will be able to quote limits.

Oleg Sobolev

@Joseph, we currently show the goal of <0.2% as a goal for Rama outliers in validation reports.

Paul Adams

@Joseph This will depend on your network link. There is some delay, but not too much for me.

Marta Martinez

Map recommended for the refinement: experimental map, sharpened map, improved density map?

Paul Adams

@Marta This is a very good question, and the answer isn’t completely clear. The map won’t really change the coordinates very much, but will impact the atomic B-factors (ADPs). A very sharp map will lead to small B-factors, a blurry map the opposite. I think this is something that we need to look at further. The density modified map might be good, but we haven’t tested this yet.

Prince Tiwari

Thanks Paul. But once I had to fill the weight option based on what coot predicted in order to get a better geometry and clash score. Is it now something we can do?

Paul Adams

@Prince The weights in coot are unrelated to any weights in Phenix, so I wouldn’t use that information in Phenix.

yuhang liu

I have a few questions around restrains: if there are several ligand in the structure, is there a way to put restrains for all of them? What is the preferred way of creating restrains at the first place? How can restrains for glycans be applied?

Dorothee

@Yuhang: You can supply cif restraints files for each ligand. Nigel will explain how to get these.

Paul Adams

@Prince I guess you might know a better weight - but I can’t think of a case where that would be true, so automatic is best

Claudia Lucía Millán Nebot

https://www.phenix-online.org/documentation/reference/real_space_refine.html

Xiaolong

I am not sure if Paul saw my question before. I will type it again. I got a symmetric map from Relion imposing C4 symmetry and I refined my model in Phenix with NCS checked. But the model out is still not exactly 4 fold symmetry. I wonder what could be the reasons. Thank you.

Dorothee

@mj: See the documentation page: https://www.phenix-online.org/documentation/reference/real_space_refine.html. The full list of parameters explains the settings.

Caillan Crowe-McAuliffe

Does anyone have experience changing the nonbonded_weight value in real-space refinement? In my experience increasing this value can sometimes really help the refinement, but I don’t know the particular circumstances where this is useful

Oleg Sobolev

@Caillan, Paul’s answer from above: “The nonbonded weight isn’t something that should be manipulated to change the refinement results. This is modeling physics and we know what the value should be for the kind of target we use. If you increase the value it will reduce clashes, but lead to very strange features in other distributions like Ramachandran.”

Pavel Afonine

“Is there a manual telling what every parameter do and affect?” — yes and no. Yes, there is a documentation. No, it may not be detailed enough to tell what happens in your particular case.

Pavel Afonine

“Does anyone have experience changing the nonbonded_weight value in real-space refinement? In my experience increasing this value can sometimes really help the refinement, but I don’t know the particular circumstances where this is useful” — This can help to minimize clashes, but the danger is it can perturb other model geometries. It’s fine idea to try but keep eye on model!

Paul Adams

Real Space Refinement Documentation: http://www.phenix-online.org/documentation/reference/real_space_refine.html

Paul Adams

nonbonded_weight value changing might help go over some barriers early in refinement like Pavel said, but later on it will be important to return to the default value - as it sets the inter atomic distances (which are physics).

Pavel Afonine

“Is there a way to refine the voxel size of the map in this refinement process?” — Do you have a real-life use case where it matters? I am happy to look into this offline with you to find what can be done. Refining voxel size never worked so far in my hands.

Pavel Afonine

“is there some resolution limit for morphing & SA?” — No. Model-to map fit is the guide here: if model needs a lot of movement to fit the map, then SA or/and morphing are warranted.

Paul Adams

@Prince depends on the map of course, but at ~4A you may see an asymmetric enough density to be confident about the position and orientation of a large ligand like ATP.

Pavel Afonine

“Thanks. So we should never fill the weight in Phenix in any situation right?” — Not quite.. You don’t need to change it most of the time. But if something indicates you need to change it (something odd about model geometry), then you may want to play with it. I’ve never seen a need for this in the past 8 years, though!

Prince Tiwari

Thanks Pavel. I tried couple of weight from 1-5 for a resolution close to 5A and I found a better stats at a certain weight/

Pavel Afonine

“Just to clarify, does phenix.reduce add hydrogens to ligands too?” — Use ReadySet to add H to ligands (and all the rest of the model).

Pavel Afonine

“From what map resolution would you recommend refining with hydrogens?” — Always use H, except initial stages of model building/refinement.

Krishan Pandey

For a new ligand (no PDB structure available), can I generate 3D structure with SMILE string in ElBOW or some other program and then fit in the EM density ?

Bonnie Murphy

When ‘automatically’ creating links to metal ions, is there a good tool in Coot to ensure that the starting model we use has a correct (sufficiently small) distance between metal and coordinating side chain?

Paul Adams

@Krishan Yes, elbow can take a smiles and make coordinates and restraints. You need to check the coordinates against what you know about the chemistry of the ligand though

Pavel Afonine

“while link_all=True also allow for links between ligands and RNA?” — it should, but outcome depends on the quality of initial model. So always check manually!

Dave Lawson

After building the ASU of a virus, is it possible to provide just the ASU and the matrices (BIOMTs) required to generate the whole capsid to real space refinement rather than generating a PDB for the whole capsid? In my case I have 240 chains and PDB format doesn’t deal with that very elegantly!

Paul Adams

@Bonnie the linking will also depend on the kind of chemistry - if it is something new and not in the library then the linking might not be correct, so checking is important as Pavel said

Mutum YAikhomba

Hi Pavel, I have been thinking - say you have a low-res map (7A) and this model has been mostly (90%) derived from a higher-resolution model (3A). Considering that your model in the low-resolution has already been minimised in ISOLDE and all model statistics look good by validation metrics, is there a need to do a refinement in phenix?

Paul Adams

@David Yes you can provide matrices. This is also something where mmCIF is probably better for the coordinates.

Oleg Sobolev

@Dave, you can do so using MTRIX, and the molecule will be expanded. BIOMT matrices are not expanded automatically.

Bonnie Murphy

Thanks, Nigel! Can you give us a short ‘best-practice’ workflow when working with metals? Is there a good tutorial somewhere?

Injae Chung

Sometimes three-letter-code ligands made in Coot doesn’t have the same atom names as those in phenix’s ligand library. What’s the best way around this?

Pavel Afonine

“Hi Pavel, I have been thinking - say you have a low-res map (7A) and this model has been mostly (90%) derived from a higher-resolution model (3A). Considering that your model in the low-resolution has already been minimised in ISOLDE and all model statistics look good by validation metrics, is there a need to do a refinement in phenix?” — You still need to refine using phenix.real_space_refine to achieve most optimal fit to the map, IMO.

Robert Rose

is there a good way to refine nucleotides manually in a structure? It’s hard to rotate around angles in coot.

stutisharma

What is the best way to covalently make an S-linkage between a peptide and a lipid molecule?

Mutum YAikhomba

Pavel, on that note, can one say that we are happy that the model is most *optimal* with a model minimised in ISOLDe and skip phenix, even if not real-space refined in phenix? I can see clashes going up as I do my real-space refinement.

Tristan

@yaikhomba In that sort of situation (i.e. very low resolution, tightly restrained), the output from ISOLDE and that from Real Space Refine are likely to be functionally equivalent. The AMBER forcefield in ISOLDE is extremely strict on van der Waals interactions, so residual errors tend to get pushed into angles and bond lengths. Real Space Refine is a bit more permissive on clashes and stricter on bond lengths/angles. But if you actually compare the models directly the differences between the two will typically be tiny.

Tristan

Still, an important point is that ISOLDE still doesn't attempt to refine ADPs - so if nothing else, you should run Real Space Refine for that purpose.

Mutum YAikhomba

Hi Tristan - I am slightly confused. I just heard ADP is a very valid metric for refinement anymore at low resolutions (7A-8A). So, will it be still worth doing a RSR in phenix at low-resolutions (5A-8A), considering the model has been minimised in ISOLDE?

Mutum YAikhomba

Hi Pavel, is it normal for the disulphide bonds to be not linked after real-space refinement? My input model had the disulphide bonds recognised. However, After refinement in phenix, the disulphide bond links are missing, though I can see that the bond distance between these 2 sulphur atoms (from the 2 cysteines) is still reflective of a covalent bond and this close distance is not reflected as clashes in phenix. Do you manually have to annotate the disulphide bonds after phenix real-space refinement?

Mutum YAikhomba

I had other restraint .cif files for lipids. I didn't supply a .geo file as an input in the refinement. Would that be a problem?

Oleg Sobolev

@Mutum, “bond distance between these 2 sulphur atoms (from the 2 cysteines) is still reflective of a covalent bond and this close distance is not reflected as clashes in phenix” is indicative that the bond was actually established. What leads you to believe that it is missing?

Dorothee

@Mutum: .geo files are generated by phenix.refine and related programs. They are very helpful for debugging issues with geometry restraints.

Mutum YAikhomba

Sorry. I forgot to add. I open the real-space refined model in ChimeraX and see that the disulphide bonds are missing

Tristan

I've seen (and reported) something like this before: when running phenix.refine via the GUI (but strangely, *not* via the command-line) sometimes the SSBOND cards were not written to the output PDB despite the geometry suggesting they were clearly present during the refinement.

Pavel Afonine

“Sorry. I forgot to add. I open the real-space refined model in ChimeraX and see that the disulphide bonds are missing” — Check this out: http://phenix-online.org/newsletter/CCN_2016_01.pdf#page=10

Tristan

Running via command-line with the .eff output via the GUI still generated the SSBONDs... just looking back through my email history, I'm not sure if it was ever tracked down.

Vincent Chen

generally most of the graphics jane is showing is using kinemage files and the KiNG software

Paul Adams

@Ilona the images probably came from KiNG. Similar images can be made in Coot, but not all of the things that Jane is showing right now.

Tristan

Yaikhomba, to fix your immediate problem: the new ISOLDE version provides a menu item: ISOLDE/Model Building/Disulphides/Make All Reasonable Disulphides - will reinstate all disulphides indicated by the geometry.

Mutum YAikhomba

Ok. Is it sufficient if I apply this on the output real-space refined model or on the input model before refinement? I was asking this because they (disulphide bonds) were annotated in the input but simply missing from the output of phenix RSR

Yahui Yan

from my experience of refining crystallographic structures, your input file should have correct cys geometry and Phoenix will recognise the S-S automatically. please correct me if I am wrong.

Pavel Afonine

“from my experience of refining crystallographic structures, your input file should have correct cys geometry and Phoenix will recognise the S-S automatically. please correct me if I am wrong.” — True

Olivia Pfeil-Gardiner

Another restraints question: We are working on a protein containing iron-sulphur clusters. The .geo of the refinements lists S-Fe bonds but these are not displayed in chimeraX (only pseudo bonds are displayed) and I cannot find any mention of them in the pdb file. Is there a way to make the refinement write these bonds to the pdb file?

Tristan

@Olivia: this is actually a ChimeraX thing. The LINKs will have been written correctly to the PDB file, but ChimeraX automatically converts bonds involving metals to pseudobonds when loading the file. Drop me an email and I can help you with strategies to deal with this.

Olivia Pfeil-Gardiner

@Tristan Within. In a similar structure from the pdb, there are CONECT entries for these bonds in the pdb file. I was hoping to get something similar from the phenix refinement..

Mutum YAikhomba

Can ChimeraX/ ISOLDE display live cABLAM outliers? So that we can correct them on the spot?

Tristan

Hmm... you could try instead opening the mmCIF output from Phenix. When opening mmCIF, ChimeraX fetches the templates for each residue type from the CCD so it should fill in the ligand connectivity.

Pavel Afonine

“I hope one does not have to go through refinement phenix to figure this out” — IMO, you can’t publish an unrefined model.

Francisco Murphy

After the talk by Jane, it seems we should be really careful with structures downloaded from the PDB with a resolution worse than 3 A. What to do in that case? PDB-REDO then phenix.refine (or phenix_refinerealspace) and then go to the validation section? or how to deal with these wrong structures?

Tristan

@Pavel I suspect he meant he hopes you shouldn't need to go through a full refinement just to get a new CaBLAM table? Yaikhomba: you don't - see https://www.phenix-online.org/documentation/reference/cablam_validation.html for instructions on how to run CaBLAM validation from the command line.

Mutum YAikhomba

Ok. A slightly different note. If you have built small molecules, say Magnesium ion, ADP, ATP (adenosine Triphosphate) into the well known ATP synthase at low resolution 7A, what type of validation metrics should one be looking at - CC-ligands?

Paul Adams

CaBLAM is also part of the comprehensive validation - so you can run this in the GUI to get the results

Tristan

@Francisco honestly, right now there's no substitute for just taking the time to inspect (and, ideally, fix) it yourself. For PDB_REDO, most low-resolution errors are garbage-in, garbage-out - it won't really attempt any serious rebuilding. The only way to be really confident of a model is to see the details in the map by eye.

Mutum YAikhomba

On the previous note of my question, what's the recommended cut-off for these cross-correlation metrics (map-model)?

Paul Adams

@Francisco - there isn’t a PDB-REDO for cryo-EM yet, but for crystal structures it isn’t a bad thing to try (see what Tristan said though). In general, it is always important to check any structure from the PDB if you are going to use it yourself (this is true at any resolution, but especially at low resolution). Note that these usually aren’t wrong structures - just structures with some local errors. Some are worse of course.

Pavel Afonine

A comment: At low-res you can’t have outliers because it is unlikely you would be able to confirm them by limited resolution data.

Dorothee

@Mutum: One important thing to look at is the map. Does the map have peaks for MG ions? Is there clear density for ADP or ATP?

Dario

Just for an overall idea, what would be a good, reasonable/acceptable, and bad CaBLAM score for a 3-4A map?

Gayathri

Download of today’s chat would be really helpful. I hope this is still one of the files that’ll be uploaded? Many thanks!

Francisco Murphy

In X-ray crystallography there is the (un)famous table one... is there something similar for cryo-EM?

Pavel Afonine

“In X-ray crystallography there is the (un)famous table one... is there something similar for cryo-EM? ” — what Phenix GUI comprehensive validation reports is Table 1. See also: https://pubmed.ncbi.nlm.nih.gov/30198894/

yuhang liu

What is the suggestion regarding removing uncertain residues/glycans if by doing so making the statistic better?

Tristan

@yuhang: take care when doing this. If it's a sidechain pointing into space with no or very few interactions, there's no real harm in cutting it. But if it's pointing into a space where it *should* interact with other atoms, then removing it will give the backbone and surroundings unnatural freedom to overfit. In the worst cases, removing sidechains can "cover up" very serious errors (e.g. out-of-register stretches) by removing the major clashes that would otherwise flag them.

Francisco Murphy

Quick question: Is there a decent and modern book to learn about cryo-EM? any suggestions?

Pavel Afonine

In the begging H may add more clutter and impede refinement. But definitely use towards the end!

Zoom Transcript