Friday, July 1, 2016

NMRlipids III: Preliminary observations

[updated 19.10.2016]

I have started to assemble the NMRlipids III manuscript. The project is initiated from the discussion started by Peter Heftberger (see Data Contributions page, discussion from December 1, 2014). Heftberger et al. compared x-ray scattering form factors with results from CHARMM36 simulations and concluded that the results did not quite agree (see the report). This was slightly surprising since simulation literature typically reports good agreement with experiments for form factor data, also for CHARMM36 with cholesterol. In the subsequent discussion we decided to take a simple two component lipid bilayer to check how accurately lipid mixtures can be described by atomistic resolution molecular dynamics simulations. For such system we chose the POPC/cholesterol mixtures up to 60 mol% of cholesterol. Extensive NMR data for this systems is already published, extensive simulation data sets are already collected in NMRlipids I and extensive scattering form factor data was reported in this blog by Georg Pabst et al. (July 3, 2015, comment in About page).

I have now started to collect this data in the manuscript (available from GitHub according to the current workflow of the project). All the data is not there yet, but some preliminary results can be already discussed. Some results for order parameters and form factors, and preliminary observations are shown here:

Figure 1: Acyl chain order parameters from experiments and  different simulations for pure POPC and POPC with 50mol% of cholesterol.
Figure 2: Form factors from experiments and simulations with different cholesterol concentrations in POPC bilayer. The Experimental form factor amplitudes are not scaled to match with simulations, thus mainly the locations of minima should be compared.





Preliminary observations:

CHARMM36: The order parameters are too large in both simulations, with and without cholesterol. The form factor gives minima at too small q values. This is not expected from the parametrization publications. However, the form factor result is pretty similar to the one report reported by Heftberger et al. (simulations ran with NAMD). More detailed discussion about reproducing some CHARMM36 results is in GitHub issue. [Update 19.10.2016] as shown in the GitHub issue, the overestimation is largely (but not completely) due to the usage of older version of Gromacs in the original results. The results from Gromacs 5 are now used in the manuscript.

Berger/Höltje: The order parameters in pure POPC are in very good agreement with experiments, as often reported in the literature. However, form factor has minima at too high q-values. With 50mol% of cholesterol the order parameters are too high, as reported before. However, the form factor minima are surprisingly good with 50mol% of cholesterol. My preliminary interpretation is that the model has too large area per molecule in pure POPC and too strong condensing effect of cholesterol. This leads to good agreement in form factor with large cholesterol concentration but order parameters are then too large. This would also indicate that the relation between order parameters and packing density would not be right in this model. However, more careful analysis is needed.

MacRog: Order parameters are not in good agreement in neither of the systems, with or without cholesterol. Form factors still to be calculated.

Things to be done:
  1. I would like to plot the electron densities from simulations together with the results from the SDP model used for the scattering data. The SDP profiles are available in the delivered data set, but only in *eps format. Would it be possible to deliver these also in ASCII format?
  2. In addition to the simulation data we already have, at least Slipid and Lipid14 models have compatible cholesterol models. At least some data from these models is almost necessary for this project. The corresponding simulations are already ran for Lipid14 and we could take the order parameters from the publication. However, for form factors we would need the simulation trajectories (or someone delivering the results). [Update 19.10.2016] Slipids data are already delivered. Lipid14 data is in progress.
  3. There is also experimental form factor data available for other binary mixtures than POPC/chol. The data for DOPC/CHOL and DPPC/CHOL mixtures was delivered by Heftberger et al. (see also report) and there seems to be more in the literature. If someone delivers corresponding simulation data, the comparison would be a good complement for the extensive POPC/chol set.
  4. The order parameters for POPC (preferably in 303K or close) from CHARMM36 simulation ran with NAMD or CHARMM would be useful to check if they are the same as from Gromacs. More discussion in GitHub issue. [Update 19.10.2016] NAMD data and much more is delivered throught the discussion in GitHub. The conclusion seems to be that Gromacs 5 gives more or less consistent results with other codes.
I will go now for summer vacation and continue working on this in the beginning of August.

Monday, May 30, 2016

Toward submission of NMRlipids II publication (lipid-ion interactions) (2)

The first attempt to submit NMRlipids II publication was postponed due to the extensive discussion and new contributions. The manuscript is again quite close to the submittable version (in my opinion). The results and discussion section has been rewritten but the conclusions and introduction are not essentially changed.

In the previous submission attempt I suggested the submission to Chemical Science for the same reasons as we submitted NMRlipids I to JACS (see discussion in Towards first submission to journal post). However, the publisher policies have changed quite a bit within one and half year so the discussion seems to be outdated. The RSC policies seems to officially allow this kind of work, thus there is no reason to test this anymore. For this reason, I suggest that we will submit to Physical Chemistry Chemical Physics (PCCP) where the acceptance is more likely.

There is still a ToDo list in the manuscript which mainly relates to missing technical information.

At some point we also need the Table of Content Graphics. I have opened the GitHub issue on that, please let us know if you have any suggestions.

If there will be no objections or major comments, I will proceed with the submission to PCCP within two weeks from now (13th of June 2016). Please, give all the comments and contributions before that.

Wednesday, February 24, 2016

We almost got reviewers' comments published

Our goal from the beginning has been to immediately publish all the scientific content related to the project. One relevant part of the content are discussions between reviewers and authors during the peer review process. We have now published two peer reviewed articles: NMRlipids I and NMRlipids V. In both cases we have asked from the editor if we can publish also the reviewer's comments since everything else is public. As expected, in the NMRlipids I case the Journal of Physical Chemistry staff replied that this is not possible. However, editorial and publishing teams of BBA Membranes'  were positive about publishing the referee's comments in the case of NMRlipids V publication. Both referees were also sympathetic to the idea. However, one of them declined stating permission to make comments available should be asked a priori, at the same time referees are invited to review a paper. This is an important learning point from this experience.

Tuesday, January 19, 2016

Does the glycerol backbone structure depend on initial structure?

For model by Ulmschneiders it does. This was observed while finalizing publication from NMRlipids II project.

First it was observed that two independent simulations ran with Ulmschneiders model (http://dx.doi.org/10.5281/zenodo.13392 and http://dx.doi.org/10.5281/zenodo.30904) gave different order parameters for glycerol backbone. It turned out that the essential difference between simulations was the initial structure, other one was dowloaded from Lipidbook and other one was generated with CHARMM GUI. More specifically, the difference in O-g3-g2-g1 dihedral in the initial structure. The results from different initial configurations are shown in Fig. 1 and the whole discussion is here.


Figure 1: Order parameter results for POPC headgroup and glycerol backbone from model by Ulmscneiders simulated with two different initial configurations (Lipidbook and CHARMM GUI). The segments labeled 4-8 belong to glycerol backbone and segments 0-3 belong to headgroup.


Importantly, superficial comparison to the experimental results in NMRlipids I publication (Fig. 2 in publication) indicates that the result with CHARMM GUI initial structure gives order parameters closer to experiments. However, in NMRlipids I publication the results with Lipidbook initial structure was used (the other one was not yet ran) and the force field quality (Fig. 4 in publication) was ranked based on these results. In conclusion, using more realistic initial structure, especially for O-g3-g2-g1 dihedral, gives better structure with Ulmschneiders model and increases its ranking (Fig. 4 in NMRlipids I publication).

Now, the question arises if the glycerol backbone (or other) structures depend on initial configurations also in other models? This has been somewhat discussed in the project but not systemically tested, see About glycerol conformations post and comments in Accuracy of order parameter measurements post. My argument has been that based on comparison to the NMR relaxation data, the typical simulation times are sufficient to sample full structural phase space. If this does not happen, the simulation is too rigid. Also, independent simulations have been ran with several models tested during the project (at least with CHARMM, Berger, Slipids, GAFFlipid) and there has never been significant differences in order parameter results between different runs. However, it is not clear if in all of these simulations the initial structures have been really independent.

In conclusion, it seems likely to me that in most models the glycerol backbone structure do not depend on the initial structure, but in Ulmschneiders model it does. Further tests would be needed to confirm this argument. I think that this is relevant topic for users and developers of atomistic resolution MD simulations.

Also another question arises; should we update the results in NMRlipids I publication? At least we could update the version in GitHub. However, science always goes forward and this will not be the only issue to be updated. It may be reasonable to update more changes at once. Traditionally, we could submit a correction to a journal or make a new article. However, due to Internet we have the possibility to update outdated information directly into the article without losing history; but I am not sure how general readers would realize that updated version is available. While waiting for journals publishing modifiable articles, we have to use some temporary approaches. Any ideas?