Coarse-grained / Multiscale Simulation Software (NAMD), including validation
Sep 5, 2023
Coarse-grained / Multiscale Simulation Software (NAMD), including validation
Emad Tajkhorshid
that we are interested in, we are interested in describing sort of larger scale kind of
phenomena at the level of coarse-grained representation, but then very often you are interested in
knowing atomistic details in those coarse-grained or in those molecular systems.
We will not be able to describe large structural changes at an atomistic level, that's why
we go to a coarse-grained model, but once you have seen something that might be relevant,
you have seen the aggregation, you have described how these four things come together, something
that you cannot do atomistically, you want to be able to go back and do atomistic representation
of those molecular systems and describe and understand what might be happening in those
systems.
So that's actually exactly the case, so let me actually skip this one, this is not really
relevant, just to give you an idea.
So even at the level of coarse-grained description, we have different representation for the side
groups, for the side chains of these four things, and we can see that interaction between
certain type of polar side chains, these green guys, is the main drive for the structure
of the aggregate that we see in here.
So we can describe certain phenomena, for example, insertion of the helix into the membrane,
which is very difficult to describe atomistically because it's a slow process, is nicely captured
in these simulations, again, timescale should be taken with a large grain of salt here in
this case, because if you do a 120 nanosecond simulation, you wouldn't see this transition
to these timescales, or not really a real timescale associated with the transition at
the real world, but we can qualitatively describe them, you can see the helix actually goes
across the membrane, comes to the surface, and actually in some cases goes back, which
nicely agrees with experimental data suggesting that you have a mixture of different states
for these four things in the membrane.
So we can see actually how it comes to the surface of the membrane, we even have the
separation of this from the membrane, insertion into the head group region, and complete insertion
into the bilayer, completely described at the level of coarse graining, we have transition
between these two, we can start looking at the kinetic properties, again, to the degree
that you can trust these timescales, but at least you can get some qualitative behavior.
Some of the data that we got, of course, it's very important to compare the experiment and
see how much these models mean, and are reasonable, so based on these simulations, actually for
the first time, you know, the classical model in the literature was really nice before there
were bundles of these antimicrobial peptides, whereas the model we observed, if you go back
and look at here, you have a very disordered kind of system, in some cases they form aggregates,
not very regular, but in most cases they have all sorts of states in the bilayer.
And that seems to be more consistent with what is known experimentally, if you look
at the tilt angle of this guy, for example, this helix, and look at the angle of the axis
of the helix with respect to the membrane normal, experimentally you have such a distribution
based on NMR, and maybe if you look at the result that we got from our MD simulation
to reproduce that kind of experimental property, you can see a disordered kind of configuration
of this molecule explains, or captures the experimentally observed property very, very,
very closely.
So after this work we actually suggested that what people think is not really true, most
of the time we have a sort of tilted and highly disordered structure of the peptide in the
membrane, but again the question is, how does such a configuration in the membrane can kill
the cell?
So in this model people suggested we form a nice regular channel, now the membrane potential
which is composed of proton gradient and ion gradient across the membrane will be killed
because these ions can dissipate through the pore form like these guys.
How can we explain the antimicrobial effect of this peptide based on the sort of disordered
configuration that we characterized using our coarse-grained simulation?
So that was the next step, and that was something that we could not really do at the level of
coarse-grained.
These are some snapshots taken from the coarse-grained system in terms of how the aggregates come
together and how they interact, but we didn't really see anything going across these membranes
and these peptides.
So what is it?
How can we perturb the permeability of the membrane by this irregular configuration?
So that's exactly the point, that's one to motivate you, why do we need to go back to
atomistic representation sometimes?
Because at this level of coarseness you might not be able to see detailed properties and
phenomena including water permeation across the membrane.
Coarse-grained waters are simply too big and they are not charged, that's another problem
associated with them, and we are not able to see any water permeation or pore formation
in this kind of description.
So what we needed to do was to go back to an atomistic description.
So now this is the second half of your tutorial.
So going from an atomistic description to a coarse-grained description is rather trivial
because you know that these four atoms are going to be grouped into one coarse-grained
bead.
So you go, you reduce the representation, it's very simple, you find the center of mass
of this guy and that would be the position of your coarse-grained bead.
When you want to come back you have one coordinate and you need to grow four different atoms
into that coarse-grained bead.
So it's not a trivial problem, there are multiple solutions for that, so you have to have some
ways of coming up with these atomistic structures for that kind of transformation.
So there are lots of different kinds of tricks of doing that, and one of them is something
that you will do today.
So we talked about the coarse-graining, so now we want to have the reverse coarse-graining
or fine-graining if you wish, and as I pointed out there are multiple ways of placing four
water molecules into this bead.
So we have to have a way to facilitate that to make sure that the chemical structure of
the system will be reasonable for our atomistic simulation, and that's one of the protocols
that we will use today.
So just as a, again, for this particular system we are going to, for those elements whose
structures are well-defined and we want to preserve those structures, we do reverse coarse-graining.
That's the case for the protein, how they aggregate it, you want to preserve those configurations
and do reverse coarse-graining for protein and liquids, but for the case of ions and
water you usually don't need to reverse coarse-grain these.
You can rehydrate and re-solve it and add salt to your system again, so you don't usually
reverse coarse-grain these more fluid, less structured components of your simulation system.
Okay, so this is sort of, this is why capture is the way we did the reverse coarse-graining,
so all of the protocol that was done in this study, we started with an atomistic description,
we coarse-grained the system, then we were able to do microsecond timescale simulations
to see changes in the configuration, in the structure, in the conformation of peptides,
how they aggregate.
Once we had captured those large-scale, slow phenomena in the simulation, we were interested
to know how these particular configurations characterized by coarse-grained simulations
would behave in an atomistic system, so that's when we started to do reverse coarse-graining.
As schematically shown here, you will initially, when you map them back or project them back
into an atomistic description, you might have really bad structures for this, so the
initial mapping doesn't give you a very nice structure, and at this stage, using different
protocols, we try to refine this structure into a chemically sound and reasonable structure.
Once you do that, after what we are doing here is it's kind of simulated annealing
of this bad atomistic structure in order to relax it into a reasonable shape.
What we are doing here is actually to constrain, so for example, this atom is going to be mapped
to these four or five atoms up here, so we are going to constrain the position, the center
of mass of these four or five atoms here to the position of the coarse-grained bead
while we are doing the initial simulated annealing, so we have all these constraints acting on
these groups of atoms, keeping them close to the center, to the position of the coarse-grained
bead that represented them in the coarse-grained representation during the initial phase, so
that they don't go anywhere, they stay where they are supposed to stay, but we allow the
other degrees of freedom to relax in order to have a reasonable structure.
That's one protocol of doing it, and that's what you will be doing today.
So once you have this reasonable structure, then you remove these constraints, and then
you continue your atomistic simulations.
Another thing that we do here, of course, when you are doing your mapping, you have
to have a file that allows you to map back this into the same, you need to have a sort
of initial conformation also for the groups that you are going to build into these models.
So you know that you need to have a nitrogen, a carbon, and another carbon here connected
to the next group, so the internal conformation or intramolecular arrangement of these groups
also could be different, and in order to do, to optimize this and to have a faster
convergence into a reasonable model, what we do is that we save the conformation that
we use initially for this particular group, coarse-grained bead.
Before the coarse-grained bead, we save this particular conformation of these 4 or 5 atoms
here, and we use the same conformation, and we build it back into the coarse-grained model.
So that way, you can also have a faster convergence.
So that's another step that you will, you might notice during your tutorial this afternoon.
Okay, so that's actually what I described on that slide, so in order to map back coarse-grained
beads, and then re-solidating the system, we do 5,000 steps of immunization, and then
simulating annealing for about 20,000 steps, allowing the system to relax.
During this step, of course, we are constraining atoms to the position of their representative
coarse-grained beads, and that has been pretty successful.
Okay, so you go from a coarse-grained model, again, you run a coarse-grained simulation,
and you take snapshots from your coarse-grained simulation, where you think major things have
happened, and then you examine them using a finer pair of glasses, using atomistic simulations.
And then you can start seeing that actually, now you can see there's a lot of hydration
in the membrane region, so here you don't see any penetration of water, there's one
exception here, but otherwise you don't see water going into this defective membrane,
whereas when you come back to the atomistic simulation, you see a lot of defects.
So these are some examples of water defects that are introduced into the membrane by these
antimicrobial aggregates, so again, we can say they do not matter, but again, this is
one example of the kind of phenomena that you might be interested in, how is this membrane
defective as far as water permeation is concerned, but you cannot see at the level of coarse-graining,
and you need to go move back and forth between coarse-graining and atomistic simulation in
order to capture that.
Not a real multi-scale description, but by using different scales, you can really get
the best of both, in one case, you describe very slow phenomena, large-scale kind of structural
changes, and then you come back with more fine-grained description to see what those
aggregation or aggregates might, how those aggregation or how those structural changes
might become mechanistically important.
So these are water defects that are formed using very irregular kind of conformational
peptide, which is again, completely against what was in the literature for a long time.
I think immediately after this, there were some experiments suggesting that there is
a large degree of disorder for these peptides, which agrees with what we observed.
This is just to highlight conformational variability and disorder of peptide in the membrane.
In this case, and here you can see a nice way of showing how much hydration you get
into the membrane due to the presence of the peptide.
Anyway, so after this work, we are suggesting that actually the main effect of this peptide
is really introducing, in a non-specific manner, defects into the membrane, and that's how
it affects the cell, so non-specific general operation.
Okay, so today, you will use a lipoprotein example.
This is what you use in your tutorial, so that's the main part of the tutorial.
But in addition to this, you have a couple of additional examples in the tutorial, one
of them dealing with coarse-graining a simple membrane, another one dealing with a transmembrane
protein inside the membrane.
So if those cases interest you more, you might want to play with those rather than the lipoprotein.
So lipoprotein is a sort of little patch of membrane inside a belt of a protein.
This is, with this lipid bilayer, it could include cholesterol and other kinds of lipids
in your blood.
That's how they are carried in your blood, and there is a large interest in knowing and
characterizing the structure of this guy, and the size of the particle here, in this
case, and how they assemble.
But of course, it's not an easy thing to crystallize experimentally, that's why modeling has been
used a lot here, to test and to come up with models, structural models, for these particles.
So you see, we have an atomistic representation here, and a coarse-grained representation.
And you go back and forth between these, and you do some simulation on these today.
Okay, so now, I think we sort of covered most of the theory, and how really this is done,
but one thing that is important, I think, is to know what files are involved.
This is something I wanted to talk about a little bit yesterday, very briefly looking
at the files that you're using for your simulation, because these are exactly where atomistic
description, coarse-grained description, parameters are safe.
You need to know where the information is safe, in case you want to change, in case
you want to use your own coarse-grained description, or whatever.
So that's why I would like to spend a few minutes just to look at the major files that
you used yesterday for your exercise.
Okay, so let me start with the structure.
So that's, just to give you a brief, in case you haven't looked at this yourself.
So this is a coordinate file that you usually use for your simulations.
This is so-called a PDB file.
We have, for each atom, we have, in a PDB file, you have the coordinate of the atom,
XYZ coordinates, you have the residue they are associated with, and you have the atom
names.
Okay?
So in case.
So these guys don't have any chemical meaning.
We only know that there is a metionine residue in this protein that has atom names of this
sort, and these are the positions of those particular atoms.
So you need to have a topology file in which you connect, or you derive the sort of chemical
properties of these individual atoms for that particular residue in order to be able to
use the right parameters.
So yesterday you remember that you need to specify all the bonds, all the bond angles,
all the dihedral angles for each kind of interaction you need to provide the right parameters.
And that's something that is not present in the PDB file or in your coordinate file.
So you can see for all the residues in your protein and later for the water.
This is a ubiquitin in water that you use for your example.
These are water residues in which I specified as T3P, for example.
So this is your PDB file.
So that's a sort of simple case, easy case.
Then what is important also for your course name kind of tutorial today is the topology
file.
That's where you specify the topology of the system.
So here, I mean, this is an example of the CHARM topology file.
It doesn't matter what you are using.
So you have the masses of all the atom types.
Now we are introducing the concept of atom types.
This is actually a pretty important distinction between atom names and atom types.
Atom types are chemical types for which you need to have the right correct mass.
You need to have the correct parameters as far as the interactions are concerned.
These are the interaction sites.
So every atom name in your PDB file should be mapped onto one of these atom types in
order for you to be able to derive the right parameters.
So the masses are here.
And then later, for each residue, you can see, for example, for the case of alanine,
we can say there is a residue called alanine.
This guy has these atom names.
And every atom name is connected to an atom type.
And it's partial atomic charge.
So if you want to change something in the file, this would be, for example, if you want
to neutralize all alanines in order to do a test simulation, the topology file would
be the place to do this.
So you have atom types, and you have the charges here.
And then later, for this particular residue, you have all the bonds.
Essentially, the connectivity within this residue is defined here.
CP is connected to CA.
N is connected to HN, et cetera, et cetera.
So your connectivity is defined here.
Now, all the bonds are defined here.
Angles and dihedrals are generated automatically, usually, in this particular file here, auto
angles, dihedrals.
And by default, the first and last residues will be capped because it's assumed to be
a protein.
This whole thing was evolved around protein simulation.
That's why everything is called a residue, including water is a residue, lipid is a residue,
and a ligand you have is a residue.
So the connectivity is defined here.
So you have all of the information about the connectivity here.
So if you are using a standard residues, you don't have to do anything.
If you want to have your own ligand, then you have to define this ligand here.
So now, instead of using a commissive description, you can have a topology for alanine that has
only two atoms.
One is a backbone atom.
One will be a side channel atom, or beam, I should say.
And then in your PDB file, instead of having 10 or 15 atomic coordinates, you would have
two sets of coordinates for these two beads.
That's the only difference.
That's the only place that you sort of essentially change the representation that you use for
the system.
Of course, the force constant parameters are going to be different as well.
But as far as the topology and the representation that you use, you will be changing your atomic
representation, because you are now looking at the center of mass of groups of atoms,
and you will be changing the topology of that.
You will have different number of atoms and different quantity.
So this is for different side chains.
So you have all this information stored here, and that goes on and on for different kinds
of residues in the protein, for water, for ligands, heme, and other kinds of molecules
that you might be interested in.
And of course, as I pointed out, if you have your own molecule that is not in here, you
have to give this topology information, so that you can use it to generate the protein
structure file.
So for those of you who might be actually interested in doing protein simulation, there
is a, let me go back.
This one thing I think is very important, because it also connects to the fact that
you don't have polarization, and you are using fixed charges.
So this is a particular side chain that can have different protonations.
We have quite a few of these side chains, actually.
And as we discussed yesterday, we have a fixed representation for the charges in our classical
description.
This is not the problem only for proteins.
In general, when you have a molecular system, you might have polarization in general, and
protonation, deprotonation of certain sites.
And that's something that is unfortunately not implemented in our classical description.
So if you have a change in the protonation, it's in any charge state of part of the molecular
system.
We will not have it included in here.
But in some cases, so you have to make up your mind what is the most representative
form of that particular residue, of that particular chemical group that you want to use.
You want to use a positively charged amino group there, or you want to use a neutral
form.
So this is usually done based on pKa calculations for these side chains.
In most cases, you have clear pKa's.
For acidic residues, you have pKa's of 3, 4, meaning that at a neutral pH, where you
usually simulate your system, these are going to be charged, deprotonated.
And for basic side chains, we have very high pKa's, 10, 12, meaning that they are going
to be charged, too, at a neutral pH.
But for certain groups, this becomes really tricky.
For histidine, for example, the pKa of this side chain has two hydrogens, two nitrogens
in its side chain.
It's pretty close to 7.
It's actually 6.5.
So it becomes really difficult to decide what is the representative protonation state.
And even if you know the neutral form has a slight preference, or sort of a slightly
higher population, you don't know where to place the hydrogen.
You can place the hydrogen here, or you can place the hydrogen here.
So it has two tautomers, so to say, two tautomeric states.
So for that, you have different topologies, different configurations.
Based on the local environment of that histidine, you can decide what tautomeric state or what
protonation state might be the most representative one.
So there are quite a few of these that you have to keep in mind in case you have them
in your molecular system.
But that's a good example of the problem associated with fixed charge or fixed protonation state
So we have three kinds of histidines.
HSD has a proton at the delta nitrogen.
Then we have HSE down there here, which has a proton on the epsilon nitrogen.
And then later we have the HSP, which is a protonated histidine.
It has two hydrogens, two protons, and it carries a negative positive charge.
There are many residues of this kind that you can decide which one is the best one to choose.
OK, so then you have the topologies here.
You see there is no information about the coordinate, but all the connectivity information
and all the charges are defined here.
So you get your PDB file.
One more point about water, which is important in case you are considering different force
fields.
So water is also defined as a residue with three atoms.
And this is the so-called TIP3P water.
So it's a three-point model for water.
There are more than 100 different models for water, depending on what you want to describe.
If you are after diffusion constants and diffusion properties of water, this is not the best model.
If you want to have the dielectric effect of water, you have to use a different water model.
But in our limited description, we need to go with one form of water.
So in CHARM, they are using TIP3P water.
And everything else is fitted to interaction of different groups with water.
So you cannot use a better water model in CHARM if you are running your simulations
on a protein with CHARM.
For example, there is another water model.
I think yesterday someone was asking SPC, which has better properties as far as diffusion
is concerned.
But you cannot mix it with CHARM.
Because in CHARM, all the energetics are fitted to interaction with this particular water.
So if you introduce a different water model, maybe your diffusion properties are better
for bulk water.
But all the interactions with the protein are going to be messed up.
This is important for the particular case of CHARM case.
But I can imagine there are other similar considerations that you are trying to choose.
The right force field, I was trying to mix and match different kinds of force fields.
You have to always keep that in mind.
In some cases, even the conditions that you use for your simulations should be the same.
For example, all of the testing in CHARM has been done using a certain cutoff, 12 angstrom,
10, 12, and a certain kind of time step, 2 femtosecond time step.
So in order to stay consistent with the conditions in which the force field has been optimized,
you should try to use similar conditions.
Don't just change it and say, oh, I'm doing better.
My cutoff is now much better than 12.
I'm using 20.
So my simulation should be performing better.
No, not necessarily.
You might actually even use a different level of theory when you are deriving parameters.
They have used Hartree-Fock and different levels of MP2.
Maybe you can say, I can do much better coupled cluster kind of calculation or much larger basis sets.
Yeah, so in principle, you're right.
But then you become inconsistent with the rest of the force field.
We have to try to stay consistent with the rest of the force field when we are developing new parameters here.
But again, sometimes, again, this is what has been done, has been tested.
We try to stay consistent.
But in some cases, you know that because the electronic structure is so affected or so wrong at this level of theory,
I cannot calculate it using the standard protocol.
I have to go into higher levels.
If you have really strong justification for doing that, good.
But otherwise, we should try to stay as consistent as possible.
OK, so we have the topology in here.
We have the coordinates.
We have the residues.
And then what you did, one of the steps that you did yesterday, hopefully, was to generate the protein structure phi.
The PSF was generated using the PSF gen for NAMD.
So what you do here is to look at your PDB file and say, I have this residue called methionine.
Let me look at the topology file, see what information I have for methionine.
And you dump that information into this PSF file.
Then you go to the next residue, get the topology for that residue, dump it into the PSF file.
So the PSF file will contain all the topological information, connectivity, the number of bonds,
atomic, atom names, atom types, and charges for the whole molecular system.
So in the beginning, you have 7,051 atoms.
They are all described, what segment they belong to, what residue they belong to,
what is their atom name, and what is their atom type.
And you will see that we will use this atom type information to derive the correct parameters
for interaction between these atoms and other atoms in the body.
So these have chemical meaning, whereas these are arbitrary atom names.
That's going to be the partial charges for all the atoms, and these are the masses.
So now, if you want to artificially, and I sort of ask you to do so in your simulation,
try to put it under extreme artificial conditions to test hypotheses.
For example, if you want to change the mass of some of the atoms in your system,
the PSF file would be the place to work with.
Maybe you can make this atom 100 times heavier so it doesn't move or see what happens.
Or you can actually, initially you have the information about individual atoms.
You can see this is the first block.
Later, you have the information about the bonds.
You see for each atom, we define an index here, and then later in the next block,
we use these indices to define the number of bonds and number of angles that are present in the system.
So you can break one of those bonds very easily by just coming to the PSF file.
So this is the bonds block where you define the bonds.
There are 5,117 bonds here.
Atom 1 is connected to 5.
2 is connected to 1.
So you can delete a bond easily by just editing the PSF file.
Or you can add an artificial bond here for yourself, really testing something.
If you introduce new bonds, of course, you need to have the correct parameters for that particular bond.
Right?
So that's how, that's the only thing NAMD needs.
You generate this PSF file.
You have the coordinate in your PDB and all the information about the bonds,
later angles, dihedral angles are stored here.
So the only thing NAMD needs, NAMD doesn't need the topology file.
It needs this one, this PSF file.
So it comes here and says, OK, you are telling me atom 1 is connected to 5.
Goes up there, finds what is the atom type for atom 1, what is the atom type for atom 5.
So I have these two atom types interacting with each other through a bond.
I need a spring force constant and I need an equilibrium value for this interaction.
If you give me that in the parameter file, I continue.
If you don't give me that, I stop.
That's how NAMD works.
So now after defining all the bonded information here and then the last file you need is the parameter file.
There are the parameters for the potential energy terms are stored.
So this is, they usually come in pairs.
The parameter file, this one is for all atom description, version 27, protein, lipids.
And it usually comes with a topology file that has a very similar name.
So we usually use this in pairs.
So here, as you can see, again, for every pair of atom types, see, that's a bond part.
So we have the, this reminds you of the shape of the potential.
You have the harmonic potential here.
K is the force constant.
B0 is the equilibrium value, the units.
And then for each pair, you have the force constant and equilibrium value.
And this goes on and on.
As you can see, the force constants are on the order of hundreds of kilocalorie per mole.
This is a very stiff degree of freedom in a molecular system.
Later, you will have the section for bond angles in a very similar fashion.
That's your potential for functional form.
And then for each three groups of three atoms that are connected through a bond angle,
you have force constant and equilibrium value.
I don't want to confuse you with a very badly sort of additional term here.
Let's not worry about that at this time.
So the force constants are on the orders of tens of kilocalorie per mole here.
Radian squared in this case, softer degrees of freedom.
And as you can imagine, you have a similar set of parameters for dihedral angles.
So dihedral angles are described using a cosine function here.
And here, for each dihedral angle, you need to have three parameters,
force constant, periodicity, and the phase.
And this is exactly what you have here for each set of four atom types.
OK, so that's it.
I mean, the only difference between dihedral angles and bond angles and bonds
is that because you have a cosine function here,
you can have a series of these cosine terms
really describing complicated, complex forms of dihedral potential for you.
So you can have multiple definitions and multiple lines
for the same set of atom types in this case.
And you can have wide types.
Sometimes you have a star there.
That means that it's not a star.
It's an x, actually.
So that means a wide type.
So if you don't have any explicit definition for the dihedral angle
for that particular dihedral angle, the program
might actually use this wide types to get this something.
So here, for example, this is all wide types.
Too many columns.
I'm sorry.
So that's it.
And then at the end, you have the information
about non-bonded interactions, Lennard-Jones.
The charges are already stored in the PSL file.
You have them.
And the Coulomb potential is pretty straightforward.
So for the, oh, sorry.
This is an improper angle, which you can use in force planarity
or certain types of shapes into centers.
But then for the Lennard-Jones, you
need to have epsilon ij and R min ij for each pair of atoms
in your molecular system.
So the way it's done, it derives these pair potentials
or parameters based on the combination rule.
You specify epsilon and R min half for each atom type.
And then by looking at the average of the two R min half
values, you get the R min ij.
So that's the parameter for that particular pair potential.
You just add this and get the average of these two.
And for the epsilon, for the depth of the energy well
for your Van der Waals potential,
you look at the square roots of the product
of the individual epsilon.
So use the combination rule to derive this.
So this is just to give you an idea of where
the information is stored in case you want to change it.
And again, as I pointed out, the only difference
that's between the sort of coarse-grained representation
that we use today and atomistic representation
is at the level of the topology, parameters, and the engine,
the kind of functional forms that we have there.
Everything is actually the same.
OK, so any questions actually for these particular files
and how we use them?
So at the end, actually, I just talked
about another sort of approach that we started, actually,
in dealing with the problem of time scale.
And that's sort of, you're taking a different angle
at the problem.
And I guess, I mean, this is, again,
the whole problem is the time scale.
The major problem here is the time scale
that you want to be able to describe.
So we talked about using bias simulation
to accelerate certain processes to excite
certain degrees of freedom in the simulation system.
We talked about coarse-graining and how
we can produce a number of degrees of freedom,
how we can use software potentials
to make things happen faster.
So another kind of approach, and I think
this will take 5, 10 minutes, because I
think it might be interesting and enlightens
you for similar other molecular systems
based on this sort of concept of fragmentation, what we call.
So one example in which we have applied the system
is a case of membrane bilayers.
We talked about this.
We have the head group.
We have the core of the bilayer.
And the distinction between corollary interaction
of the head groups with water and hydrophobic fields
that don't like water results in certain type of structures
that you have to deal with.
So oftentimes, there are lots of proteins
that need to have atomistic description of their head
molecules, because they specifically interact
in certain type of head groups.
And sometimes, actually, these head groups
are necessary for their proper function,
just to motivate you.
In your blood, you have a lot of coagulation factors right now.
But they don't form a clot.
Because you don't have the right lipid molecules
on the surface of the membrane.
They just keep flowing in the blood.
As soon as you want to form a clot,
or your body wants to form a clot,
there are certain types of lipids that come out
from the inner leaf head to the outer leaf head.
And the exposure to the surface is the signal
for these proteins to bind to the surface.
And once they bind to the surface,
the whole coagulation cascade starts.
So the whole coagulation cascade depends
on the specific interaction of this head group
with this particular protein.
And that's what we wanted to do for this particular case
to describe.
But the problem is that if you want
to look at this process in using atomistic simulation,
binding of this protein to the membrane
doesn't happen.
Times get longer, right?
The same thing that we discussed.
So we have a lot of problems of similar kind.
The same thing happens if you put things
under very large conformation changes.
And this might not happen because one problem here
is that these lipid molecules are very slow, right?
We looked at diffusion of cancer on these guys.
This is on the order of 10 to the minus 8 centimeter
per second, per square per second.
So now, one thing that we did a while ago was to use SMD.
So we keep the bilayer and we apply a force
to the center of mass of this whole object.
And we slowly pull it down.
We run fairly long simulation back then,
four years ago, combining bias simulation
and further equilibrating the system
to see the process taking place.
Sometimes it was successful.
Sometimes it was not.
But we managed to have one single nice simulation in which
we described the insertion of this protein
in the membrane by this kind of simulation.
Very nice model.
We published it.
But we never managed to collect enough statistics
on this process and to find a specific binding site
for that particular lipid.
So it was a problem.
We needed to stay optimistic because otherwise we
couldn't describe the specific interactions
of the whole processes.
So we have a lot of proteins of this kind
that are known as peripheral membrane proteins.
So then we came up with this idea
that, OK, instead of coarse graining the whole lipid,
in which case you lose optimistic details,
why not going in and sort of liquidify
the core of the bilayer?
The core of the bilayer, as far as interaction is concerned,
is just a bunch of hydrophobic molecules.
And instead of having these long lipid tails that entangle
and they make the lipid molecules very slow,
we can actually replace the core of the bilayer
with an organic solid.
Hydrocarbons is still liquid.
It's still hydrophobic.
But it gives you much more fluidity
in the core of the bilayer.
Still everything is optimistic.
You don't have to worry about coarse graining approximations,
et cetera, et cetera.
By doing this, we are able to increase
the mobility of the head groups that are still optimistic.
Right?
So the important region of the system
is still described optimistically.
But in the core, those guys are represented by fragments.
It's like you go in and chop those heads into little pieces.
They become much more fluid and much more liquid.
And by doing this, without worrying
about coarse graining or other problems,
we have been able to actually increase the mobility.
So this is what we did here.
So again, we had to optimize.
Most part of the tail is chopped into little pieces.
But we kept a few carbons of the tail.
We had to calibrate this.
That doesn't matter.
But actually, by doing this, we have
been able to see the formation of bilayer, something
that people only could do in coarse-grained description.
Now we are looking at formation of the bilayers.
And they was actually one of the outcomes of this simulation.
So we have been able to see the formation of the bilayers,
starting from a completely scattered configuration
of these short-tailed liquids in 22 nanoseconds.
And that's really the key over here.
So by doing this, we can now repeat this simulation
multiple times, collect a lot of statistics.
And do this simulation forming the bilayer on your protein
or on your object.
Look at the reference of different kinds
of head groups for your protein.
Find binding sites for the protein.
So this has been extremely, actually,
successful in terms of capturing this behavior
of increasing liquid diffusion.
Membrane-like structures.
This is the same thing that I showed you in that movie.
You can see how nicely and how fast
the liquids converge to the two leaflets in nanoseconds,
using atomistic representations.
If you look at the distribution of different chemical groups
in the head group region, it matches very closely
what you get in a full membrane representation.
The face of the liquid bilayer is very similar
in full representation and the HMMM.
Even the position of the counter ion is matching very closely.
You only have, actually, wider distribution here
because they are more flexible.
And this we have been able to fix
by adding additional strain to the system.
So if you look at the diffusion,
the full membrane representation in 10 nanoseconds,
you can see that it's just locally fluctuating,
whereas this one is our new membrane model
in one nanosecond, 10 times smaller.
But you see a lot of motion here.
Of course, you can quantify this.
So these are diffusion constants for different models.
These are the full conventional membrane models
that people have been using.
And you can see that by going to this HMMM representation,
this diffusion constant by one or two orders of magnitude.
This is a very nice scale.
So we increase by two orders of magnitude
the motion of the liquids without losing
the atomistic equilibrium.
That's really good.
And now, in the simulation, we have
to do using bias, pulling simulations,
and sometimes working, sometimes not.
We have been able to do this 10 times,
really, in our cases, seeing binding of the protein.
In equilibrium simulations, now we
are multiplying very nicely with the membrane conversion
to a very similar one.
I'm not an experimenter, of course.
OK.
So and again, I think this is another case just to show
that this system is working for different kinds
of peripheral membrane proteins that
bind only to one surface.
This is actually a nice way of a nice model
if you are interested in binding or absorption
of other objects, drugs, for example,
or any membrane-associated phenomenon
that is slowed down by slow liquid diffusion.
It would be nice to see what happens
at the surface of the membrane.
We can again actually see how membrane binding changes
the conformation of the protein.
These red residues initially form the cluster,
but once they see the membrane, they
open and insert into the membrane.
This particular protein is actually
something that is involved in integrin activation.
Again, the details do not matter of the membrane-associated
phenomenon.
So let me skip this because these are unnecessary details.
The helix insertion problem that we
did using coarse-grained models.
And I told you, using mechanistic simulation,
it would be extremely expensive.
We can describe here.
So you can see what the helix is.
Please pay attention to this.
Correct answer.
This is extremely fast for what we
think is a fairly realistic representation of the membrane.
So most copies of the helices are inserted into the membrane.
It's a very nice way of looking at the effect
of the thickness of the membrane on the tilt angle of the helices
and how they interact with each other.
What is the thickness of the membrane
to have optimal interactions?
Because everything is happening 10, 100 times faster.
And still, you are preserving the logistic details
of the head groups that are important,
that might be important for that.
So OK, so that's.
And again, as far as the properties,
something that you need to remember
is that the mechanical properties of this membrane representation
is, of course, different.
So in a full membrane, you have different compressibility.
The compressibility is high, whereas in this representation,
you have it much lower by design, because it's
supposed to be looser, more flexible,
and moving much faster.
So that's one thing that we don't agree
with the full membrane representation.
But if you look at the electrostatic properties
of the system, so membrane bipolar potential, for example,
we are able to capture the core equilibrium,
which is important, reflecting actually
what the degree of hydration of the head protein
is right on.
And of course, when you look at free energy,
insertion of different aminases, again,
becomes important for protein interaction with this membrane.
We seem to be doing very well.
So in the core of the membrane, we
have some differences between our model and a full membrane.
The core is an organic solvent.
Too much fluidity.
We have much higher entropy.
And usually, they are over-stabilized
in hydrophobic residues.
In the head group region, the initial part, they are quite.
Anyway, so that's actually that.
I thought, actually, this is a sort
of a fairly different kind of approach
toward the problem of time-scale,
describing the system using fragments.
It's still a bit mechanistic, everything consistent.
You don't have to worry about problems that arise.
And you don't have the time-scale kind of description
with different kinds of liquids.
And actually, one additional thing
that we have been implementing in lambda and EMB
is that once you have your phenomenon described very well,
using this different representation of the membrane,
you can slowly grow the scales and make it a full membrane
in a stepwise manner.
So we grow them two carbons at a time,
simulate them a little bit, relax them a little bit.
We think it's a combination of one part of selection
of the solvent in the middle.
We essentially are eating the solvent in the middle
to grow the liquids and slowly make a folder.
And then I'll have to revise a little bit
for the kind of phenomena.
I mean, this is just to give you an idea for what systems
we are using them.
The variety of transmembrane peripheral membrane proteins,
drugs, viral peptides that bind to the surface of the membrane.
So if you have any interaction of some materials
with the membrane, you might want to actually consider this.
This is really useful.
OK, so I guess we are done.

