Molecular Dynamics Software (NAMD)
Sep 5, 2023
Molecular Dynamics Software (NAMD), including validation
Emad Tajkhorshid
This is one of the major simulation engines that were used for bio-molecular simulations,
for protein and nucleic acid simulations, for membranes, water ions, etc.
I think the major advantage of this program over other programs is that it's highly scalable.
This has been actually developed, partly here, on this campus.
This has been actually highly scalable.
Look at the performance of the program and if you compare it to competing programs,
you can see that actually really nothing beats this.
This has been developed here in the Beckman Institute,
and over the last 15 years or so, we have always had quite a few computer science students
working on the program, optimizing the code for parallel performance.
And this has been achieved, actually, on any new platform that came online
at the transfer computing center.
So it's really, I think it's fair to say, it's a standard MD code
for bio-molecular simulations at the whole center.
We have the program also installed on PCC machines,
including the FORGE machine that you can access to and run your simulations.
The scalability is pretty good.
It is available, actually, on all platforms pre-compiled for Windows, Mac, essentially the link.
And on Linux, Mac, there's no problem.
On Windows, it can get a bit tricky.
We can help you get the program installed and running.
So, again, I mean, the theory and the idea and the algorithm,
to some degree, you know a little bit about it.
Now, there are specifics about how to run a NAMI simulation that's covered to some degree.
And then later this afternoon, we will set up a simulation.
So the example that we will use is ubiquitous.
Ubiquitous is a protein that does some interesting biological stuff.
You don't want to know about that.
Let's not worry about that.
So this is the potential energy function that we talked about.
I have bonded terms and non-bonded terms.
And now I need, before I can start propagating,
before I start, before the program starts propagating,
the equations calculating forces,
we need to give the program all this information.
We need to give the program the topology of the system,
how many bonds for each bond,
the force constant, equilibrium values,
and all these parameters,
as well as the coordinate of the system,
should be introduced to the program.
There are different ways of doing that.
But in NAMD, sort of language,
these are coming from the geometry.
These are the geometrical relationship
of the coordinates, distances, bonds, and angles.
Then we have a set of parameters
that need to be introduced,
as well as partial charges.
And now in NAMD language,
we are going to deal with quite a few files.
I think four different files you need to prepare.
You need to have them before you can start an NAMD run.
One is the PDB file.
That's the coordinates of the molecular system.
Then we have the protein structure file, or PSF file.
It doesn't have to be a protein,
but that's the traditional name for that,
which includes the topology of the system,
how many bonds you have, what is connected to what,
as well as the charges, atomic charges.
And then all the other parameters
are stored in a parameter file.
So if you have these three files,
plus a configuration file,
which is your control panel, controlling the conditions,
specifying the temperature, how long, what time step,
what cutoff, et cetera, et cetera,
then you have all the files that you need
to run your NAMD simulation.
So what you're going to do,
under ideal conditions,
so today you're going to have all these four files
on your laptop, set up a very short simulation,
maybe two, three, four, five picoseconds of simulation
on your laptop.
Make sure that the simulation runs, everything is fine.
And if that's the case,
then you can copy all these files into FORGE,
your training cluster.
How should we do this best?
Do they, should they submit like half-hour jobs, or?
Yes, yes.
So we have about eight nodes reserved for this class.
Eight nodes.
Yeah, there's over 128 cores.
Uh-huh.
So, yeah, they can.
Okay.
And then there's a reservation,
so that means that all these jobs
will be submitted to the reservation.
So if they are running in a short period,
then they may be able to get results faster.
Okay.
So I don't think, actually,
that everybody will have a large molecular system
that really requires that supercomputing facility.
But, I mean, I highly encourage,
especially those of you who are here,
to learn about protein simulation
because you have a particular protein,
or any other molecular system that you want to run.
Start working on those systems here
to learn how to run them and to run into trouble
so we can help you.
Three of my group members are here
that can help us with the lab sessions
today and tomorrow.
So this is a great opportunity to get things going
and to learn the program.
So if you, for most cases,
you will be fine with your laptops.
I don't think we need to have the FORGE system.
But if you have a fairly large system,
you want to run it longer,
then we should copy the files over to FORGE,
run the simulations there,
and then we get the trajectory and output file
back on your machines,
and then we visualize the results locally.
In a few minutes, I'll write the path for the...
Oh, excellent, excellent.
I also asked, actually, my group members
to have a sort of run batch type of script
so you can use as an example to submit jobs.
But...
There's also a run batch.
Oh, perfect.
Okay, excellent.
Then we are all set.
Okay, but again, you will see
that we need to provide these files.
And now let's look at, actually,
we will look at these files tomorrow a little bit
so you know what information is there,
how to change them,
and that actually also is a good transition
into coarse-graining,
because coarse-graining is, I mean,
you are essentially using essentially the same terms,
but the definition of atoms and bonds
and things of that sort are different.
So everything is essentially done
at the level of the topology and parameters.
Coarse-grain, protein, atomistic protein,
everything is the same.
You just change the topology and parameter file.
Instead of having a metal group described with four atoms
interacting with another metal group with four atoms,
you have two beads that interact with each other
like two particles.
They are bonded to each other.
You will have a bond there,
but the force constant is going to be different.
The parameters that you will use for coarse-graining
is going to be different.
So let's look at the configuration file.
That's, by the way, again,
I sent you an email about this tutorial.
This is a PDF file that you can open on your laptop
and then go through it.
We don't want to do the whole thing.
This is usually for two or three days of exercise,
but we want to just cover the basics of NAMD
in this tutorial as well as basic type of analysis.
So you learn how to run the program,
how to get the simulation, and how to do basic analysis.
There are advanced kind of examples in that tutorial,
and you're more than welcome,
if you feel comfortable already with NAMD,
you can use those advanced examples
and run those simulations.
So, okay, so let's look at the configuration file again.
The files that you need, the structure file is specified.
Here you can actually use both TCL commands
like sets, loops, and things of that sort
as well as NAMD-specific commands.
You specify the structure.
The coordinates are in this file.
You set the temperature.
That's your target temperature in Kelvin,
and this is used to randomly assign the temperature.
Anything that is set using the command set can be used later
as a variable using the dollar sign.
So you set this variable to 310 and then use
that to specify the temperature of the system.
You can have other variables.
For example, the file name is input files, all the input files.
If they have the same kind of file name,
you can specify that as a common variable here using the
command set and then use it in other commands.
That's a sort of TCL kind of scripting language.
If you are starting a brand new simulation, of course,
you have to specify the temperature,
and that temperature will be used to assign velocities.
If you are continuing a previously run simulations,
you already have some velocities from the previous step, right?
Always calculate the new velocities, new coordinates,
forces, new velocities, new coordinates.
So you already have a set of velocities and a set of coordinates,
so you can use those restart files to continue a simulation.
You want to run a microsecond, half hour on FORGE.
You can only run five nanoseconds.
You want to be able to restart the simulation exactly from the
point that you end it.
So there is a way of continuing the simulations that is
specified in this block.
Again, don't worry about the syntax and the details.
Just know that these capabilities exist.
So that's the main idea of showing you a little bit,
pieces of this configuration file.
You can also specify the output file name,
how frequently you want to save the restart files that you want to
use to restart your job later,
how frequently you want to save the coordinates.
You are running every femtosecond.
You are calculating the coordinates, forces, velocities, positions.
So do you want to save the trajectory of this huge protein
every femtosecond?
Probably it gives you a very large file that you don't want,
so you can specify how frequently you want to save the coordinate.
And actually related to that, this is also specifying the size
of the system because it fluctuates.
You want to save the size, the extensions of the system at the
same frequency as the DCD.
But the DCD file, that's the file extension for trajectory showing
the position or monitoring the position of the atoms in your
simulation system.
That is something that you will use later for your analysis.
We usually don't save the velocities unless you really need them
for, for example, velocity autocorrelation or other kinds of analysis.
So usually we just monitor the position of atoms to see how the
protein or how the molecule is fluctuating and moving.
The same applies.
You can output the energies.
For all the terms that we described in the potential energy function,
you can have the energy at each time step or a certain frequency,
as well as you can see how fast the simulation is running.
You don't want to monitor this too frequently because that slows
down the program.
So about every thousandth step or so, you can specify that.
It tells you how fast and how efficiently it is running,
so you can monitor that.
So as far as the simulation conditions, excuse me, is concerned,
this is the name of the parameter file.
So I told you you need to have a parameter file.
Excuse me.
This is the so-called all atom char 27 parameter file for protein
and lipids that we are using.
If you have different molecular systems,
you can use different parameter files.
Some, okay, so that's something I forgot to mention.
So when you are doing a molecular simulation,
I told you all non-bonded interactions are calculated
between every pair of atoms.
But if these atoms are bonded to each other, of course,
you don't want to recalculate non-bonded interactions
because the interaction between two non-bonded,
two bonded atoms is described by a different set of rules.
So bonded interactions are excluded essentially
from non-bonded calculations.
So now you do this for, of course,
atoms that are immediately bond to each other.
You do it for atoms that are connected through two bonds
because they are connected through the bond angle.
But when it comes to the bond,
to the atoms that are connected through three bonds,
forming a dihedral angle,
now the question is whether these two atoms in space
should interact through non-bonded interactions or not.
And the way the force field is set up
is based on certain type of inclusion or exclusion
of this 1, 4 kind of interaction.
So that's what you specify here,
whether you want to have any scaling involved
and any switching,
any scaling or exclusion and inclusion.
You specify this based on the force field.
So cutoff, what is the cutoff distance that you use
for calculating non-bonded interactions?
Mostly for van der Waals,
but also for some part of your electrostatic calculation.
Switching distance, you start switching your function,
start going towards zero at 10 angstrom distance.
So by the time you reach the cutoff,
your interaction is completely gone,
so you don't have this continuity
in your energy and forces.
So you usually, when you specify a cutoff,
you usually also need to have a switching distance.
Otherwise, if an atom appears and disappears
from your cutoff distance,
you will have discontinuity of your energy and forces.
Exists, doesn't exist completely,
so you want to have a smooth function
connecting these two regions.
That's specified here.
Pairless distance, so again, you want to see,
you want to only calculate the interaction
between every atom and atoms within 10 angstrom distance,
but just calculating what atoms are
within 10 angstrom distance is costly.
You don't want to repeat this kind of calculation
every time step,
because you have to evaluate all the distances.
So in order to avoid that,
you would define a region that is larger
than your cutoff, 14 angstrom.
You find all the neighbors within 14 angstrom
for each atom,
and then this is not going to change
in the next 10 or 20.
So if you just evaluate atoms
that are within 14 angstrom distance
of your central atom,
there is a very small chance
that you miss any of the atoms
for which you want to actually calculate
the interaction energies.
So by doing this trick,
you will limit the evaluation
of 12 angstrom distance atoms
to a smaller region,
and the likelihood that an atom
moves three or four angstrom
coming from the outside region
all the way into the interacting region
would be very small.
So by this trick, you can evaluate
the neighbor list.
Instead of every time step,
you can do this every 20 or 30 or 40,
depending on your molecular system.
So these are some conditions.
Each time step is specified here.
I told you one femtosecond
is usually the regular time step,
but if you rigidify some of the bonds,
including the hydrogens,
then you can take longer time steps.
So that's what is specified here.
Then you can have multiple time stepping.
For example, nonbonded frequency.
Every time step, you are going
to calculate nonbonded forces,
and non-PME part of nonbonded forces.
Full electrostatic energies,
which are calculated using PME,
will be done only every other time step.
So this is the multiple time stepping
allowing you to make your simulation
even faster.
So this is about the Langevin dynamics.
So we want to have...
We turn on Langevin dynamics.
That allows us to control the temperature.
So we want to...
The target temperature is the same temperature
that we initially specified, remember,
using the set command.
We use that one also
for our Langevin dynamics algorithm.
This feature doesn't affect hydrogens.
Hydrogens are too light, usually.
So you usually don't want to mess
with hydrogens too much.
If you are doing bias simulation,
if you are subjecting atoms
to some sort of random noise
or other external type of forces,
you usually don't want to include hydrogens.
They have a very small mass,
and their velocity can be affected significantly
if you add forces to them.
So we usually leave them alone
and work with heavy atoms of the system.
The Langevin dynamic damping coefficient,
that's gamma that we talked.
This should be actually 0.5.
Please actually...
Even if you have 0.5...
Even if you have 5 in your example,
please change it to 0.5.
This is historically...
I don't know if that's been corrected,
but we don't want to damp the system too much.
5 is way too much.
You see, I mean, we have a problem with time scale.
The system doesn't move enough
for us to see anything significant.
Now you go in and damp it significantly,
you're not going to see anything.
So we usually try to minimize this
just to have it enough to be able to control
to some degree the temperature of the system.
So please correct this to 0.5
in case it's still in the configuration file.
So I'm going very quickly over this
just for you to have an idea.
If you have more sort of specific questions
and you want to change this,
we can discuss it more in the afternoon.
So we usually use different kinds
of periodic boundary conditions.
And I think for that,
let me show the next slide first.
So this is our simulation.
Let's say this is your unit cell.
This is what you want to simulate.
But of course, if you don't have anything around it,
all these atoms or molecules
are going to evaporate into the vacuum, right?
So you want to have something
that helps you maintain the correct density of the system.
And then at the boundary,
these molecules don't want to interact with vacuum.
So you should have something real here.
So in order to avoid this problem,
this boundary problem,
instead of just simulating a simple system like this,
we are going to copy the unit cell.
The actual atoms only exist in the unit cell.
But then we have images of the unit cell
to the left, to the right, up, down,
front of us, in the back, in all three dimensions.
And these images, all these atoms,
follow the position of the atoms in the central unit cell.
So by doing that, you will have actually...
It looks like you have another box sitting
to the right, to the left, to the up and down.
All the boundaries of the system will be occupied.
For example, for a solvent,
you will have solvent all around the unit cell.
So here at the boundary,
you will be interacting with other solvent molecules.
So it's much more natural,
especially for a bulk system,
to be treated using this boundary condition.
We call this periodic boundary condition.
This is also required for PME,
particle mesh evolved,
calculating electrostatic without truncation.
So you can see actually there is a lot of advantage
without really increasing the size of the system.
Significantly, we have taken care of
having real boundaries for the system.
This is...
The size of the boundary system is specified here
using three vectors, x, y, z dimensions,
and that's the origin of the cell.
So there is...
This block is interesting because...
So what happens in periodic boundary condition is that...
So these atoms...
Let's say you start with a box of argon atoms,
and they start diffusing around.
After a while, the initial box
that looked like a very nice rectangular box
will look like a cloud in space.
So if you want to monitor that, visualize that,
things become a little bit messy.
So in order to avoid that, what we do,
we say with this ramping command,
what we do is say, okay, if an atom leaves this box,
that means its image
from the neighboring cell would be coming in.
So we stop monitoring the copy that left the box,
and we start monitoring the position of the image.
So that way, all the focus will be always
on the unit cell in the middle,
and the shape of the molecular system
stays sort of rectangular.
It's nicer to visualize it,
and certain type of analysis might be easier,
et cetera, et cetera.
The only problem is that if you want to really monitor
the actual diffusion of these particles,
then this ramping becomes a problem
because you artificially are reducing
the total displacement of this particle
from its original position.
You're just bringing it back always.
So in that case, you have to unwrap your trajectory
or you turn off this feature.
So this is what is specified here.
So these are really just really the most basic features
that are specified in the configuration file.
You can change them.
You can, of course, use the default values
except for that gamma.
Please check that.
Otherwise, you can use the example
that is provided in the tutorial files.
So now, in addition to periodic boundary condition,
this is, I think, the best way of doing things.
It allows you to control the pressure as well
because you have a well-defined volume.
You can actually change the volume
to reach a target pressure.
But it can be expensive
because you have to include a lot of,
for example, if you have a protein,
you place a protein in the middle
and you have to fill the box with solvent or other things.
So that increases the number of atoms in the system.
So for some cases, people have been using
a different kind of boundary conditions.
For example, you can have a spherical boundary condition
in which you will have solvent around your protein,
but then you are essentially cutting the corners
from a rectangular cubic kind of cell.
So this way, you reduce the number of atoms.
Of course, things are more difficult.
It's a little bit difficult to control the pressure here,
and you don't know if you have the right density of water
within this sphere or not.
But that can actually significantly reduce
the number of atoms,
especially if you have a sort of irregularly shaped protein,
sort of like an extended protein.
You might ask why we are including solvent
in this kind of calculations.
So there are lots of what is known as hydrogen bonds
within proteins and between proteins and water
that are very important for their function,
for their structure.
And if you don't have those hydrogen bonds
described properly, then the behavior of the system
might not be realistic.
So that's why, because biology happens in water,
we have to have a representation of the solvent in the system.
So before we had access to large computers
and running large simulations became easy.
Many people used implicit solvents,
like just having dielectric kind of effect
present in the system.
That actually reduces the number of atoms
in your simulation system,
but again, explicit hydrogen bonds
can be very sensitive to the presence of water.
So nowadays we have really access
to large computer clusters.
The size of the system is not the major limiting factor anymore.
So you can easily go to 100 or 200 processors
if you have a larger system.
Including explicit solvent in the simulation
has become a sort of common practice,
which is more accurate.
So there is another problem associated
with explicit solvent simulations.
Anybody wants to guess?
So I talked about the effect of hydrogen bonds
having a dielectric effect
or having explicit water molecules
that can actually replace a hydrogen bond for you.
Helping you break some hydrogen bonds in the system
makes a big difference.
But there is another major difference
that takes place when you introduce explicit solvent.
Anybody wants to guess?
There are lots of them, actually.
But another one that might make a big difference
in terms of what you see in your simulation,
how much motion you see in the simulation.
Friction.
So your protein, if you run it in water,
because it has to displace water
when it wants to undergo conformational changes,
becomes much slower.
You run 220 nanoseconds, your protein,
with explicit water, nothing happens.
You do it with implicit water,
suddenly things move much, much faster
because you have less friction.
There is nothing to cause friction for you.
And if you want to have the right hydrodynamic behavior
in some cases, because your system is so huge,
there is no way you want to include
explicit solvent in your calculation.
Then you can go back and take advantage of the gamma,
the friction, damping coefficient that we talked about.
So you can increase gamma to a degree
that gives you, actually,
the effect of explicit water in your simulation.
So that's something that, again,
especially in coarse-grained simulations,
when you go to really, really, really coarse representation,
it doesn't make sense to have water explicitly in the system.
But if you want to have the mechanical properties
and effect of water in your simulation,
you can take advantage of gamma,
increase the gamma to a value that represents
the correct sort of water environment for you.
And this is, again, the block in which you specify.
You turn on PME, and you specify the grid sizes.
So there is actually a rule here.
You are trying to have a product of prime numbers here
that gives you the closest value
to your x, y, z dimensions of the unit cell.
Okay.
So some of the additional functionalities
or things that you can specify in the configuration file
that becomes sometimes very, very useful
is to constrain and fix atoms.
You can turn on fix atoms
and then specify what atoms are fixed
by flagging those atoms in this file,
in this particular column.
We can talk about that a little later.
You can fix atoms, or you can constrain atoms.
Anybody wants to guess the difference?
It's almost obvious, but...
What's the difference between fixing an atom
and constraining an atom?
Fixing means it doesn't move at all.
Constraining means you are adding
another harmonic potential on top of the force field
that you have.
It can still move and fluctuate,
but it's not going to be able to move too much away,
depending on the strength of the harmonic potential.
So these have different functions
and different effects
on the speed of the calculation
and many other properties.
So when you fix an atom,
it doesn't make sense to calculate
the interaction between two fixed atoms, right?
Because they are not going to move anyway.
So there is no reason to calculate
forces acting between two fixed atoms.
So if you fix a large part of your simulation system,
you can significantly increase
the speed of your calculations.
Make sense?
For example, you have a large protein.
I have an ATPase protein,
and I'm only interested in the binding site
to see what happens if I mutate
a couple of residues in the binding site,
how many water molecules come in.
You can fix the rest of the system
and only allow the binding site residues
to undergo dynamics.
So this way you can significantly increase
the speed of your calculations.
If you constrain atoms,
you are calculating additional terms.
It becomes more expensive
because on top of your force field
now you have these harmonic potentials
acting on those constrained atoms.
So there is a fundamental difference
between fixing and constraining.
Usually constraining is better
unless you really want to exclude
a large part of the system.
As I pointed out,
initially you want to minimize the system
in order to reduce bad contacts,
high force regions, high temperature regions,
however you want to call it,
to minimize it, to bring it to a sort of
local minimum configuration
before you start giving temperature to the system
in order to have a stable simulation system.
So that was the essentials of the input file.
So in the output file
you have the references, etc., etc.,
how many processors you are using.
That's actually a pretty good line here
to make sure that you are taking advantage
of all the cores that are available.
You have the energies, as I pointed out.
So this whole thing is one line.
This is the energy title, time step,
bonded energy, angle energy, dihedral, improper,
electrostatic energy, van der Waals.
All the energy components are listed here.
You have the pressure, volume of the system,
as well as the temperature.
So you can plot these nicely
and monitor the behavior of the system.
I'm just wondering what the system is
because we don't have any bonded energy at this point.
So then you see the trajectories
are written at certain frequencies.
This is the performance information,
how fast it is running.
You can get an estimate of how long it takes you
to run a nanosecond or two nanoseconds
or whatever you want to simulate the system for.
The example that you will be actually working with
is the ubiquitin protein.
You can solvate it in a water sphere like this
or in a water box.
I think you have example files for both systems.
If you have problems setting up the system,
you can actually make use of the files
that are already in the tutorial files.
These files are there.
You can actually take...
The way this tutorial is set up is that you have steps.
So you have 1-1, 1-2, 1-3, 1-4, 2-1, 2-2, 2-3.
So essentially you can do each step on your own
and prepare the files for the next step
or you can take the input files from the directory
that is in the zip file.
I sent you the link to the tutorial PDF file
as well as the required files.
They come in a zip file or a gzip file.
You can download.
You will have the whole directory structure.
So you can go to that directory and get, for example,
a fully prepared, ubiquitous water box
because you are not interested in setting up the simulation.
You just want to run the simulation
and get the results and analyze it.
You can skip some of the steps if you want.
So these are the two examples that give you the idea
of how to run a NAMD simulation.
Then in terms of analysis,
these are the equilibrium properties of the system.
We want you to be able to look at RMSD of individual residues.
You can actually extract the Maxwell-Boltzmann distribution.
This is another example, 2.12.
You have all the files.
You don't even have to run the simulation.
Again, if you just want to run the analysis,
we have pre-calculated the trajectory for you.
You can get the trajectory and do the analysis on those files.
But I highly recommend to go through all these steps
so you feel more comfortable next time you want to use NAMD on your own.
Analyzing energies, temperature distribution, and specific heat.
These are actually some of the examples that are in the tutorial.
There is a sort of advanced example on calculating the echo effect,
but it's way too advanced.
And again, you can do this on your own later.
I don't recommend that for this sort of short two-day school.
So just one note on the membrane system, especially since you were asked.
So we usually start with the protein structure,
and then we try to provide the environment.
The environment could include membrane.
Usually we have included water to have the correct environment of the system,
and that results in a sort of significant increase in the number of atoms
that you have to simulate.
But again, that is really the correct environment of the simulation system.
So for membrane-protein systems,
in case there is somebody who is interested in membrane-protein systems,
we have a separate tutorial for preparing proteins in membrane
and how to simulate and analyze them.
So you can use that tutorial as well.
So you will be preparing your molecule, getting the PBA and PSF file,
minimize it to relax it.
Then you heat it to the desired temperature,
and you run equilibrium simulations under different ensembles,
under different conditions.
You can have constant temperature or constant energy.
You can have constant volume or constant pressure.
Again, depending on the ensemble,
depending on the conditions that you would like to study your system.
Get your data.
So far, this is trivial.
This is what you learn today and tomorrow.
No problem.
Or later, you can continue with these tutorials.
But then you have to analyze these results and make discoveries.
That's the absolutely non-trivial part of this business.
But again, that's something I don't think we can teach you here,
but hopefully you will manage to do that.
So tomorrow we will take a quick look at what is inside these files.
I think today would be too much.
And we talk about parametrization a little bit.
But let me actually jump to,
so we will also talk about bias simulations tomorrow
and see what if we want to apply forces
and we want to induce certain molecular events.
A couple of tricks that you can use.
But for today, I would like to actually,
very briefly in the next five or six slides,
just give you one example of application of the method of MD
to a biomolecular system and to demonstrate hopefully
that this, in general, molecular simulation
could be an extremely powerful method
if you really take advantage of what is possible in this system.
So this guy is an example.
The details do not matter here.
So this is a protein that acts as a water channel.
It sits in the membrane.
If you don't have this protein in your kidneys,
instead of one or two liters of urine,
you will have 20 liters of urine, which is pretty inconvenient.
So some membranes need to be transporting a lot of water.
And if you don't have this protein, you will not have this efficiency.
So this protein essentially increases water conductivity
of certain membranes in humans and in many other organisms, actually.
So this is actually the result of five nanosecond MD simulations
that we did now almost 10 years ago.
So again, this is what I showed you earlier.
So you can see due to fluctuation, thermal fluctuation,
of water on both sides of the membrane,
you have these transient surges inside the protein,
how water molecules move.
And we could describe this behavior.
For the first time in a membrane channel,
we are looking at the full transport without any bias
based on the approximate potentials that we described.
So for all those crappy force fields that we had,
still we could actually see something interesting.
So this was nice.
And we watched this simulation,
and we thought actually that the behavior of water molecules
in this particular water channel was very strange.
Because again, this was an example in which we watched
and we saw something, and then we quantified that.
So if you look at the water molecules here,
you can see this is already now 10 years old.
But if you look, you can see that water always comes in
with oxygen sort of heading in first.
In the middle, it turns 180 degrees and goes out like this.
So really, it comes to the middle like this
and goes out like that.
So we observed this in a couple of snapshots.
Maybe it's random.
Maybe it was just those snapshots.
So if you analyze this, once you find this feature,
you go in and look at the average behavior.
This is what you see, distribution of oxygen
and hydrogen molecules over the whole 4 nanosecond production
run.
And you can see that this confirms oxygen in the middle,
hydrogen's on the side.
After this, you have oxygen, hydrogen, oxygen, hydrogen,
oxygen, hydrogen, really confirming
that this is a preserved feature in this channel.
So we thought we had discovered something very nicely.
This is actually something unique.
Water doesn't like this bipolar distribution.
This is now, if you include also the fluctuation,
you can see that in the two half channels,
it has opposite orientation.
This is a dipole moment of water.
It fluctuates, but it never flips
to the other orientation.
So this is unique for water because water
likes to have a polarized chain, all the water
molecules in the same direction.
If you look at it in nanotube systems, that's what you see.
It's either like that or like this.
So something in the protein is causing this.
We thought actually we had a very interesting feature,
so we wanted to go publish it in science.
In fact, we did publish it in science.
But instead of stopping here, so we
decided to also characterize what in the protein
is causing this effect.
And that's where the power of simulation comes in.
So you can change a lot of things
in your simulation system.
You can mutate side chains.
They do it experimentally.
We have to do it too.
You can mutate backbone residues.
You can assign completely unrealistic charges.
You can have a negative sodium ion in your simulations.
Right?
Just change a number in your parameter file
and you have negative sodium ions.
You can make a system extremely artificially light,
artificially heavy, additional constraint.
You can really design simulation systems
that address cool hypotheses.
And I think that's, again, going back
to the discussion about the prediction of the simulation.
So you can really take advantage of this tool
as an exploratory tool to test different hypotheses.
So what we did here is said, OK, so let's
see what in the protein is.
And that's actually my take home message,
another important take home message.
So you have an advantage here that there
are conditions that experimentally cannot be
realized.
You can do anything.
You can play God in your simulation.
And if you don't take advantage of this important, powerful
aspect, then you are not taking 90% of what's there for you.
So you do basic simulation, RMS, OK, that's it.
So let me give you an example.
Try to use it.
Try to be playful in the beginning.
Don't worry too much about having the correct conditions.
Let's go to extreme conditions.
If something happens there, we can step back and do it
more carefully and see if you understand that phenomenon.
If nothing happens at 500 degree Kelvin,
nothing is going to happen at 100 Kelvin.
So take it to extreme conditions.
Try to be playful and take advantage of your power
over the conditions here.
So what we did is say, OK, so this is the y-type behavior.
Now I'm speaking in experimental biological language.
The y-type behavior was water was flipping.
Then it goes from one half of the protein to the other.
So let's start turning off electrostatic interactions
coming from different parts of the protein.
And we started doing that, monitoring
the behavior of water.
The first mutation didn't cause much effect.
The second one was significant, but not enough.
But when we did them together, when we mutate those effects,
then you can see that water sort of recovers
this normal configuration.
So through this computational mutagenesis experiment,
we proved that, we showed that actually at least
within the simulation approximations,
within our model, what part of the protein
is responsible for this behavior.
And this is something that I don't think
can be experimentally tested even.
So it's not reachable, these kind of modifications
that we are doing here.
So this really made a big difference
in terms of the significance of our finding.
It wasn't like just we saw something,
we don't know how it is happening,
we don't know where it comes from.
It's interesting.
So we really can show what part of the protein is responsible,
what part of this complex molecular system
is responsible for these effects.
Actually, after this work, this generated
a lot of additional work.
People have been trying to design
the same kind of electrostatic patterns in nanotubes,
for example, to see the same behavior.
And it made a lot of important impact.
So that's, please remember that this
is a power you have in your hand and take advantage of it.
OK, good.
So I guess, so we can have, let me see actually
if we have questions at this point.
Yes?
Were there experimental observations of this?
For this?
There is no experiment.
So they.
Just the observation of the water molecule.
So there is no experiment that can measure this.
So this is at this stage, I mean,
they can see the position of this protein structure
was done in the presence of glycerol
at a very high concentration of glycerol.
They didn't know where the water molecules
inside the protein were at the beginning.
After we did this work, our collaborators
actually increased the resolution of their structure
and removed the glycerol.
And they could see the oxygen atoms of water molecules
inside the channel, which matches very nicely
with our peaks of density taken from the simulation.
But they don't see the hydrogens.
It's really sort of one of those cases
in which experiments cannot measure it, at least not now.
Maybe in the future.
Actually, one comment is that the simulations can give you
insights that experiments may not be able to give something.
Probably important to mention.
The simulations can give you insight that?
That experiments don't have.
Absolutely, absolutely.
Yeah, that's again, I mean, sometimes it's completely wrong.
We know all the approximations.
We know that there are lots of shortcomings.
But you know, if I get five of these,
and even one out of five is correct,
we are still doing great, something that cannot be done
even experimentally.
And we are getting better and better
in terms of understanding where things might fail,
binding affinities, or where actually things are lost.
In this particular case, I forgot the sort of recipe
I gave you, just to make sure that this is not
the effect of bad charges and things like that.
We scale charges down and up for water.
This behavior is sort of robust.
So this bipolar configuration of water
is not an artifact due to bad charges.
People did these simulations with different force fields,
with GROMACS, using a different set of force field parameters.
They saw the same behavior.
So it seems that this is one of those phenomena
that is robust, insensitive to the choice of force field
or approximations in the force field.
So that's one example also of that sort.
Good.

