While very impressive, I'm not sure about their goal:
> Modeling genes at the atomistic level is the first step toward creating a complete explanation of how DNA expands and contracts, which controls genetic on/off switching.
That does not work in isolation like this: McGuffee and Elcock did some fantastic work in 2010 where they showed that protein stability is quite dependent on that the cytoplasm is insanely densely packed. The same is also true for the nucleus of a cell. You can't make any serious claims of a bottom-up explanation of gene regulation and expression unless you can model the entire nucleus, which is out of our reach for years and years to come.
And that is before I start on whether this is an actually useful way to research gene expression in general, which I'd very much argue against.
Depending on how they model the simulation's fields, long range intra-atomic forces, and boundary conditions, I'd still say this kind of molecular dynamic simulation is better than doing nothing and waiting to simulate the whole nucleus. It could have accounted for these sorts of effects in a way that is cruder than whole nucleus simulation, but I haven't read any papers about this simulation.
Fascinating discussion! Could any of you knowledgeable folks say a bit more about 1. why whole-nucleus simulation is so hard and out of reach, 2. what role simulation currently plays in understanding gene expression, 3. to what extent will simulation be important for understanding how genes translate to phenotypes over the long term?
I know next to nothing about this field or the relevant biology, but would love to learn more!
Imagine you're doing a straight-forward computer simulation. You have a state-of-the-world, a time-step, and rules for how the world evolves during that time-step.
You might write your own orbital simulation; the positions of the various planets and asteroids are your state, you might use a time-step of a minute or an hour or a day, and your rules for how the state evolves during each time step might just be Newton's laws of gravity and motion. You can probably update your state in a fraction of a second, even if you are simulating hundreds or thousands of planetary bodies, and you can simulate eons in hours.
The problem with simulating biology on an atomic level is that your time-step needs to be in the femtosecond to picosecond range, because atoms can move fast, and small differences in their position make big differences, but interesting things like protein folding can take minutes of real time.
So to simulate anything interesting, you need to step through quadrillions of time steps.
Oh, and you're running this on more than one processor, so you need to do smart things so your simulation isn't trying to synchronize all of its processors on every step; oh, and we can't tell you the exact position of every atom in a cell to start with, or even necessarily all of the different types of molecules that are present. Oh, and all of these cells -- or pieces of cellular machinery -- behave in ways that are dependent upon the environment they are immersed in, so you also need to simulate that somehow.
There's a reason Randall Monroe described the difficulty of protein folding as "we may one day find a harder problem".
I think that the entire enterprise is a red herring.
It's nigh-impossible to understand this system by working from the bottom up. I don't see what kind of insight you'd get from this that you wouldn't get from a more understandable simplification in a minute fraction of the time.
For example: yeah, if a gene is contained in densely packed nucleosomes it'll be harder to express. But knowing that is enough; the exact details of how that works atomically is not the issue here. There are much larger issues with gene expression we don't understand which can much more easily give us much more insight into the workings of the system.
>And that is before I start on whether this is an actually useful way to research gene expression in general
Haven't we already reproduced existing bacteria/DNA from atoms which has allowed us to more easily study gene expression?
Moreover, we take a bacteria with 200 genes ( we understand a lot of the gene functions/expressions, but we don't understand what all these 200 genes do). Thus, we synthetically reproduce the bacteria from the atoms up while removing nearly half of the genes that we determine are not needed to sustain the living bacteria, and the resulting synthetic bacteria we created in the lab allows us to more readily isolate and study the individual genes we didn't previously understand.
Of course gene expression in bacterial isn't the same as humans...obviously when building synthetic life from atoms up we aren't going to just reproduce a human in our first effort, we start with a simple bacteria, and because of the complexity (and our lack of understanding of the gene expression in even simple bacteria) we purposely reproduced the bacteria with about 1/2 the genes for the sole purpose of studying those genes.
But maybe I am wrong and this is in no way helpful for studying gene expression.
>pruning its genome came to something which was minimally viable
That sounds something like CRISPR or gene modification of an existing organism genome, maybe even taking genes from one species and inserting them into another (taking bioluminescent jelly fish and putting certain genes into cats or plants to make them glow). But, I'm talking about more recent successes of actually reproducing the DNA (synthetic DNA) from the atomic level up not modifying existing "natural DNA".
Sorry, I don't have enough time to search for recent studies, but here is old article about first "synthetic DNA" we created and inserted into cells that took off and self-replicated back in 2008: https://www.wired.com/2010/05/scientists-create-first-self-r...
The link seems to reference the first success of fully synthetic DNA, but I have recently been listening to podcasts about more recent experiemnts/studies, specifically reproducing bacteria with full synthetic DNA (which is stripped down as much as possible to maintain viability) for the purpose of studying gene expression and hopefully begin to identify what the currently unknown genes sequences do.
you can get 95% of the effects of macromolecular crowding by just shoving PEG into the system (the very first paper we covered in grad school showed that proteins in packed cells have elevated kinetic rates and free enegry of ATP hydrolysis isn't the value measured at STP, by using PEG as a proxy).
If I were to attempt to do gene regulation studies by MD (I share your skepticism!) I would just put a bunch of PEG in as a water replacement.
In fact that's a paper right there- "Effect of increasing macromolecular crowding on the accuracy of DNA simulations for regulation prediction" :)
This is the "hard part" with atom-level biological simulations. You need extremely tiny timesteps -- like, sub-picosecond, even -- but many interesting molecular events happen on the timescale of minutes.
Game of life doesn't have long range interactions, and doesn't have to deal with electrostatic interactions (which, with parallel B.C., diverge and cannot be calculated in real space).
when I wrote my thesis in 2001 I got 1ns/week on a much smaller system (1M atoms running on a T3E w/ ~256-512 processors).
It's not really clear whether MD simulations like this are true contributions or whether they will ever produce useful results compared to well-parameterized neural nets.
Are NNs producing useful results in this area? I'm on the density functional theory side of things, so much smaller scales than MD.
To me, it's not clear whether we are good at quantifying when a neural net is well-parameterized in atomistic simulations.
The limitations of MD simulations and all the built in assumptions are well understood. We can reason about where our simulations might fail, even for very large systems. It's always a question about whether the underlying model can capture all the physics that occurs.
With a NN or other black box model, there's no way to reason about it except for benchmarking test/validation sets that (hopefully) capture the physics that you care about, but you still cannot really reason about how well the model will extrapolate to multiple interacting physics with different magnitudes at different length scales.
I don't think anybody has shown that NNs produce better predictions of protein models (yet).
The limitations of MD aren't well understood. We don't know the implications of using a polarizable water model other than to say "it should be more accurate".
All the statements you make about NNs apply to MDs since MDs are basically feedforward ASTs with the same level of complexity and non-linearity as NNs.
in 2011 I was building 40 000 atom models on a 2001 SGI Fuel simulating them on a single 8 core node at 1ns/week -> 8000 atoms/core. They're about 7 times faster then me 8 years later.
So from a computation point, meh.
From a tools point this is cool. It isn't trivial to build or analyze a model of 40 000 atoms, never mind 1E9 atoms.
Biologist's perspective: this paper is not about the biology. The simulation performed here has zero biological interest - the point of the paper was to show how efficient and scalable their software is. This article about the paper is terrible, but honestly I feel like they should be given a pass - it's hard to justify to a lay audience that understands neither DNA nor memory bandwidth why you would choose to study something that is of no use to the field.
There's an inherent tension when doing a method development paper - if your result is too fantastical it's hard to know whether it's an artifact of your (potentially faulty) technique, and if the thing you're studying is well-understood then it serves as a good control, but it's less interesting. I suspect they chose the latter path since it requires no validation via existing methodologies.
I'm not familiar with this method, but NAMD and CHARM are a great way to inspect biological systems at greater resolution than experimental methods can measure. There was really cool work on fibrinogen about a decade ago showing how it provides elasticity.
One engineering model is to put the code on github under an open source license so people can use it. But please dont just put it out there and walk away. If you want to develop a useful piece of software, find a way to maintain it.
Another model is to offer it for sale as a product. Either way, engineers dont really care about publishing, we care about making useful things and Shari g how-to.
Most engineering fields have journals and publish their research in the same manner as scientists. Whether they care about publishing or not is related to whether their position is in academia or industry.
I agree this paper isn't really useful for biology. LANL maintains team of pseudo-biologists who scale up MD simulations and publish papers like this periodically (https://www.lanl.gov/projects/karissa/Ribosomes.html shows some previous work).
I wish this was a field where you could just apply more CPU and get better results (I say that with multiple papers published under my name which claim exactly that) but I'm skeptical. The force fields we use are far too approximate (point charges, assumption of transferrability) and it really does seem like we don't actually need most of this information to answer the questions we want to answer- those can be answered using more parsimonious techniques.
I don’t see how this would be of “zero biological interest.” Is it because presently it is not that useful or because it will never have any applicability?
34 comments
[ 3.8 ms ] story [ 35.3 ms ] thread> Modeling genes at the atomistic level is the first step toward creating a complete explanation of how DNA expands and contracts, which controls genetic on/off switching.
That does not work in isolation like this: McGuffee and Elcock did some fantastic work in 2010 where they showed that protein stability is quite dependent on that the cytoplasm is insanely densely packed. The same is also true for the nucleus of a cell. You can't make any serious claims of a bottom-up explanation of gene regulation and expression unless you can model the entire nucleus, which is out of our reach for years and years to come.
And that is before I start on whether this is an actually useful way to research gene expression in general, which I'd very much argue against.
I know next to nothing about this field or the relevant biology, but would love to learn more!
Imagine you're doing a straight-forward computer simulation. You have a state-of-the-world, a time-step, and rules for how the world evolves during that time-step.
You might write your own orbital simulation; the positions of the various planets and asteroids are your state, you might use a time-step of a minute or an hour or a day, and your rules for how the state evolves during each time step might just be Newton's laws of gravity and motion. You can probably update your state in a fraction of a second, even if you are simulating hundreds or thousands of planetary bodies, and you can simulate eons in hours.
The problem with simulating biology on an atomic level is that your time-step needs to be in the femtosecond to picosecond range, because atoms can move fast, and small differences in their position make big differences, but interesting things like protein folding can take minutes of real time.
So to simulate anything interesting, you need to step through quadrillions of time steps.
Oh, and you're running this on more than one processor, so you need to do smart things so your simulation isn't trying to synchronize all of its processors on every step; oh, and we can't tell you the exact position of every atom in a cell to start with, or even necessarily all of the different types of molecules that are present. Oh, and all of these cells -- or pieces of cellular machinery -- behave in ways that are dependent upon the environment they are immersed in, so you also need to simulate that somehow.
There's a reason Randall Monroe described the difficulty of protein folding as "we may one day find a harder problem".
It's nigh-impossible to understand this system by working from the bottom up. I don't see what kind of insight you'd get from this that you wouldn't get from a more understandable simplification in a minute fraction of the time.
For example: yeah, if a gene is contained in densely packed nucleosomes it'll be harder to express. But knowing that is enough; the exact details of how that works atomically is not the issue here. There are much larger issues with gene expression we don't understand which can much more easily give us much more insight into the workings of the system.
Haven't we already reproduced existing bacteria/DNA from atoms which has allowed us to more easily study gene expression?
Moreover, we take a bacteria with 200 genes ( we understand a lot of the gene functions/expressions, but we don't understand what all these 200 genes do). Thus, we synthetically reproduce the bacteria from the atoms up while removing nearly half of the genes that we determine are not needed to sustain the living bacteria, and the resulting synthetic bacteria we created in the lab allows us to more readily isolate and study the individual genes we didn't previously understand.
But maybe I am wrong and this is in no way helpful for studying gene expression.
I don't think so. I think they just took a small existing bacteria and by pruning its genome came to something which was minimally viable.
That sounds something like CRISPR or gene modification of an existing organism genome, maybe even taking genes from one species and inserting them into another (taking bioluminescent jelly fish and putting certain genes into cats or plants to make them glow). But, I'm talking about more recent successes of actually reproducing the DNA (synthetic DNA) from the atomic level up not modifying existing "natural DNA".
Sorry, I don't have enough time to search for recent studies, but here is old article about first "synthetic DNA" we created and inserted into cells that took off and self-replicated back in 2008: https://www.wired.com/2010/05/scientists-create-first-self-r...
The link seems to reference the first success of fully synthetic DNA, but I have recently been listening to podcasts about more recent experiemnts/studies, specifically reproducing bacteria with full synthetic DNA (which is stripped down as much as possible to maintain viability) for the purpose of studying gene expression and hopefully begin to identify what the currently unknown genes sequences do.
If I were to attempt to do gene regulation studies by MD (I share your skepticism!) I would just put a bunch of PEG in as a water replacement.
In fact that's a paper right there- "Effect of increasing macromolecular crowding on the accuracy of DNA simulations for regulation prediction" :)
Still a bit to slow... I guess it is progress :)
It's not really clear whether MD simulations like this are true contributions or whether they will ever produce useful results compared to well-parameterized neural nets.
To me, it's not clear whether we are good at quantifying when a neural net is well-parameterized in atomistic simulations.
The limitations of MD simulations and all the built in assumptions are well understood. We can reason about where our simulations might fail, even for very large systems. It's always a question about whether the underlying model can capture all the physics that occurs.
With a NN or other black box model, there's no way to reason about it except for benchmarking test/validation sets that (hopefully) capture the physics that you care about, but you still cannot really reason about how well the model will extrapolate to multiple interacting physics with different magnitudes at different length scales.
The limitations of MD aren't well understood. We don't know the implications of using a polarizable water model other than to say "it should be more accurate".
All the statements you make about NNs apply to MDs since MDs are basically feedforward ASTs with the same level of complexity and non-linearity as NNs.
in 2011 I was building 40 000 atom models on a 2001 SGI Fuel simulating them on a single 8 core node at 1ns/week -> 8000 atoms/core. They're about 7 times faster then me 8 years later.
So from a computation point, meh.
From a tools point this is cool. It isn't trivial to build or analyze a model of 40 000 atoms, never mind 1E9 atoms.
There's an inherent tension when doing a method development paper - if your result is too fantastical it's hard to know whether it's an artifact of your (potentially faulty) technique, and if the thing you're studying is well-understood then it serves as a good control, but it's less interesting. I suspect they chose the latter path since it requires no validation via existing methodologies.
Reference: https://www.ks.uiuc.edu/Highlights/?section=2008&highlight=2...
Another model is to offer it for sale as a product. Either way, engineers dont really care about publishing, we care about making useful things and Shari g how-to.
I wish this was a field where you could just apply more CPU and get better results (I say that with multiple papers published under my name which claim exactly that) but I'm skeptical. The force fields we use are far too approximate (point charges, assumption of transferrability) and it really does seem like we don't actually need most of this information to answer the questions we want to answer- those can be answered using more parsimonious techniques.