Apparently the interface between molecules uses the Chemical Master Equations (CME) and Reaction-Diffusion Master Equations (RDME) both of which I'm unfamiliar with: http://faculty.scs.illinois.edu/schulten/lm/download/lm23/Us...

“Lattice Microbes is a software package for efficiently sampling trajectories from the chemical and reaction-diffusion master equations (CME/RDME) on high performance computing (HPC) infrastructure using both exact and approximate methods.”

Or, in other words, it doesn't solve the Schrodinger equation at all, but uses well known solutions for parts of the molecules, and focuses on simulating how the molecules interact with one another using mostly classical dynamics.

As far as I can tell from their model, molecules don't interact with each other ~at all~ through classical dynamics. Rather, they define concentrations of various molecules on a voxel grid, assign diffusion coffecients for molecules and define reaction rates between each pair of molecules. Within each voxel, concentrations are assumed constant and evolve through a stochastic Monte-Carlo type simulation. Diffusion is solved as a system of ODEs.

This is a cool large scale simulation using this method, but this is a far cry from an actual atomic-level simulation of a cell, even using the crude approximations of classical molecular dynamics. IMO it is kind of disingenuous for them to say 2B atoms simulation when atoms don't really exist in their model, but it's a press release so it should be expected.

Yes, this is not the standard "force field" pairwise stuff you're used to when you heard "simulation" of biomolecular systems. I don't know if it's quite disingenuous, just not what we expect based on what the vast, vast majority of the field does! It does represent that many atoms. We shouldn't include atoms for the sake of having them, right? It should depend on what questions we're asking of the system.

I like seeing other (simulation or analytic) methods get attention. Lattice methods --(HP models[0], for hydrophobicity[1], lattice-boltzmann even. field theoretic (see polymer physics, melts, old theory[2], new theory, and even newer simulation[3]). Even the simplest shit like springs[4]!

[0] https://scholar.google.com/citations?view_op=view_citation&h... [1] https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.11... [2]https://www.google.com/books/edition/Renormalization_Group_T... [3]The Equilibrium Theory of Inhomogeneous Polymers (International Series of Monographs on Physics) https://www.amazon.com/dp/0199673799/ref=cm_sw_r_apan_glt_i_... (numerical scft) [4] https://en.m.wikipedia.org/wiki/Gaussian_network_model

It also represents even more electrons and even more quarks than that. I think it would be silly to characterize this system by the number of constituent quarks, but that's just me. To me, the important number is the number of degrees of freedom within the model. In a force-field model this scales linearly in with the number of atoms. In the model presented in the paper, this depends on voxel resolution and number of molecular species. Sadly, this is omitted in the press release.

Thanks for your links :) I work in inorganic materials but I really should understand more about models for more complex systems.

I don't believe there is any case where a molecule in one voxel knows about molecules in other voxels. DNA and RNA is coarse grained where a 'molecule' might represent a specific sequence of interest rather than a full chain. Transcription and translation are modeled, essentially by saying if the neccessary ingredients are present in a voxel (DNA/RNA sequences, enzymes, raw materials, etc) there is some chance of forming mRNA or a protein as a function of the molecular concentrations present in the voxel. DNA and RNA reactions are treated with somewhat different equations than the rest of the molecules, I think to handle the coarse graining.

Do you see a future for this simulation method with increased computing power, or do you think limitations of the method might still limit its applicability. Maybe this is naivete coming from atomistic perspective, but it seems to me that the inability to model reactions that aren't explicitly predefined would be a significant challenge.

In practice, none of this matters though and we can still get very useful results at the resolution we care about.

There will never be true knowledge of both a particle's location and momentum a la uncertainty principle, and will always have to be estimated.

And, like, there is still uncertainty about the position of the "center of mass" pretend particle, as well as for the position of the "displacement" pretend particle.

(the operators describing these pretend particles can be constructed in terms of the operators describing the actual particles, and visa versa.)

I don't know for sure if this works for many electrons around a nucleus, but I think it is rather likely that it should work as well.

Main thing that seems unclear to me is what the mass of the pretend particles would be in the many electrons case. Oh, also, presumably the different pretend particles would be interacting in this case (though probably just the ones that don't correspond to the center of mass interacting with each-other, not interacting with the one that does represent the center of mass?)

So, I'm not convinced of the "nuclei have to be treated as frozen, otherwise electrons don't have a reference point" claim.

People have done simulations with quad precision (very slow) but very few terms in molecular dynamics would benefit from that. In fact, most variables in MD can be single precision, exceptt for certain terms like the virial.

If the universe is discrete, how does one voxel communicate to the neighboring voxel what to update without passage through ‘stuff in between’ that doesn’t exist? Heh

It seems physics is going the opposite way with infinite universes and multiple dimensions to smooth out this information transfer problem and make the discrete go away.

I don't think your "voxel" intuition can be right because it's a small jump from that to (re)introducing an absolute reference frame.

That kind of reminds me of the 'aether' that was once hypothesized as a medium of transmission for light and radio waves [0].

Also, voxel's communicating sounds an awful lot like a higher-dimensioned cellular automata.

[0] - https://en.wikipedia.org/wiki/Aether_theories

How much smaller?

But photons resulting from the same event but with different energies arrive at detectors an appreciable distance away to all intents and purposes simultaneously, something that would not happen if spacetime were discrete at a level close to the Planck length. So it would have to be quite a big difference for an effect not to show up as a difference in time-of-flight.

What accounts for the expectation that they do not arrive simultaneously in a voxel-based universe?

As in, photons that leave the surface of the sun always strike those specific points in space-time which are at a zero spacetime interval from said surface. If you take the described geometry seriously, then "spacetime interval" is just the square of the physical distance between the events.

(And any FTL path has a negative spacetime interval. If that's still the square of the distance, then I think we can confidently state that FTL is imaginary.)

It's certainly not that -- that's a hideously difficult algorithm with exponential complexity.

https://en.wikipedia.org/wiki/Full_configuration_interaction

https://en.wikipedia.org/wiki/Verlet_list

But that's for classical simulations. Full configuration interaction is effecftively computing the schrodinger equation at unlimited precision, in principle if you could scale it up you could compute any molecular property desired, assuming QM is an accurate model for reality.

i never did any qm work beyond basic parameterization

i'm guessing you are/were also computational physics guy :)

Pretty cool, what're you up to now?

i saw some of your other comments about being at google. did you touch jax-md at all?

https://github.com/google/jax-md

My work mostly predated tensorflow and was much more about massive-scale embarassingly parallel computing, and produced some interesting large-scale results from MD and protein folding.

https://www.nature.com/articles/nchem.1821 https://onlinelibrary.wiley.com/doi/full/10.1002/pro.2389

I'm very familiar with the first paper, the second author was on my committee.

so what does a cloud migration at a biotech company mean?

is it sort of a standard orchestrator + warehouse/lakehouse + distributed compute + cicd tools stack?

Ideally, yes, exactly. Except there are 100 orchestrators, 100 small local warehouses, and CI/CD is mostly jenkins.

Some things get forklifted over. I'm not trying to push people to adopt cloud native practices, just move them off physical onprem resources. Even that is a challenge because of data gravity.

Larger proteins (a few alpha helices and beta sheets), the folding process can be studied if you start with structures near the native state.

None of this means to say that we can routinely take any protein and fold it from unfolded state using simulations and expect any sort of accuracy for the final structure.

Other people use ab initio very differently (for example, since you said "level of theory" I think you mean basis set). I don't think something like QM levels of theory provide a great deal of value on top of classical (and at a significant computational cost), but I do like 6-31g* as a simple set.

Other people use ab initio very differently. For example, CASP, the protein structure prediction, uses ab initio very loosely to me: "some level of classicial force field, not using any explicit constraints derived from homology or fragment similarity" which typically involves a really simplified or parameterized function (ROSETTA).

Personally I don't think atomistic simulations of cells really provide a lot of extra value for the detail. I would isntead treat cell objects as centroids with mass and "agent properties" ("sticks to this other type of protein for ~1 microsecond"). A single ribosome is a single entity, even if in reality it's made up of 100 proteins and RNAs, and the cell membrane is modelled as a stretchy sheet enclosing an incompressible liquid.

I also agree that QM doesn't provide much for the cost at this scale, I just wish the term ab initio would be left to QM folks, as everything else is largely just the parameterization you mentioned.

I am more than happy to use "ab initio" purely for QM, but unfortunately the term is used widely in protein folding and structure prediction. I've talked exdtensively with David Baker and John Moulton to get them to stop, but they won't.

In inorganic materials ab initio means you actually solve Schrodinger's equation (though obviously with aggressive simplifications e.g. Hartree-Fock).

Having developed a few newtonian force fields, calling them "derived from QM" is very, very generous :P

This is the only real update of the year: https://foldingathome.org/2022/01/03/2021-in-review-and-happ...

SARS-CoV-2 has intricate mechanisms for initiating infection, immune evasion/suppression and replication that depend on the structure and dynamics of its constituent proteins. Many protein structures have been solved, but far less is known about their relevant conformational changes. To address this challenge, over a million citizen scientists banded together through the Folding@home distributed computing project to create the first exascale computer and simulate 0.1 seconds of the viral proteome. Our adaptive sampling simulations predict dramatic opening of the apo spike complex, far beyond that seen experimentally, explaining and predicting the existence of ‘cryptic’ epitopes. Different spike variants modulate the probabilities of open versus closed structures, balancing receptor binding and immune evasion. We also discover dramatic conformational changes across the proteome, which reveal over 50 ‘cryptic’ pockets that expand targeting options for the design of antivirals. All data and models are freely available online, providing a quantitative structural atlas.

I looked at the various approaches and sided with folding@home. At one point I had 1 million fast CPU cores running gromacs.

but, as I pointed out elsewhere, this would not be particularly helpful as it would use an enormous amount of resources to compute something we could probably approximate with a well-trained ML model.

It also wouldn't address questions like biochemistry, enzymatic reactions, and probably wouldn't be able to probe the energetics of interactions accurately enough to do drug discovery.

Now, the question is, does it matter? When do you EVER need to know the exact atomistic, yet alone electronic, trajectory of a single protein starting from a given position within a cell surrounded by waters in a given configuration?

It doesn't really matter. This is the beauty of noise and averages and -- dear to me-- statistical mechanics. At finite temperatures, AKA most everything we experience as living things, quantum details (or precise classical trajectories for that matter) aren't that important for the vast majority of questions we tend to have about a system.

Why am I saying this? Because I thought the same as you and it took me 20+ years to realize MD doesn't affect medicine at all.

This research is something like a pixar movie, or one of those blender demos with a lot of balls :P