Comment by timr | Hacker News Reader

Comment by timr | Hacker News Reader

Parent Full thread

Wow. I'm floored that computational structural biology work is getting a Nobel. I spent a lot of time reading their papers in grad school...they're totally foundational for the field, but the field is still really young:

Martin Karplus' group is behind CHARMM (http://www.charmm.org), which was the first(?) molecular dynamics package (you can think of it as a precursor to folding@home, though it's still under active development, so that isn't a totally fair statement).

Michael Levitt has done a bunch of things, with no one big software package, but he was one of the earliest people trying to do ab initio protein structure prediction. He was also one of the first people to really start categorizing protein structure in a way that allowed for computational modeling -- back in the 70s and early 80s.

Arieh Warshel is probably best known for bridging the gap between quantum mechanics and the (relatively quick-and-dirty) molecular mechanics work. He's done a whole bunch of work modeling enzymatic reactions, coming up with better electrostatic models, and other things where quantum mechanics does a good job, but is way too slow to be used on giant molecules like proteins.

Yes, but I am also not surprised given this is pretty fundamental work, even if it hasn't quite hit the sort of revolutionary goals many of us hoped for. If Peter Kollman had been alive, I suspect he would have shared this Nobel.

This Nobel is hits very close to home. Much of my PhD work was done in CHARMM and I come from the Karplus lineage, and at one point in time I had pretty much read every paper Levitt and Warshel had ever written.

Why no Harold Scheraga? Seriously...

I thought the same thing.

According to press accounts, a maximum of three people can share the Nobel prize.

Computational methods are powerful, but not respected enough in many fields. I think it's harder to publish computational work in a respectable journal, but rather easier when it's wet lab work. I think it was a good choice of the Nobel committee that may change scientists' view of computational work.

The beauty of CHARMM (at the time) was that it was the first practical combination of a forcefield and MD engine, that made it practical to simulation proteins. While other packages have left it behind on perf and capability, the forcefield itself remains very high quality, and the breadth of functionality is still as good as it gets. Unfortunately the implementation could do with an overhaul since it still looked like something from the early 80s (I haven't seen the CHARMM code in 5 years)

> the forcefield itself remains very high quality, and the breadth of functionality is still as good as it gets.

It's kind of weird that forcefields haven't improved all that much. Implementation has improved drastically (LAMMPS, Gromacs, etc.) and QM level theory has too (FCIQMC, exploiting symmetries to eliminate the sign problem), but MD forcefields don't seem to have made a whole lot of progress from CHARMM, AMBER, and UFF. I suppose some of the reactive potentials (AIREBO and Tersoff) are a step up, but still...

Maybe that's true for biological compounds, but for the rest of the periodic table new potentials are constantly being developed (and some may even be improvements). Semiconductors, transition metals, heavy elements, ferromagnetics, you name it... its being worked on.

Actually even for biological molecules there are still advances being made. Water is an excellent example seeing how there is a new potential released once every 6 months or so. Its difficult to say whether these actually show a quantifiable improvement over old potentials, but its still an area of constant effort and progress.

> whether these actually show a quantifiable improvement over old potentials

Yeah, that's what I was meaning with regard to force-fields in general. What are we, like TIP23P for water now? ;)

It's still mostly fortran :)

Since you seem to know about this field, any chance you can share you opinion on the working going at DE Shaw Research? Does developing custom hardware for molecular dynamics give them any sort of qualitative advantage? Are the systems they're simulating significantly larger or more useful than what other groups work on?

I know a thing or two about the field. Custom hardware for molecular dynamics allows one to simulate longer timescale biological processes, namely the folding of larger proteins. More generally, in the study of any size protein, countless molecular dynamics papers are simply plagued with statistical error and cannot be trusted. I think it's unanimous that DE Shaw Research is doing a great service and I'm excited to see the performance of their next supercomputer, Anton2, http://dl.acm.org/citation.cfm?id=2451175

But as in all things, Anton makes a controversial set of assumptions and approximations w/r to long-range electrostatics so it's anyone's guess whether their approach is valid or not.

All that aside, Anton is a wicked cool piece of hardware able to accelerate their approximation of traditional molecular dynamics by a factor of 50 over GPUs and up to a factor of 500 over CPU clusters, hitting milliseconds of simulated protein time in a single trajectory.

And since it's the only hardware capable of simulating a single monotonic trajectory out to that timescale, it's challenging to compare it to other approaches.

Agreed. DE Shaw's hardware work is very cool, but I'm in the camp that believes that the force fields suck (in general; not just theirs), and bashing through some multi-millisecond trajectories will only highlight their limitations.

You should check out Lane et al. 2013 Fig. 1: http://www.sciencedirect.com/science/article/pii/S0959440X12... All computational protein folding times match experimental results within 1 order of magnitude. If folding time isn't a good measure of forcefield accuracy then I don't know what is. Sampling error is a worse offender than forcefield accuracy in my opinion.

" If folding time isn't a good measure of forcefield accuracy then I don't know what is."

Folding kinetics are surprisingly insensitive to detail. For example, it's been understood for a while that even ridiculously simple models can predict transition state energies for simple proteins:

http://www.ncbi.nlm.nih.gov/pubmed/10322214?dopt=Abstract

http://www.ncbi.nlm.nih.gov/pubmed/10500172?dopt=Abstract

But to answer your question, I'd say that "predicting the correct structure of a protein" is the gold-standard benchmark of forcefield accuracy, and MD forcefields are really bad at it.

(You could reasonably add other great measures, like: "does the simulation tend to fly apart without hacked-up pseudo-physical constraints?", but that feels like piling on.)

I'm biased since I do simulations but you're being too harsh on force fields. All the proteins in the paper I mentioned had their structure correctly predicted starting from an extended state. Correct structure and folding time! That's a serious accomplishment compared to the decade-old papers you're citing.

"That's a serious accomplishment compared to the decade-old papers you're citing."

What? It's not like the "decade-old" paper has become less correct over time. I mean, it's great that MD is finally close to predicting something that could be predicted a decade ago with much simpler methods, but that isn't saying a lot. Plus, as the other commenter pointed out, there's a complicated cross-validation problem that biases the results you're reporting.

In truly blind tests (i.e. CASP), MD force fields just aren't that good at predicting structure from sequence. The usual counter-argument is that they aren't meant to work with non-MD methods (fair enough, I guess), but you don't have to try very hard to find reasons not to trust them. MD simulations have always been very finicky things, requiring lots of manual intervention to get "right".

To be fair, the Pande Lab's data has involved a degree of human intervention during the generation of the multi-state trajectories upon which they're based. And that's a failing that has only recently been addressed. IMO it's indeterminate whether that human intervention could have unwittingly influenced those trajectories.

Similarly, a lot of the folded proteins were used to develop the same force fields now used to simulate their folding. I'm not dismissing this data, but I am saying I think the jury is still out on the models until we have a true test set/training set dichotomy, a separation that was an absolute revelation for ab initio structure prediction algorithms in the 1990s.

That said, I think we both agree that undersampling is the biggest offender. And it only gets worse with results published for larger systems simulated at the same timescale as much smaller ones.

Don't forget Karplus' colossal contribution to NMR

I had a whole other paragraph about that, but took it out because I figured that most people on HN wouldn't even know what NMR is, let alone NMR structure prediction.

These guys have collectively done so much stuff that it's impossible to give a fair accounting of it all in a brief post.

yeah, I'd say that is way more important to science. I have never used and never would use CHARMM, Karplus' J-coupling values - indispensable.

Reply on news.ycombinator.com