From bull horns to under the lens of Anton


In 1989, a young computer scientist named David Shaw was working at Morgan Stanley, one of the first Wall Street firms interested in using computer algorithms for trading. Shaw was an expert in parallel processing, speeding up calculations by executing them in a parallel process over multiple processors. Previously he had been a computer science professor at Columbia University had tried to sell his computer skills to a number of companies, but only Morgan Stanley was genuinely interested. As Shaw started working at the company, he began to think not just of programming strategies but of creative ways in which they could be applied to trading. In a meeting where he was supposed to talk only about his algorithms, he went one step beyond and described better methods for trading using these algorithms. Eyebrows went up in the room. Shaw was essentially seen as overstepping his bounds as a programmer. The higher-ups told him clearly that his job was simply building the computer architecture. He could leave the trading to them. Shaw quit and started his own company. Ten years later, it was one of the most successful hedge funds in the world and Shaw was a billionaire. One can only speculative how much the Morgan Stanley executives cried over the loss they had suffered when Shaw left.

But now D E Shaw is a totally different animal.

One of the most anticipated talks at the ACS meeting was by this Wall Street mover turned pure scientist. He is a remarkable and brilliant man. What other Wall Street hedge fund manager who made billions using mathematical algorithms for trading (and was known as “King Quant” at one point) basically retires from the dizzying world of finance to fully engage himself with computer simulations of proteins? Well, Shaw has done this, and is blazing his way toward some potentially revolutionary research. At the very least it is inspirational to see men with money actually care about basic scientific research.

He heads D E Shaw Research, a company totally separate from the financial powerhouse that has as its long-term goal, a fundamental transformation in the process of drug discovery. As the story goes, Shaw got somewhat bored of making millions and wanted to attack scientific problems that could benefit from the application of advanced computer algorithms. He got his old job as computer science professor at Columbia University and started looking around for the right problem. Fortunately for the field of biochemistry, Shaw started having discussions with a friend of his, the well-known physical chemist Richard Friesner at Columbia who is also the chief scientific advisor for the computational chemistry company Schrodinger. Friesner piqued Shaw’s interest and started giving him little problems in computational chemistry and biology which Shaw solved during his spare time. Finally he realized that MD simulations of proteins which had previously been typically restricted to the nanosecond time range stood a chance of being truly and very significantly useful if they could be expanded to the 10 microsecond-millisecond range, since this is the time scale on which most interesting biological motions such as large conformational changes occur.

Shaw started D E Shaw research and collected a team of highly talented chemists, biologists and computer scientists to tackle the problem. After a decade or so, these efforts have manifested themselves as Desmond, a protein MD program that has vastly accelerated computer simulations of proteins. Desmond essentially relies on many ingenious methods to simplify the calculation of forces and velocities involved in a typical MD computation. It especially calculates the non-bonded forces- the sheer number of which constitutes the bottleneck in these kinds of calculations- with unprecedented efficiency. What is even more remarkable is that Shaw’s group has designed ‘Anton’, a 512 node state-of-the-art machine, a special purpose machine explicitly designed for protein MD and named after Anton van Leeuwenhoek, the legendary 17th century Dutch scientist who trained the microscope on the microbial world and unearthed a wondrous universe teeming with life. Just like the 17th century Anton probed the events of the bacterial world, the 21st century Anton seeks to probe the molecular-level events of the protein world, The machine does only MD, and it does this using a razor sharp scalpel.

To give an idea of the kind of quantum leap Anton provides for MD simulation, Shaw gave some numbers, and I can swear I saw some people who were almost nodding off suddenly become wide awake. According to Shaw, the fastest supercomputer which does parallel processing today can crunch about 200 ns/day for a typical sized protein. Anton surpasses this number by two orders of magnitudes and spews out 17,400 ns or 17 microseconds per day. Such numbers would have been unthinkable a decade ago; until Desmond appeared on the scene, the world record for long protein MD simulations had been held by a group from the University of Illinois, with a total time of 10 microseconds.

So what’s the significance of being able to simulate in this time scale? Tremendous. It’s like the difference between nuclear weapons and the biggest conventional bombs previously used. When nukes arrived on the scene, some politicians like Winston Churchill shrugged them off by thinking that they were “just bigger bombs”. But as the old saying goes, quantity can have a quality all of its own. Nuclear weapons heralded a completely new era of warfare because of the ability of a single weapon to raze a whole city. The basic unit of destruction changed from a human being to entire cities. Desmond and Anton promise such conceptual transformations. As mentioned before, breaking the 10 microsecond barrier is a real turning point since most interesting physiological events happen on time scales of microseconds-milliseconds.

Entering the world of millisecond simulations is like unlocking the door to a rainforest with millions of exotic species that you suspected existed, but which you had no way of viewing and studying. In the last few years, Desmond has been used to study highly significant conduction events in ion channels, has been used to reconcile experimental and conceptual contradictions in the structure of GPCRs, and has been used to study very large conformational changes in kinases. All these events are very slow with respect to conventional MD. Shaw showed some spectacular examples of proteins actually folding and unfolding multiple times. In some cases his group has obtained quantitative agreement with kinetics and NMR experiments.

I think it was the end of the talk which made a few jaws drop. When you have a protein structure and want to find out a small molecule which can modulate its activity, one of the key goals is to first find out where the small molecule binds. With the kinds of time scales available, Shaw can achieve this with a devastatingly straightforward simulation. In a video that appeared a little surreal, he simply let the molecule roam all around the protein surface and find the binding pocket. Like a curious dog sniffing around for the buried bone, the little guy went in and out of crevices and gullies, lingered for some time outside the binding site, and then, with a little hesitation, finally ensconced himself firmly in his cozy home, having surmounted all the challenges of entropy and desolvation that he had to face.

This may not always be the best method to find binding sites and MD admittedly is not going to transform the process of drug discovery by itself, but what we witnessed in that room on Thursday was a different ball game. One in which the ball had been hit out of the park. More surprises should follow.

10 comments:

  1. My application to DE Shaw Research was rejected or ignored three times so far in my life. I'm actually proud of that rejection...it meant they at least looked at my CV.

    Still, glad to hear Anton and Desmond are working so well. Amazing work.

    ReplyDelete
  2. There are at least three reasons why D is not going to be the Final Solution for classical MD: a) unresolved questions surrounding the quality of force fields, especially long-term stability and accuracy of PESs far from equilibrium, b) hardware optimization means lack of flexibility in trying out new functional forms, eg polarizable FFs and c) the totally unaddressed issue of how to cope with the higher rate of data generation and whether or not human labor is up to the task of analyzing it, even with the help of computer programs. How fast can human insight extract chemistry, physics, biology and materials science from petabytes of dynamics trajectories?

    ReplyDelete
  3. These are good points. Were you at the talk? I ask since Shaw actually discussed the first two. He mentioned that force field development was an important part of his research. Plus, there are many problems for which classical force fields do suffice as has been demonstrated in so many applications of MD. If you want to see large conformational changes like helix rotations, the current force fields are not bad although they can be improved. I think it's only when you get to events like proton hopping that you may need to worry about the limitations of classical force fields. As for the second point, at least Shaw claimed that Anton was flexible enough to handle new functional forms...don't know about polarizable FFs though. The third point is an interesting one and I think the answer is open-ended. Until now, human beings have been able to extract relevant information from several trajectories. Only further development can testing can tell us whether human labor is up to the task of analyzing this huge amount of data.

    ReplyDelete
  4. yes they've discovered that long timescale MD simulations basically unfold proteins. They aren't stable. haha. Also, NVIDIA and the AMBER team just debuted hardware at the ACS last week that beats DE Shaw by 1000 fold. Sorry dude, your press release is over.

    ReplyDelete
  5. oh god i just read your penultimate paragraph.... you realize there are much much faster ways of doing that. To screen new drugs you have to screen millions... DE Shaw can't claim to do that.... this article is all hype and no science.

    ReplyDelete
  6. For all your talk about this being all hype and no science, where's your science? How about some actual substance rather than just rhetoric? Where are the papers? How about at least a link? Benchmark comparisons? Some numbers? And you seem to know nothing about biochemistry if you think that unfolding proteins is not interesting in terms of the details.

    Also, D E Shaw is not in the business of screening drugs so you are barking up the wrong tree. And hiding behind the cloak of anonymity makes your statements sound so much more credible.

    ReplyDelete
  7. Would this make the task of finding and making positive allosteric modifiers (PAMs) and negative allosteric modifiers) (NAMs) easier? See Proc. Natl. Acad. Sci. vol. 107 pp. 14943 - 14944 '10. They sound like exquisitely specific drugs (if only we could find find them). If we really understood proteins, this would be a snap. I'm going to have a post on this in the near future.

    Retread

    ReplyDelete
  8. That's an intriguing point. Allosterics often bind with weak binding affinities so finding them through simulation may not be very easy. That being said, some allosterics also interestingly act by influencing protein dynamics. This kind of effect is not easily discernible through experiments and I think this is where MD might be useful.

    ReplyDelete
  9. Dear Wavefunction,

    Thank you for the blog, which is fun and informative.

    I don't mean this in a snide way at all, but why do you believe in MD? The algorithms rely on gradient-based minimization, and these algorithms are unreliable. Only local minimization is solid; everything else is not. Not to mention inaccurate force fields etc.

    Also MD people do not seem focused on making testable predictions. E.g., a group will propose some MD-derived detailed functional mechanism involving a residue x, and then the experimental check will be 'we mutated residue x and the protein was no longer functional.' But that is a long way from being a test for what was proposed. If I worked on MD, I would go all out to develop experiments that would test my methods and predictions.

    Could you recommend any references on protein MD for the skeptical, which address some of these issues? So often papers preach to the converted.

    Thank you,

    Curious too

    ReplyDelete
  10. Oops, forgot to come back. I've previously met Shaw himself, as well as some of his scientific team. Shaw's answer to C) was "I have people working on it." The problem is that they've decided to do hardware optimizations for the computations. Some stuff is undoubtedly useful and universal, like hardware-optimized square roots. However a lot of the architectural details like deciding where to spawn threads and data pipe bandwidths are pretty much going to work just for whatever model they decide to work on, and the moment you decide to use a different model, a lot of the optimizations will be ineffective because the hardware is no longer being used optimally. It's an interesting idea to optimize at the hardware level - very rarely is a machine built to solve a specific problem - but that speed gain comes with a concomitant tradeoff in the lack of generality. Polarizable force fields require the computation of additional terms in the FF, and often involve solving for more dynamical variables. Do Shaw et al. plan to make a custom Anton for each generation of force fields? Certainly some of what they learn will be transferrable, but I just don't see how this strategy can be generalized easily.

    ReplyDelete

Markup Key:
- <b>bold</b> = bold
- <i>italic</i> = italic
- <a href="http://www.fieldofscience.com/">FoS</a> = FoS