Skip to Content

Chemical News

Relay Calculates Its Way Through

Bloomberg has a feature on Relay Therapeutics, who are just a few blocks away from me (and where several former colleagues of mine work). It’s a nice writeup, and also features a (relatively rare) spotlight on David Shaw of D. E. Shaw research. He’s one of those guys that you’ve likely never heard of unless you’re pretty into Wall Street stuff, but he and the firm he founded are a good example of how important things don’t always make the newspapers (see below). If his name (or that of Jim Simons) rings no bells, that’s probably how they like it.

First, Relay. As the article says, their big thing is molecular dynamics. That’s easy enough to explain in concept, but in practice it’s. . .nontrivial, as we applied-science types like to say (while exchanging significant glances). Even many people outside of drug discovery, chemistry, or science altogether have heard of the idea of using a computer-aided model to “dock” drug molecules and see which ones fit better than others. This has been discussed in the popular press since at least the early 1980s, with the result being that some people think that’s how all drugs are discovered.

Far from it, though, as practitioners will tell you: docking-and-scoring has come a long way, and can be useful depending on the circumstances, but it’s far from a royal road to wonder drugs. There are so many things that are tricky about evaluating the energy gained and lost as a small molecule interacts with a protein (and with the surrounding solvent) that the error bars can start to creep up on you rapidly – and what’s worse, you many not even realize that they’re doing so. And that applies even when you have the right three-dimensional conformations for the protein and for the small molecule (and you don’t, not always, nor do you necessarily realize it when that’s happening, either).

Note that in that case we’re talking about a relatively static situation, which in most cases just isn’t realistic. Small molecules move around, and solvent molecules do little else but move around. And Lord knows that protein binding sites can be flexible, too. The old “lock and key” model suffers from our experience of locks and keys being made out of rigid metal. If they were made at least partially out of pizza dough we’d have a better mental picture. So bringing in molecular dynamics (MD) has the potential to help out, because that explicitly tries to calculate all these motions across the whole binding event. That way, you can pick up (in theory) situations where Side Chain X rotates out of the way, which allows Water Molecule Y to be exposed and more easily be displaced as Ligand Z backs into the spot like someone learning to drive an 18-wheel truck, with plenty of halting and readjusting, etc.

But getting that to work is another thing entirely (as this and the comments sections to this older post will make clear). One problem is that if static modeling has its problems, what happens when you assemble frame after frame of static modeling into a movie? “That’s not what we’re doing” is the response from the MD people, but the point about errors possibly propagating through the simulation remains. The only way to do dynamics right (in my own view) is to throw a really ferocious amount of computing power at it, make as few simplifying assumptions as possible, and just calculate until you can’t stand it any more.

And that’s pretty much what Relay is up to. The company’s founders knew very well what they were up against, and recruited David Shaw’s organization to help out. Shaw, as mentioned above, led an extremely successful quantitative-trading company for many years, using computational methods to exploit small inefficiencies in the equity markets. But more recently, he’s been using his resources to fund D. E. Shaw Research, doing fundamental work in computational chemistry and biology with as much computing horsepower as can be obtained. (Here’s Shaw talking about that work in a lecture to the Biophysical Society). Even a few years ago, the hardware/software combination didn’t really exist to take on the long binding simulations that Relay’s founders wanted to go after, but that’s recently come into the realm of possibility. (Update: Mark Murcko of Relay (a longtime reader here) has left a comment pointing out that the company has plenty of experimentalists, structural biologists, etc., and that they’re far from a pure computational play. That’s true, of course – that 1981 story linked to above showing steely-eyed drug designers making drugs right off the screen is just as much of a dream now as it was back then).

And they seem to be getting somewhere. I’ve known some of the folks at Relay for years, but they’re not going to tell me any secrets (I work for the competition!) But it looks like they’re taking a lead molecule into the clinic next year if all goes well, and it’s my understanding that they’ve gotten where they are by using an unprecedented amount of computing power to guide their synthetic chemistry. I’m very interested to see where this leads. One the one hand, molecular dynamics, done right, could indeed provide insights that are both important and difficult-to-impossible to obtain any other way. On the other hand, though, done wrong, even slightly wrong, it could just be a very expensive way of making yourself believe that you’ve got a handle on things, even if you don’t. (I can assure you that Relay’s people are keenly aware of that danger, which is a good thing). And on the gripping hand, there’s always the chance that producing a good molecule using MD will just allow you to move a layer deeper and find the real problems with your therapeutic hypothesis, as so often happens in drug discovery. But that’s still progress.

From what I can see, other companies are taking a wait-and-see attitude towards all this, not least because “Use David Shaw’s custom-built supercomputer” is not a technique that scales well for now. But people are watching, and watching with interest, to see if this actually is the future.

26 comments on “Relay Calculates Its Way Through”

  1. Billy says:

    Even if computational design is successful, there’s still that next hurdle: “is there in vivo efficacy?” For those who aren’t sure what I’m talking about, see Derek’s post about the amyloid hypothesis yesterday. It’s a tough business for sure.

    1. anoano says:

      and the question, could they have done the same (find this drug) without those super computers?

    2. MB says:

      Agreed. All computationally selected candidates are going to have be put through a biological efficacy filter. And if that filter is poor at predicting human-relevant clinical translation, no amount of front end work can solve that.

  2. luysii says:

    Not many people have the brains (or the money) to design a supercomputer from scratch for the protein folding problem, but Shaw did. It’s worth a look at https://en.wikipedia.org/wiki/Anton_(computer) — but there’s a lot of jargon in it.

    1. Demis says:

      Sounds like it might be able to play Minecraft on full settings!

  3. Rhodium says:

    I think of David Shaw as the closest we have to Iron Man. I guess it is standard in the computer business, but when you have to decide chip component locations based on what gives you the fastest computation, that’s advanced design.

    1. Karl says:

      Actually, component position sensitivity is something of a rediscovered art. Years ago, I was briefly associated with a company that had a number of CDC machines for engineering timesharing; 6600’s and at least one 7600. I was assured by one of their field engineers that quite a few reliability fixes involved lengthening or shortening various wire-wrap connections by as little as a quarter inch…

  4. Mark Murcko says:

    Hi Derek – One clarification – 90% of the scientists at Relay are experimentalists doing cutting-edge research. The readers of this blog know better than anyone that you can’t think your way into a drug. (I love Kevin Kelly’s term for that mistake, “thinkism.”) Relay is a drug discovery company with the goal to find better drugs against hard but well-validated targets by understanding and exploiting protein motion. This involves many elements: novel and complex methods in biophysics / structural biology; really creative chemistry; elegant biology — in short, a team of really smart experienced people from many diverse backgrounds all contribute to this effort. So, yes, the collaboration with DESRES is fantastic — absolutely essential in helping us to understand how to benefit from MD — but in the end drugs come from experiments and having a tight integration of computation and experiment is core to our approach. Best / Mark

    1. Derek Lowe says:

      Thanks! Absolutely; I don’t want to suggest that you guys are just sitting back watching the screen. But the MD stuff, and how it informs structure-based design, is definitely something that you folks have that others don’t. I mean, this is Cambridge – we’re all doing elegant biology and novel methods, amirite? (sarcasm mode off!)

    2. D.E.Shaw says:

      You both are fools! Absolutely don’t know what you’re talking about. Didn’t you get the memo? The whole MD stuff is about sucking the money from gullible VC guys. That’s the genius in getting rich…

  5. long ago and far away says:

    How does the amount of computational power that Relay Therapeutics is applying to MD compare with the amount of computational power being applied to mining Bitcoins? Sorry if this sounds like a facetious question.
    Comp Sci is way outside my field of expertise, and I’ve heard tale of monstrous computer farms being set up near huge sources of cheap electricity; so I’m curious how these two very different applications compare.

    1. Mister B. says:

      I have been told that Bitcoin mining requieres only brute force calculations. The more transactions you verify, the more bitcoin you get.

      At the contrary, you have to design as well as possible your MD experiment to avoid unreliable data. To me, the main difference between the two process is the human sitted between the chair and the screen.

      The amount of computionnal power could be less important if better used.

    2. chimaera says:

      Any computational effort is brute force: it’s just churning through calculations. Where simulations differ from Bitcoin is the size of the problem.

      An individual Bitcoin transaction is relatively trival and can be carried out quickly: throwing more computing power at it increases the rate at which you can complete calculations and thus collect more bitcoin.

      As with most mathematical descriptions of the real world, molecular dynamics is formulated with partial differential equations. Typically the set of PDEs is unsolvable using analytical techniques for any problem that is close to reality, so numerical techniques are used. This involves breaking up the continuum into discrete elements and converting the PDEs to a system of linear equations which can be processed using a computational algorithm. If you’re familiar with Simpson’s Rule for integration, the principle is the same.

      The computational effort arises from this discretisation on two fronts. First, resolution in the results depends on the size of the mesh: if you want more detail, you need a finer mesh. There is also an upper limit on mesh size: too coarse and your solution becomes unstable and/or unreliable because it can’t capture phenomena that are important to the problem. The size of the computational effort scales with the inverse of the mesh size. Second, the size of the domain: a bigger physical space increases the number of mesh cells needed at a given mesh size, which again increases the size of the computational effort.

      The limiting factor in these analyses is often RAM, particularly for models incorporating a lot of detail from the physical problem. The solution methods are iterative, and during an individual iteration, all of the data for the iteration must be held in RAM. In my own work (CFD in chemical engineering) it’s not hard to come up with a problem requiring over 64 GB of RAM, and I suspect it’s similar here. If you want to reduce memory usage you need to simplify the model, and a certain point the simplification results in a model that is no longer reliable.

      Having studied the history of CFD and how it has developed in capability, and taken advantage of advances in computing power, I can see molecular dynamics becoming a useful tool in time as computers emerge which can handle the detail needed for it to be useful. If quantum computing can be realised on the ground, simulation work will be transformed. As things stand, we can expect that Moore’s Law will continue to be pushed back and incremental advances will bring more problems within reach.

  6. Uncle Al says:

    How fragile are our rock-solid fundamental understandings versus a lab bench rather than a peer vote? Matter diffraction is the beating heart of quantum mechanics – and it never fails,

    . 1,298.7 amu, DOI:10.1038/nnano.2012.34
    10,122.8 amu, arXiv:1703.02129

    Any attempt to localize the process, tracing pathways, collapses diffraction – no dissipation. Diffract a resolved chiral molecular beam and Hund’s paradox emerges, DOI:10.1103/PhysRevLett.103.023202, DOI:10.1088/1361-6455/aa5115, DOI:10.1039/C2CP40920H

    A resolved chiral molecular beam racemizes when diffracted, or does not diffract and QM is no longer a sure thing. 2-trifluoromethyl-D_3-trishomocubane is one step beyond literature synthesis of the 2-carboxylic acid (SF_4). It has no accessible path to inversion. Nanogram samples are vacuum downstream enantiomeric excess observed to high confidence by 3-wave mixing rotational spectroscopy.

    Theory fears Pyrex, the patent system rejoices in it. Grind it out, look, think better…net retained earnings. Ya gotta look.

  7. MoMo says:

    Please-Spare the talk. Show us the molecules.

  8. I am the Chem God says:

    From my time at Relay, I can tell you that many of their oncology targets are quite a bit of a stretch in their feasibility. Their MD platform is a bit wishy washy and is left to be seen if it can hold much water. They try to cuddle up this uncertainty with pampering the employees with catered lunches. So take what you want with that…

  9. Curious Wavefunction says:

    Generally speaking it’s pretty hard to do structure-based drug design using structures from MD; you’re taking the system farther and farther away from a “real” crystal structure (notwithstanding the problems with crystal structures), and it’s not easy to know whether the motions you’re seeing are artifacts or correspond to a real, exploitable parameter. For instance, there’s a rather long history of failed attempts to design new ligands using “ensemble MD” in which you dock molecules to hundreds or even thousands of snapshots from an MD simulation rather than from a static structure; these efforts were confounded both by well known inaccuracies in scoring functions (which aren’t going to go away because of MD) and a high amount of noise in the system which obscures useful real details.

    There’s also the issue of observing rare events; flickering transitions in side chain movements and loops that might dictate ligand binding. I remember a publication from a few years ago in which they demonstrated that even a long MD simulation was unable to sample an energy basin that was separated from the original one by a high energy barrier. Presumably the massive simulation trajectories from DESRES combined with other techniques like replica exchange and high temperature MD have that particular horse tamed.

    So, Relay is attempting something genuinely tough in my opinion, but all this being said, if there was ever a team that didn’t fall for hype, knew the strengths and limitations of the methods and knew how important it is to couple simulation with experiment, they’re the ones who have it, so I wish them all luck and am hopeful.

  10. David Edwards says:

    I have a tiny insight to offer here, which is probably relevant.

    One of my past coding projects involved trying to construct a toy model of proteins. Just getting the 3D geometry to work alone was hard work. Especially when such factors as ligands spinning about bond axes were introduced into the code. Basically, if you have a single bond from a carbon atom, to a functional group, that functional group can rotate about the bond axis in space, and when it does so, it can result in intermittent steric hindrance of a previously open site available for a reaction, if the functional group has the right shape. Which means that a site available for reactions could, occasionally, be closed off. Writing code just to take this feature into account in a toy model of proteins, took a huge amount of effort. Now imagine transferring a problem like this to a realistic protein model, and life becomes one or two orders or magnitude harder still.

    That’s before you move on to folding, which is a very hard problem from the computational modelling side.

    Of course, real proteins don’t worry about all of this. They simply respond to electrostatic forces and do their stuff. But if you’re going to model this, you need to work out what those electrostatic forces are, what vectors they’re acting along with respect to a particular part of your molecule of interest, and all of this changes frame by frame as the model updates its element positions to take account of those forces. In short, you’re dealing with the electrostatic version of the N-Body Problem, with all that entails. Even a toy model of this starts becoming horribly complicated to code (and debug!). A realistic model of such a system, well … you’re in supercomputer territory.

    Then you have to add in rules for chemical reactions. Even in a toy model, a non-trivial exercise. In a realistic model … if I could do this, I’d probably be able to write my own pay cheque.

    Then, you have to see what happens when you toss in your molecules of interest, and see if what you’ve modelled bears any relation to the data obtained from years of organic chemistry grunt work. If it doesn’t, you have all the fun of finding out which of the many levels of interaction involved you’ve buggered up in your model, and working out how to correct your previously flawed understanding thereof.

    Needless to say, this toy model insight, gives me a lot of respect for people who can solve problems in this field while working on realistic models.

    1. anon the II says:

      I’m intrigued by your notion of “a toy model of proteins”. I don’t think there is really much of a difference between the underlying math of your toy proteins and the realistic models that they’re using at Relay.

  11. Daniel Barkalow says:

    I’m kind of more worried that you’ll find a small molecule that really does bind to a particular protein, but that protein turns out not to be useful to bind to after all. If you’re spending all that effort developing the tech for the molecular chemistry, it would be a shame if you were on the hook if your well-validated targets didn’t actually help.

  12. Useless Molecule says:

    “However, it hopes to begin its first clinical trials next year (…)”

    Q: Do they have a molecule that made it through tox yet? A: Very unlikely

    “The timeline that Relay envisions, if it succeeds, would shave years off the typical drug-discovery process.”

    The only to shave years off the typical timeline is to be involved in a patent-busting scheme.

    Q: Is Relay involved in a patent-busting scheme? A: Quite possibly

  13. MoMo says:

    Useless Molecule –They all say that- “We will shave years off” and we will “disrupt” then “create 100s of new drugs using existing drugs”, our “platforms work in 3-D” and our people “are driven and unique” see? Look at their Bios. One ate a whole box of chocolate covered ants once now runs our computational team.

    But the genius here is DE Shaw himself, making billions on “inefficiencies” in the stock market and maybe he found “inefficiencies” in Pharma?

    He wouldn’t have to look hard.

    1. Bond Covalent Bond says:

      Finding “inefficiencies” in the stock market is like running from a bear. You don’t have to win, merely run a little faster than the other guys.

      It’s unclear to me how this maps into pharma success, especially coming from a guy who measures people’s intelligence by their SAT scores (well known Shaw thing).

  14. hn says:

    I don’t know how Shaw contributed to the research other than funding it. It was usually Ron Dror that presented in meetings.

  15. Relay Founder says:

    Relay’s strategy will not work long term and we’ll all witness the demise of this wretched company in 5 years time. Mark my words!

  16. Inquisitive says:

    Any further word on this company?

Leave a Reply

Your email address will not be published. Required fields are marked *

Time limit is exhausted. Please reload CAPTCHA.