Today's speaker is David Baker, from the University of Washington, what we'll do tonight, I think that worked quite well last time is they will give his talk people please drop in questions of any type in the chat. And then David, what we did last time is the speakers just read the questions out of the chat. I think that's quite bad. And that's the easiest way to do it rather than me, me reading them. Because if there's massive comprehension issues, there will be having to understand what the question means. I'll add to that is when if you have asked a question and David's answering it and you want to get more information, we've got plenty of time today. So you can unmute yourself and have a conversation and do that that's absolutely fine. And we're not going to police the questions or anything like that. So we're not gonna prioritize anything, just we'll just do them in chronological order. And that should be fine. And then at the end, I think David's thought about having a discussion, I think we'll do that based around the framework of the questions that we originally asked people to fill out on the Google Sheet related to molecular machines, directions of research in this area. And with the extra time that we have today, that'll be quite useful. And then we'll lead eventually into what Allison mentioned about, we're talking about the band seed brainstorm is that I've described it I can never, never recall, it's alliterated, which is lovely. So that'll lead quite nicely into that. And then we'll move on together afterwards. So I've done enough talking. So David, the Okay, or is yours?
Yeah, and let's see, like I said, Oh, no, it's the dis host disabled screen sharing. Maybe it'll just be a conversation.
You have you have full power now.
So you owe a full power. Oh, my God. I hope I use it wisely. Yeah, okay, here we go. Um, so, let's see.
Sorry, you present.
So does that work?
Okay, so, um,
right, so I'm gonna talk about de novo protein design, I'll give an intro and then maybe I'll give you some choices, I'll give you a menu of what I have to choose from. So okay, so the way that we've been thinking about protein design for many years, actually, up until very recently, is, is really based on the principle that the folded states of proteins are likely global energy minimum for their sequences. So each protein has a unique amino acid sequence. And that amino acid sequence will have a lowest free energy state, which depends on that sequence. So in this picture, if you're trying to predict the structure of a protein from a sequence, you need some way of computing energies of you know, all the of inter atomic interactions, you need to sample through all the different possible structures, and you search for the one that has the lowest energy. So in this talk, though, I'm not going to talk about protein structure prediction, I'm gonna be talking about protein design. So we're making a brand new sequences that will fold up into new structures. And here, we kind of do the opposite, you start you make up your, you make up your structure somehow. And I'll give you lots of examples of that. You make up your backbone, and then you have to find an amino acid sequence whose lowest energy state is that backbone. So um, so over the years, we've been developing a program called Rosetta. And so what this does is it calculates the energies of any particular configuration of a protein. And it samples through the space of possible structures, and the space of possible sequences. And that's how we've been approaching protein structure prediction and protein design. So very recently, we, deep learning is infiltrating quickly. And so we are starting to use deep learning for for both of these problems, I can comment on that there's interest.
that the problem that that we've really been focusing on for the last five or more years is, is the problem of building proteins completely from first principles. So sequence space is really, really big. Even if just use the 20 naturally occurring amino acids. There are for 100 residue protein, you have an astronomical number of possible sequences. And the number of proteins which have existed on Earth, it's just a tiny, tiny fraction of this. So there's this amazing, there's this enormous space of possible proteins that evolution hasn't explored. So that's what I've depicted in this gray gray box here. And so in red would be not to scale but naturally occurring sequences which clump into families because of his evolution proceeds incrementally. And almost all protein engineering has consists of making small changes to proteins, which already exist, which we affectionately refer to as a Neanderthal protein design. And the what we're doing is trying to build proteins completely from scratch that are unrelated to any proteins of known structure. So we've been doing a lot of this over the last five years. So, um, and I, I, let's see. So what I, what I have here and this, maybe I should just skip this entire section. Maybe I'll just summarize in words on this slide. So for the last year, obviously, we got we have been like the rest of the world very preoccupied with the Coronavirus. And so we use this sort of de novo protein design, to design brand new proteins that block the virus from entering cells, that's antivirals that sends the virus diagnostics and induced immune response against the virus. And I could I could sort of run through this quickly, or I should could just, maybe I'll maybe, I don't know, I'll just give you a sort of give make this protein design problem more concrete. I'll just go through the first one about that. So as you know, the the the Coronavirus gets into cells by way of this h2 receptor protein. And so this is sort of a depiction of a there's a classic protein design problem. You're given the structure of a target like the spike protein of the Coronavirus. And you the Tech Challenge is designed proteins which bind to it. So here we generated in silico, very large numbers of virtual protein structures, and then dock them against the surface and design the interface for four really high affinity interactions. And these are these the blue one here and the orange one, here are some of the solutions that came out from this. And they're kind of remarkable proteins. We've made them in the lab, this one is only 55 amino acids. So it's very short for a naturally occurring protein. It's rock stable, it doesn't melt. And then and then of course, there's the question of Did, did this calculation actually get the answer right for the right reason. And so we can, what we did was to solve the structure of how this protein actually binds to the Coronavirus by cryo em in collaboration with collaborators here at the UW. And so this is the spike tremor. And this this little thing is our computing our design protein. And so you can see three of them bind each one of these binds to each RBD on the spike. And we can now compare this to the model that we had made with Rosetta, when we were designing it remember I told you we are making up a protein that would design your protein that would that would bind tightly. And what's really quite remarkable is that the structure of the of the design protein is nearly identical to the actual experimental structure. So in this case, the sequence we designed really folds up to exactly the structure we wanted. And beyond that, it actually makes the almost exactly the interactions that we were aiming for with the target with the with the Coronavirus, which is down here. So the methods are good on a on a favorable day to actually design the sequence completely from scratch that folds up to a new structure very precisely the bind to target in exactly the way you've designed. And we have used this been using this procedure to make small proteins like this that bind to many protein, sort of proteins on the surface of cells, trying to make new types of signaling related therapeutic proteins. So I'm going to kind of skip through all of this data except there's just basically shows that these proteins actually do work therapeutically, they prevent animals from, from getting sick from the virus. And
I think I will just go through so so this is probably the most relevant slide. So there's all these escape variants that people are worried about now. And we we buy by the actually, maybe I'll go back one slide, if I can able to even not able to do that, by the way we can make we can make versions of these that that are travaillent that match the three rbds on a spike. And these versions, neutralize all of the all of the variant virus South Africa, UK, Brazil. And so we're pretty excited about these, we can make them in large amounts and bacteria. And we are they're headed for clinical trials, which is kind of cool. These are just completely computer generated proteins. So that's the antiviral, and maybe I'll so the this is this is now maybe a little bit more what you're interested. This is like design of a molecular device. So we want to use this for sensing. And so the idea was we could design a system which had two states, shown here and this, this there'll be this off state and then on state and the presence of the virus and then these two would be in thermodynamic equilibrium, and the Presence of the virus would pull the equilibrium to the on state. And that's sort of depicted here. And, and so, um, and the way that and so when this thing opens up, it allows reconstitution of an enzyme called luciferase which generates light. And this works really well. So if you take this molecular device, and you add, it doesn't, it doesn't emit light, but you add the virus, or the the RBD, or the spike, and you get this rapid increase in in luminescence. And we're now so we can design these these multi state devices that are emit light and one state we've we've actually, maybe I'll just skip through this part, we've used this to make basically use this method to make devices that sense a wide variety of different compounds. And so for example, this is a proponent which is involved in heart attacks. This is botulinum toxin. And in each case, we have, we take all these sensors, we add the compound only the sensor that was designed to bind to example for botulinum toxin binds to botulinum toxin.
on the vaccine side, maybe I'll just show talk about this. So we've been designing methods for making self assembling nanomaterials. And this was one of the very first examples where we take a protein with a five fold symmetry and one with three fold symmetry. And we we design interfaces between them, such that the five fold and the three fold sit on the five folds, and three fold axes and Mikasa hedron. And using this approach, we've been able to make many, many different types of very homogeneous nanoparticles. So you make the two proteins mix them together, and they form these beautiful assemblies. And so my colleague, Neil King has been making vaccines out of these by putting the RBD on the nanoparticle. And these turned out to be extremely potent vaccines, they're in clinical trials now, and they seem to be considerably more potent than what's in the mRNA vaccines. So I think I'll skip that. So I'll give you now a few examples. And then I'm going to get to molecular machines, which is what I'll spend most of my time on. So this week's issue of science has another cool example of a design nano structure. So here, this, this yellow, what we realized is that, for these symmetric materials, the basic principle is we match axes of symmetry of, of the protein building block with a point group point group or space group symmetry operators. So in this case, we have a pen tumor that sits on the fivefold axis of an icosahedron. And on the to fold axis, we place antibodies and the nice, the nice thing about antibodies is that they people made antibodies to just about anything, everything. And so by by by using the antibody to fold axis, we basically have that we that it's the constant part of antibodies of a constant part. And the variable part is what gives them all their different binding activities. And we the constant part essentially becomes structural part of the nanoparticle. And so we made so the point of this was we could really unite form and function because we could use the antibody itself as sort of a structural building block of the material, and then have the the sort of the business ends open for interacting with with the target. And we made things with many different symmetries. So here's a dihedral version with two antibodies, here's a, this one I think is tapped is I can't see whether it's cubic or tetrahedral. So you make different by changing the arrangement of these arms and how they interact with the antibody, we can make things that form different types of of nanomaterials and here's sort of a list of them all.
And the protein version of Ned Siemens dealt with DNA.
That's right. It's exactly that. So you know, in a lot of what we've done here is just sort of, it's, it's, it's sort of following in Ned's footprints, but but footsteps but but with proteins, exactly. And then what's interesting, the nice thing about proteins, they have all these functions, so so when we assemble antibodies into these nano cages, they have all sorts of new biological activities, which I think I'll skip because Okay, so here, here's another example of design nanomaterials. So here, we've taken two protein building blocks, one that has d3 symmetry, one that has D two symmetry and designed interactions between them, that dry formation that are predicted to dry formation of a hexagonal lattice. And what's really cool is when you mix these two proteins together, you get this really, really extensive very regular hexagonal lattice. And this is electron electron micrograph. And if you look at it, if you if you average, you can actually, it's quite close to the design model the here's one component. And here's the other component. And so if you ever need of protein chain mail, you know, let me know, what well we've been doing with this is attaching binding domains for cell surface receptors, and we get the cells to undergo all sorts of interesting transitions. But obviously, sort of one idea is self assembling circuitry. A lot of you know, a lot of, and I know this is this is this has been a theme with the foresight Institute for a long time a lot, a lot of you know, manufacturing is top down. and here we can proceed bottom up these everything I've shown you, you know, those those nano nano cages of all the different shorts and these kind of hexagonal lattices. These aren't patterned from the top down, these are made by programming very specific interactions between atoms.
we can use, we can actually use these kinds of devices I described earlier, to do calculations on the surface of cells. And so like one of the challenges in cancer therapy is good, you have little machines that go in and kill the tumor cells, but lead leave normal cells fine, untouched. And that's hard problem. These cancer cells don't often differ very much from normal cells. But if you could do calculations in the body on the surface of cells, then then then you could then you could make more subtle discriminations. So, so here, we're This is sort of a diagram of how, how we discriminate cells that do and logic on the surface of cells. So recognize cells discriminate cells that have two markers on their surface, opposed to just one or the other. And this, this really works, I don't think I'll take you through the details. But we can actually make do one A and B, not c logic. So the basic way, Well, okay, roughly speaking, how this works is we have sort of a closed, one kimono, which isn't a closed state, if another component gets brought to it, it opens it up, and we get a signal which, in this case, car T cells can use to kill the cells. But then we also have a third component that we can target that gets targeted, there's a third marker present, say, a healthy cell marker present on the cells. And that shuts down this whole thing. And so you turn it off, so you only get, you only get response on the cells that have markers, one and two, but not markers, three. So in terms of building devices, one of the really interesting types of proteins in biology are channels, things that allow ions to go through. So of course, we've been interested in building those from scratch. So here we have here, a crystal structures of different types of cylindrical pores we've developed. And we can actually incorporate these into membranes. And when we do, it turns out they are very specific channels. So this one, for example, is a specific potassium channel. So it's kind of interesting about this is, people have been arguing about why potassium channels and nature specific for potassium for a long time. And here, you can just sort of, so a lot of biology is descriptive, you observe things, then you try and build up explanations for it. But here, we can just sort of prove by construction. And so here, we just we build something from fairly simple principles, and then end up with a fairly specific potassium channel. And I think I'll skip this one. And we've also designed larger pores here that allow larger molecules to go through. And then there's so what I've described so far are our helical transmembrane proteins that have rings of alpha, Hela, C's, but some of the really interesting transmembrane proteins that are, for example, using single molecule DNA sequencing are beta barrels, and I sit in the membrane. And we've recently learned how to design these proteins from first principles. And so now we are designing a series of proteins that look like this with larger and larger pores, and that we think we'll be able to functionalize for all kinds of molecular filtration processes as an addition, more sort of new approaches to single molecule analysis, because you can look at the way the same way that single molecule DNA sequencing works is you can look at changes in conductance as things go through the pores. We have a game that where what that sort of an adjunct to what we do were sort of just people in the general public can go in and design proteins and they can pull the proteins around sort of as illustrated here. You can change the amino acid sequence and we've been been testing a lot of the designs in the lab and getting a lot of interesting ideas from from all the very smart people out there who aren't from Rational scientists. Okay, so here this is this the I can go in a little more detail here. This is now actually our steps towards designing molecular machines. And this is the work of a super talented postdoc in the group, Aleksey Kobe. And, and so, and these are his slides, which I showed which I stole as well, you all recognize the quote at the top. And so his idea was to design machines with out of rings, and basically rings and axles and try and assemble rotary systems at the protein level, so this is kind of at the mechanical level, he's trying to assemble things like this that have like, kind of an end, cap on the end, here's a, here's an axle, and then here's a rotor that spins on the axle. And so the basic ideas, design these parts, the rings and the axles, then assemble the ring on the axle, and then incorporate hydraulic chemical reaction at this interface to actually drive rotation.
And I'll show you where he is on this. The as you can tell, he's a little more colorful in his slide construction than I normally have. Okay, so the first class of materials are these axles, which are basically rods, that that generally have something that we built, that Alex's built something on the end, that sort of to sort of hold the rotor on the ring on. And these are different types of constructions, these are electron density maps. And he's got quite a few different shapes. And what's interesting is that they have, they have different symmetries. And so and so this matching or mismatching of symmetry will be an interesting thing as I go through. And he can extend these to make to make it fall easier to follow by imaging. So those are the those are the rods, and then here's the rings. And we again, he has got quite a few different flavors of rings, these are all like everything in my talk. So far, these are all completely de novo design proteins. And, and so so so Okay, so we've got these rings that are structurally validated these rods that are structurally validated. And now the here's just some more structures of the rings, just so you can see what that rings in rods, these are the ones that he has the highest resolution structures up. So you can see we want to put these various types of rings on these different types of rods. And I actually thought that was gonna be pretty hard, but it turned out that Aleksey was able to, to thread the rings on the axles pretty straightforwardly by by by doing some primarily incorporating electrostatics between the ring and the axle. And here are and we can play tricks like using pH or or or de sulphide reduction to actually assemble and trap the rings around the rod. So these are some of the systems that Alex's main now so they have their ones where the symmetry matches. So both the the axle and the rotor have see three symmetry, then there are others, these other ones here where the symmetries don't match at all. So the axle For example, here is d3 symmetry, but the rotor has D five symmetry. And what you see here are actually cryo em maps of the rings, the rotor, the rings assembled on the axle, and your you see some views of these. Now what Alex has done is to calculate the energy landscapes in Rosetta basically was doing that there's one degree of freedom as the ring rotates. And you can see the landscapes look really different. As you might expect, when there are when there's symmetry matched, you have a you have more three pronounced minima, because there should be three rotationally identical states. Whereas when you have symmetry mismatches, for example, B eight on c four or D three on C five, you get much more rugged landscapes. And here's an example of of one of these, this is the one that was on the left in my previous slide. And so by by cryo em, Aleksey has actually solved different states of, of this system and you can kind of see them here. So they are, here's here the what we can distinguish them because the the road they're two rings that are mounted on, on on the ring, and they have these, these things emanating from them. So you can see here they're more eclipsed, and here they're more staggered. And it's kind of neat out they correlate to minima and this rotational energy landscape that Aleksey is mapped out. So we can design the systems that undergo that have multiple rotational states. So now the question is, can we get them to undergo road directed motion. So the way that Aleksey has approached this is to build in catalytic sites. Many years ago, we started doing enzyme design with Ken how cool I was happy to see here. And so we've been, this is a reaction that Ken will recognize the retro alldis reaction with but we've put a very simple site in it, this interface between the ring and the rotor. And what we, what we know at this stage is that the the, this is an active enzyme, there's a so the enzyme turns over. And we can also chemically modify this site site with different chemicals that are sort of like suicide substrates.
Here's a picture of how the substrate fits into this binding site. And I'm
lips whoops, sorry.
So here, here are some of the compounds
so that we get enzymatic activity. And the so we have very, so obviously, what we want to know is what happens when we add the substrate so we're collaborating with Zeb Brian's group at Stanford, they get these very tantalizing single, they basically put a gold bead on the end. And of course, what we want to see is, is a rotational motion. And so this is this is sort of Angular rebel, angular motion over time. And but there's a serious data selection bias here. So we don't really have definitive data yet on whether we have wonder sort of uni directional rotation. But we're working hard on that now. These are just different trajectories. zevs group is getting in there single molecule experiments. And that's, that's pretty much where we are now with, with the design these rotational systems, I should say, we're also designing systems. We, I didn't show it in this talk. But we have we've designed a variety of protein, one dimensional fibers using the same self assembly systems. And we're now using sort of similar principles to try and design walkers that walk along the fibers. So you know, that might be that's probably a good place for me to start, I have a little bit of on on deep learning, as opposed to as applied to protein design, but maybe that's that may be of less interest here. So just to summarize, I think we understand a lot about protein folding and assembly and now we can really start building a whole new world of functional proteins which is which is very exciting. And just to quickly acknowledge the people did the work long Jim and Brian did did the design of those many proteins that block Coronavirus. We have a lot of collaborators for all the animal data the the the the molecular devices that sense the viruses work of Alfredo and Andy the self assembling nano cages those antibody Nandy nano cages, Robbie divines work that two dimensional sheet designed by Arielle Ben Sasson the membrane proteins that the synthetic potassium channel work of shunfu and pay long, Anastacia design the beta barrel, the systems were doing logic inside cells work by Mark and Scott and Gillian fold it is the work of Brian called Nick and a number of other people really all around the world now, and Aleksey Kobe is really the kind of the genius behind the rotary motor stuff. And I didn't really talk about the protein hallucination. So why don't we Why don't we go to discussion mode here? Actually, I guess I can check. I'm supposed to check the questions. Is that in the chat?
We do have a few questions.
Well, we have a few hand raises. Let's see. We have Korean and Korean maybe say one or two words about you and then we move on.
I can also talk for two minutes at some point about the deep learning stuff that people want to hear.
Well, I'm sure they want to hear about that. Um, yeah, I'm Korean. I'm a longtime foresight person and worked in nanotechnology and
just go and lovely beautiful work and then I knew net soon and all those guys back in the day. Anyway. Um, so I'd like to know maybe one level down if you don't mind. The Like under the hood of some of your modeling, to what degree? Is it molecular mechanics? To what degree? Is it electronic structure? You know, all the complicated stuff like, what about the water?
What about the reactions? Is it just shapes? Is it more than that? Or is it all? Is it all, like machine learned concepts where you don't have to worry about physics?
It's a really good question. So it's basically molecular mechanics with basically the the the main difference between the Rosetta forcefield and traditional molecular mechanics forcefield is that we treat the hydrogen bonding orientation dependence explicitly, you need to use the hydrogen bonds give you a lot of the precision and the orientation control. And when you have two surfaces coming together, you have to make all the hydrogen bonds. These those groups are making interactions with water before. So basically, but about the functional form. It's basically a pairwise additive something largely pairwise additive sum over Adam Adam interactions, there was a force field.
Wait a second, that that's great. Thank you for sharing that. That's amazing. But that's not going to make, you know, a sodium channel or a motor. Right? There's more electronics?
Yes, you're absolutely right. You're obviously right. And in, for example, the enzyme work we did with with Ken, you know, obviously, that that's clearly gets into quantum mechanics. So. So, but there is an issue of control and what control we have and what control we don't have. So. So I think it's so you can sort of think about it as so for the potassium channel, for example, we design things, we designed structures using sort of this this force field I described, that had a cylindrical aperture, that was about the right size, or potassium channel passage, but we did not explicitly model before the fact the, the, you know, the dynamics of potassium traversal. Likewise, in the molecular machines, those were rotary machines, you know, if that was, so it's sort of a combination of the sort of atomic level design, and then shape reasoning, right, so Aleksey designed the rings and the rods, but in terms of precise modeling of of an eye, you know, we had these after the fact energy landscapes, but and and then I showed you we designed catalytic sites in there, but we're not computing in advance what the, we're not choosing the catalytic site, based on what would maximally power rotation is much more heuristic than that, we put it out at the binding site. So there's a combination of sort of real atomic detail at the base design with sort of hypothesis, build design, build test, at the sort of the more meta level, because, you know, he's our calculations just aren't accurate enough at this point, to, you know, to work out, you know, exactly how to design this molecular, you know, rotary machine.
Okay, one last question. Follow up is that, um, this is beautiful, what you've described this sort of multi level analysis with, you know, molecule mechanics at one level and kind of human heuristics. Other level, tell me a little bit, if you will, about the feedback that's happening, like, presumably, the mechanics are informing your heuristics, heuristics, and then maybe heuristics and formula mechanics?
Yeah, well, there's a lot of feedback. And part of it comes from saying, you know, in a sense, my talk was probably the most misleading team you've ever showed, heard, because I showed you all these beautiful things that work. But 95% of what we make doesn't work. And so we get a lot. And so typically, there's a lot of different failure modes like, so basically, what we're doing is we're doing these design calculations, we get out a protein structure, which has a particular amino acid designed amino acid sequence, we make a synthetic gene that encodes that amino acid sequence we put into bacteria, the bacteria make the protein, and some reasonable fraction of the time, they simply don't want to make the protein at all. So there's no protein made. Other times they make the protein, but it's just, you know, you know, turns into cook at the bottom of the tube. And other times, it makes a perfectly well behaved protein, but it doesn't do what we want. So we get a lot of feedback, we're, you know, we're we're getting a lot of data back on what works and what doesn't. And so that's definitely, there's definitely been a lot of going of iteration. So we can go back to the forcefield. Look at what try and figure out what's wrong and prove the parameterization, and so forth. Also, our understanding of how to approach problems is evolving as well. So it's really what you described is exactly it is very detailed physical model with heuristics sitting on top of it. But you know, maybe what I should do is just show a couple more slides on the deep learning stuff. And you'll see that's kind of kind of, I mean, I can go through the comments more, I don't know, whatever people would prefer.
Well, there's been questions about learning. So I say we all say yes,
yeah. Okay, so So obviously, so, like I said, we're collecting a lot of data. And so there's a really interesting opportunity to use ml Let's see. So if I, I don't know how I get back to my slides,
you should be able to just if you use so cool. So if you go to the bottom and
Okay, sure. Yeah. Okay, sorry. Where did it go though?
I'm sorry, I'm just having a hard time finding.
If you go to the green button where it says share screen?
Yeah, no, I know, I'm just for some reason, it's gonna be all you have to, you have to,
you have to first click on the presenter, like you have to click on it, and then you go, and then it should be the first one that pops up on the left. But if you had a different browser on before, then then it just shows you their browser, always
open your slides and don't minimize them and then go back to to zoom.
Okay, let me try that. I thought I just tried that. But it didn't work. So if I if the time I'm in presenter mode. If I am in presenter mode, I'm just somehow it's giving me 100 zillion mail messages. Why the hell is it doing that?
Well show us that.
Yeah, you you know, I mean, okay, let me let me just let me just try. I'm gonna try exactly what was suggested. So I will, I'm now I am in present. So I go to present removed mode. But then how do I get from there to back to the zoom window to
just click on this? You don't have to be in presenter mode. Just bring that up as the last document. Don't minimize it. Okay, go back and share screen and you should see it. Okay. Yeah, I should really be there. I just don't know why it's not. Oh, here it is. Here it is. Okay. Sorry. I don't know why wasn't finding it before. Okay.
So it always happens when you talk about AI?
Yeah. Okay. So, um, so all so we've been playing around with. So basically, that the, the prelude to this is we've been working on we, you know, structure prediction is a big industry in my group. So we've been working for several years on methods for predicting protein structure from amino acid sequence. And we, like everybody have been, you know, DeepMind has been doing great work on this. And so we've, we've basically, I think we're sort of the people who actually make some of their ideas available to the world because, you know, we in court, you know, we basically try and build on what they do. So okay, so I'll talk about three very briefly three, basically, this idea of deep network hallucination. So Okay, so here's protein structure prediction, you have an amino acid sequence, you predict the structure, you get some kind of, well, what we actually predict is in this institution is distance map distances between all these different pairs of residues. And what we can then do, and in the same way, that if you have a deep network that recognizes images of cats on the internet, that you can, you know, start to feed that noise and then optimize the noise so that the network really thinks you're you it's a cat, it's looking at you that way you can use to sort of hallucinate brand new images of cats, we can use the same idea with proteins. So start off with random sequences, predict their structures, they don't look like proteins, but then we can optimize the sequence such that the network really thinks it's a protein. And so this is, this is what a this is a residue by residue contact map. And if we do this optimization process, at the end of day, we get these very featured contact mess maps. So this is residue by residue, and you get a black.if. Those two residues are close in contact. And these are are, these correspond to protein structures. And when we do this, starting from different random seeds, number seeds, we get lots and lots and lots of things out. Lots and lots of proteins out. They look like proteins. They're not related to any naturally occurring protein, and Rosetta, which in this molecular mechanics model predicts that they actually fold up to the hallucinated structure. We made a whole bunch of these, and they actually fold in the lab. And we've we've solved structures now of three of them. And actually, these hallucinated the seek hallucinate sequences, fold the loose native structures, which is which is pretty neat. One of the things that this enables is sort of a shortcoming with our sort of energy, but sort of physically based method for protein design is we're always, when you when we're optimizing the sequences, we were optimizing for trying to minimize the energy of find to find the lowest energy sequence for a structure. So that's kind of depicted here. This was everything I described so far. But there could have been off target structures that that had, that were competing that we didn't know about till the end of the day when we tried to search the landscape. But with this approach, I'd be striving we can actually explicitly maximize the probability of the desired structure as opposed to other possible structures because we're essentially using the network to predict a probability distribution over all possible structures. With the prince that was the difference that took you from the first state The second stage? Yeah, so that's a good question. So in this molecular mechanics model, right, you have the interactions between atoms and you it's very near sided, right? It just seeds the structure that you have at the constellation of atoms you have at any given point. So if you're trying to find optimize the sequence, to have to find the very low energy sequence, you're doing this combinatorial search in sequence space, trying to get very low energy interactions. This, the network instead is reasoning over the probability distributions of distances between all pairs of residues. So as you change the sequence is updating all those probability distributions. And so you're essentially, yeah, it's like, you know, it's, it's, it's one that the one that the deep learning thing essentially knows about the partition function implicitly when it's doing its probability estimation. Okay, so the last thing, and then I'm going to shut up is we can also we can start designing things that a function by constraining parts of the protein to have shapes that are relevant for protein function. And just to give you one example, so here's a big protein, that's part of the the complement cascade is called this big protein, it's got this business, then down here in green, which binds tarde here, what we can do is take this green part out, say we want a sequence of protein, that sequence folds up to a protein that has this structure, we want everything else to be free. And then the network will just hallucinate different proteins that have this region. And we can do the same thing. Like with an enzyme active site, you say, okay, we want this part to be fixed. And then the rest of it can be whatever the network thinks is best. It's like fixing the cat's ears and then asking the network to come up with the rest of a cat that would have those years.
And, yeah, so this has been the work of really talented people who are listed here. Yeah, so that's kind of the that's kind of the deep learning for sort of new protein design. And then there's also the very exciting area with generating all this data from all of our six failures. And using that to iteratively. Improve the model now. So there's, there's a, there is I'll stop sharing now. So there's now this, you can totally appreciate the issue, though. So I would I always thought we were going to be doing is we're collecting this data to improve the physical model, you know, the parameters, the functional forms. But now, you know, these deep learning models, there are no functional forms, and there's millions of parameters. And you know, they work but it's, and I see we're at the stage now, where there are this tension between the deep learning models in my group and development, and then Rosetta, and it's not and right now we're kind of going back and forth. But it's hard for me to predict exactly how that's going to evolve. It's going to be very interesting. Oh, I see a question. I'd love to hear more about the self assembled 3d porous materials.
Let's say before we before we get to that, David, sorry to go mean, dog trainer had his hand up for a while. So we'll just have him ask his question first, which I think came in before the questions in the chat did. Oh, that's all I told him. So that's what we'll do. Sorry.
No, I just yeah, in any orders. Fine.
Hi, David. This is third from NYU. David, back to that rotary machine, which I found absolutely cool. reminds me a little bit of heroes. Kelly is chemically power roller. I don't know where you cross that. So how is your spot? Exactly? If I understand this correctly, this is a retro l direction, right, which should have a positive delta H, is this purely entropically driven? Or what is the power stroke in this machine?
Yeah, that's an excellent question. And so what I can tell you, I can say at this point is we have not, so what we can do is we can assemble these systems, and they clearly are undergoing rotational diffusion. But we have not convincingly demonstrated directed motion. So there's so and it's basically this comes back to the previous comment. So we are still trying to characterize you can imagine it's a bit of a tricky problem, you need to do single molecule fluorescence, or some other measurements to see. So basically, we've been faced actually for over a year now with just technical problems and how to really monitor what's going on. So. So the idea ultimately, would be some sort of Brownian ratchet, right, that the that the binding of the substrate changes the energy landscape. And then there's actually after the hydrolysis there is you remain with the product bound, then that comes up. So the hope is that we can sort of deform the energy let the rotational energy landscape in such a way as to get directed motion. But I we have not convincingly demonstrated that now. And actually, if anybody has ideas about how to characterize these systems experimentally, you know, we are we need help, they will be really cool to run into some form of decarboxylation. But making a gas right.
Or some something along the lines and
I think that was the only question with it with a hand up. So David, you can now go into the into the chair, sorry to cut you off with Yeah,
really porous materials we're working now. So those arrays get The 10s of microns long, you know, I think they would get bigger at some point, you start using up all the protein, we actually now can design three dimensional crystals. And and we actually have some been designing hydrogels that look like non Newtonian fluids. It's very interesting. So I would say there's not really a limit on the size. It's I mean, these these are all symmetric systems. It's really just an it's they're limited by the amount of material that you produce.
And then Tad's goes hand up, so that if you want to, oh,
what's the limit on number of accuracy? nanoblock.
Thanks. Hi. Hi, David. I am enjoying this, this discussion, I a couple of questions on the protein logic that you described. And so firstly, I'm wondering what sort of what's the limit on the number and the accuracy of the Boolean operations that these devices can perform on the surface of a cell? In particular, if you start having a lot of logic you'd like to do to the devices get in each other's way. So you have to kind of space them apart? And then maybe there's errors in the implement. And the second piece of my question is wondering whether you can include delays in these logic so that you could detect those receptor one that happens, and then a little bit later, it's number two, as opposed to two in the order of two and one. So can there be delays in that it's a little bit more complicated than just Boolean logic?
Yeah, no, that's a really good question. I would say this is a rapidly evolving area. And I think there are issues about crowding that would come up. Ultimately, there's also I kind of glossed over it. But the way that we're, the way that we're doing that got targeting large logic to specific cells is through domains that bind to proteins that are expressed on the surface of those cells. And we need those to be present at a at a high enough level. So right now, we're kind of trying to optimize those sorts of things. We're also designing synthetic receptors systems, so that you can control cells from the outside, by and there, you can get delays, we're building integrating devices that sit inside the cell, that would sense the activation of multiple different receptors. And so that's sort of the way that we're trying to get at sort of timing more from the cell and trying to put these design devices into cells.
how would you? what's the what's the picture of going from machines to function? You know, I think all of you will simply I mean, so basically, I my experience has been whenever you can achieve a real sort of breakthrough in technology, there's always applications that you can't anticipate. And so, um, we have not really, you know, they're crazy things you could imagine with a rotary machine would be like, cleaning up, you know, amyloid fibrils and neurodegenerative disease, you know, because basically, if you add that they're they're rotary systems in biology that untangled DNA, for example, but we have we, we really haven't gotten delved in that much. chaperonin nascent protein interactions, we haven't really looked at that. We're really focused on de novo protein design. And most of the proteins I described, I'd say essentially, all of them fold spontaneously, so. So that hasn't been so much of an issue. And then I think I can thanks for that. Thanks for the nice comments. A lot of the failures are due to aggregation, I would say because, you know, a lot of these systems they interact via via some hydrophobic interactions. And so those can also lead to aggregation. So how does the design space work for flexible structures? Well, a lot of the design stuff now, I would say, is not completely flexible, or trying to design multiple multi state systems with multiple discrete minima. And and they're you basically just have to come up with, you know, they're the design part is harder, because you have to be multiple independent states. And as far as designing intrinsically disordered proteins, it's very easy. That's what we call failed. And then failed designs are basically intrinsically disordered. So but we are, we also have systems that that are disordered, but then become ordered upon binding. So yeah, I think that's, that's most of the questions on the list.
That's cool. Cool. Most of them in Ben's question, sort of the thing that one of the questions that this group is sort of targeting is sort of speculating on where things might go in the next like 515 or 30 years. And so I mean, obviously, there was like a very sort of specific connection that was asked about this function. Where do you have any like, because I mean, this as soon as I said, the sort of stuff you do is basically magic. This
Saturday, I would
say the overall goal is like really patterning matter at the atomic scale. So a lot The the allowing the nano space, this principle of self assembly, we're able to pattern things, you know, with these, you know, hexagonal lattices, nano cages, different point group symmetries. But if you want to pattern more extensively and asymmetrically, then I think that's where things like the motors come in, like the walkers, because you could now start moving things around. And so I think that's probably that's one way to think about sort of, like, suppose you want to assemble a circuit, an asymmetric circuit, then you're gonna need something that can do work and that you can control. So I think that the general form of the the answer to the question, I would think about these, these systems that use chemical energy is actually pattern, you know, you know, patterning matter in in highly directed ways at the atomic level.
So that's probably would you agree, this is me leaking to the soul question, you say that was an area that you feel the the community should focus on as, like, where there's maybe a lot of high value to be had from doing
that? Well, that's certainly a very general form of the problem. And
I don't know, I think it is hard to anticipate, you know, I think that's sort of the history of technology is they're they're sort of technological revolutions, and then people figure out what it's really good for after the fact. But yeah, we just be really happy to see a spinning motor, a uni directionally spinning motor or a nice Walker walking along the fiber right now.
Yeah, absolutely. I mean, if there if there are no other questions, because we're, we're, we've done fairly well for time, Allison, I suppose we can probably go to your
Yeah. Well, okay. David. I mean, we still have three minutes to view your mind. We just kept getting questions out of it. You need to see which you get to. I
don't know, but I think we went through all the questions.
Okay. Well, I
come in, okay. We ask everyone,
or whatever. Yes, sure. I forgot about that. Yeah, go ahead.
Okay, well, so is there a tool or an enabling technology or something that you really, really wish you had that would significantly advance your field? Where you like, if someone listens to this in the engineering department make me this? I want that?
Yeah, well, like I said, we're having a hell of a time characterizing these rotary rotary machines. So sort of measurement devices for measuring directed motion on the nanoscale, that that's that that's really a major limit right now.
Okay, and if you could ask people in this group, what do you think is like the number one question that people in this group should have an answer to, to drive progress and molecular machines? Like, if you were in my position, what would you ask everyone here?
Well, I think I
think it is, a lot of it is measurement methods, like I would say develop methods for quantifying the the behavior of molecular machines, because to some extent, you can't really design before you have measurement. So I think I would, I would really urge sort of thinking about, you know, how will you quantify the behavior of molecular machines as people start designing them?
All right, thank you. And then we have one from William, she's done the question, can you measure a rotary motion with a micro length lever attached to the rotor?
Yeah, that's a really good question. And let's see. So one way you could do that is if you made the arm really long, you could put a fluorescent fluorescent group on the end, and then you could watch it that was what was done for the ATP synthase. And that is where we're basically playing around with engineering ways to do it. Yeah. So that's, that's certainly a very good idea to basically that sort of transferring. That's trying to train the length scale over what you need to measure amplifying it by basically going out to a larger circumference circle. Yeah. So I think that's a very good idea. Okay,
and then let the two questions for one minute. One is what would you recommend young people in new fields to focus on now? And to do and then the other one is, what can this group to help with your work? Is there anything we could do for you after you've gracefully given us an hour of your time? What do you need? Well, I
think there's really stimulating discussion. I mean, I think, again, if people have ideas about about experiment about measurement methods, I think that would be that would be really good. I think sort of the question about what would you do with machines once you have them? I mean, at some point, this field is going to have to make an argument for why you need them. And I think your group is is very, very well poised to that. So I think that will definitely help build enthusiasm for these views. Of course, you need come in some kind of society buy in to really do something like this. So I would say measurement methods, applications. And then the you had two questions. So when I think that was the second one, what was the other one?
What would you recommend young folks like? yangdi?
Yeah, let's see. Well, I think one of the things I would recommend is to try and become a generalist. Because there's a lot of different things coming together here. You know, there's, you know, there's the physics, engineering, a measurement there, sort of the, the, you know, Physical Chemistry, Biology of machine, you know, proteins and machines. There's the sort of the computer science of deep learning. So really trying to getting a broad view, there's, there's so many different things converging in this field. So that that would be my suggestion, don't become a specialist too early.
Okay, thank you. I hope you didn't mind that. I tried to fit them in.
Oh, no, no, not at all. I would have felt bad otherwise, because I just want to remember, let's see, there's
no. Well, thank thank you very, very, very, very, very much for your time. It's really, really much appreciated.
Sure. Of course, no, no, that was a pleasure. And if people have ideas, um, definitely send me an email, particularly about, you know, like, Well, I'd say about, you know, things like measurement problems and applications. Okay. I can't talk to everyone.
Yeah, are you? Thank you, David. And everyone else would like to stay on Ted, and I have a little bit of a report out, I just want to make sure that we get all the time that we have with David. But we had a meeting on Saturday on the molecular machines bounty brainstorm. So where we pay your prizes? And you answer the questions, some of which I asked, they would know. And so the idea is that, you know, we pay you for the expertise that we have, so that we can generate a report that we can then put into the in front of the eyes of potential funders, some of which, some of whom have funded this group. So hopefully, there are some, some, yet some, actually some some potential that this funding will will get us somewhere. And Ted has a little bit of a report out and then for everyone else would afterwards just like to stay on and socialize and want to open up our gallery space in which you can meet other folks here. Okay, Ted, I'll hand it over to you and I will allow you to co host dude,
okay, am I'm on muted. All right, I'm pasting into the chat. Okay, there's a URL in the chat. So in the molecular machines group, we've been building a table, which you probably all of you know about already. Let me see if I can share my screen. See here. So this one here, okay. Okay, and share. Sorry, I'm new to this. Okay. So this is a table of devices. So people keep mentioning all these things that were in journal articles that do something or demonstrate some principle. And so I wanted to collect them all in one place. Actually, Tom Schroeder suggested that we do this because I think it will have a benefit later. And so we started on this. And we on last Saturday, we put a bunch of things in here. So all of you, it's open. Anybody who has a URL can edit, please, either take one of the articles mentioned here with nothing beside it like this one, and go look at the article and fill in the rest of the table, or have your own idea of one of these devices. That is one you know about and make a new row for it and put it in. And we would be very grateful to have this table filled in. It's really important. Okay, good. Now, I would like to now On another subject, let me get another URL just a moment here is what I mentioned,
our nomenclature for how to notate who put in an entry?
Yes, please. Yes. So the column on the right hand of that table says initials. And so if you put in a row or did most of the work on the row, just put your initials in there, and maybe your email address since we might not know exactly who you are. And so it may email address a name in that initials column would be great. Okay, back to here that's affected just one cell then they would put it there. Yes, yeah. You can put it at the end of the cell, like in brackets or something like that. If you just did one cell. Okay, I'm having trouble getting back to the chat. In zoom. Anybody want to help me here?
I would say just close. Close the recording. Sorry, not close the screen here.
Okay, so escape up. I think I've got the chat button here. No, they have got now. All right, good and, okay. And, okay, so this is On another subject. Now. Going back to this. How to get rid of the chat.
Let me just share it with you a link in the chat.
I did that come through my link shared. Okay, good. Now over here, come on. Just a second here. Oh, come on. Okay, this all right, can you see this thing. So on February 18, Dean Estonian showed us have talked about a very interesting thing. And I know he's maybe he actually in the talk, if he would like to, to say something here, that would be wonderful. The proposal is a rod with a couple of rings on it. And these rings, attach a and b, and then you change the pH and bring the two rings together, and then a and b might bond to each other, which would be great, because then you could go this way, and get rid of a bonded pair, and then go around and around and around powering this thing. And taking, but you're basically forcing a chemical reaction to happen. So it's a chemical synthesizer, you know, sort of at room temperature in a single flask without filtering. You know, this is be very, I think it would be a great product and would be the first stepping stone on a chain of you know, of commercial activity, which would be really important. And of course, we don't know exactly how to build this. I'd like to know if any groups are working on it. I don't know yet. If any groups are actually trying to do this.
Everyone just said in the chat. This there are.
Okay, let's see if I can get back to the chat here. Okay.
I didn't know, Stefan, have you. Let me just have to say something.
Okay, okay, good.
I'm gonna share the link. Yes, thank
you. I appreciate that. And I'll go look at that. Also, the, the thing that's able to make four different shapes of the same molecule that we've talked about before, that has the arm that swings back and forth, that must be able to release those products and get new and new substrate to start on. So that motion. So parts of this have been done. And of course, the the rod with the rings on it has been done. Okay, so let's see. Okay, so my main message here is that there's a little paper about it, that the URL I put in the chat leads to this paper here. And so if you're interested, look at that. And if this is all incredibly inappropriate, you can tell me, but I think we're really onto something here. And Dena Simeon has really like open Pandora's box, I think here so that's pretty cool.