Scientist Stories: Lulu Qian, Constructing Artificial Molecular Systems

Great pleasure to welcome Professor Lulu Chan to give the seminar today. So for those of you who don't know Lulu, she did her undergraduate and PhD in China, I think in Nanjing and Shanghai, before moving to Caltech where she worked with, in particular with Eric Winfrey, which is how I got, got to know her. And while she was there, they did some really quite remarkable work in a number, a number of times. It was like, oh, here's another paper, and it's massively pushed the boundary of what people can achieve with nucleic acid nanotechnology. Again, really, um, nucleic acid nanotechnology is a field where people, um, will say, look, it has the potential to do XYZ, and XYZ sounds really cool, but then people will actually demonstrate something really quite simple by comparison. And then I always felt like Lulu and Eric actually believed the hype and therefore would actually try to demonstrate the really impressive big systems with it. And so a series of those experiments, and then they have continued after a brief trip to the Wyss Institute in Harvard. Lulu's now set up her own lab where she continues to do pretty much the same thing, generate amazing papers that really blow everyone else out the water. So I'm really looking forward to hearing her speak today. As we go along, if anyone has any questions, put them in the chat box. If I think it's particularly pertinent, I will interrupt, otherwise we'll return to them at the end. But without further ado, Lulu, please take over.

Sounds good. Thank you, Tom, for that very nice introduction. I'll just go ahead and jump in. If I can switch my slides. Okay, so machines can be programmed to carry out mechanical tasks and especially roles that humans cannot do well or do not wish to do. And these tasks exist at all scales. And I am particularly fascinated by how to make machines out of molecules and program them to perform tasks at the nanometer to micrometer scale. And to do that, I like to first think about the engineering principles that are desired for building molecular machines. So, one could use complex mechanics to build simple behavior, which is a fun challenge, or to do the opposite. We can use simple mechanics to build complex behavior if the right building blocks are wired together in the right ways. And because molecules are pretty difficult to manipulate, It is probably easier to focus on simple mechanics that could be composed together for interesting functions. Then remote control vehicles allow humans to direct the machines to do just the right things. But with a combination of embedded computation within the vehicles, they can perform some tasks on their own and handle unforeseen situations in an environment that humans don't necessarily have the full knowledge of. And because a molecular environment could have lots of unknowns, it will be desirable to build molecular machines that are capable of performing autonomous tasks. Now, machines built for a specific task can be optimized for just that task, but modular machines that are reconfigurable could be used for a variety of different tasks and adapt their behavior to changes in the environment. And because it will be difficult to build an army of molecular machines, each optimized for a specific task, and to have them collectively make a decision for which ones should be in action in response to a changing molecular environment, it will probably be desirable to build molecular machines that can reconfigure themselves to efficiently perform different tasks. And so overall, I'm interested in developing algorithms and architectures for molecular machines that utilize simple mechanics, embedded computation, and task-specific reconfiguration. So today I will tell you one example story for each of these three principles. And first, how do we use simple mechanics for complex behavior? Well, there are three important types of motors in natural molecular machines. Motors such as kinesin perform transportation tasks to deliver molecules to where they're needed. And motors such as actin network perform shape reconfiguration tasks to allow mobility and adaptability of cells. And then motors such as RNA polymerase perform assembly and disassembly tasks to build molecules when they're needed. Accordingly, in the field of DNA nanotechnology, we have seen all three types of motors, and some of them are called robots. So here, I like to focus on motors and robots that perform transportation tasks. So, this is a large warehouse in China. So, on one side of the room, you can see that the incoming packages are being Notice—

excuse me, we seem to have some audio issues. We can't hear you very well.

Sorry, that music louder than I thought.

It might be worth maybe replaying the, um, the slide and then speaking on the slide after it so that everybody can see what happened in it.

Ah, is there a delay in how soon my slides are showing up?

Um, no, they just seem some issue with the audio, you talking over it. So if you would just let the slide play and then the audio can play and then you can speak after the audio finishes.

Okay, I think that might be the only one that has a sound, so it should be good to move forward. Sorry about that.

Thank you.

Okay, so we developed the DNA robots to perform a similar task, and now these robots can pick up molecular packages on one side of the room on the surface of a DNA nanostructure, and then drops them off to the other side of the room. And the drop-off location is determined based on a DNA barcode, like a zip code, tapped to each molecular package. So, I started this project when I was a postdoc working with a team of undergrads, with help from Niranjan Srinivas, Damian Woods, and Eric Winfrey. But we didn't get— and when a new— joined my lab, she took over the project and led it to a beautiful completion. And so the history of building DNA molecules that can walk around started over 10 years ago. And their first steps were powered by adding DNA strands to the test tubes one at a time. And then soon they were able to walk autonomously with the help of enzymes or deoxyribozymes. And if we look at the numbers of the steps that they can take, And that number has been growing exponentially over the years. And then DNA walkers that use just the base pairing were also developed with consistent progress over the years. More recently, DNA walkers become capable of walking while performing additional tasks. And probably we were able to build a DNA robot that bumps up the complexity of this particular category to approximately 300 steps. So I believe that two strategies that played an important role in building this cargo sorting robot is to use simple algorithms that are compatible with simple and modular building blocks and explore molecular processes that are energy efficient. Okay, so let's imagine if we were to design a macroscopic robot to perform a cargo sorting task, we probably consider the following algorithms. The robot would systematically explore the relevant area, like your Roborock does, and if it bumps into a cargo, pick up the cargo, recognize the cargo type, and choose a path directly to the destination for this cargo type and drop it off at the destination. And then the robot repeats the process until all cargoes are sorted. Now, to implement this algorithm, the robot should have a memory for where it has been the ability to recognize different types of cargoes, and a procedure for choosing after pass desired destinations. Obviously, it will be very hard to program all that intelligence into a single molecule. Instead, we pursued a much simpler algorithm. The robot performs a random walk. If it bumps into a cargo, It picks up the cargo and continues the random walk. And if it bumps into a destination that matches the type of the cargo, it drops off the cargo and continues the random walk while searching for other cargoes. Now, the implementation of this algorithm is very simple. The robot should be able to walk around in random directions, pick up a cargo, and drop off a cargo, and that's it. So now it is conceivable to build this much intelligence into a single molecule and more specifically, a single DNA molecule in our case. Now, the mechanism that we use to create motion is called DNA strand displacement. As you know, DNA strands can bind to each other if their sequences match, meaning A pairs with T and C pairs with G. If there is an invading strand with just the right sequence, it can bind to the partially bound strand,, and the two strands will then compete for the complementary strand, and the previously bound strand will be released. So the process is initiated by a very short single-stranded domain that is called the toehold. Now this animation shows the dynamic process. The invading strand has a toehold colored in red, on which— hanging off from the double helix. That starts a competition with the blue strand. And then base pairs are being opened and closed very quickly, moving back and forth until the blue strand is released and the invading strand now becomes part of the double helix. Then the techniques that we use to build a testing ground for DNA robots is called DNA origami. This is invented by Paul Rotemund at Caltech. So this technique folds a long single-stranded DNA into any desired shape, such as a smiley face, with the help of hundreds of short DNA strands. And because each short strand has a unique sequence and occupies a unique position in the structure, they can be extended to attach to other molecules and create any desired molecular pattern by simply selecting which locations are extended. And for example, by laying a track made of substrates for a deoxyribose, a robot with deoxyribose legs can be guided to make a turn while walking on the surface of this DNA origami structure. So, in our case, we wanted to program a DNA robot for more sophisticated tasks than not just walking. So we developed 3 modular building blocks that can be composed together to perform a task. And on these diagrams, squiggle lines indicate short domains of DNA strands that are just a few bases, and the straight lines indicate longer domains that are over a dozen bases, but still relatively short. Now, the first building block is for random walk. So when the robots, which is just a single strand of DNA, is bound to a DNA origami testing ground by one leg domain and one foot domain.. It won't fall off in a million years. But it has a single-stranded free foot that can bind to a complementary foot domain on a nearby track strand. And then the two track strands will compete for binding to the robot's leg in a strand displacement reaction. When the robot is only bound to the previous track just by the short foot domain, the the foot can spontaneously dissociate, allowing the robot to make one step from a track location to another. And this is a conceptual animation of how the robot makes a single step. It first forms a few base pairs with the adjacent track strand, followed by strand displacement, and finally dissociation of a few base pairs. So note that the geometry and the total number of base pairs are exactly —same before and after the robot takes a step, and therefore no energy is being consumed in this walking process. And then adding a hand and an arm domain to the robots allows for cargo pickup. So the cargo molecule can be a DNA strand itself or any other molecule that is conjugated to DNA. So in our case, we used a fluorescent molecule which allowed us to monitor where the cargo is. And the robot should be able to walk around just as before. And when it gets close to a cargo molecule, which is also attached to the surface of the DNA nanostructure, now the hand domain will bind to the complementary hand domain on the cargo strand, and strand displacement will occur within the arm domain until the cargo molecule is transferred to the robot. In this case, a few more base pairs formed within the hand domain. So this step is energetically driven forward. And then the robot will just continue to walk around while carrying our cargo. So finally, for the cargo to be dropped off, we added a short cargo domain on the cargo strand, like a delivery address of a package, and created a goal strand that has the complementary cargo domain together with a hand and an arm domain. So when the robot carrying a cargo gets close to the goal location, the goal strand will actively grab the cargo away from the robot. And this is similar to the process of cargo pickup. Okay. Now to sort multiple types of cargoes, we only need to give each cargo domain a unique DNA sequence, like a unique zip code for each type of package, so that cargo 1 can only be recognized by goal 1, and cargo 2 by goal 2. And with this design, let's find out if the robot can indeed perform cargo sorting. So here's what the testing ground looks like. We have 2 types of cargoes, which are fluorescent molecules with 2 different colors, and they are initially at unordered location near one edge of the DNA nanostructure. And then the 2 separate goal locations are near the other edge. And we place the robots in the center of this testing ground. So to monitor where the cargo is, we used a quencher molecule at each goal location. So when the cargo is successfully dropped off at a goal location, we expect to see a decrease of fluorescence signal. And it's like turning off a molecular light bulb. So from this data, you can see that cargo 1 was successfully dropped off at goal 1. And then combining 2 sets of experiments together, it was clear that the robot was able to sort 2 types of cargoes with multiple cargoes per tack. Then with microscopy images, we can look at individual molecules and convince ourselves that the cargo molecules were indeed delivered to their destinations, one type near the bottom left corner corner of the testing ground and the other type near the bottom right corner. The next, we placed multiple robots on the same testing ground to collectively perform a sorting task. And we expected these robots would just wander around the DNA nanostructure surface and occasionally they will bump into each other, but mostly they will be doing the same cargo pickup and drop-off as when they were alone. And in this data, you see that multiple robots can sort the cargoes more effectively and more quickly. And with a very simple model, we were able to semi-quantitatively reproduce the experimental data. And now you probably have noticed that the robot is very slow, and it takes about 5 minutes to make a single step. So it's natural to ask the question, can it be faster? Well, a study from Nils Waters' group at the University of Michigan looked at the same reaction mechanism, but with smaller feet of the robots and a reversed strand orientation so that rather than walking upright, the robot walks head over heels. And both of these changes promoted faster motion, we believe. And now from the experimental measurements, this robot walks at 1.4 seconds per step. So this is still about 500 times slower than protein motors. But it is already 200 times faster than the cargo sorting robot. So, there's a hope in terms of better understanding of the mechanism in order to build better robots. And another strategy is, of course, to take advantage of the fast protein motors and integrate with DNA or RNA components to improve their programmability. And examples of these motors include one from Andrew Tuggerfield's group at Oxford and another from Zev Bryant's group at Stanford. Okay, to summarize our work, I believe it is important that the same robot design can be used to sort more, uh, sort more types of cargoes and more cargoes per type. And this is because the correct drop-off is simply encoded in the sequence-based recognition between cargoes and their destination. And so after dropping off a cargo, the robot will simply start searching for another cargo. And there is also no constraint on the initial locations of the cargoes, so the robot performs a random walk to search the entire testing ground. And except for picking up and dropping off, there is no energy consumed for walking, so the robot could potentially work on a much larger testing ground. Okay, so we managed to build DNA robots for me. More sophisticated tasks than previously possible.

Just to interrupt, we've got a question that's probably relevant, so, at this point. So, Christopher Rowlands asks, what's the performance advantage of mediating the, the search for the, the targets via the robot? Why not just let the— why not just release the strands into solution and let them find their binding partners just by diffusing?

That's a great question. So, you can imagine that if we let the molecules find the destination by diffusion, because of the hybridization will still occur, or strand displacement will still occur, you do have this behavior of the molecules landing at the destination. However, if you were to do that, then the entire test tube will be performing exactly one task, and only capable of performing that. Exactly one task. But now if you have this nanostructure testing ground and you have this robot sorting the cargoes facilitated by random walking on this 2D surface, then you can imagine it's potentially possible. And we demonstrated some simple cases in our paper as well. You could have a large number of testing grounds in parallel. For example, each testing ground would have a different number of robots. Depending on the difficulty of the task, it will have a slightly different number of cargo types or goal types. It could have different, very different cargoes and goals on different testing grounds. So all these individual cargo sorting tasks could be different, yet the entire test tube could be composed of millions to even billions copies of different tasks taking place. Advantage of parallel between those different tasks that is not possible if you simply just have diffusion without any geometry at all.

Does that make sense? I think that's a good answer. Yeah, I guess Chris can— we could carry it on at the end if Chris isn't satisfied. And there's one other question which you might as well answer now because I think you're about to move on to a new topic, which is, have you looked at different origami shapes and/or locations of the drop-off points? And how about biasing them randomly?

Random walk with different foot affinities? That's definitely a nice thing to consider. So biasing them for certain tasks would definitely have an advantage if you already have certain information about how you want to bias the random walk process. But if you have no information at all about what the initial distribution of the cargo molecules would be, then maybe if biasing the random walk could actually make it less efficient because it might get into, let's say, some dead-end structure, and because it's biased forward, so it won't be able to efficiently back out of that dead end. And thus, in that process, it would even take longer for the robot to get the job done. So my answer is yes, it will be very interesting to consider a biased random walk, and depending on the task, you might want to consider such mechanisms very carefully.

Great, thanks.

Okay, excellent. Okay, so, um, right, so I was summarizing a few properties we believe are the desired properties of this particular system. And, um, so now I want to point out that, so even though that we were able to build this particular DNA robot that perform more sophisticated tasks than previously possible, but we do not yet have and knowledge about how to build general-purpose molecular robots that can perform a wide range of tasks by composing different building blocks that are readily available to everyone. And to get there, we need more simple algorithms and more modular building blocks to perform more collective behaviors until there are enough understanding for the development of programming languages that will work in practice. So Namita Saref, a graduate student in my lab, is now working on a maze-solving robot to develop another simple algorithm and a new building block that will hopefully get us one step closer towards the general-purpose molecular robots that we're hoping for. Okay, yes, moving on. Next, how do we embed computation within molecular systems? We know that from bacteria's capability to follow simple chemical gradients to the brain's capability to distinguish complex odor information, specific patterns of molecular signals is an essential type of computation in biological organisms. And inspired by the biological neural network in the brain, computational models have been developed to create artificial neural networks that perform sophisticated pattern recognition tasks. And one particularly desired property of neural network computation is that the same network could be reprogrammed to perform different pattern recognition tasks. So 10 years ago, working with Eric Winfrey and Shuki Brook, I made a DNA-based neural network that can recognize 4-bit patterns. One or two clues that are encoded in short DNA strands were sufficient for a test tube of DNA molecules to figure out which of the 4 remembered 4-bit patterns is most compatible with a clue. Obviously, that was not smart enough for many tasks that you might want to program a molecular machine to do. A question that motivated my lab's research is, how much intelligence can be programmed into artificial molecular machines? When Kevin Cherry joined my lab as a graduate student, He was determined to make neural networks much smarter, and now he has successfully scaled up DNA neural networks from recognizing 4-bit to 100-bit molecular patterns. So in this work, we implemented a very simple competitive neural network model called winner-take-all, which is inspired by the lateral inhibition and competition observed among biological neurons in the brain. So here's the function. If a— for a number of inputs that are real values, a binary output is 1 if and only if the corresponding input is the largest, which means the output indicates who the winner is. And then adding a layer that computes the weighted sum of a number of inputs before the winner-take-all layer allows pattern recognition. So here's an example. Let's consider a circuit that is designed to distinguish a letter L from T. And we can simply use the two target patterns as weights. If an input pattern looks like a letter L but one bit— And now the two weighted sums are calculated as 4 and 2. And because 4 is greater than 2, the winner-take-all layer should compute that y1 is 1 and y2 is 0, which means the input is recognized as L. In this example, the weights are binary, and the function is equivalent to computing the smallest Hamming distance between inputs and weights. But the model also allows analog weights, which are more powerful than just binary. For example, if the input pattern is heavily corrupted, it may result in a tie and thus cannot be classified. But if the weights have analog values that basically show which bits are more important than the others in each memory, then highly corrupted input patterns may still be successfully classified. And because of how simple the circuit function is, there exists a set of also very simple chemical reactions that can implement the winner-take-all computation. So I'm not going to explain all these reactions for the interest of time, but I will tell you about one reaction, the pairwise annihilation, that allows all molecular species, in our case DNA strands, to compete with each other until there is only one winner left. So here there's a mostly double-stranded molecule that we call an annihilator, and it can bind to a single-stranded DNA signal that's called Si, and because the DNA sequence of the orange domain in the signal molecule is designed to be identical to the orange domain in the annihilator, so they compete with each other for the complementary strand. Now, if Si is present, this process is reversible, so the signal strand will not be consumed. However, if signal Si and Sj are both present, then Sj can bind to the annihilator from the opposite side, and when both signal strand branch migrate to the middle, the annihilator will be split apart, and then both signal strands will be turned into waste products that cannot interact with other molecules. So this process ensures that Different signal molecules will continue to kill each other until there is only one winner left. Similarly, all other functions involved in a winner-take-all neural network can be implemented using a set of fairly simple DNA molecules. As you can see from the diagram here, they're just short single strands or short double-stranded molecules. And now once we mix all these molecules in a test tube, they can perform surprisingly complex tasks. And in our study, we chose a well-defined and understood task, which is handwritten digit recognition. But in real-world applications, these molecular handwriting could represent any information of interest, such as a gene expression profile, and the recognition results could represent any desired classification task, such as differentiating viral and bacterial infections. So the digital recognition task allows us to evaluate basically two things. One is the complexity, and the other is the diversity of molecular patterns that a DNA neural network is capable of recognizing. For example, the complexity will tell us how many unique microRNA or mRNA species can be simultaneously detected, and then the diversity will tell us how different an individual's gene expression profile can be tolerated compared to a common profile. So we went ahead and took 100 example 6s and 100 example 7s from the MNIST database, which is often used to test machine learning algorithms, and then we averaged them to determine the weights. Now, these weights are analog, and so in this diagram, it shows darker bits indicate that they are more commonly used in the examples. Now, each bit in the memory corresponds to just two DNA molecules that can perform weight multiplication function, and the concentration of these molecules correspond to the value of the weights. When the neural network is used to recognize corrupted test patterns, each bit in the inputs correspond to the presence or absence of a specific DNA strand. So in the database, there are a total of over 14,000 example handwritten digits 6 and 7. If we look at where the example patterns are in the weighted sum space, it basically tells us how easy or hard the winner-take-all competition should be because further away from the diagonal line is easier and closer to the diagonal line is harder. And about 2% of of these patterns are on the wrong side of the line, so it is impossible for our circuit to recognize these patterns. And 8% of the patterns are fairly close to the diagonal line, which we expect to be experimentally difficult. And 90% of the patterns are far enough from the diagonal line, which we expect our circuit to recognize correctly. So then we chose 36 examples from these patterns based on a uniform distribution and use those to perform the experiments. So the perfect patterns, which are the same as in memories but in binary, were correctly recognized. As you see here, the green fluorescent signal going high indicates a 6, and the yellow fluorescent signal going high indicates a 7. But more importantly, patterns was up to 30 bits of deviation from the memory, which means very sloppy handwriting, were also recognized. And then to scale up the number of memories, we took a grouping approach and divided 9 memories into 3 groups in 2 ways, shown as these 3 rows and 3 columns here. So each y output indicates which row the memory is in, and each z output indicates which column the memory is in. Then we determine the weights for each group using an average-then-subtract method. Take the average of 800 examples for each in-group digit and subtract the average of 100 examples for each out-of-group digit. You might be wondering, we can use much more sophisticated machine learning algorithms and figure out the weights by computing it on a laptop and to improve the performance of the neural network. That would do a lot better than this very simple algorithm. Called weight average and subtract, and that's absolutely true. But we intentionally chose a very simple one so that learning can later be implemented within the molecular system itself rather than being done on the laptop, which I'll touch on just in a bit. Okay, so now our experimental data showed that indeed all 9 digits can be simultaneously recognized for a total of 99 representative examples. And Kevin developed an online software tool. It converts a user-defined set of memories and test patterns into program code describing a DNA neural network, which can then be used to simulate the kinetics of the network behavior. And it also provides sequences of the DNA strands required to experimentally construct the DNA neural network. And this is what the web interface look like. And so we used this software tool in a class, and even undergrads and graduate students with minimal prior background were able to design and create their own DNA neural networks with very complex behavior in just a few hours. Okay, so now the ongoing work is more exciting to us because instead of us computing the values of weights and encoding them in the concentrations of molecules that we pipette into a test tube, we're now building DNA neural networks that learn from their own environments. And Kevin is exploring supervised learning, while Tianqi Song, a postdoc in my lab, is interested in unsupervised learning. And Kevin has built an adaptive memory where concentrations of weight molecules can be adjustable in response to training signals. Well, Tianqi has come up with an elegant solution that allows the weight molecules to be restored after each round of computation, which is critical for unsupervised learning. Okay, so I've showed examples of simple mechanics in DNA robots and computation in DNA circuits, but I haven't discussed how these DNA circuits could be used to control DNA robots? Well, imagine that the output of a DNA circuit, which is often a single strand, could be used to activate or inhibit the motion of the DNA robots. Now, this mechanism should be relatively straightforward to engineer in a test tube environment. But on the other hand, if we want to build a molecular machine for exploring a natural environment, and possibly traveling a long distance, then the free-floating circuit components may simply get lost. So, alternatively, DNA robots could be controlled by DNA circuits tethered to the surface of the same DNA nanostructure. For example, DNA origami can be used as a circuit breadboard, and each circuit component can be tethered to a specific location on the breadboard. And then the spatial arrangement of these components will simply determine the function that a circuit computes. Imagine an artificial cell whose membrane is made of DNA origami nanostructures, and on the inside surface of each origami, there are a DNA circuit, and they will be able to orchestrate the inner world of the artificial cell. So if we had such an artificial cell, how do we make it reconfigure itself for different tasks? For example, could the artificial cell equip itself with the right tools at the right time? Could the decision be made in response to a changing environment? Okay, at this point you might be thinking this is too much dreaming. We don't even have any molecular machine that looks like a submarine and is of size similar to a cell. Well, There have been recent advances that scaled up the complexity of DNA nanostructures of sizes similar to a small bacteria.

So a challenge is[No speech] Hi everyone, sorry, we seem to be having a technical issue at the moment. I think Luli might have dropped out of the call. Bear with me, I'm just going to see if I can find her and bring her back in. Apologies, we seem to have lost Lily. Bear with me, I will try and find her.

Oh, looks like I was muted.

Now can you hear me?

Yes, we can hear you now. Okay, excellent. I don't know what happened, seemed like I just got kicked out or my computer was doing something. Okay, now I started sharing screen again. Are you able to see my screen? Yes, we can see it. Okay, wonderful. Okay, we're back. Sorry about that. Um, okay, I'm not sure when I got cut out, so let's just start over from this one slide. I was talking about there was a challenge. So the challenge is to develop a mechanism that allows information-based autonomous reconfiguration in these DNA nanostructures. And ideally, we want this mechanism to be able to use to build to develop building blocks that are compatible with general-purpose reconfiguration tasks. So for example, we want to be able to include the kind of tasks that allow structural components at any desired location to be reconfigured at any desired order. So this is what Philip Peterson and Greg Tikhomirov did. They discovered this mechanism that we now call tile displacement. When they tried to understand a mysterious result on nanostructures self-assembled from DNA origami tiles. So I don't have the time to tell the discovery story today, which was a fun story, but I will briefly show you what tile displacement is and what it can do. So you already know that a strand displacement— what strand displacement is. An invading strand, which is colored in orange shown here, can bind to a double-stranded complex by the short TOCO domain. And then branch migration occurs when two strands are competing for binding to the same complementary strand until the previously bound strand is released and the remaining strand becomes unbound. So now this tile displacement follows the same abstract principle but occurs at a much larger scale between interacting DNA origami structures that we refer to as tiles. And different from binding by a string of nucleotides, the The DNA origami tiles bind to each other by a set of stacking bonds and very short sticky ends. So now you show that invading tile is colored in orange, and the DNA origami tile have a good degree of structural flexibility between adjacent helices and along the seam between the four triangles composing each tile, so that the invader tile can bind to the two-tile complex via a domain that has just a few staples, and then branch migration occurs between this perpendicular occurs perpendicular to the direction of the adjacent DNA base stacking when the two tiles are competing for binding to the complementary tile until the previously bound tile is released. And so this is just showing that two different viewpoints where you can see the same process where the tiles can just bend a little bit to allow displacement. Okay, so we measured the kinetics of tile displacement with varying number of staple strands and varying lengths of sticky ends in the toe-hold domain, and demonstrated that we can control the reaction rates over about 5 orders of magnitude. And then we show that a few example reconfiguration processes, including this one that is a 2x2 tile array, first reconfigured from a frown to a grimace and then to a smile. And then another case of frown that can only reconfigure itself if two types of invaders are both present. And we also showed that tile displacement can take place at any location of an array, including the interior location. And then these reactions at different locations can also take place in any desired order, as shown in this tic-tac-toe game. So there are practical issues that need to be resolved. For example, the yield of reconfiguration is these reconfigured structures, they decreased quite a bit with an increasing number of cascaded steps. But we hope that tile displacement has now opened up many possibilities for future developments in reconfigurable micromachines. And from a fundamental point of view, tile displacement has established displacement as a generalizable concept that not only occurs between small DNA strands, but also occurs between DNA nanostructures at a different scale and in a different dimension. And now the information encoded in edges of the 2D structure could potentially be extended to surfaces of 3D DNA nanostructures, and that will probably allow even more general forms of structure displacements. So as we now know, tile displacement is particularly good for complex structural reconfiguration, while strand displacement is particularly good for complex computation. So integrating these two mechanisms across scales will allow us to combine their unique advantages to create even more sophisticated behaviors in reconfigurable molecular machines. And to do that, Dominic Scalise is working on demonstrating how an input strand is released to a larger output circuit used to activate a tile. And at the same time, Sam Davidson is building strand displacement circuits on tile surfaces. When specific output signals are released, they can be used to initiate part of a circuit to be swapped out for updated circuit function in response to to a changing environment. Okay, to summarize, I hope that with effort, the combination of simple mechanics, embedded computation, and task-specific reconfiguration will allow us to build smart and efficient molecular machines that eventually approach the sophistication of life itself. Similar to how programming language have enabled everyone to write their own software, molecular programming languages will eventually allow everyone to write their own molecular apps and enable embedded control within chemistry, medicine, and materials. Okay, so the focus of my lab is to establish the foundations for future molecular technologies, and many of these things that we do may or may not turn out to be useful. But I feel very lucky to have a group of excellent students and postdocs who are excited to work with me on crazy ideas that may one day change the world. So I'll end with this lovely photo when we were still able to stand next to each other within 6 feet. And I want to say that the tallest person among us is starting as an assistant professor at Berkeley. So he has open positions for graduate students and postdocs. And please contact him if you're interested. He's a wonderful scientist and a wonderful person. I'm sure you will love working with him. Okay, that was my little bit of advertisement, but now I'm ready to take questions.

I'm muted.

Yeah, I'm unmuted.

Ah, now we can hear you. Yeah, yeah, yeah, sorry, so weird, right? Everyone's trying to press the unmute button to give a clap at the end of the talk and it's like this weird silence. Thanks, Uli, that was great. We've got a couple of questions that were brought up during the talk itself, which I'll go back to, and then I guess we'll throw the open. So the first one, well, was a bit far-reaching, but perhaps the, the most directly applicable part of the question is— sorry, my thing's scrolling quite badly because people are adding new questions. As I remember, it was something along the lines of Is the programming approach very hardware dependent in this, in DNA nanotechnology, and is that a problem?

It is absolutely hardware dependent. That's why I feel like it is important when we develop circuit architectures, we have this programming language notion in mind, so we develop the circuit architectures that can support support versatile programming languages. For example, when we choose molecular building blocks, I was emphasizing at the very beginning that we want to choose really simple molecules because they work more robustly in a test tube environment. However, we also want to simultaneously be thinking, can this building block enable us to build a variety of different functions that we are interested in? And if so, then this particular kind of building block block could potentially serve as a general-purpose building block for general-purpose molecular programming languages. So we can't basically think about the molecular programming language only after we already have decided what molecules we're going to use to build things. We have to think about it at the very beginning when we think about those two questions simultaneously. And there are a lot of challenges involved, of course.

Yeah, I think it is fair to say that there are lots of efforts, including from people like people at Microsoft, to abstract this process as far as possible. And in particular, DNA strand displacement— if you're building a basic DNA strand displacement network, you almost don't need to think about the hardware at all. You, you, you, you take it as a given, and you're designing at a much higher level. But if you're trying to engineer a new, a completely new molecular functionality, you know, like, for example, your tile displacement, that, then you need to think about what's actually going on chemically again, I guess. Yes.

Yeah. Well, even for the first case of scenario, when you can compile arbitrary chemical reactions down to DNA implementations using strand displacement, combining all the efforts that the community has been making, wonderful efforts, but we still kind of have to think about the low-level mechanism, even if theoretically we know how to convert an abstract function to DNA strand displacement system. But when you design a system and following that particular general principle and mix everything in the test tube, like what Nianjun did in this wonderful example of oscillator, strand displacement oscillator, there were still years of effort put into debugging the system at the sequence level. And so that the oscillator would work robustly. So we're not, we can't claim that we actually already have a high-level programming language that will do the job we want to build a variety of things. But I think we're learning in this process of combining theory and experiments. We're trying to learn what is an easier approach that will get us there faster.

Yeah, I mean, oh yeah, I totally agree. Otherwise, I probably wouldn't have a job, right? But yeah, certainly, certainly if you want to, if you want to program A plus B goes to C, that's relatively straightforward. If you want to have precisely the kinetics needed to achieve some complex functionality, then, then that's definitely a separate question. Okay, so there was a question from Regina Agüeta Carreño— sorry for butchering your name— which is, is the random walk the most efficient method of sorting, or is there some other type of walk that can be studied, although it may have certain disadvantages in other aspects? So referring to the first part of your talk, So she's thinking, she's thinking in terms of things like chemotaxis. Um, you know, chemotaxis goes down a gradient rather than, than, than just following a random walk.

So yeah, yeah, that's an excellent question. So one of the main reasons that we chose random walk rather than more sophisticated ways of thinking about a more efficient algorithm is the, um, is the focus on energy efficiency. Because currently, which I didn't go into the detail, but many of the strand displacement systems that were able to build consume energy in a way that they would only be able to show a temporary kinetics before the system go to equilibrium. And this is because they are powered by a number of DNA strands, in the test tube, and once they make more base pair, they drive the reaction forward. But then when those molecules are used up, the system can no longer perform interesting functions. So, there are separate efforts on trying to build chemical batteries that can be used to make the system run much longer. But before that is successful, when we try to think about an algorithm for cargo sorting, where we don't really have an energy supply that's available, that we simply can't do much if the molecule required for take each individual step actually requires energy to do that. So that is why we chose the random walk mechanism, which does not require energy and which means it's not time efficient. But then we're hoping that once we start from there to try to explore what is possible with random walk as a very basic mechanism, and then we start bringing in other thoughts about how to build chemical energies to make the system have more sustained dynamics, then that will be a very good point for us to start thinking about more efficient algorithms, more time-efficient algorithms.

Yeah, I guess, I guess, I guess, I guess I'm echoing, I'm echoing. Yes, we can hear you. I was just saying that If you want to walk in a systematic direction, you need to use fuel, which means you need to get the fuel from somewhere and maintain it. But you also need to design the coupling mechanism whereby the fuel only gets spent if it goes through your engine, which is also like a whole world of complexity to add into the challenge of selectively picking up things, putting it down, I guess, as well.

Yeah, but that is something that the community is actively working on towards future molecular machines. So there's hope. It's just at the time we demonstrated the cargo sorting robots, we're not quite there yet to have those mechanisms available that allow us to do that.

Okay, so I, I think that's a good answer to the question. Regina, if you're not happy with that, please do chip in. So Philippa Samela asked, can artificial DNA-bound cell activation be modeled to replicate replicate T cell and B cell activation?

Okay, that is something that I don't think I have a good answer for, because one of the things that we do is when we try to program these artificial molecular machines, we are— except I'm speaking for my lab, particularly in this case— we're not really specifically thinking about any biological applications. So which means the system we develop are not directly integratable with all the biological systems that one might care about. But there are a large body of work in the field that have demonstrated the compatibility of artificial DNA molecular machines being integrated into a cellular environment and interact with the biological system. It's just that there's a lot and a a variety of very different challenges that needs to be overcome before that could be possible. So I think if the question is to say, can we directly use our machine to program those kind of biological systems that we're interested in? And the answer is, it's not immediately possible, but there are ways that could allow us to get there eventually.

I mean, I think that's one of the key long-term goals is to have like clever, clever, clever doctors in your that you can inject into your blood, right? Or something like that. So John Gertz asked, if I understand correctly, the neural network systems you've developed all have linear activation functions. Have you thought about ways to implement nonlinear neural networks?

Yes, we have. We explored a few different activation functions in the lab during some kind of summer projects, but we didn't get very far there, mainly because There are definitely possibilities to tune the activation function, but then based on our preliminary study, we did not actually see substantial advantages there because the linear function— essentially, when we say it's a linear function, it's in the ideal case a linear function. But when you actually look at the simulation of the system we're looking at, for example, in the previous network we demonstrate, it's a perceptual model. So we say that we have this comparing the weighted sum to a threshold, but nonetheless it's not really a hard threshold because at the molecular implementation level it was a competitive mechanism where there's a faster reaction pathway and there's a slow reaction pathway in order to create this thresholding. So in that sense, the thresholding itself is already a softer, more or less a softer thresholding mechanism, and you can potentially tune the slope of that function by changing the reaction rates of those two competitive pathways. So there are ways you can do that, and I think it's definitely interesting to explore all those possibilities, but we haven't got very far with that.

Great, OK, so the next one is a bit more bread and butter from Wei Shi Sim, which is what temperatures do you run these systems at? Oh, they are all at room temperature. So I guess, I guess some of the origamis, you presume you fold the origamis with a temperature ramp, but then you run the, run the dynamic experiments at room temperature.

Yes, so when you prepare the structure, you need to put them in a thermal cycler, heat it up and cool it down so the structure forms nicely without much tangles. Cool, but then once you have the structure ready, all the dynamic behavior were programmed at room temperature. All the interactions took place between the structures happened at room temperature.

So there's a couple more technical questions, but while we're doing that, could you possibly put up the slide with the contact details for Gregory's position? Yes. And there we go. Is that visible to people?

Can people see that?

Yes, we can see it. Great. I think my computer's really slow for some reason. Um, uh, okay, so last couple of talks, there's a question: how is the breadboard built, um, and how are the components inserted? I guess, which I guess the question is, how does DNA origami work? So I guess, yeah, we just would refer that person to— I will send a link to that person on DNA origami, because that's a whole extra topic that wasn't really a topic of this talk. And then Jurek asks, hi Lulu, you mentioned the SD system could be potentially used to monitor and combine gene expression levels. Can you tell us more about the potential biological applications and what are the limitations?

Yes, so that work was actually done by Georg Selig's group at UW Seattle. So in that experimental system, they took a simple—

We've lost her again.

Hello? Sorry, we seem to have lost her again. I will just see if I can find her, ring her back into the call.

Yeah, just about, just about. The technology is Just about. Whilst we're doing that, I'm just posting a link into the chat about how there's a review on DNA origami from 2017 in Cell, and that's a discussion of how these sort of tech— how to make these breadboards essentially, and do other things with them. Not everyone wants to just use them as a breadboard, so the person who asked about that, hopefully that will answer their question. They are really cool. So yeah, introduced by Rothermund about 14 years ago, and ever since then people have loved making funny things out of them.

And Lulu's back. I am back. Sorry, I have no idea what just happened again, but I'm back. Okay, I wanted to go back to the, uh, the question. So this, I think, was the slide that I was trying to refer to. So this work was actually done by Garek Selig's group from UW Seattle. And in this system, they took a linear classifier built from DNA, which follows a fairly simple— a fairly similar principle compared to the basic building blocks of DNA neural network. Networks that we showed. So then, but they look at this kind of circuit and then use the gene expression profile as the input of the circuit. In that case, they look at mRNA signals and see if those could be used as recognized as part of the input signal to the circuit. And that's why you can look at a disease profile when you look at multiple of those mRNA expressions. So that was a very beautiful work. So if they— I forgot who asked the question, but you could potentially look up this paper and then you will see more details in that particular publication.

Great. I think that actually gets to the end of the questions that we've got in the chat. If anyone feels that they've been skipped or harshly done by or not properly answered, do— well, it's time to get to the point where you could probably probably actually unmute and ask yourself, or type in the chat. Lots of people saying great talk, thank you very much, so again I'll reiterate that. One final thing—

oh, go on.

I was just going to say, yes, thank you, excellent talk, very interesting and inspiring, I enjoyed it, and thank you for the link as well. OK. Great, so I had one question. Was the really the DNA, the tile displacement, is that published or is that currently?

Yes, it is published. Let me go pull that reference out.

It looks just— you went through it quite quickly, but it looked like the rate, the limiting rate was about the same as the strand displacement limit.

Was that correct? It's about 10 times slower, actually.

Oh, about 10 times slower. But okay, so not that much slower.

Okay.

Yeah, this is the reference. Okay, okay. I somehow missed— I somehow managed to miss that. Okay, my bad. All right. Yeah, okay. So that's probably everything from me. Again, thank you to everyone for joining us and asking very good questions. Lulu, you need to have breakfast, I guess, or something. I'll let you, I'll let you get away, and hopefully see you soon.