Scientist Stories: David Liu, Founder of Prime Medicine & Pioneer in Gene Editing & Chemical Biology

All right, thank you. It's a great pleasure. It's actually a special pleasure to give a talk to a general public audience. And it's fair to say that I actually spent more time trying to get the slides completely free of lingo, which hopefully will make it clear for scientists and non-scientists alike, while preparing for this lecture. And it's a great pleasure to be a new member of the core faculty here at the Broad Institute. And for those of you who intersect a lot with the Broad, I'm looking forward to increased collision frequency with all of you. Okay, so— Our lab is really in awe of evolution. Evolution is an amazing problem solver. It is responsible, of course, not just for the rise of life on Earth, but also the evolution of higher organisms, including man. And the power of evolution has created a number of incredible problems for mankind as well. For example, the rise of antibiotic-resistant bacteria, is a worldwide challenge that's a direct consequence of evolution. And indeed, even cancer is the consequence of the evolution of cells in a person to replicate in the face of a wide variety of signals that would normally stop a cell from replicating. So the power of evolution is responsible for this incredible complexity, yet if you Boil down evolution to its molecular essence, it's conceptually very simple. Evolution begins with a population of information carriers. This is typically a population of genes, of DNA. These genes are translated using biological machinery into corresponding proteins. And the proteins differ in their structure because the genes differ in their sequence. As a result of differences in their structure, some of these proteins have a better or worse ability to cause the cell that hosts this information and this protein to survive. So following a Darwinian selection, a subset of this population is more fit and survives preferentially. And that survival enables the information carriers, enables DNA to do what DNA does best, which is replicate. And of course that replication is imperfect. Every now and then mistakes are made, and that results in mutations. But the point is that the second generation of genes that survives one generation of evolution are also a population of genes, but they're all related to the genes that encode the molecules that caused the survival of this organism. OK, so that's really how evolution works. And so then this population is translated into a second generation of related proteins, et cetera. Now this cycle of translation, selection, replication, and mutation has been harnessed by researchers for the last several decades. It's been used in laboratories to create a wide variety of proteins, nucleic acids, DNA and RNA with different properties, in some cases very useful properties. But one of the real limitations of laboratory evolution is that arranging this cycle to take place is labor intensive. It requires a lot of babysitting by the researcher. And so as a general rule of thumb, I like to tell my students, if you're evolving a protein in a laboratory, to do one evolution generation, to do one complete turn of this cycle, takes about one week. Sometimes quite a bit longer, but one week is a good ballpark. The consequence of the time and labor intensive nature of evolution ends up imposing a variety of constraints on what one can evolve in the laboratory. And to perhaps best illustrate this, we'll turn to a construct that evolutionary biologists love to use called a fitness landscape. So you can think of evolution as traversing, climbing a fitness landscape. A fitness landscape is this hypothetical landscape in which elevation represents fitness. So if you're evolving a protein to have, let's say, faster activity, then the faster the protein's activity, the higher you are in the elevation of its fitness landscape. The x and y-axes, the space the location of where you are on this fitness landscape, irrespective of elevation, is the sequence of the protein. So as you mutate a protein, you end up populating different regions of this fitness landscape. And some of those regions have higher or lower fitnesses. And then when you do a Darwinian selection, you cruelly decide that everything below below a certain fitness threshold goes extinct. So that leaves us with these variants. And then in the second generation, you diversify them through mutations to create a new population of genes, which get turned into a corresponding population of proteins. Again, there's some kind of selection. You do a third round. And again, there's some kind of selection. And at the end of 3 generations of laboratory evolution in this hypothetical fitness landscape, you're left with a variant that has substantially higher activity than the starting point. So you've increased the activity, the fitness of that protein. But if you look more broadly at the fitness landscape, you'll realize that this is far from an optimal solution. And that's most frequently what we observe in laboratory evolution. In contrast, natural ecosystems evolve continuously over many generations with very large population sizes. And so exploring a fitness landscape this way might look something like this. Of course, it takes an incredibly long period of time, and if the generation cycle of the organism that's evolving is many years, as it is with humans, then this can take hundreds of thousands or millions of years. We wondered whether there was a way that we could evolve molecules in the laboratory in a way that would explore a fitness landscape more efficiently, like this, rather than using classical evolution techniques, as I demonstrated earlier. So we turned to a piece of biology that's called the filamentous bacteriophage. So a bacteriophage is a virus that infects bacteria. This virus, like most viruses, basically consists of a small number of proteins that form a capsule, and a piece of genetic information. In the case of this filamentous bacteriophage, it's DNA that sits within the capsule. And the DNA encodes all of the proteins that the phage needs to form progeny phage, to form babies. When you incubate this filamentous bacteriophage with a host cell, and this bacteriophage happens to infect common E. coli cells, The phage docks onto the E. coli cell, injects its genome, and then this genome does what viruses do very well. It hijacks the cell's biosynthetic machinery to produce the proteins that end up making new progeny phage.

Okay?

One of the amazing features of the filamentous bacteriophage life cycle is its speed. So this process can occur about every 10 minutes. So from the point of infection to the point that new progeny phage are made can be as fast as 10 minutes. So a very creative and ambitious, brilliant student, former student in my lab, Kevin Esvelt, who's now a professor at MIT, had an idea that perhaps we could map the laboratory evolution cycle onto the filamentous bacteriophage life cycle. And in the process, we would transform these steps shown in black that require a lot of researcher intervention into steps shown in green that the bacteriophage does for us. So instead of having to transform or introduce genes into cells, we'll rely on infection to do the job for us. Instead of relying on on researchers to take genes out and amplify them, will rely on phage replication, phage propagation to amplify the winners. So this simple insight boiled the problem down of mapping laboratory evolution onto a much faster time scale down to this question: how can we do this selection? How can we make a Darwinian survival of the fittest selection in which the ability of a bacteriophage to propagate a gene depends on the activity of the gene of interest. So it turns out that virtually all genes in a viral genome are important for that virus's replication. A virus has about the most svelte kind of genome of any life form, or most people would consider a virus to be a pseudo-life form. And one of those genes is called gene 3. Gene 3 in the viral genome encodes a protein called p3, and the job of p3 includes, among other things, allowing the bacteriophage to dock onto the outside of a host E. coli cell. So if the bacteriophage doesn't have p3, it can't dock onto the host E. coli cell. And that means it will be unable to infect any more E. coli, and its sort of lineage is at an end. So the cell has to make PIII for the bacteriophage to end up with PIII, which is necessary for them to produce progeny phage that can go on and infect new host cells. So no gene 3, no infection, in subsequent generations. So we used that simple fact, that necessity of gene 3 in the bacteriophage life cycle, to produce a system that enables proteins to evolve continuously in the laboratory for the first time without researcher intervention. We call the system PACE, or Phage-Assisted Continuous Evolution. And the evolution itself takes place in a fixed volume vessel typically a small flask that we call the lagoon because that's where the evolution is happening. That lagoon is constantly diluted by the host, by a culture, by a liquid culture that contains host cells. So host cells flow into and out of this lagoon continuously. Now these host cells are E. coli cells just like I've shown you, but they've been engineered in a couple important ways. One is that they have a piece of DNA that we call the mutagenesis plasmid. It's a circle of DNA that allows us to control the mutation rate of the bacteria, and therefore to control the mutation rate of the phage that come out of the bacteria. And second, it has this important, this essential phage gene, gene 3, taken from the phage genome and put onto a piece of DNA that the host cell provides. That piece of DNA we call the accessory plasmid. And the accessory plasmid is carefully designed so that gene 3 is not expressed always, meaning it doesn't get turned into gene 3 protein always, and instead requires the gene that you're trying to evolve to be active. Okay, so to kick off a PACE experiment, you seed this lagoon with some selection phage, what we call SP. And these selection phage are just like the phage I showed you in the previous slides, except we've stolen gene 3 from their genomes and we've replaced it with the gene that we want to evolve. That can be a gene encoding an antibody or a protease or a different enzyme or a DNA binding protein, whatever you want to evolve. You've replaced gene 3 with that gene. Now the job of the accessory plasmid is to link the activity of the gene that you are trying to evolve to the expression of gene 3. So after these phage infect the E. coli cell for the first time, and they get a free lunch once because we've grown these special seed phage under conditions that provide them with unlimited P3. So they're maximally infectious, but that only lasts once. After that, they're on their own. They inject the SPE, their genomes, into the host cell. And then if the gene you're trying to evolve is active, then by definition it turns on the expression of gene 3 in the accessory plasmid. And that causes the host cell to produce P3 proteins. As a result, the phage that come out of this cell can go on and infect new cells, as we'll see. If, on the other hand, the gene variant that was encoded by this phage is not functional, then by definition it's unable to turn on the expression of gene 3 on the accessory plasmid. And the phages that come out of that cell are uninfected. So if you turn on gene 3 expression, you make P3, and the baby phage that come out of this cell can go on and happily infect new cells. If on the other hand, the gene that you're trying to evolve is unfit, then it doesn't turn on gene 3 expression. The phage that come out of this cell are uninfectious because they lack that key P3 protein that allows them to dock onto new host E. coli cells. And eventually they just get washed out to waste. So because we can regulate this flow rate, we can make sure that the average time that an E. coli cell spends in this lagoon swimming around before it gets sent out is too short for the E. coli cell to replicate. Because under these conditions, the E. coli requires an hour or two to replicate. But the residence time is long enough for the phage to replicate, because again, they replicate very quickly, every 10 minutes or so. So this has created an interesting situation where bacteriophage are constantly replicating under conditions that that require that they survive some kind of Darwinian selection, that allow them to replicate, and that give us control over the mutation rate of their DNA. So they have everything needed to complete that evolutionary cycle. They can replicate, they can mutate, and they're under some Darwinian selection. Now, if you drop some phage into a lagoon and run this experiment and see speed up more and more, faster and faster, how quickly the host cell culture is flowing into and out of this lagoon, you can actually come to the conclusion that bacteriophage can replicate fast enough that the average phage surviving 24 hours of continuous dilution can evolve through 50 or more generations in one day without the researcher having to do anything. One of my students will set up one of these experiments, walk away from it, come back 24 hours later, and the phage, the genes in the phage will have undergone up to, say, 50 or 70 rounds of evolution, generations of evolution. And that really is about 100 times faster than we could evolve proteins in the past. Now the key to this is that the only piece of DNA, and indeed really the only molecule that can persist in this entire slide is the DNA encoding the selection phage that has the gene you're trying to evolve. If you think about it, everything else gets replaced faster than it can replicate. So only the phage DNA, the SP, is able to survive and accumulate mutations. And this is important because this system provided one of the first ways that in a practical time frame you could perform 1,000 rounds of evolution. And normally, if you perform 1,000 rounds of evolution, aside from the fact that it would take you 1,000 weeks, you would end up with highly mutated genomes of your cells and of different pieces of DNA that were able to somehow cheat your selection, that were able to bypass the selection that you've set up and end up producing solutions that are uninteresting. So replacing everything on this slide except the viral DNA ensures that it's much more difficult for the system to cheat the selection. So what does it look like? This is a photograph taken not too long ago from one of the warm rooms in my laboratory. So this room is a balmy 37 degrees centigrade, your body temperature. And it has a couple big host cell cultures. So these are E. coli cells that are just happily growing away. They get pumped into these lagoons, and these needles constantly pump liquid out of the lagoon once it reaches a certain volume. So the liquid never goes past— the volume never goes past a certain amount and goes to waste. And so the waste is just a big jug, which at the end of the day, we autoclave to sterilize and get rid of. OK, so what can you evolve with PACE? There's an obvious constraint. That you can only evolve what you can link to the expression of gene 3. But that has turned out, somewhat to my surprise, to be not as much of a constraint as we thought. For example, the first class of proteins we evolved continuously were RNA polymerases. These are enzymes that turn on gene expression. They turn DNA into RNA, which then becomes protein. So it's very easy to imagine how you can link an RNA polymerase gene an RNA polymerase activity to the expression of gene 3. Just put the key DNA sequence that signals for the RNA polymerase to start transcribing DNA, called the promoter, upstream of gene 3. And that works, as you'll see. You can also evolve DNA-binding proteins. Some of you may be familiar with zinc fingers or tail proteins or CRISPR. These are proteins that bind DNA in a programmable manner. And as a result, they have tremendous interest for use in modifying our genomes in ways that could end up providing treatments for genetic diseases. You can set up a situation where binding to DNA causes the recruitment of an RNA polymerase that turns on the expression of gene 3. Proteases are very important, useful enzymes that cut proteins. How can you link a protein cutting activity to the expression of gene 3 so that it can be paste? Well, in this case, we took an RNA polymerase and made it inactive by tethering it to a natural protein that shuts down its activity. And if the protease is able to cut the sequence that we put in the tether, then that causes the RNA polymerase to become active again. And its activity causes gene 3 to be expressed and causes the phage to propagate. And it turns out that there's lots of other ways that we've been able to link protein activities, a wide range of protein activities to the expression of gene 3, and therefore to PACE. So it turns out that the answer to this question is you can evolve a lot of different enzymes, even ones that are as sort of out there as enzymes that produce bioplastics, using this system simply by linking that activity to the expression of gene 3. Okay, so we've used PACE to do a number of things. And I'll tell you just two stories. One story is really a basic science story where we used PACE to answer a question experimentally in protein evolution that would normally be quite difficult or time-consuming to answer using conventional laboratory evolution techniques. And then second, I'll tell you about our use of PACE to solve an important problem facing modern agriculture. OK, so the first application is really inspired by a question that Stephen Jay Gould, the late, great evolutionary biologist at Harvard, famously asked. If you wind back the tape of life, that is, if you imagine replaying the long evolutionary trajectories that gave rise to life on Earth, how different would the outcomes be? If you replay evolution, many times, do you end up with many different outcomes, or do we all end up more or less as we are now? It's a fascinating question. And as you might imagine, this question has inspired a lot of debate. So Stephen Jay Gould argued that replaying the tape of life would tend to lead evolution down pathways that are very different from the road that was taken historically. Other evolutionary biologists like Simon Morris argued the opposite. In his book, really the title says it all, Life's Solution: Inevitable Humans in a Lonely Universe, which I always found a little bit depressing. So PACE allows us to perform hundreds or thousands of rounds of evolution. So we can experimentally test this question, at least for a particular experimental system. So we decided that we would answer two related questions inspired by this debate. The first is, how path dependent are evolutionary outcomes? So frequently in evolution, the following situation arises: you have a population of organisms and genes, they split into two populations. Those two populations end up under somewhat different selection pressures, and then those two populations end up facing the same selection, same future selection pressure. In this case, it's what kind of seeds a bird's beak has evolved to be able to handle. And the question is, even though these populations, these two populations end up converging on the same, faced with the same selection pressure, how do their evolutionary histories permanently imprint themselves on the outcomes of how they deal with a common selection pressure. Now one answer could be that they evolve the exact same answers, and it doesn't matter what your history is. Or it could be that making an evolutionary pit stop in Island B forever locks you into a certain set of solutions for Island D, and no Island C birds will ever evolve those solutions. It's an interesting question. How can we study these questions of how reproducible is evolution and how pathway dependent is evolution, taking advantage of our newfound ability to perform hundreds or thousands of generations of evolution? We took a protein that's very easy to evolve with PACE, RNA polymerase. The protein that we started with is called T7 RNA polymerase. It's named T7 RNA polymerase because it recognizes a piece of DNA called the T7 promoter, which has this sequence shown in black. It binds to that DNA sequence, and it initiates transcription. So it turns DNA into RNA, which can eventually become a protein. Now, T7 RNA polymerase recognizes the T7 promoter, but it doesn't recognize the T3 promoter. It doesn't bind the T3 promoter or initiate transcription from this promoter. And it doesn't bind or initiate transcription from another different promoter called the SP6 promoter. And so throughout this part of the talk, the T3 promoter is in red, the sp6 promoter is in blue. Now, there's something really convenient about the substrates for this enzyme. Because the substrates include a long piece of DNA, we can tweak those substrates to be as similar or dissimilar to the starting substrate as we want. So that provided a nice way of testing evolutionary reproducibility and evolutionary path dependence. So to test evolutionary reproducibility, we challenged an evolving population of T7 RNA polymerase genes to recognize this promoter, and then the T3 promoter, and then this promoter, and then a highly mutated promoter that we call the final promoter, shown in green, in which more than half of the bases of this promoter have been changed from the bases that T7 RNA polymerase normally recognizes. We did this evolution over about 200 generations. It took about a week per path. And we set it up in fourfold replicate. So we had four lagoons. The lagoons never mixed. They were set up, fed by the same host cell culture, set up to be as identical as humanly possible, evolved in the same room side by side at the same time. Okay, so the question is, well, how do the answers differ at the end of this process? To study path dependence, we decided to do a second evolutionary experiment, also run at the same time, in which we evolve this T7 RNA polymerase enzyme through a different pathway to recognize this SP6 promoter before challenging it to recognize the same final promoter as the first pathway. OK, so we have 4 lagoons running through the left pathway, 4 lagoons running through the right pathway. None of the lagoons mix. And the question is, how different are the outcomes? OK, so if they're all able to reach this final goal of recognizing this very different green promoter, how different are the solutions? How pathway dependent are the solutions? And how different are the sibling lagoons, are enzymes that evolved in sibling lagoons that were treated identically? OK, so here are some answers. I'll first show you some answers from the midway, halfway point of the experiment. At this point, we've diverged the populations. One has been evolving on the T3 promoter. One has been evolving on the SP6 promoter. So you'd expect the outcomes to be quite different. And indeed, they are. So here's 4 sibling lagoons that went through the red pathway. Each x represents the activity level. Of an individual clone that we pulled out of that lagoon at that time point. There's a lot of Xs, so we just average them to these red lines. So what this plot says is that all 4 lagoons were able to evolve pretty significant activity on the T3 promoter. So they were able to solve this problem of recognizing this very different DNA sequence. Plotted on this scale, 100 is the activity of the starting natural T7 RNA polymerase on its natural promoter. So 100 is about as as good as nature chose to evolve the T7 RNA polymerase. And you'll see that if you assay, if you combine all these clones and assay them against the other promoters, that just as you would expect, these enzymes have lots of activity on the red promoter, but they have no activity on the blue promoter or on the green promoter. They're below the yellow background lines. Likewise, the blue pathway has evolved activity on the blue promoter, So it's evolved really good wild-type-like, that is natural-like levels of activity on the blue promoter, and little or no activity on the red or the green promoters, just as you would predict because these two pathways have been kept separate. OK, so after 100 generations, they evolve different activities and really with no sort of cross-activity evolution. Now what if you look at the end of the experiment? So after 200-some generations, what does it look like? So both pathways, both the red lagoons and the blue lagoons, end up evolving activity on that final green promoter. They differ though. The pathways that— the lagoons that went through the blue pathway ended up with quite a bit more activity than the ones that went through the red pathway, even though they were challenged for roughly 100 generations on the same final promoter. Okay, so that was interesting. There's a pathway difference in their overall activity. And then if you look closely at the activity of the different lagoons, you find that some sibling lagoons have quite different average activity. So for example, SB6 pathway lagoon 3 and lagoon 4 evolved quite a bit more activity, in this case an order of magnitude more activity than sibling lagoon 2, even though they were treated identically. So we became very interested. What's the molecular basis of this path difference? How did an evolutionary pit stop in the history of these populations permanently affect their ability to solve this problem? And why is it that sibling lagoons that were treated identically, how could they have evolved different outcomes of replaying this tape of life? So we looked at their DNA sequence. And the first thing that struck us— so this is just Lagoon 1 from the T3 pathway, Lagoon 1 from the sp6 pathway. The first thing that jumps out is that the two different pathways evolved quite different mutations, even though the final task is the same, recognizing this green promoter. So shown in this protein structure, color-coded are the amino acids that mutated from the blue pathway versus the ones from the red pathway, and in purple are mutations that evolved in common. So there are some mutations that are in common, but most of the mutations that evolved were unique to either the red pathway or the blue pathway. Now if you look at all 4 lagoons, you can see that the red pathway lagoons tended to be somewhat similar to each other. The blue pathway lagoons tended to evolve somewhat similar mutations. But there are some notable, stunning exceptions. For example, this mutation uniquely evolved in Lagoon 2 of the T3 pathway, and not in really any other lagoon. This mutation uniquely evolved in Lagoon 3 of the blue pathway and was really not represented significantly in any other lagoon. So how is it that sibling populations created in sort of parallel universes at the same time came to different answers? Okay, so I'll answer the first question first. And there's many examples. You can read this paper. It was published a few years ago if you're curious and want to dig into the data. If you're very bioinformatically oriented, I've sent the gigabytes of data that came out of these experiments to quite a few other researchers, and I'm happy to send it to you, in case you have nothing else better to do, like watch a presidential debate or something. But I'll answer the first question. So what's the molecular difference between two pathways? So here's two mutations that were different in the two pathways. And I'll tell you right now that both of them help activity on the final promoter. So both of them help the enzymes recognize the final promoter. So the difference between the purple mutation and the yellow mutation is that the red lagoons seem to evolve either one of these mutations. The blue pathway lagoons really only evolved the yellow mutation. And they seemed to not evolve much at all of the purple mutation. Why is that? So there's something about the different histories of these two pathways that caused them to make this choice, even though both of these mutations help with the final answer. So we dissected what these mutations do by taking them in and out of different DNA clones that evolved at different stages in this experiment. And what we found is that the purple mutation helps with recognizing the red promoter activity. Okay, so it helps with the intermediate task along the red pathway. And as I told you, it helps with recognizing the green promoter. But it's really bad for recognizing the blue promoter. So that means the blue pathway populations had to avoid this purple mutation for the first 96+ hours, that is 100 generations or so, of their evolution. What about the yellow mutation? The yellow mutation, like I said, also helps recognition of the final promoter. But it turns out that if you ever try to combine the yellow mutation and the purple mutation together in the same enzyme, you end up with a dead enzyme. In other words, the presence of one of these mutations severely impairs the ability of the enzyme to accept the other mutation. This is called negative epistasis. And this makes sense because if you look at where these amino acids are in the protein structure, they're right next to each other. So it would make sense that if you mutated one of them, you might affect the ability of the other mutation to be beneficial. And in fact, in this case, you cause the presence of both mutations to result in an enzyme that has virtually no activity, shown here. So this answers the question. The reason that the blue lagoons had to choose the yellow mutation is because their histories necessarily had to avoid the purple one. And then this epistasis once their purple mutation dominated their population, epistasis prevented that purple mutation from emerging in a background that was dominated by the yellow mutation. So that's an example of how an evolutionary history forces two populations to make a different choice that permanently affects their ability to solve a future problem. All right, so what about irreproducibility? So an interesting An example of this is this mutation highlighted in orange here. So this mutation is remarkable. It's really good for solving the final problem. In fact, it is the reason why that third lagoon of the sp6 pathway evolved better activity than any other lagoon. So why is it that it only emerged in that lagoon and not in any of the other lagoons? So we introduced that mutation into lots of other genetic contexts that we found dominating the history of all 7 other lagoons. And what we found is that this mutation is very helpful in the one genetic context in which it evolved, in SB6 Lagoon 3. But in all the other genetic backgrounds of what we tested, it does at best nothing, and at worst it's very bad for activity. So how can that be? Well, it turns out that this mutation occurs at this amino acid shown in blue. Which is in a hotspot of 4 amino acids that all affect the activity of this enzyme. So it just happened by random chance that only in SP6 Lagoon 3 did a constellation of mutations emerge at these 3 positions that supported this mutation, this orange mutation of phenylalanine becoming leucine, of that becoming very beneficial. And in all the other contexts, it's a harmful— It's a harmful outcome. So what this tells us is, in this case, evolution was irreproducible for this mutation because random chance created a situation in which only a rare combination of mutations allowed this mutation to be beneficial. And only rarely did a population happen to evolve that background, that constellation, so that it was receptive to this mutation. Resulting in an increase in activity. This is called sine epistasis. It's another form of epistasis. So the answer to these two questions are actually quite related, that whether by random chance or by a forced divergence of the evolutionary pathways of populations, these differences through epistasis permanently affect the ability of these populations to accept future mutations. Okay, so what about using PACE to actually solve a problem? So I thought we'd go from Stephen Jay Gould to Janet Jackson. What have we done with PACE lately? And the most recent project that we've completed and published with PACE is a really interesting problem that has to do with insect control. So it turns out that in many countries, including this country, a majority of important crops, cotton, corn, soybeans, are grown using plants that have been engineered to express a protein that kills insects. This protein is called BT toxin. It's named BT toxin because it comes from Bacillus thuringiensis. BT toxin is not to be confused with Botox. So Botox kills people, BT toxin kills insects. In fact, the mechanism by which Bt toxin kills insects has been exploited because of the fact that it's very difficult to do harm to a non-insect creature. It requires activation in an alkaline insect gut. Your guts are acidic, not alkaline. The activated Bt toxins then bind to an insect gut cell protein. That binding event triggers a complex series of molecular changes that result in the Bt toxin protein forming a pore. And a classic way to kill a cell is to form a hole in it, because cells really don't appreciate allowing the free flow of what's inside of the cell and what's outside of the cell to exchange. So that results in insect death. You can go online and see all these amazing pictures and videos of of crops that either have no Bt toxin gene or that make Bt toxin. And they'll drop a single larvae that eats the plant. And the normal plant gets ravaged by the larvae very quickly. The Bt toxin expressing plant has a few bites taken out of it, and then the larvae rolls over and dies. So it turns out that this strategy for pest control has resulted in a large increase in agricultural gains, something like $100 billion of increased agricultural gain over the past 15 years or so, resulting from higher yields and lower use of chemical insecticides. Now, as you might imagine from this picture, binding of the BT toxin to the insect gut protein is absolutely essential for this mechanism. And so if you were an insect population and your farmer kept giving you as your major food source crops that had BT toxin in them, you would probably evolve a way to get around the BT toxin. And the easiest way that the insects have found to do so is to mutate this receptor so that it no longer binds the BT toxin or simply downregulate it, meaning make less of it. So that the Bt toxin is not able to— enough Bt toxin is not able to bind to the insect gut cell, form a pore, and kill the insect. So most or all of the major Bt toxins currently in use have witnessed the rise of insects that are resistant to them. So Bt toxin resistance in insects is considered the most serious threat to sustaining the gains from transgenic crops. Okay, so we wondered whether we could use PACE to continuously evolve BT toxins that would bind to a different insect protein in the gut and thereby overcome this mechanism of insect resistance. We chose an insect called cabbage looper, which is, I realize, unfortunately cute. Or T. nee. So this is an agricultural pest that has evolved BT resistance, and its resistance mechanism has been very well studied. So BT toxin normally kills cabbage looper by binding to a protein called ABCC2, and mutations that downregulate or alter the structure of ABCC2 have given rise to cabbage loopers that are BT resistant. So what we decided to do was to comb through the cabbage looper genome and look for proteins that were similar to proteins that we knew Bt toxins could bind in other species. And we found such a protein called— that we called TnCAD. This protein is not bound by Bt toxin, but if you squint and look sort of at the ones that are bound by the Bt toxins, you could say, well, this is kind of related. And just as I did— I showed you with the T7 RNA polymerase, you can't use PACE to directly evolve a BT toxin that binds to this final TN-CAD target because it's too far away from the proteins that BT toxins can normally bind. So instead, just as we did with the T7 RNA polymerase story, we had to use an evolutionary stepping stone. So we made a protein that was sort of halfway between the final target and proteins that we knew Bt toxins would bind. Now you need a selection that links binding of a Bt toxin to this target protein or to the stepping stone to PACE, to turning on gene 3 expression. And so we constructed such a selection by taking advantage of what researchers like Ann Hochschild at the Harvard Medical School and others have developed, which is commonly called the bacterial two-hybrid system. It basically says if you're trying to evolve a BT toxin to bind to a target, let's link the target to a DNA binding protein so that it stays put. Let's link the BT toxin to a subunit of RNA polymerase. So if the BT toxin binds to the target, then the RNA polymerase is brought to the region of the promoter that's upstream of gene 3, and that results in turning on the expression of gene 3, which results again in phage being able to produce infectious progeny. So that's the simple protein-protein binding. And we validated that this selection works by using antibodies that bind to certain antigens and showing that we could take a dead antibody and evolve it into an active antibody using this selection. And then we applied this selection to evolve BT toxins. So Ahmed Badran, the talented student who led this work, did 22 days of PACE, so it's about 500 generations, what would have taken about a decade to do by conventional protein evolution. And for the first half of the evolution, for the first 11 days, he challenged BT toxin to evolve on the stepping stone target. And then for the last half of the evolution, he challenged it to evolve on that final target. And throughout the evolution, he increased the mutation rate, and he increased the flow rate through the lagoon, which allows us to play with the selection stringency, because we want the selection pressure to be strong enough that the populations are always moving towards our goal, but not so strong that we get washout. Washout means no phage can find an answer, so you come back the next morning, you check your lagoon, and there's no phage left in there. That means that the task was too hard. And one of his first attempts actually went all the way through, and the phage population survived all the way to the end. One of the beauties of being in an environment like this one where high-throughput DNA sequencing is readily available is you can just take the waste line from these lagoons and sequence them every day, every half day, and piece together the complete evolutionary trajectory of how the Bt toxin genes evolved when challenged with this problem of how to recognize this different protein. So I won't go through the details, but we were actually able to put together a plausible evolutionary trajectory, a history of how natural Bt toxin ended up evolving through many mutations. The average Bt toxin gene at the end of this process has 23 non-silent mutations to get to a final answer that's represented by this blue curve. So we can actually observe the rise and falls of little empires of Bt toxin genes. As they are challenged by our different selection conditions. And you can see in this table, the details don't matter, but the colors tell you that there's an emerging consensus of different mutations that eventually take over the lagoon and are represented by these late blue species. Okay, so how do these evolved Bt toxins— what do they do? First, we mapped them onto the structure of the BT toxin, and we found that these mutations all tend to cluster on the face of the BT toxin that binds to the target protein. So that was good. And then we assayed how do these BT toxins behave. Now, the lowest bar we could imagine asking of these BT toxins is that they turn on our gene circuit. That's what we directly selected them to do. So how does the natural BT toxin compare with our evolved BT toxins? With respect to their ability to turn on this gene circuit. This gene circuit used to drive gene 3 expression. We replaced gene 3 with an easily assailable gene called luciferase. It's the same gene that makes fireflies glow. And this just allows us to assay how good they are at turning on gene expression. And what we found is that the evolved genes do an amazing job turning on the expression of this gene when you put the stepping stone as the bait or when you put the final TnCAD protein as the bait. Okay, then we just purified these evolved BT toxins and measured their ability to bind to the target proteins. And what we found is that natural BT toxin had no detectable activity on our final target, but the evolved BT toxin had evolved very good binding activity. Its dissociation constant is 11 nanomolar, so that's a dissociation constant that signifies the binding is strong enough to resemble that of some some drugs, for example. So it's a very tight binding interaction with the target, even though the natural starting BT toxin had no detectable activity. Okay, so next we wondered whether these BT toxins could kill cultured insect cells that we had rigged to express this TNKad target. And the answer is yes. So in this experiment, we— remember I told you that BT toxins work by creating holes in insect cells. So the holes allow a fluorescent dye to come into the insects, and that causes the cells to light up green. So this is a kind of odd set of data where green cells, which normally we think of as happy cells, are actually dying. Okay, so green means the cell is dying because it's got holes in it. So if you take cultured insect cells and genetically equip them with the ability to express the target receptor, you find that our evolved BT toxins kill them. But the natural BT toxin can't, which makes sense because I told you the natural BT toxin can't bind to this target. If you take the natural receptor that the natural BT toxin binds and express it from these insect cells, the natural BT toxin can kill these cells, and our evolved BT toxin still does as well. So it hasn't lost the ability to bind to the original target. And this is just sort of graphical form showing that the evolved BT toxins are much more potent at killing these cultured insect cells that we've rigged to express this target. Okay, but the real, the real test is how do they kill actual insects, actual cabbage loopers? So we started a collaboration with a group in Cornell that has colonies of these cabbage loopers. And we assayed the ability of the natural or evolved Bt toxins to kill two types of cabbage loopers. The sensitive ones, these are ones that are not resistant to Bt toxin, and the resistant ones. Sensitive cabbage loopers are very potently killed by natural Bt toxins. So less than 0.1 part per million in their diet will kill sensitive cabbage loopers. Our evolved Bt toxins are a little bit better or comparable to the natural Bt toxin. And it's tempting to speculate that maybe they're a little bit better because they've evolved two different mechanisms by which they can now kill this cabbage looper. They can bind to the original protein or they can bind to our new target. Although that's just speculation, we haven't shown that. But then the real question is what happens when you challenge resistant cabbage loopers with our evolved Bt toxins? And we sent the samples blinded, so they weren't decoded for our collaborator. And our collaborator, Ping Wan, called me about a week after we sent them and said, David, I think you guys messed up the concentrations of the proteins because everything is dead. And I said, no, no, that's good. Keep diluting. Just keep diluting until you get some life. And this was the answer. So the natural Bt toxins are about 1,000-fold worse at killing Bt-resistant cabbage loopers than the normal cabbage loopers. Cabbage loopers, but our evolved Bt toxins are up to 350-fold more potent at killing the Bt-resistant cabbage loopers than normal Bt toxin. So they've restored almost all of the activity that they lost by the cabbage looper turning down the expression of the ABCC2 natural target. And so this established, for the first time, that binding to a different insect gut protein could actually restore the potency of these toxins and overcome BT-resistant insects. Okay, so to summarize, I apologize for going slowly enough that I took longer than I thought I would. We co-opted the filamentous bacteriophage life cycle and turned it into a system that allows us to evolve proteins continuously in the laboratory with a minimum of research intervention. And that allows us to evolve proteins at a speed that's about 100 times faster than we could do before. Those capabilities allow us to answer some fundamental questions about evolutionary biology that would be difficult or impractical to answer otherwise. We found, for example, that at least for T7 RNA polymerase, evolutionary path dependence and irreproducibility arise from historical choices. That are made either because of the difference in path or by random chance between interacting amino acids that are close together in the protein structure. And then epistasis, the ability of one mutation to affect whether another mutation is beneficial or harmful, permanently preserves those differences. And then finally, I showed you an example of using PACE to solve a problem in biotechnology. We evolved BT toxins that had high affinity for a new protein that's not normally bound by BT toxins. And as a result, these new evolved BT toxins can overcome BT resistance in an agricultural pest. Okay, so that leads me to thank a very talented and hardworking group of students and collaborators. So PACE was developed by a former graduate student, Kevin Esvelt, now a professor at MIT, working with Jacob Carlson, now a postdoc at UCSF. The Tape of Life work was led by Brian Dickinson, a former postdoc who's now a professor at the University of Chicago, working with Aaron Leconte, who's now a professor at the Claremont Colleges and a former undergraduate in the lab, Ben Allen. BT Talks and Pace was led by Ahmed Badran, a current graduate student who's about to become a fellow at the Broad Institute, and was a really terrific collaboration with scientists at Monsanto, including Victor Gusev and his coworkers. And as I mentioned, Ping Wan was our collaborator at Cornell who had the colonies of cabbage loopers, and his coworker Wendy Kane performed the bioassays on the cabbage loopers. I'm grateful to the NIH, DARPA, HHMI, and Monsanto for support of this research. And finally, thanks for your patience, and I'm happy to answer questions. Oh, sorry. You're supposed to go up to the microphone or something. We'll pass around the mic.

Excellent work, David. A couple of questions. As an academic, the question that you posed at the beginning of the thing, whether there's a path dependency in evolution, you answered the question. But as a practical matter, when you look at the BT cotton, and if you think that in an actual field, the Cotton crop will suffer from a variety of stresses. Insect is one of the stresses, insect pathogen. And therefore I think what I call what you did, not a genetic engineering but an evolutionary engineering, so to speak. And that one did, if that is the problem that you're solving for a BT cotton, you showed successfully that you could create, evolve the BT toxin and then potentially can do that. However, I've been to many cotton fields. What I found is that in the course of doing that, there are many other stresses the cotton plant experience and they lose the competence to fight that one. For example, these are the insect-related stresses, but abiotic stresses the plant experience and they lose the competence to deal with them. One question is that this is interesting way of fast-forwarding multiple generations to find the BT toxin. However, if you go and look at all the cotton crops available around the world and different kind of cotton seeds that are available around the world, I'm sure there are some of the seeds, if you sequence the genome of that one, they have a natural tendency to resist some of the pests. And would that be a faster, more effective way? Because in the sense that they evolved through Mendelian selection as opposed to doing it in laboratory selection. So I like your comment on that.

Yeah. Yeah, so I think the question has a couple parts. One part is to point out that it's not necessarily pest resistance that is the rate-determining step, the bottleneck in the growth of a cotton crop, for example. That it could be that drought tolerance or heat tolerance or tolerance to other environmental insults could also be important. And the second part of the question is, it's probably the case that some cotton plants somewhere in the world have evolved insect resistance, perhaps through a different mechanism. And would it be as fruitful or more fruitful to try to harvest the genes that have evolved naturally to convey insect resistance rather than using laboratory evolution? And I think those are both interesting questions. Good points. You know, we were specifically aiming this application of PACE to address the problem of BT-resistant insects. If you had a molecular mechanism by which drought or heat tolerance or an environmental assault could be narrowed down to the activity of a protein, and if you could link that protein to gene expression, then perhaps you could use PACE to try to address that bottleneck as well. But I completely agree with you that even though I lack almost any farming experience, that it's likely that in some fields insect resistance is not the big bottleneck. And I also think, of course, harvesting resistance alleles from nature is a major strategy that should be looked at and is looked at. And in fact, Bt toxin itself came from nature. It came from a realization that this bacterium, Bacillus thuringiensis, kills insects naturally as part of its reproductive mechanism. It reproduces in the insect that unfortunately has a bunch of blown-up gut cells, and that harvesting the genes that it evolved to kill those insects, the Bt toxins, was an effective way to do pest control. So I actually agree with you on both points. A few really quick ones. For example, one is, Is that gene 3 encoding for the P3? Does the P stand for protein in that? Yes. How original. Yes. And the G stands for gene. And yes, all the genes in the phage are Roman numerals. And it's all incredibly boring nomenclature. When you had the diagrams of the proteins that were linked together, how did you make the linkers without sort of interfering with the protein activity, or is that really— or is that those linkers not to scale? Yeah, so it's a little bit of an art. And sometimes you guess right, and sometimes you guess wrong. In the case of linking protein activities to gene 3 expression that required the interaction of two proteins, we really benefited from the history of people working in the so-called to hybrid community. So there's a whole sort of subculture of chemists and biologists who have come up with all sorts of creative ways to get two proteins to interact with each other, and by virtue of their interaction, get stuff connected to them to interact more frequently. So we benefited from a lot of the past research that had already shown that certain proteins are very amenable to this kind of modular tethering. I'll say that if you do a simple back-of-the-envelope chemical calculation, you can actually convince yourself that you don't need to have a tether in a very specific orientation or length in order to dramatically increase the collision frequency of something that is connected to molecule A with something that's connected to molecule B if A and B get together. So if you imagine a solution as occurs in a test tube or maybe in a cell— it's a little more crowded in a cell— stuff is really far apart unless it's tethered together. So just the mere action of making two molecules one molecule through tethering already dramatically enhances the collision frequency of partners that are tethered together. So David, can you speak to, or could you speculate a little bit about other uses for PACE, for example, in space? Yeah, so we've used PACE to evolve some genome editing proteins. Many of you have probably heard of CRISPR-Cas9 and maybe TALE proteins as proteins that allow us to make programmable changes in genomes. One of the main challenges facing the use of genome editing proteins is that they have some off-target activity if you just take them from nature, meaning that they will hit your target, but they'll also hit other targets that aren't your target. And generally, it's not a wise idea, either for research or therapeutic purposes, to modify semi-randomly in the genome. So we used the PACE system to evolve genome editing proteins that were much more specific for maintaining their on-target activity, but getting rid of their off-target activity than the natural ones. And we did this by using positive selection PACE to constantly enforce a Darwinian selection pressure that they had to bind to the target DNA sequence. And using negative selection PACE, something I didn't have a chance to talk to, to get rid of their ability to bind to the off-targets. So P3 is this essential protein that the phage need to replicate. There is a sort of dark parallel universe bizarro world cousin called P3 neg that poisons phage infectivity. And so you can put some undesired activity, you can link that undesired activity to the production of P3 neg. And then it's a trade-off between how much good activity does the protein have and how much bad activity does it not have. That determines its ability to propagate. So we can select for good traits, and we can select against bad traits as well. And so we use that to do genome editing protein evolution that has greater DNA specificity, less off-target activity, as an example. And we've worked on quite a few other projects, many of which are published and others of which are still ongoing in the lab. Great. OK, last question. I'll hang around afterwards if people want to ask more questions since— So it seems that the goal of the PACE is to have some very fine-controlled and very fast evolution in the lab. So I'm wondering, how does this compare to, say, simulation of evolution on a computer? Yeah, so a lot of progress has been made. So the question is sort of how does PACE compare in speed with evolution in silico, these computational evolution, which has become its own subfield that has made a tremendous amount of progress. So researchers can model evolution by having kind of artificial genes in a computer program and put them through some artificial selections. So I'll say that computational evolution, for the most part, is not able to predict with any reasonable accuracy what changes in a protein sequence are necessary for certain activities. Instead, what computational evolution is good at is more answering questions like the tape of life questions, where you simulate genes with traits, you connect them through kind of arbitrary designations, and you study how mutation rate or different selection pressures affect the outcomes. But we're not good enough, unfortunately, at being able to use a computer to predict what mutations we need to evolve in order for a protein to have desired activity, in part because we're not even good enough to say forget about evolution, how should I change this protein with one amino acid change to have this new activity? Proteins are macromolecules. Macromolecules are very hard to predict purely computationally exactly how they'll change in response to changes in sequence. Although a lot of progress has been made, and that continues to be an area where I think there's a bright future. Translation state? It's really that the complexity of relating the sequence of a protein to its activity. Is too hard to predict purely computationally. So you can make general statements about the effect of migration in evolving populations, or the effect of mutation rate, or how reproducible evolution is. But to actually reduce it to experimental reality is more difficult. And there's sometimes not a great correlation between what you predict in silico and the experimental results. Thank you, David. Great. Thank you.