Scientist Stories: David Schaffer, Directed Evolution of Next-Gen AAV Vectors for Gene Therapy

Hi, my name is David Shafer. I'm a professor of chemical and biomolecular engineering, bioengineering, and molecular and cell biology at UC Berkeley, where I'm also director of the Berkeley Stem Cell Center. In addition, I'm co-founder of a company called 4D Molecular Therapeutics that I'll be talking about today. So the focus of the work that I'll be discussing with you today is our efforts over the past 20 years to use a technique called directed evolution to be able to create new viruses for therapeutic gene delivery. So the motivation for this work is that there's serious unmet medical need within our society. And depending upon which tissue you look at, you can typically see two categories of diseases. So taking the nervous system in the upper left-hand corner as an example, we have conditions that affect large numbers of patients where the genetics is not entirely clear, and it's likely a combination of both genetic as well as environmental influences, such as Alzheimer's or Parkinson's disease. And we also have more rare conditions such as, as Storger's disease, Rett syndrome, Huntington's disease, amyotrophic lateral sclerosis, which I'm sure you otherwise know as Lou Gehrig's disease. And these are conditions where you can pin the blame on specific genetic changes, specific mutations within particular genes of the human genome. So regardless of which category of condition you're looking at, however, there's a number of situations in the nervous system where people have developed treatments that can alleviate symptoms for these horrible conditions. But there are currently no cures that are available. So for all of these conditions on this slide, it's quite clear that we need new treatments, new therapies. So if we take a step back and consider the basic biology that underlies all biological function, we can see that, of course, DNA encodes all the genetic information on the planet, and this information can then be transcribed within organisms into RNA, and then that in turn becomes translated into a protein. The flow of information from DNA to RNA to protein, of course, is called the central dogma of biology. From the point of view of somebody who's trying to develop a therapy, we can view each one of these levels as a potential drug target. So the majority of pharmaceuticals that are currently in use target at the level of proteins. Small molecules float around, typically bind hydrophobic pockets, such as the active site of an enzyme, and inhibit or alter the function of that enzyme. Monoclonal antibodies float around and bind to protein targets floating on the outside of the cell. Such as a growth factor or a growth factor receptor. If we move back one level, you can begin to develop therapies that work at the level of RNA inside of the cell. So these include RNA interference or antisense RNAs, whose goals are to knock down the production of proteins from RNA, or messenger RNA itself can actually be a therapeutic molecule. But moving back one further level, at the level of DNA, this has really been the focus of my entire career. So gene therapies, the delivery of new genetic information to a cell, genome editing, the changing or the editing of existing genetic information inside of a cell, or cell therapies or stem cell therapies where the goal is to deliver an entirely new— So the power of targeting DNA to be able to affect a therapeutic outcome is the fact that DNA, unlike small molecules or proteins, are actually —tissue. So if you can add new DNA to a cell or tissue, that can potentially persist for the life of a patient, which raises the possibility of single administration, long-term patient benefit. So just to illustrate this idea, we have here on the x-axis the number of years that an individual is affected by a particular condition. So typically post-diagnosis, Parkinson's patients are on the order of another 16 years or so. So congestive heart failure between diagnosis, ultimately when a patient passes away due to CHF is on the order of 10 years. Age-related macular degeneration itself is not a condition. It robs people of their vision, but typically does so towards their lifespan. So they typically die of other causes after around 10 years of diagnosis of AMD. So if we consider or compare these to the lifetimes, the pharmacokinetic half-lives of therapeutic modalities that are being developed to be able to treat these conditions, we have a small molecule with a pharmacokinetic half-life on the order of hours. Monoclonal antibodies last days to several weeks. However, DNA in the form of a gene therapy can potentially last years or even decades. So here are a couple of examples of gene therapies, one of which I'll be discussing in more detail. Where a single administration back in the 2000s led to, in some cases, over a decade of therapeutic benefit from the patients who were still enjoying the fact that they received this gene therapy many years ago. So let's all get on the same page and actually define what gene therapy is. It's the delivery of genetic material to the cells of an individual patient for therapeutic benefit. And so while this idea was originally proposed back in the 1960s, we've had to wait for 50 years in order for it to begin to come to fruition. And the reason is this word highlighted in yellow here, delivery, which is the fact that DNA, from the point of view of a cell or an organ, is a very large molecule. It's difficult to get it past the cell membrane and into the nucleus of a cell. And in addition, on its own floating around in the bloodstream, DNA is not particularly stable. It's degraded by nucleases that are present. So it's been hard to be able to develop delivery vehicles or vectors that can carry DNA inside of enough cells to begin to move the needle on treating a disease. So the way that this has been achieved in the field increasingly over the past couple of decades is that people have been identifying viruses, very effective viruses, that can carry DNA inside of cells. So AAV, or adeno-associated virus, is a particularly interesting one. It has risen to be the most promising virus within the field of gene therapy for a couple of reasons. First and foremost, which of course is a major factor to discuss today, it's safe. It's not a pathogen. Infection, and in fact, the majority, on the order of 80 to 90% of the human population, has already been infected with this virus and never even noticed it because it's not associated with human disease. It's a harmless, non-symptomatic virus. In addition, it's among the smallest of viruses. So if you take a look at its genome structure here, it has two genes called rep and cap, and the rep encodes some enzymes that mediate the replication of the viral DNA. The cap has really been the focus of about half of my career because the cap encodes some nonstructural, sorry, some structural proteins, around 60 copies of which self-assemble to form this particle here on the right-hand side. And the particle, as you can see on the top side of it, there's a black circle. It contains a pore at what you call the fivefold axis of symmetry. And once this particle is assembled, one copy of the viral DNA will snake its way inside of that pore to be able to load the virus with its own genome. Only around 4.7 kb in size and the particle is around 26 nanometers in dimension. Importantly, a number of natural variants, well over 100, have been identified and contain a lot of sequence variations in the amino acid compositions on the surface of the virus. So those white peaks, for example, that you see emanating from the core of that viral particle. And these changes in the surface composition of the virus have been correlated with changes in the delivery of the pathway and the bio-distribution and the cell preferences of the virus, which raises the possibility that by engineering the surface of the virus, you can actually create highly efficient targeted viruses to carry DNA medicines to particular locations inside of the cell. So the way that the virus works is that after being in tissue, for example, the bloodstream or the lung, it makes its way past a number of barriers, such as bodily fluids. It finds its way to the surface of a target cell. Binds to a receptor on that target cell surface, undergoes internalization in a receptor-mediated fashion, escapes from the endosomal network. It actually contains a domain of the viral protein that enables its way to sneak, enables it to sneak its way past that lipid bilayer that surrounds it, and then eventually finds its way into the nucleus, uncoats the viral capsid, and releases the viral DNA, which can then go on to carry out subsequent steps in the viral life cycle. And the way you turn this into a therapeutic is you quite simply take the rep and cap genes, delete them from the viral genome, place them onto a separate piece of DNA, and this results in two outcomes that are advantageous. One is that it liberates space for the insertion of a gene of interest between those flanking sequences, those ITRs, and as a result, when that viral particle on the right-hand side is assembled, that gene of interest becomes internalized into the viral particle your virus then with a piece of medicinal or therapeutic DNA to treat a disease. In addition, as a result of the fact that the rep and the cap no longer make it inside of the virus, this is no longer a replication-competent virus. It simply deposits the gene of interest into the nucleus of a target cell, and that DNA can then persist for long periods of time and be able to produce a therapeutic product such as an RNA molecule or a protein molecule to treat a disease. So this resulting recombinant virus has a number of advantageous properties. It's very safe. Like I mentioned, 90% of the human population has already been exposed to one or more natural variants of AAV. It's somewhat efficient, the natural variants are, and have been used in over 200 human clinical trials involving delivery to muscle, liver, lung, brain, and retina. The product from that genome after delivery is very stable and persists at least in non-dividing cells for years or a decade and counting., and it's possible to manufacture these at high yields. And as a result of all these properties, the virus is enjoying success with an increasing number of human clinical trials, resulting, for example, back in 2017, in the first FDA approval of a gene therapy for a rare disease. And this therapy was based upon an AAV virus, and it was to treat a blinding condition called Leber's congenital amaurosis type 2. So this is currently being developed and commercialized and sold by Roche Therapeutics. Roche from Switzerland. So we've seen an increasing number of successes within clinical trials. That LCA disease that there was an FDA approval for in 2017. There was an approval for spinal muscular atrophy from Novartis to be able to treat a rare condition which results in fatality and progressive paralysis and fatality of young boys. Hemophilia B has been has advanced to a pivotal trial, a Phase 3 clinical trial, and shown strong promise. Hemophilia A is not that much further behind, and we have a long list of successes— well, not that long yet, but increasingly long list of successes that have resulted from the gene therapy field. However, these successes have resulted in some pretty savvy clinicians who have identified diseases where the proper— good immunologic efficacy. But even so, these viruses were not designed as pharmaceutical gene delivery vehicles. And as a result, these investigators have had to really brute force the delivery, had to use very, very high dosages of virus, or have had to use invasive routes of administration to place the virus right next to the target cell. And the field knows that to be able to build upon these initial successes and extend them to the majority of human diseases, We quite simply need better viruses. So in other words, gene therapy is starting to work, but many diseases are still beyond the capabilities of current delivery technologies, and we need new AAVs. So what are the, some of the existing challenges with AAVs? The first one I've indirectly alluded to already, which is the fact that 90% of us have already been exposed to one or more AAV variants, and as a result, we've been accidentally vaccinated against the virus. Virus and high concentrations of antibodies that will neutralize natural versions of AAV. And so that's a significant barrier within the field. It's in fact the most common reason that patients are turned away from clinical trials. In addition, after the virus potentially makes it past those antibodies, it inefficiently distributes to most target tissues inside the body. It tends to go very well to the liver, but not as well to many other tissues. There's inefficient spread with those tissues. Once the virus is able to reach the surface, it can't penetrate deep inside that tissue. If it reaches near the target cells, it has the ability right now to target delivery to some cell types but not others within that tissue. And even if it reaches the surface of an intended target cell, it's often taken up very poorly by that cell. So these are, after all, respiratory viruses, and it's a lot to ask of them to be able to very efficiently infect other target cell types such as a neuron or a cardiomyocyte. So all these shortcomings arose from a very simple fact, which is that nature never evolved viruses for our convenience to use as medicines. And so around 20 years ago, we developed a hypothesis that if we'd like to improve the properties of these viruses for therapy, we need to re-evolve them or re-engineer them at the molecular level. So the toolkit or the methodology that was emerging in the field that we saw promise in enabling us to do this was the product of an investigator by the name of Frances Arnold. She actually grew up in UC Berkeley back in the '80s, and since that time she's been a professor at Caltech, and she won the Nobel Prize in Chemistry in 2018 for the concept of using evolution as a tool to design novel function into biomolecules. So she did this specifically with industrial enzymes, and we became the first to do it with viruses like AAV. So this approach of directed evolution, and I'll explain how it works in just a second, we've been conducting it the past 20 years with AAV, and we have multiple examples in which we've used this approach to be able to create new AAVs that have overcome some of these barriers I talked to you about before with a variety of cell and tissue targets. So the way that this technology works is it's based upon evolution, and evolution has two basic steps. The first is to create a gene pool, and the second is to select the fittest. So as a gene pool, we take several natural versions of this virus, and we perform a number of steps to genetically diversify them. And specifically, I'm not gonna go into this in great detail, but we've created some libraries in which we've generated chimeras of several of these parent viruses and generated new variants that contain pieces of individual parents. We've used bioinformatics and computational approaches to reconstruct family trees of AAVs and began to predict sequences of these viruses that may have existed long time ago and built libraries around those, a process known as ancestral reconstruction. And in fact, we have those two libraries and 35 others that we've utilized to introduce diversity into that viral coat protein in a variety of different ways. So that's the gene pool where we then package these into libraries of particles where every single variant or every single particle is composed of a variant protein shell, a variant capsid, contains inside of it the DNA that encodes that capsid. So we'll then, as the second step, play Darwin, select the fittest, and I'll describe to you some selections. And by spinning around this loop iteratively, we can end up evolving or selecting variants that have desirable properties. And at any point along the way, pop open the particles and sequence the genomes to recover the information, the sequences or composition of those capsids that mediated the efficient gene delivery or the efficient targeting to a given a specific tissue. And once we have those new engineered capsids, we can then begin to utilize them to deliver DNA medicines. So, just to outline today, I'm going to show you a couple of topics that we've been working on. One is engineering AAVs for the retina to be able to treat blinding conditions. Another is for the central nervous system. And then I'm going to switch gears a little bit afterwards and talking about our efforts towards using this approach to be able to solve problems problems in process engineering for AAVs. So as I mentioned to you, the very first FDA-approved gene therapy for a rare disease involved an AAV for treatment in the retina, and this is the structure of your eye. At the back of the eye, you have a number of layers of neurons that are painted onto that surface area, and all the way at the back is the so-called ONL, or outer nuclear layer, and this contains the photoreceptor neurons, cells that sense photons. Once a photon hits one of those cells, it fires an action potential, and it communicates that information to various inner neurons in the INL, internuclear layer, and GCL, the ganglion cell layer. These GCL neurons will then send information along their axons through that white line that lies at the— or sorry, yellow line at the back of the eye, which is the optic nerve. The eye then communicates that information to the brain for further processing. So the place you would like to be able to— oh, one final point I'd like to make is that in the back of the eye in the center, there's a region called the fovea that contains a high concentration of cone photoreceptors, and this is what's responsible for high visual acuity. So if you're reading a page and your eye scans back and forth, what your eye is doing is keeping the light from a page focused in on the fovea because you need that high acuity vision in order to read the page. So, this is an important region of the retina to protect. So, unfortunately, mutations in over 200 different genes can result in the death of photoreceptor neurons in the eye, and other conditions, such as macular degeneration, can also kill off this cell layer. Many, many patients, age-related macular degeneration affects over 1 to 2 million patients within the United States. So, the place you would like to deliver a drug to be able to treat these conditions is through into the vitreous of the eye, this fluid or gel-like substance in the inside of the eye, because if you can do an injection there, it's very simple. It takes around 4 seconds, local anesthesia only, and this gives your drug the opportunity to be able to diffuse throughout the eye and to come into contact with the full surface area of the retina, and therefore protect the majority of cells inside the retina. And in fact, this intravitreal injection is conducted several million times a year in the United States to deliver protein drugs produced by Genentech and Regeneron to be able to treat age-related macular degeneration. So, unfortunately, the cells of the INL and the GCL represent a barrier where you would like your AAV to deliver the DNA medicine to the photoreceptor neurons—that's the target for the therapy—but, unfortunately, no natural version of AAV can actually distribute or transport through the several hundred microns of dense tissue it would take to be able to make the trip from the vitreous fluid all all the way back to the outer nuclear layer. So as a result, natural versions of AAV have to be injected right next to that target cell. So in that FDA-approved product, for example, which uses AAV2, natural serotype 2 for AAV, they have had to take the needle, actually puncture it through the retina, and inject fluid between the outer nuclear layer and the underlying retinal pigment epithelium,, and this is problematic for a couple of reasons. So, first of all, it detaches the retina, and at the injection site, this retinal detachment can actually cause some damage. And second, the injection of the fluid behind the photoreceptors results in the spread only in a small bubble or bleb, where the fluid ends up causing that retinal detachment. So, this mediates delivery of AAV to those target cells, and only around 5 to 10% of the surface area of the retina. So in that FDA-approved product, they're actually sacrificing around 90% of the cells to the progression of the disease. So we asked the question, can we do this? Can we actually evolve AAVs to be able to reach this target cell deep inside the tissue, which is representative of many problems in gene therapy where you would like the virus to spread throughout a tissue to be able to reach inaccessible cells? So the way that we did this is that there was no cell culture model that was capable of recapitulating the complex structure of the retina for a long enough period of time for us to conduct the evolution, so we did this in vivo. We took our library of viral particles, and in sum total at that stage, our library size was around 100 million different AAV variants. We then administered these initially in a proof of concept study into a mouse retina, waited a certain amount of time for the winners to be able to traffic back to the photoreceptors shown in the bottom in green. We then, using fluorescence-activated cell sorting, recovered the AAV genomes that made it to those cells because those particles carried the DNA that encoded the viral particle along for the ride. You could think of them as being barcoded with their own DNA. And then repackaged the library, readministered it, and went through 6 rounds of selection so that that initially very rich pool ended up converging just down to a small handful of AAV variants shown there on the bottom left-hand corner with one particularly dominant variant which comprised around 3/4 of the pool. That particular variant, which we called 7M8, has a structural cartoon shown on the bottom right-hand side where we see in green there were some changes in the amino acid composition on the viral surface. Specifically, a peptide was inserted into that region of a very specific sequence. And you can also see some dots in red. So we had some point mutations as well within that variant. We wanted to test the ability of that variant to start carrying recombinant DNA inside of cells. So we initially loaded it with DNA encoding the green fluorescent protein fluorescent protein and began to test its delivery properties in the mouse eye compared to a parent virus, specifically AAV serotype 2. So shown there on the bottom left is AAV2. The vitreous fluid is on the bottom side. The photoreceptors are on the top. See why in that FDA-approved product, they have to use a needle and a subretinal injection to place the AAV right next to the photoreceptors, because AAV2 from the vitreous can't even reach the photoreceptors. So this compares to the variant that we created, 7m8, where you can see that we're carrying GFP, we're carrying the DNA to every single, roughly, cell population, all the way from the bottom, the retinal ganglion cells, to the top, the retinal pigment epithelium. And furthermore, we're doing this across much larger surface area than 5 to 10% of the retina. We're actually delivering DNA to the full surface area of the retina. So this was an initial proof of concept study within mouse, and we wanted to test how well it translated to larger animals. And we didn't necessarily anticipate it would translate well to, for example, to a primate, simply because evolution teaches you, you get what you select for. We selected in the rodent, we got a really good rodent vector. And the structure of the rodent eye is quite simply different from that of larger mammals. In a couple of ways. Shown here is size, that as you move from a mouse to a dog to a monkey, a cynomolgus macaque, to a human eye, there's an increase in the size of the retina, where a mouse has on the order of 100 microliters of vitreous fluid, whereas a human has 4,000 microliters of vitreous fluid, and macaque is closer to 3,000. And there's some structural differences as well. Just wanted to remind you that this fovea is the region of the back of the human eye that is responsible for high acuity vision. It's a particularly important region to protect. Mice don't even have a fovea, so you can't study the effects of gene therapy on protecting the fovea very well within a mouse. In addition, there is a matrix barrier, extracellular matrix barrier, that lies between the fluid of the eye and the retina, the neural retina, and that's called the inner limiting membrane. And it's been shown in a number of studies by ourselves and others that this poses a significant barrier to gene delivery. And that ILM, that layer is quite simply thicker within larger eyes, such as monkey and human. So we wanted to test how this mouse variant did within macaques, within animals, monkeys rather. And so we tested using a number of natural versions of AAV, AAV9, 2, 5, and 8. And what we see is this transduction pattern on the lower right-hand side, you can see the reason for it, that the inner limiting membrane is pretty thick across the majority of the surface area of the retina. It thins out right around the fovea, and in that region, it begins to expose some of those superficial neurons, those retinal ganglion cells that kind of encircle or surround the fovea region. And so you end up with a ring of transduction around the fovea, a ring of retinal ganglion cells. But on the bottom of the bottom right-hand side, you can see those cones, those important cells you'd like to protect,, and if these variants were delivering DNA to those cells, you would see white in the interior of that circle. You can see there is none, because those natural versions of AAV aren't reaching those cells. So this is the variant 7m8 that we created. We can see there's better transduction across a number of puncta across the retina, and those turned out to be regions where the vasculature within the retina actually resulted in a thinning of that inner limiting membrane. And in addition, we can see in the middle of the eye that fovea, we do have some transduction, some delivery to cone photoreceptors. So, it's better than the natural versions, but we felt we could do even better. We're not getting broad, widespread, uniform transduction across the retina. So, however, this variant was good enough to be able to treat a particular condition, age-related macular degeneration. And a company down in Menlo Park known as Adavira, I wanted to mention that I have no affiliation with this company, but they licensed this variant 7m8 from UC Berkeley and took it into a human clinical trial back in November 2018 to carry DNA encoding an anti-vascular endothelial growth factor inhibitor to be able to treat patients with wet age-related macular degeneration. And they've reported more recently, as recently as January now, that the first cohort of 6 patients who were dosed back starting in November 2018 have now for, for a period of approaching a year and a half, been off the protein therapy that I mentioned to you earlier. There are anti-VEGF protein therapies that are injected into the eye on the order of 10 times a year, drugs that were manufactured or produced by Genentech and Regeneron. So, these patients have been off of those monthly injections of the protein therapy for almost a year and a half now, and their vision is stable, and the structure of the retina appears to be stabilized, as well. And this is just showing that that press release specifically calls out that variant, 7MA. However, we felt that we could do better for the gene expression pattern. We could get more widespread transduction across the retina. So, we began to move towards larger models of the retina, ones that bore significantly higher resemblance to the human eye. And to do so, we started with the dog. So I mentioned to you that FDA-approved product for a blinding condition, Leber congenital amaurosis type 2. Earlier I mentioned that particular gene therapy, and it turns out that dog played a major role in the clinical development or the preclinical development of this therapy, specifically because there is a natural strain of dog which has a mutation in that same gene called RPE65, whose mutation causes that disease within human beings. So, as a preclinical step, an investigator at the University of Pennsylvania, Jean Bennett, treated these dogs and showed that she could get correction of vision within these animals prior to moving on to human beings and showing therapeutic efficacy within humans. So, in collaboration with William Beltran, also at the University of Pennsylvania, we began to conduct directed evolution within the dog retina. And so, we repeated the directed evolution, except within dog, and after a number of rounds of evolution, we began to to isolate some variants of AAV. So one difference between our initially doing this within mouse and more recently within dog is that sequencing technology has gotten a lot better. And specifically, we began to use Illumina sequencing to characterize convergence of our library over time. And you can see here A through F, these are a number of different libraries within our pool. And specifically, these are libraries in which we made changes to a specific small stretch of the surface of the capsid, for example, by inserting a random peptide sequence into those regions. And on the Y-axis, you can see the number of independent variants, which ranges typically from 1 million to 10 million variants for each one of those libraries. And you can see on the X-axis progressive steps in the evolution process, where the first line is plasmid, that's the original library as we cloned it. AAV library means we've packaged it once, and packaging into recombinant virus plays or exerts a selective pressure where not all of the changes that we made into the capsid are going to result in capsids that can stably fold. So we see a selection there for packageability, if you will. And then rounds 1 through 6 were progressive rounds of selection for the ability of those variants to reach target cells inside the eye. We observed something interesting in the original library to begin with. This is the plasmid library. And ordinarily, you would ideally like to see that every single recombinant variant had an equal shot on goal, had an equal opportunity to be able to show off, to be able to reach the retina. But some of these had actually stacked the deck. And on the left-hand side, you can see that there's some where there's a significantly higher abundance relative to the majority of variants, which are equally represented as indicated by by the fact that most of those curves have a flat line towards the right-hand side. But on the left-hand side, you can see some variants due to, we believe, artifact associated with oligonucleotide DNA synthesis from DNA synthesis companies where we ordered random sequences from those companies. There was some bias that led some variants to be represented as much as 100 or 1,000 times higher in frequency relative to the rest of the members in the library. So to be able to make sure those variants didn't have an unfair advantage, we conducted evolution in which we sequenced the distribution of variants emerging from the Darwinian selection every single round along the way. And this is the selection. And I'm gonna run through this— sorry. Where on the Y-axis we have the percentage of each one of those variants. Each variant within the pool is represented as a single dot in that dot plot, where the y-axis shows what percentage of the pool those dots comprised after it was packaged in that first plot, after round 1, 2, 3, 4, 5, 6. On the x-axis, that is compared to the percentage or the frequency of that variant within the original plasmid library. So what you'd like to see is that some variants should, of course, stay in the same position on the x-axis because they had a particular presence, a particular percentage or frequency within the original library, but you'd like to see some rise up progressively higher and higher on the Y-axis indicating that they're being selected for, they're increasing in frequency as a result of the evolution process. Furthermore, the ones that had an unfair advantage to begin with are shown there in green, and the ones that ultimately are the winners after round 6, you can see in the lower right-hand side shown in red, those are the most frequent variants within round 6. You can see where they lay within the original rounds. They were low frequency variants but progressively climbed the y-axis over time. And so the ones we really focused on were the ones that were red but not blue, not the ones that were the most abundant within the original pool. And those were the variants that we wanted to then test. So the way that we tested them is that each animal subject is quite precious, both for ethics as well as economical reasons. So we took a number of a number of different variants on the order of 20, and we used each one of those capsids to package a particular genome that contained the green fluorescent protein and also contained inside of it a small barcode, a molecular identifier, where the identity of that barcode shown in the upper left-hand corner there of that plot, there was a one-to-one correlation or relationship between the sequence of that barcode and which of those variants shown at the bottom right-hand corner in that circle, which one of those red variants was used to add a barcode. So we packaged 20 different barcodes with 20 different capsids individually, and then pooled all of them together, administered, mixed them together, injected them into the retina of a dog, and then both visualized the retina. So you can see here that we have a large number of cells that are expressing the green fluorescent protein, which means that at least one or more of those variants of AAV, the new variants that we created, were capable of carrying their DNA deep inside of the retina. However, we can then isolate the photoreceptor neurons and use DNA sequencing to recover the frequency of those barcodes within that tissue. And that tells us which capsids were capable of carrying DNA very efficiently inside of that tissue and which variants were capable, as a result, of mediating the expression of high levels of GFP, high levels of messenger RNA. Within those cells. So we took our entire pool, we applied it into the mouse just as a control, and then we applied it within the dog. And we can see a number of interesting features here shown on the right-hand side. You can see that regardless of whether we apply it to mouse or dog, AAV2 is near the bottom, indicating that that natural version of AAV is not very effective. On the right-hand two columns, we can see that the variants that we call canine, canine number 9 and number 16, were very effective at being able to mediate delivery and gene expression in the dog. But if you take a look within the mouse, K9-9 and 16 are a little bit lower. This tells us that there's sequence specificity for which variants are optimal. Again, telling you that nature teaches us, yeah, what you select for in an evolution. So if you want to do the— if you want to create variants that are optimal within a larger animal model, that tells you you can't do the selection within the mouse. You have to do it within the larger animal. And you see the converse as well, that the top variant within the mouse, K9-12, is actually lower on the list within the dog. So we took the top variant, K916, and we then tested it individually within the eye, and we actually generated two genomes just to test its ability to mediate delivery both within photoreceptor neurons, as well as the other target cell within the eye that I haven't talked about too much yet, which is the retinal pigment epithelium. So we specifically loaded two different genomes into that variant, one that contains GFP under the control of a photoreceptor promoter, promoter, and the other one of which contains tdTomato, a red fluorescent protein, under the control of retinal pigment epithelial promoter. We compared this to the variant within the literature that had been reported to be the best for delivery to the dog eye from the vitreous, a variant called AAV24YF+TV. And you can see our variant here, the K916, at various levels of magnification is mediating very, very high transduction, very high levels of delivery and gene expression. Both within photoreceptors shown in green, as well as within retinal pigment epithelium shown in red. So ultimately, we, of course, were very interested in translating this into the clinic, into human gene therapy to be able to treat the 200+ blinding diseases that rob people of their vision. So I'm going to show you a couple of slides now in which we conducted the evolution within nonhuman primate, within cynomolgus macaque. And this was work that was done at a company that I co-founded called 4D Molecular Therapeutics. It's now an 83-person company located in Emeryville, California that is taking these products into human clinical trials. So, this shows on the y-axis sequencing results or the percentage of the pool of variants that were recovered after every single round, or at least rounds 3 through 6, the pool that came from photoreceptors, from the inner nuclear layer, and from the retinal ganglion cell layer, as well as the retinal pigment epithelium layer. And there was one particular variant where you can see that that regardless of which one of these layers you're looking at, it was the dominant variant comprising in some cases over 80% of the variants that were emerging from that region. We then took this variant, used it to package the green fluorescent protein, and began to conduct studies that were IND-enabling. And shown here is a fluorescence live fundus image of, in this case, several nonhuman primates. And you can see that there's widespread expression at the back of their eyes. This is actually 6 months after delivery of that AAV carrying the green fluorescent protein into these animal eyes. And we took these animals post-life and did literally thousands of histological sections. And you can see that some regions are brighter than others, but across the entire retina, even whether it's the regions that are white or gray in the upper left-hand corner, we see high level of expression of the green fluorescent protein across multiple layers within the retina. So based on this, we've actually filed two INDs to the FDA. These are now open INDs where we have authorization to initiate human gene therapy clinical trials for X-linked retinitis pigmentosa as well as for choroideremia. When things calm down around the world, we're going to be initiating those human clinical trials. In addition, our company 4D Molecular Therapeutics has extended this evolution approach where our pool of variants now is 1 billion, so 37 different libraries, 1 billion different AAVs. And we've done selections not just in the retina, but in the central nervous system, lung, liver, heart, et cetera. So actually 14 selections to date that have resulted in 300 novel variants of AAV that we've created. And the next target tissue that we're going to be clinically pursuing is actually muscle, both skeletal and heart muscle. And we're soon going to be this month filing an IND to be able to initiate clinical trials to treat Fab rays. So I'd like to talk about a different mode of delivery for gene delivery in the central nervous system, and it's actually a different mechanism of spread of virus into this tissue, a process called retrograde transport. So whether the tissue is the eye, or the brain, or the heart, we have some fundamental problems, which is that these are relatively large tissues and can be very delicate surgically., and it's difficult to get AAV to spread to a large amount of that tissue. So one of the clinical trials that illustrated within the brain this concept or this problem was to treat this lysosomal storage disease called Sanfilippo B, and to be able to try to infuse AAV through a majority of the volume of the brain, they actually drilled 8 holes into the skull and did 16 AAV injections and still didn't restore high enough levels of enzymes to be able to modify the disease in that condition. We co-discovered a mechanism for AAV transport within the brain called retrograde transport. So specifically, we know neurons have a cell body where the nucleus is, and then they extend, extend axons, in some cases millimeters or many centimeters inside the body. And we found that if you deposit an AAV right at the end of one of these axons, the axon can actually internalize the virus and traffic it back retrogradely along that axon for very long distances in some cases, infect remotely, if you will. And this was utilized subsequently by Brian Kaspar in a study back in 2003 in which he injected AAV into the muscle of a mouse, and the virus was internalized by the axons of motor neurons, and the motor neurons then retrogradely trafficked the AAV back to the spinal cord. And this is a way of being able to mediate high efficiency, or relatively high efficiency, delivery to the spinal cord of that mouse after a simple intramuscular injection, sneaking the virus brain barrier. And use this to mediate delivery of a protective factor into the central nervous system of a mouse model of amyotrophic lateral sclerosis in this extended animal lifespan. So a company began to develop this retrograde transport technology and unfortunately found out that the efficiency of retrograde transport was too low. It worked reasonably well within a mouse, less well within a rat, and not well at all within a non-human primate. And so it became too— it should scale up for human clinical use, even though this could be a very effective approach to be able to get spread within the central nervous system, as I'll show you in a couple of examples. So we began to conduct AAV evolution for retrograde transport, and the system that we utilized is shown right here. This was initially done within a mouse, and the cortex of these animals sends projections over a long distance to a centralized region called the pons. So the idea is that you could do an injection of the library inside the pons, and some cortical neurons would would end up internalizing that virus and retrogradely transporting it back to the cortex. And as a result, with a single focal injection, you could get the virus to spread through a large volume within the cortex of the brain. So we repeated this evolution multiple times and ended up with one dominant variant emerging. We then loaded it with a transgene that would end up resulting in expression of tdTomato. And you can see a single injection into the pons resulted in very widespread expression across across the brain. So we wanted to quantify the level of increase in this retrograde transport, so we ended up loading several different AAVs with a cargo that resulted in expression of GFP just within the nucleus. This is tdTomato across the cytosol of cells, so it's kind of hard to count individual neurons. This was a nuclearly localized cargo, so we can see that upon injection into the pons in the lower left-hand side, you should get expression across the cortex. In the upper right-hand side, there's none in the case of AAV2. In the case of this particular variant called DJ, this came out of Stanford, you can see at least in the bottom panel that there's some neurons in the upper right-hand side in the cortex expressing GFP, but it's not comparable to this variant, the AAV2 retro that we created, where you can see widespread transduction across those neurons. When we quantify the efficiency of delivery, you can see that compared to AAV1, 2, 8, and 9, we have 50 to 100-fold higher levels of retrograde transport, and it compares quite well to a number of other approaches for retrograde transport, including even this fluorescent dye in an organic molecule called fluorogold, which is being utilized for retrograde transport labeling. We then wanted to see strategically with a couple of different injections how we could maximize spread across the brain., and you can see here if we do 3 different injections with 3 different AAVs of 3 different colors, red, green, and blue, into the superior colliculus, thalamus, and pons respectively, just these 3 focal injections gets you widespread expression across large volumes inside the brain. And we've since utilized this to effect genome editing within the brain, as described in that article on the bottom right-hand side. So I'd like to move on quickly to the final chapter of the work I wanted to describe to you today, which is going to be some work that carries on the same concept of solving problems through libraries and evolution, but now applies it towards a different problem, which is AAV production. So there's been, as gene therapy has been increasingly successful, there's been a bottleneck within the field related to AAV manufacturing, as shown here in this relatively old article from the New York Times. And this problem is illustrated here. So AAV serotype 9, which is the serotype that's used for that FDA-approved product to be able to treat spinal muscular atrophy. So from around a 500-meter squared bioreactor, you can create on the order of 10^16 viral particles. So you can then translate in this into how much surface area you need to be able to treat individual patients who are suffering from certain conditions. So in the middle here, we have one SMA patient, and they typically treat, this is Novartis AveXis, they treat patients typically on the order of weeks of life. So they're treating young, young patients just postnatally. And as a result, you don't need as much virus because simply, you know, young patients aren't as big yet. So even so, you need around 50 square meters to treat a patient. And AAV9 is also in clinical trials to be able to treat Duchenne muscular dystrophy, And in that case, you need 500 meters squared. In other words, in some cases, a single manufacturing run per DMD patient treated. So how does this translate to the incidence and prevalent populations? You need around 20,000 meters squared to treat the annual incidence of spinal muscular atrophy type 1. And you need around 7.5 million square meters to be able to treat the prevalent Duchenne muscular dystrophy population. So this is actually hundreds of, around the order of 1,000 soccer ball fields worth of surface area. So they're quite simply, it's going to be challenging to have enough manufacturing capacity to treat these conditions. So one way in which you could ease the manufacturing burden is by creating an AAV that's much more targeted and efficient at being able to reach a tissue. So let's say you created a directed evolution AAV that's 100-fold better at getting there. Well, that's going to alleviate the manufacturing burden burden, but another way in which you could do it is to actually engineer the manufacturing cells to pump out more AV, and we began to do that using a— based upon— investigators within the field have endomide expression perturbations where in some cases they're actually using a version of Cas9 fused to transcriptional activator system. And then as a result, guide RNAs will guide this transcriptional activator, this effectively synthetic transcription factor, to, you know, depending upon the guide, many different locations of the genome. And so you can develop genome-wide activation libraries. Or you can use Cas9 that has an active nuclease and using a guide RNA target that Cas9 towards, let's say, the first exon of every gene within the human genome and generate genome-wide knockout libraries. And so these guide RNA libraries are what direct these different systems, either a transcriptional activator or a knockout system, towards particular regions within the genome. So we ended up embedding these guide libraries into AAV vectors so that we could select for guides, or specifically for genes whose overexpression or whose knockout could increase AAV production. And we did this within 293 cells, which is the dominant manufacturing cell line to produce AAVs. So the rationale for this was that just like we'd anticipate that nature has never evolutionarily rewarded an AAV capsid for the ability to reach a photoreceptor neuron deep inside the eye, you know, just, it's not part of the AAV life cycle, likely nature has never evolutionarily rewarded a 293 cell, which is a a transformed cell line that was derived from embryonic kidney, human embryonic kidney, is never evolutionarily rewarded for the ability to pump out AAV. So it was the opportunity to engineer the cell line to be able to better do that. So the way that we did this is that we created libraries of guide RNAs inside of AAV where we cloned in all 70,000 guides present within this initial system, which contains, the 70,000 contains 3 guides per gene. For each of 23,000 genes within the human genome, or the GEKO, the knockout library, contains 3 guides for each of 19,000 genes, so 58,000 guide library. So we cloned those inside of AAV and packaged it along with the RepCap helper and an adenoviral helper so that we created libraries of AAV vectors where every single variant on the right-hand side has the same capsid gene in this case, but contains a different guide RNA inside of it. We can then take these and infect cells with them at a low ratio of virus to cell. So each cell took up only one virus, and as a result expresses only one guide RNA. And these cells now have the CRISPR-Cas system, which can take the guides and either overexpress or knock out a particular gene, depending upon the nature of the AAV library. So we then gave this cell around 48 hours to perturb gene expression through that overexpression or that knockout. And then transfected in the helpers so that we could then use that cell to manufacture AAV. And the logic is that if a particular overexpression of a given gene in the human genome or a knockout of a given gene in the human genome ended up changing the cell in such a way that it could better produce AAV, then that cell is going to pump out more AAV. And just to remind you, that AAV genome is inside of it, the barcode, the information that tells us which gene perturbation resulted in the better production of AAV. So we can then harvest that library, use it to reinfect the cells, repackage, and as time goes on, the AAV variants that, or the guide RNAs, if you will, inside those AAVs that result in better packaging are going to get exponentially amplified within the population, and at the end, We can then do sequencing, next-gen Illumina sequencing, to sequence the guides and tell us which genes whose perturbation would result in an increase in AAV production. So we have a number of different hits, and two particularly interesting hits emerged from the overexpression library. One is called ITP-RIP, which is inositol triphosphate receptor interacting protein. So this modulates calcium release inside of cells in response to, to signals that release, that result in ITP production. And SCOT2, spindle and kinetochore-associated protein type 2, which is involved. So we took cells, 293 cells, and overexpressed these two proteins initially by overexpressing the guide and Cas9 system. And then we did it by making DNA encoding these proteins and overexpressing them inside of cells. And we found out that upon overexpression of these proteins, we get around a 2 to 3-fold increase increase AAV production from these cells, and this is both at the level of vector genomes produced as well as infectious units produced, and furthermore there was the possibility that these two genes could be acting independently within those cells and therefore by combining them we could get synergistic increases, and we see that upon overexpressing both cells we get around a 4 to 5 fold increase in vector production, and furthermore this was not limited to AAV2 but extended to other AAV serotypes such as AAV6 which is notorious for being a particularly difficult AAV to produce, and we see a 2 to 4, around a 3 to 4 increase in both viral titer, genome titer, as well as infectious unit titer. So we have since been investigating mechanisms by which overexpression of these genes ends up resulting in better AAV production, and we're seeing evidence that these genes actually perturb cell cycle in a way that is advantageous for AAV production. And we look forward to going into detail into that mechanism soon. But I'd like to close now with a few bullet points from the work that I have presented. The first is that gene therapy is working. In addition, recombinant safe viruses like AAV can represent very potent therapeutic pharmaceutical gene delivery vehicles, but of course nature didn't optimize them for use in human therapies, and therefore it's up to us to reoptimize these viruses to be safe, effective, efficient therapies. So these particles are very complex products of millions of years of evolution, and it's very difficult to rationally engineer changes into them. However, directed evolution ends up being a very powerful search algorithm where we can search billions of potential viral sequences to be able to isolate the variants that have highly optimized properties for delivery in principle to any cell or tissue target through an optimal route of administration.. And we use this approach to create AAVs that are highly efficient, can target delivery, or in work that I didn't talk about today, can also get past that preexisting antibody problem within the human population. So one of the AAV variants created has already been in the clinic for almost a year and a half, and several additional variants are going to be entering into the clinic this year. So I'd like to acknowledge the people who actually did the work. The work within the retina was the product This is mouse retina, the product of Leah Byrne, Denise Dahlkhara, and Tim Day, in collaboration with John Flannery at UC Berkeley. The work in the dog retina was the product of, again, Leah Byrne at UC Berkeley and John Flannery, in collaboration with Welling Beltran, University of Pennsylvania. The work in the nonhuman primate retina came out of 4D Molecular Therapeutics, and I especially like to call out David Kern, my co-founder, and our CEO. And the work on retrograde transport was a product of Tom Guy and Bumyul Wong in collaboration with investigators at Genelia Farms. And finally, the work on the development of better AV manufacturing cell lines was the product of Chris Barnes and David Ojala. So thank you so much, and thank you for your time in listening to our work.