Scientist Stories: James Dahlman, CRISPR Delivery & In Vivo DNA Barcoding in Nanoparticle Selection

So, so thank you for coming to TriLink Spotlight lunch presentation. So it's my distinct pleasure to introduce our two speakers, two up-and-coming speakers in the field of nucleic acids. So we have The two talks today, the first one is by James Dahlman at Georgia Tech. That will be— is entitled Advantages of In Vivo DNA Barcoding in Nanoparticle Selection: Quantifying How Thousands of Nanoparticles Deliver Nucleic Acid In Vivo Using DNA Barcodes. So, with that, I'll hand it over to James. Good afternoon, everybody. I hope you guys can all hear me. First, I want to thank TriLink and the organizers for giving me the opportunity to share some of the work that we've been doing in my lab. And the first time I gave this talk, I was at a company, and somebody— the very first question I ever got about this material was, I think you misspelled your title slide. And I understood why, because this title says Testing Thousands of Nanoparticles in Vivo. Using DNA barcodes, and I think we're all accustomed to testing thousands of nanoparticles in vitro. But what I'm hoping to convince you today is that we can actually now test thousands of nanoparticles in vivo. My lab to date has tested about 2,000, and I've only been a faculty member for about 18 months. So to set the stage a little bit, my lab focuses on non-liver delivery. The reasoning for this, I think, after hearing the on-island presentation today hopefully is relatively obvious, which is that hepatocyte RNA therapies are working really robustly in patients. This is obviously a modified slide, and it's kind of hijacked from on-island, but this is the slide that I use to show just how exciting the RNA therapeutics field is right now. I won't go over this for too long, but this is patient data, which number one, I think it's really cool that the field's starting to publish you know, not just in Nature and Nature Biotech, but actually starting to get publications in New England Journal of Medicine. That's a really fun place to be. But the other thing I want to point out here is that you— this is days after injections, and you can now get, you know, silencing for months. As Alnalan suggested, you may be on a quarterly schedule with a subcutaneous injection for hepatocyte siRNA silencing. So this stuff is really starting to move into the clinic. But again, all these things are— I don't want to say limited to because that's the wrong word— but all these things are focused on hepatocyte They still have all these clinical programs. Cas9 is the new kid on the block. Cpf1 is there. Cas13a is there. Cas13d is coming along. There are a lot of these therapies, but if you kind of look at what's coming down the pipeline, the first two papers for systemically administered gene editing with a particle were both focused on hepatocytes and published by Dan Siegwart's lab at UT Southwestern, then Yi Zhao Dong's lab at Ohio State. Since that time, there have been two very impressive publications, one from Intellia that came out in Cell Reports, and the one from Dan Anderson's lab, of which I'm an alumni, in Nature Biotechnology. But again, in all four of these cases, not a limitation, but they've all been focused, as many of our colleagues may predict, on liver delivery. So my background delivery is in endothelial delivery, and specifically on delivery to the heart and lung. And if any of you were in the presentation for Professor Langer today, I came out of the Langer lab and some of that endothelial work was highlighted by him a little earlier. I'm not going to rehash that, but I do want to just highlight one thing, which is that the nanoparticle that I developed as a PhD student with the help of Carmen Barnes at Alnylam called 7C1 has been extensively validated. So since 2014, that particle has effectively delivered siRNAs to endothelial cells in 20 labs across the United States. I can now say that this thing has silenced target gene expression in endothelial cells by at least 80% in non-human primates at tolerable doses, and it is being considered for clinical development. So we have a little bit of history developing these non-liver vehicles. However, the stuff I'm going to be talking about today, to be honest with you, excites me way more. And this is stuff we're doing at Georgia Tech. And the way I kind of characterize my work at Georgia Tech is that our lab focuses on what I call technology development for nanomedicine. And this marries my background. I did my PhD in the Langer lab where I learned high-throughput chemistry and focused on RNA delivery. But right before I graduated, I started working with this young gun who was at that time 31 or 32 years old at the Broad Institute named Feng Zhang. And this is before Feng got trendy, so I think I'm a hipster member of his lab. But in Professor Zhang's lab, I really took away one thing. Which is how to use molecular biology and how to use big data. So I'm a chemical engineer that basically learned how to run PCRs and learned how to design primers the proper way. And a lot of that interface, the interface between big data and nanotech and delivery, I think is a very exciting space. And my lab's really trying to push this boundary a little bit, so we call this technology development for nanomedicine. This is the most important slide in the deck by far. I'm gonna embarrass Corey. Corey, please raise your hand. Corey's right there, and he has been a lead scientist in my lab since I started in basically fall of 2016 doing this work. And a lot of the work that you're gonna be seeing today has been driven by Corey, Kalina, and Melissa. We have excellent students in the Georgia Tech BME department, just world-class students. Okay, so what have we been doing? Well, a few weeks ago I was asked to go to the NIH for a small meeting about gene editing, and I was asked to to think about 4 fundamental problems that we all face as drug delivery people. So whether you're a polymer person or you're doing CRISPR or you're doing siRNA or whatever, or ASOs, I think these, these are universal problems. They are that it's hard to and expensive to measure delivery in vivo, that we still don't know the biology of delivery. So I'm not saying that there's a master regulator, but if there is a Myc or p53 for drug delivery, we don't know if it exists, and if it exists, which gene it is. The biology delivery is unknown. That limits, I think, all of us to these observational experiments devoid of any really strong biological hypotheses. That also makes it hard to relate a mouse— delivering a mouse to a rat to a primate to a human because we don't know which genes actually matter. This is a really big inefficiency in industry, as you guys are probably well aware. The other thing I want to point out is that because We don't know how to screen or test these things in vivo. It's very hard to design nanoparticles a priori. So we're stuck kind of throwing things at the wall, and unfortunately almost nothing sticks, and everything that does stick goes to the liver. Okay, so enough complaining. What are we trying to do about it? It's pretty simple. We're trying to develop ways to study thousands of nanoparticles directly in vivo, and some of the datasets I'm going to present are kind of exciting. Today, like I said, we've done about 2,000 particles in a typical experiment using FACS. My lab has gotten pretty good at FACS. We will, you know, put a few hundred particles into a mouse and then FACS sort 20 to 30 cell types, and you get these very large datasets all in vivo. Okay, so the one point I want to make here is that nanoparticle libraries are really easy to synthesize. I'm okay at chemistry. People like Dan Siegwart at UT Southwestern, people like Yi Zhao Dong at Ohio State, Chris Alabi at Cornell. I could keep naming— Jordan Green at Hopkins. I can keep naming people who are great at chemistry, and I think that we're all aware that if you're a nanoparticle person, if you want to make a big nanoparticle that looks like this or a small nanoparticle that looks like that, you can do it. As Professor Langer alluded to earlier today, you can make these diverse libraries. This is just a cartoon of a library. This is 600 different compounds I synthesized. As a PhD student, each vial is a different compound. So this is one chemical structure, that's another chemical structure. Fun digression, when you're in the Langmuir lab, it's pretty big. And as a PhD student, I synthesized all these things by hand, and then I went and presented this in lab meeting, and it turned out that there was a robot down the lab, but it was in another room I'd never been to, so I didn't know at that time. But I got really good at weighing out amines, I promise. Okay, once we make these diverse libraries, we all do the same thing. Whether you're doing polymer chemistry, whether you're doing ATRP, whether you're doing click, whether you're doing lipidoid stuff, it doesn't matter. We make these big diverse libraries and then we say, okay, we have 10,000 nanoparticle candidates, so let's pick one. And the way you do it is the way you're forced to do it, which is you test it in vitro. If you're lucky, you might be able to do it in primary cells, but realistically most people do it in some immortalized cell line because that's the only thing you can grow to scale. And so you say, okay, I want to deliver something to the heart or the lung, so I'm gonna try to predict this. And the way I'm gonna do is I'm gonna stick some A549 cells in a dish and see if this thing predicts that thing. The problem is that the hurdles that prevent delivery from occurring in vivo— liver, kidney, spleen, immune system, complement, pulsatile blood flow— differential vascular structure. You don't have that in this dish. It makes it very difficult to predict. Okay, so let's say we have 1,000 particles. We want to test them all in vivo. You have two options. One is do a 4,000 mouse experiment. The other one is test a bunch of particles per mouse. And as an academic, although we're happy about the NIH budget increase, we still don't have the money to do 4,000 mice. So we test them all in one mouse, and this schematic is really simple to understand. You know, you take the, the siRNA or the nucleic acid that you were going to deliver therapeutically, and you replace it with a DNA barcode. However, I can attest, and if Kevin Kaufman is in the room— Kevin Kaufman, yes, Kevin's right there. Kevin and I can both attest that that this is about 2.5 years of very frustrating engineering to get this to be practical and for this to actually work. So I like to say that this is the easiest schematic to understand but the hardest to actually perform. What you can do is if you have access to high-throughput microfluidics, then you can, you know, make nanoparticle 1, which would be this yellow nanoparticle, to carry barcode 1. Nanoparticle 2 carries barcode 2. Nanoparticle N carries barcode N. My lab is doing a few hundred a day now. We have a big quality control here, so this arrow is actually also kind of complicated, but we had to develop the engineering to make sure that we're only putting in good particles. So some of the particles get cold in this arrow, but you take the stable particles, you administer them all simultaneously to whatever you'd like, whether it's a cell line, primary cell, or a mouse, or whatever you want, and then you isolate the cell types of interest, and then you deep sequence everything. To simultaneously measure where everything went in a single animal. The data that we generate— and this is what you're going to be seeing on the y-axis— what we call normalized delivery. Here's just a very simple cartoon. Let's pretend we had 3 barcodes that were delivered, and we took, let's say, sample 1, and we got 2 gray barcodes, 2 green barcodes, and 2 red barcodes out of it. Then the normalized delivery for all 3 particles would be 33, 33, 33. And if in sample 2 you get 4 greens, 1 red, and 1 gray, then you get this normalized delivery. This is just analogous to counts per million and RNA-seq experiments for those RNA-seq people in the audience. So we're not doing anything fancy here. I'm about to show you a heat map. It occurred to me the other day that these heat maps are not obvious. I just want to have a very simple cartoon of the data you're about to see. So in this heat map, basically red means high normalized delivery and low means— blue means low normalized delivery. So in this cartoon, we would say that LNP1 beat the other LNPs in cell type A. So this would be— sorry, this would be LP1 here. And it— I'm sorry, it was the other way.

Whatever.

Blue is good and red is bad. So if you see a blue dot, then basically the particle did well. And if you see a red dot, then the particle didn't do so well. Okay, here are the data we generate. I just want to highlight— this is one experiment. I just want to highlight the fact that this is— each row here is a particle and each column is a cell type. Or tissue type. I also want to highlight the fact that typically when you publish a delivery paper— I've been in delivery space since I started in Bob's lab and Professor Langer's lab. I started my PhD in 2009, and I've been delivery person since that time. This is the data that you need to typically publish a paper. You know, you have two particles, you test a bunch of tissues, and you have your paper. And this is what we're able to generate in a single experiment. The DNA barcodes are rationally designed, so I just want to highlight one or two things here. When people think about a DNA barcode, they think, oh, every nucleotide is different, and that's actually the opposite of what you want. If every nucleotide is different, then you don't have a way to perform this experiment. The barcode part of the barcode is actually in the middle. So if you have 8 nucleotides here, that gives you 65,536 combinations, potential barcodes. You don't need to have a big long DNA barcode. In fact, what you need is you need a universal site here and a universal site here so that you get unbiased PCR amplification. So if you run into a DNA barcoding technology that does not have universal primer sites, run away, because you can't guarantee that the PCR amplification is going to be even. Your data are going to be biased and everything's going to come out dirty. We've actually improved upon this barcode design pretty significantly, and I'll talk about that in just a second. We run a bunch of experiments. Actually, the paper I published in PNAS with Kevin is a co-first author on that paper, was dedicated solely to control experiments. If you want to see a paper where there's nothing interesting, it's all control experiments, please, you can characterize our system this way. But this is just one very easy control experiment to sort of explain quickly. In this case, what you see is normalized delivery for something like 104 particles, I think. So every barcode is delivered by 104 particles. Those are all the black dots, and then we spike in a naked barcode., which shouldn't be delivered as efficiently. And when we sequence, sure enough, the naked barcode is the only thing that drops out. And we can actually do this across, you know, 20 different samples, and the naked barcode falls out, falls out, falls out, falls out. So that's one control. The other thing I want to point out, another control that's going to be important for— you'll see why it's important in just a minute. Make sure I'm not going over— is whether or not one system will predict another system. So what you're seeing is normalized delivery in one cell type in vitro. This is an immortalized aortic endothelial cell.

It doesn't matter.

This is an endothelial cell, cell line. And what you see is normalized delivery for 100 particles when the total DNA dose in vitro was 20 nanograms. So this is 0.2 nanograms per particle. And when you plot that against in vitro delivery, same cell line, same day, same particles, same everything, but just a slightly higher dose, 100 nanograms total, you see exactly what you would expect, which is that it's exactly predictive. So the particles that do really well at 20 nanograms also do really well at 100 nanograms.

That's this guy.

The particles that do really poorly in the 20 nanogram dose also do really poorly in the 100 nanogram dose, just like you would expect. That's this guy. And then this is actually the naked barcode. Okay, so this gives us confidence to say that, you know, does in vitro delivery predict in vitro delivery? And our assay says yes, which gives us confidence to ask the question, does in vitro delivery predict in vivo delivery? The way we did this is, I think, is a pretty simple experiment. We don't do any experiments that are overtly complicated in the lab. We try to keep it pretty simple. But in this case, we had a barcoded nanoparticle library. We did several of these experiments, but here's the workflow. You take your barcode and nanoparticles, you make them all on the same day, you administer— everything gets made on the same day and administered the same day. So if you're thinking in the audience, well, what if the particles are mixing? Even if that's the case, And we've done experiments to show that that probably isn't the case. But even that is the case, the cells and the animals are still getting the same mixture, even if that is the case. So you take 100 and so particles at a time at this time, and then you administer them either to an endothelial or macrophage cell line, a bunch of primary endothelial cells, or you administer them to a mouse. And then you will FACS sort the endothelial cells from that mouse at the same time point. We also FACS sorted macrophages, and then you compare them. Simple. So these are the data. This is delivery in vitro in endothelial cells, immortalized aortic endothelial cells. So this isn't even cheating. This isn't like a HeLa cell line. This is freshly isolated heart endothelial cell plotted against delivery for heart endothelial cells in the mouse. Again, same particles, same day, same everything. This is a collection of 3 or 4 experiments totaling about 400 nanoparticles. We don't cherry-pick data in the lab. I was actually terrified that this was the result. You can imagine this is the first paper I published as an assistant professor. I was terrified by this. I thought I'd get kicked out of the academy, so to speak. But we kept seeing and kept seeing and kept seeing it. So I would encourage you to read Colleen and Corey's paper. This is Figure 2A through F of that paper where we did a bunch of different in vivo cell types to be rigorous. And this is the endothelial stuff. This is the macrophage stuff. And then if you go into the supplement, we varied time point, condition, I mean everything. Our supplementary figure 2 goes A through double Z or something, and I think some of the panels have like 9 subpanels within a panel. So we've done this a lot and we see the same thing every single time. So I think I've convinced myself at least that in vitro delivery does not predict systemic in vivo delivery. I want to correct Nature Biotech. After about an hour after our paper, the preprint came out, Nature Biotech tweeted this thing. I'm just gonna put an asterisk here. They tweeted, in vitro systems are a poor predictor of in vivo efficiency of nanoparticle-mediated nucleic acid delivery. And I think that that can be true for systemic delivery, but I want to put a giant asterisk on this. And the asterisk is this: if you have a process where the particles are going to fail via splenic, renal, liver, liver, and immune clearance, and your cell culture plate doesn't have those things, then this statement is probably true. But if you have a condition like in cystic fibrosis, where the limiting factor is this impermeable rubber-like mucus, and you can mimic that in a plate like Justin Haynes and Mark Saltzman do, then this statement may not be true. So the question I would just encourage us as a field to ask is pretty simple: if X prevents the nanoparticle from working, does the system actually model X? In some cases, like systemic delivery, the answer is probably going to be no. In other cases, like CF, the answer may be yes, and your own systems may vary somewhere between that spectrum. The other thing I want to point out, I can't get into the barcode design because I don't want to run over, but we have redesigned the barcode so that we can get highly sensitive readouts. So we can now read out delivery in a very linear way. This is in vitro, but in vivo we can also improve sensitivity quite a bit at 10 to the negative 20th grams per well in a dish. So this is between a billion and a trillion times more sensitive than fluorescence. The fact that we have much more sensitivity then fluorescence does matter in vivo. And again, I don't have time to get into that story, but I promise you it does, and we're trying to get that story out right now. But you can really think hard about the types of, you can think hard about how your DNA is designed and get sensitivity like this. So I would encourage you guys, when this paper comes out, give it a glance, 'cause we're really excited by these data.

We call this thing QUANT.

Yep. Last thing I want to say is that, you know, I don't have time to go through all— we're trying to work on all these problems, but I don't have time, so I just want to highlight this one. It's hard to design nanoparticles a priori, and I just want to say that although we don't have time to show a bunch of data because they're pretty fresh, this is the workflow that we use. It makes a bunch of particles, we screen them directly in vivo, we select the winners, and then we iterate. And using this system with another one of Corey's papers, we have identified nanoparticles that that when you administer them systemically, preferentially edit genes outside the liver more so than they do hepatocytes, which is pretty cool. So just to wrap it up, my lab does technology development for nanomedicine. I really do feel like there's a bunch of space at the intersection between big data and omics and nanotech and delivery, and I'm really optimistic that we're going to be able to combine these things to do some pretty exciting stuff. Then finally, I just want to re-highlight the students here who have done the majority of this work and reiterate just how lucky we are to have amazing students in the BME program at Georgia Tech. Then before I go, I just want to highlight one thing. I do want to highlight one thing I did. I had a 1-year postdoc with Feng, Professor Zhang, but we did get a big paper out of it and I just want to highlight this. If you guys are doing any signaling stuff, this is unrelated to delivery whatsoever. But if you're doing signaling stuff and you need to turn off one gene and turn off another, turn off one gene and turn on another gene in the same well, so do orthogonal regulation, you can actually turn on a gene with an active Cas9 by fiddling around with the guide RNA. So right now you turn on a gene with dead Cas9, you can do it with an active Cas9, and because of that you can put in a regular old Cas9, a regular old guide to A, and cut A, and then put in that same well you have regular old Cas9, And then what we call our dead guide to B and get the gene turned on with B. And so you can do gene A and B like this in the same well really easily. And our results were recently recapitulated by a group in a Cell paper that came out this year in vivo using the Cas9 mouse. So with that, I'll take any questions. I'd like to highlight again Cory, Kalina, and Melissa, and thank our sponsors who have allowed us to do this. And of course, please feel free to contact us with any questions. We do work with industry., and we're big into translational stuff, so we want to see our stuff reach the clinic, and we know that we need industrial partners to get us there. So thank you very much for your time. Yes.

So you showed that the in vitro at one concentration versus another concentration had good correlation, and you didn't get good correlation between in vitro and in vivo. What about just a repeat of the in vivo versus in vivo?

Do you get good correlation there? Yeah, so we have done that experiment, and I would say that the top particles perform as the top particles. So when we did, we did a 120-particle experiment, and then we did exactly the same 120-particle experiment on a different day. And if you rank the top 10 particles, I think 6 out of the top 10 were shared. Now when you get to the bottom, that's not within the sensitivity assay, and so you're going to get scatter, but we're not overtly concerned with that. The other thing I'd like to point out is that I do think the sensitivity assay outpaces our ability to consistently make particle formulations that are the same, the same day in and day out. So any time where if we were to see a situation where you do an experiment on Monday and you do an experiment 2 weeks on a different Monday and it's not the same, To be honest with you, my guess is that that actually wouldn't be the system. It's because you have particles that you're making differently. I think all of us have gone through colloidal systems and seen variability day to day. And so the one thing that can be lost here is if you make 300 particles, you don't have 300 optimized particles.

And so that's an important asterisk, but that's how we're thinking about it right now.

Yes.

So right now your endpoint is uptake? Yes. What is your approximate relationship between successful uptake and successful knockdown?

We all know we can get uptake.

Yeah, sure. But we don't know that we can get release from the endosome and get knockdown.

Yeah, so I didn't have to go into that, but I've been obsessing about this for years. And we have a solution to that. Yes. So we have a screen that I don't have time to talk about today that doesn't read out distribution.

It reads out what you're describing. And that's under review. I'm sorry, in submission. For your in vivo screen, do you have a control for the nanoparticle concentrations in your blood?

We haven't done blood stuff. We could, we just haven't. But I should say that all the data that you see are always normalized to the input. So we take, you know, we make a vial of it, we inject most of that vial into the mouse, but we save some, and we sequence what we save so that way we're not just assuming that every barcode is being administered evenly. And what we found is that You have to do that. And that's, if any of you have done gecko stuff or whole genome screening stuff, that's not going to be a surprise to you at all. It's the same problem. You get this distribution of input that you always get. One last question. Yes, Ken Woo.

[Speaker:KEN_WOO] Oh yeah, really beautiful talk. I really enjoyed it. And my question is related to time dependence because we test like thousands of nanoparticles in our system and do one particle at a time.— that works the best 1 hour after injection? Is it the best one that you see work the best like 3 days after injection? And is there any correlation there, or it fluctuates?

Yeah, so I'll say that in the PNAS paper, we did that, and we saw some fluctuation. But because these are all normalized delivery, because they're all being plotted relative to one another, it's impossible to interpret. The broader question of which time point would you want to choose to actually be predictive is something that we have to figure out. And the size and scale of that in vivo experiment would be brutal because then you have to do 100 individuals. So I hear you loud and clear, and it's something that I wonder about as well. Thank you.