Founder Stories: Harlan Robins, Adpative Biotechnologies & ClonoSEQ

Thank you. Thanks for the invitation. Thanks, David. It's an easy place to start when we've had 3 talks ahead that have described a bit of the background. So first, I'm going to tell you a little bit about Adaptive Biotechnologies, and then I'll talk about the questions that really need to be addressed, in my opinion, in order for high-throughput sequencing to sort of replace and/or supplement flow cytometry in widespread. So first, Adaptive Biotechnology was spun out of Fred Hutch in 2009. It was created in my lab. The technology was created in my lab and two other colleagues as well. In April, we took in a $120 million investment. And we've basically doubled in size every year since 2009. So by the end of this year, we'll have 100 employees. And we'll have 200 by the end of next year. We've sequenced over 100,000 T or B cell repertoires at this point. We have 80 publications and counting. There's at least one a week coming out with Adaptive Biotechnologies data. And in terms of— I don't have that much time to talk here, but we have about 30 ongoing studies in the MRD space. So I'll just give you a few anecdotes so that you can get— try to focus the flavor on the things that are, I think, instructive as opposed comprehensive. Okay, so there's, there's some burdens that we have as introducing a new technology. As we've already heard from Steve Hunger, this is— MRD is, is standard of care in ALL. So if we want to, if we want to replace the flow cytometry standard of care, we have some burdens on us. And first, we have to prove that our assay works quantitatively, and I'll tell you what that means in a second. And second, we need lab-to-lab standardization. I know, I mean, that was another topic already discussed, and I think that's vital to going forward, and I'll give you some data on that. And then, of course, biological utility. Are the differences between high-throughput sequencing and flow, are they clinically meaningful, and is this something that we can convince payers to pay for? Okay, so, of course, there's other considerations: speed of turnaround, throughput, cost. But all of these are, I consider these logistic. I mean, we're, The inherent cost of sequencing is extremely low. The throughput— it's high-throughput sequencing— I mean, it's extremely high. And the speed of turnaround is really improving very, very rapidly. And it's already, you know, these MiSeq machines can run, you know, in a few hours. I mean, there's no— I don't think these are inherent burdens. These are just maybe present stopgaps. So— but let's get into what I consider to be the bigger issues. And one is quantitative accuracy. And so we're looking at BALL, for example. And BALL, the receptor that we're using to track the B-cell receptor, this receptor is a neutral marker. It has nothing to do with the disease burden. It just happened to be the one that got rearranged. Well, there's a really large number of ways you can create a B-cell receptor. So we're tracking something in a given cancer, and what we're tracking is different in every single cancer. So in order for us to have a viable assay, we need to prove to you that it doesn't matter what cancer that you had, what that rearrangement happened to be, that I'm going to give you the same answer. And so I'll tell you how we did that first, okay? So, and what I mean by that— how do I point here? Does that work? Okay, yeah. So just as a schematic, you get a— and this is a schematic of T-cell receptor beta locus. One of these Vs of 54 or so, one D and one J get rearranged next to each other and to create this T-cell receptor gene. And since you have different Js and different Vs, there's no viable technology that can literally use the same, say, primers to amplify this V and J. So in order to do that, we use a multiplex set of primers. We have a set of forward primers and a set of reverse primers, which means that we have this pool of primers in the same tube. But we don't know a priori what the receptor is. So only one of those is going to be used to amplify the clone. So we need to prove to you that it doesn't matter what it is, that we've balanced this thing in the right way to get the right answer no matter what. OK, so how did we do that? What we did is we actually created a full synthetic immune system with every possible VJ combination.— and we do this for all our assays— such that we can independently measure how much of each of those receptors there are. And then, which is, you know, we take that set of templates, this synthetic immune system, and we hit it with our assay. And of course, since there's no good way— we're not good enough to design primers perfectly— we get this mess. In other words, we get huge amplification bias. But since we know what the amplification bias is, we can correct it. And what we did in our process is this rounds of iteration to improve our assay. We modify primers that don't work, and we titrate primers that are either under- or over-amplified to the point where— it's kind of hard to see, but we're getting, instead of a 250-fold variation across any VJ combination, we're at about 20% across the whole range. And that is well within, I think, what would be considered viable clinically. So just as a— to give you an example, this is peripheral blood, and we just binned all of the— we took this whole whole set of T-cell receptors and bin them according to their V or J, just whatever V or J combination you had. And if you take peripheral blood before controlling for bias using our synthetic assay, you get somebody who— you get this huge clone that looks like these people have leukemia, right? You also get all these other guys that are basically dead. In other words, if your leukemic clone was one of these guys, you're not even going to see it. So you get this huge variation. And what you don't want is the problem that It depends on that your assay only works with some subset of people quantitatively, with some subset of people who have leukemia. You want it to work with everybody. So we then fixed it, and now you get this repertoire that this is what the real repertoire of this person's blood looks like. You do have differences between— real differences between usage of different Vs and Js, but, you know, this is a huge difference between what we— before and after this quantitation. Okay, so that's problem 1 that we solved. And it turns out that the solution for problem 1 actually really helped us with Problem 2, which is lab-to-lab variation. How do you standardize this between labs? Because that's obviously the end result. And so what we did is we created a kit. And we created a kit that we have beta tested by sending it out to 9 different labs, including Johns Hopkins, Memorial Sloan Kettering, Vanderbilt, 9 different groups all over. And we sent them also— the same set of samples. We sent every lab the same set of samples. And the way we standardized the kit is that we use this synthetic template as our inline controls. It's actually spiked into every single sample. And what does that give you? Well, first we can— this is a comparison. So I'm going to pick just 2 labs that we sent this to, although I can tell you that we have this 9-by-9 grid. And all of them, every combination looks the same, right? Which is that you get almost no— you get very little variation in just— these are the synthetic templates. This is for T-cell receptor beta. And each dot is one of these synthetic templates. And you can see that you get very good correlation. But it's not perfect. The nice thing about it not being perfect is that we know what it is. So we can correct for— in that lab, in every given lab, we can correct exactly what we need to do to get the right answer. And so then when we compare a blood sample between these two labs, each dot here is a— is a T-cell receptor, we can then get as tight a quantitation as we get with between literally doing the same— a blood draw, splitting it in two, and running them both in our own lab. We get that tight a correlation between, you know, Johns Hopkins and University of Washington lab medicine running the samples. So I think that's a huge advantage of high-throughput sequencing is that effectively, in some sense, it's digital. It's sequencing. Standardizing it is much easier, in some sense, than standardizing, say, any array-based technologies. OK, so great. But then biologic utility. You've been hearing quite a bit about this. So our first standard of burden, I would say, is showing that we can reproduce the flow results. And I think David, as well as Malick and Aaron, have really showed quite a bit of data that we reproduce what flow gives. We almost never get a false negative, meaning that if flow finds residual disease, we find residual disease. There's this whole other case of where we find residual disease and flow doesn't that hasn't been fully fleshed out yet. But the first burden of proof is just to show we can reproduce what flow does. And I think that we've crossed that threshold. So I'm going to run you through one— clinical case. So, sorry, I should say, I'm going to talk then about, OK, clinical value of the increased sensitivity and additional information that we get. So what do we get in addition to flow? And I'm going to run through a clinical case. I should say, I'm not a clinician. I've never done this before. So if this comes off awkward, I apologize ahead of time. So this was a 23-year-old female who had already failed a couple of times. She actually had CART19 therapy and failed that, and then went through transplant. And her bone marrow, she was hypocellular pre-transplant. And we have this series of blood samples. OK? And we used a diagnostic sample to find— this is B-ALL, I should say. But she did not have a rearranged IGH receptor. But she did have a rearranged T-cell receptor gamma loci. I should tell you that— Every cell that has— or almost every cell that has rearranged gamma loci have both receptors rearranged, so we get to track both at the same time. And if your quantitation works right, you better get about the same answer for both, because there's one copy in each cell, right? This is a very good way of determining biologically whether your assay is quantitative. And this is about the error we expect, you know, 20%— not even— this is even less, but about that. So here's just the case of the sample. So I know it's a little hard to read, but just— All our samples are about 1 month apart, okay? So we're serially tracking this patient. And the interesting thing is— the things that I want to point out first is that this patient was hypocellular right before transplant, so flow didn't pick out the cancer, although we were readily able to identify the clone. You don't need many B cells in here to determine a clone. If you have 100 cells and 98 are one clone, it's a clone. It's pretty easy to figure out. OK, but then the next thing I want to point out is that we were actually tracking in blood, not in bone marrow. And I think the shift in what the increased sensitivity— I know Aaron already mentioned this a little bit, but I'm going to show you a specific example of how this applies. The ability to be able to track in blood allows things— first of all, obviously, you don't have to do a bone marrow, which saves the patient pain and money. But then in addition to that, you can also do serial tracking, which you can't really do in bone marrow, at least not at a 5 months in a row. And so what we found is that even at the first month after transplant, we found some very low-level residual disease. And we— as we followed this patient month after month— by the way, flow was done at all these points, and it's negative, um, at all these points— we start to see this person's, uh, disease burden just rise. And then finally, at the last time point, um, they clinically relapsed, and flow did find the, uh obviously found the disease as well. So, you know, unfortunately this woman died, although I should say that we got this great note from the doctor saying, look, we were able to tell this woman that her disease was coming back and the transplant hadn't worked at 4 months prior. And she ended up having— she got engaged. There was like a whole like bucket list of things. It was a great note to get. It made us feel really good about what we do, I think. And then, OK, so that's the case study I want to talk about. And then I'm going to talk about one other example. I only had about 15 minutes, so I'm going to just do one other quick example. And this also, I'm going to tie it into something that Erin was talking about, just so that we can go quickly. I don't have to set up the whole background. And that is that if you have— these ALL patients that are going into transplant, there's been evidence that having residual disease burden prior to transplant predicts outcome. Okay, I think in their study they didn't have enough patients to show this, but, um, I'll show you the data that we got. And just to explain what this is— part of a COG trial. It was, um, ages 1 through 21, so probably borders the, um, young and the young adult. Um, and then they all received an allo transplant, and they were in more morphological remission when they had the transplant. So, we had 41 patients that met the criteria for analysis. I should say that Adaptive was blinded to this analysis. This was done by the COG statistician. And we found 22 MRD-negative cases and 19 MRD-positive cases. And for clonaseq from Adaptive, none of the 22 patients that were negative relapsed, and 9 of the 19 who were positive relapsed. So, obviously, that's exciting, but as we were saying, the burden isn't against nothing. The burden's against FLOW. There is a standard out there, so we have to compare to that. So, how do we do that? So, in the same patient population, there was— FLOW was done as well, and you can see that, you know, about 20% of the FLOW-negative patients ended up relapsing, and none of the high-throughput sequencing negative patients relapsed. Patients relapse. And I think this really does show the sensitivity. It doesn't— it, it, it, in a, in maybe a different way than we were thinking about it, you know, in, in the first few talks, which is that it's not necessarily only the low-level MRD-positive patients— what do you do with those patients? But there's this other standard, which is, let's say, if, if we can show you're MRD-negative by, by high-throughput sequencing, that means that you don't even have a cancer cell at 1 cell in a million. You have very— if you have any cancer at all, you have very, very low amount of cancer. And when we find MRD negativity at that level, the number of patients that end up relapsing from— in, in, in across many, many different studies is extraordinarily low. So I really think that, that threshold— now maybe it's not the perfect threshold, maybe it's 1 in 100,000 whatever it is. But at the 1 in a million or below threshold, um, we've really— can you— can be pretty confident that the cancer is gone in a high, high percentage of the time. In this case, even in the— in the— not a single patient relapsed that was MRD negative. Um, okay, so then just, uh, I'm just about done, so I'll just, uh, show— you know, the other thing is that that's been mentioned, and I think this is something that, um, as a company Adaptive, we put a lot of energy in. We have a 10-person coding team that's run by one of the senior executives that we stole from Microsoft, and he is leading the charge where not only will you get back a clinical report, but you'll have access to a full suite of tools and your full dataset. So you can do sort of analysis of, you know, VJ breakdown, you can look at the clones, you can track the clones over time, you can, you know, graph the distribution in— you can parse this data in a zillion different ways. You can really do clinical studies. How does this come into play? For example, let's say you're looking at CLL and you want to know, you know, somatic hypermutation in the V region is a predictor of relapse. Well, we can— you can click a button and it'll tell you how much— show you directly where the somatic hypermutations are and tell you exactly how many there are. So really, we can leverage a lot of this information. And we're trying to build the tools to allow you to do that as clinician researchers. And then we modeled, obviously, how you order this thing after Amazon, because they're— although they're annoying me because I have to drive by their construction on the way to work every day. OK, so I'd just like to thank the group, specifically Brent and David, and then, of course, the group that did the transplant that I was talking about was Steven Grupp and Michael Pulsifer, and then Brian Langholz was the statistician who did the analysis on that one. And then, of course, the Adaptive Group and my kind of co-founders, Hudi and Chris, from Adaptive. Thank you.