Founder Stories: Robert Plenge, Impact of Human Genetics on Drug Discovery Productivity

Thank you all for having me, and Dan for the invitation. I'm going to tell a slightly different sort of story, and what I want you to think about is not necessarily genetics to actually predict some of the phenotypic variability in drug response, but rather genetics is actually a starting point for drug discovery. And so David actually talked about it in his talk, the more that we learn about human genetics, the more that we're reaching these therapeutic dead ends. I think the story that I want to tell you today is about how we can actually use human genetics to start that drug discovery process. So, so let me kind of start by framing this in maybe a slightly different way, which is think about yourself not necessarily as a human geneticist, but think about yourself who's doing drug discovery. And the problem that we actually face in the drug discovery industry is the following. First, the cost of getting one drug approved continues to go up and up and up. And there are various numbers that are actually cited, but the bottom line is it's expensive, upwards of $2 billion per approved drug, and the costs are continuing to go up as shown here by nearly 70% over the last 7 years. And so I'm going to refer to this as the attrition problem, and I'll spend a few minutes on it. But it's even worse than that, that the drugs that were actually developed, that we're delivering into the real world, are delivering less and less value to patients. And one way you can actually quantify that is actually the peak sales of the drugs that are actually approved, and these continue to go down and down and down. So the cost of drug discovery is getting more and more expensive, but there's this innovation problem as well. And I wanted to spend a few minutes on the attrition problem because I think that's fundamental to the argument that I want to make. So think about where things go wrong in the drug discovery process. What you're looking at here off to the far left is the very beginning of a drug discovery campaign. You have a target, you have a pathway, and you want to make a drug against it. Let's say about 30 molecules or 30 targets start that particular journey. And then as you go closer and closer to the clinic and to, say, Phase 1, you whittle that down to about 10 that actually start a Phase 1 clinical trial. And then over the course of Phase 1 to Phase 2 and Phase 3, you have various levels of success, ultimately leading to one approved drug. And so about 1 out of 10 drugs that actually starts in a clinical trial program at Phase 1 ultimately makes it to approval. And what's important about this is, is not necessarily just the numbers, but actually where those dollars are spent. So about for every $1 that's spent in the preclinical stage, you spend about $3 in the clinical phase. And that's actually working on fewer and fewer programs. And so it's the expense, actually, and the failures in Phase 2 and Phase 3 that are really drivers of this cost. OK, so then if you actually dig deeper into that problem and ask, why are things failing in Phase 2 and in Phase 3? And if you step back and think about it, despite 5 to 7 years of preclinical modeling, a tremendous amount of information that actually goes into designing a Phase 1 study, still over 50% of things that are actually failing in Phase 2 are failing because they just don't work. That number goes up even more into Phase 3 failures where it's Nearly two-thirds of drugs that are failing in Phase 3 are failing because they just don't work. So it's a reminder of the complexity of human biology that we're actually facing. So the key point here is that the decisions that are made at the very beginning of this drug discovery process, the very beginning of this journey, are really having a very strong influence on the success in the later stages of drug discovery. And this is where I think human genetics can actually have a very, very substantial impact. So there are a number of different ways to think about genetics and drug discovery. I'm going to provide you with one model, which is a solution to how to think about this problem. And I'll actually build on some of the discussion points that David had in his previous talk. And so the way I like to think about it is that it used to be, as we thought about a drug discovery campaign, we relied very heavily on preclinical animal models to select the targets that we wanted to pursue. Maybe there was a knockout animal that led to a particular phenotype and we decided to make a drug against that particular target. And then we actually delivered drugs into patients that were really unselected. And so this is where some of the themes that you've touched upon here already in terms of pharmacogenetics will actually play out. So this is the way that much of drug discovery was actually done. But I think you've already heard quite a bit about this over the last day and a half, that this This paradigm is changing. And increasingly we're using humans as the model organism of choice. We're using humans, whether it be from genetic observations or other causal human biology information, to make choices as to which targets we should pursue along this drug discovery journey. This is not to say that preclinical animals are not important. In fact, you actually heard very good examples from David just now about how we should actually use these preclinical models to understand the biological consequences of human genetic mutations, and also to actually understand the pharmacology of the molecules, and to guide dose and schedule before going into humans. But the other important part of this diagram is that actually as we go into humans, we shouldn't be going into unselected patient populations. We should be going into very specific subsets of patients that are defined by the underlying genetics of the target, or the underlying biology of the target. Target. And so I think an important theme that I want to get across to this particular group is that as we think about pharmacogenetics as a community, I think the biggest impact that genetics is going to actually have, and ultimately pharmacogenetics, is not going to take drugs that exist in the real world today and finding genetic factors that tailor it to the right patients. That will happen. We should always look for it. That's going to be important. But I think the biggest impact that genetics will actually have, and ultimately pharmacogenetics, is when we start the drug discovery journey with a very specific genetic hypothesis, and that's carried all the way through into the biomarkers that we choose, in terms of the clinical trials that we actually design, and ultimately the patients that we treat in the real world. So let me just build on this model a bit more, because I think lots of people say, well, we'll use human genetics, we'll find these drug targets, and then voilà, we'll actually have this wealth of new drugs. And of course, it's far more complicated than that. Than that. And I think one way to think about it is putting on a graph function and phenotype, where human phenotype is along the Y-axis, gene function is along the X-axis, and you can think about this very much in the same way you think about a dose-response curve. Now, if you think about how we actually do a genetic association study or a genetic discovery program, you know, the first step is we pick a human phenotype, and I think you have to pick a human phenotype that is a good surrogate for drug discovery. You heard yesterday from Matt Nelson about his wonderful work and how to think about genetics in drug discovery. And in his paper from a few years ago, he looked at a number of different phenotypes that are actually quite good surrogates for efficacy in a drug discovery study. And I won't go through them here, but a list is actually shown on this slide. So once you believe that a phenotype is actually a good drug discovery phenotype, You then need to find those alleles that are influencing those traits. Ideally, you'll find not just a single allele, but you'll actually find multiple alleles. Those alleles could be common, they could be rare. The frequency is less important, but as long as you can actually make a really unambiguous association between the genotype and a clinical phenotype, and ideally have a range of effect sizes for those variants itself. But then you have to do the dirty work of getting into the lab and understanding the consequences of function. And this, I think, is really one of the major bottlenecks that we're actually facing today. So understanding, are those variants gain-of-function, are they loss-of-function? What are they truly doing to the target? What is that mechanism that we need to modulate with a drug to achieve a pharmacologic effect? And this will then not only have a range of effect sizes in terms of phenotypes, but also a range of effect sizes as it pertains to biological function. But I think genetics has a very important property which I think is often underappreciated. You can actually take those same alleles and understand the pleiotropic consequences, much in the way that you actually can think about the side effects, the on-target side effects of drugs. And so you can now use human genetics to assess pleiotropy as proxies for adverse drug events, and I'll talk a little bit more about this in a few minutes when I describe PheWAS. And again, if you have a target with a range of alleles, a range of clinical effects, and you understand what that— those clinical effects are, not just only on the phenotype you think is of interest for drug— for a clinical trial, but also potential adverse effects, this again is a starting point for a drug discovery journey. And so a point that comes up again and again is common versus rare, and I just want to say that I don't think it's common or rare, it's really common and rare. It's really about having a deep understanding of the variants, the influence on clinical phenotype, and influence on gene function. I think rare variants are incredibly important, because often those effect sizes are larger, and I think human knockout phenotypes in particular are incredibly valuable, because it tells you the maximum effect that you can actually have on a drug target and the phenotypic consequences. But I also think common variants are actually quite important as well, because it often gives you more kind of subtle— a more subtle understanding of biological perturbation of a target and the effects that you actually might see in the clinic. And as you'll also see as I describe in a couple of examples, the ability to select indications for a drug of interest. So I'm going to go over one example in human immunology, and then I'm going to talk about two emerging resources, Mendelian randomization and protein quantitative trait loci, and then phenome-wide association studies, or PheWAS. So the first example is an example of this allelic series model. It's for a gene, a TIK2. TIK2, as I'll show you, is an intracellular signaling molecule that's downstream of another— a number of cytokine receptors. If you knock it out completely, and there are human knockouts, this is associated with immunodeficiency and an increased risk of infection. There are also common protein coding alleles that influence TIK2 function, and these have been shown to be protective for a variety of autoimmune diseases. I'll focus on psoriasis, but they're also associated with protection from RA and lupus and inflammatory bowel disease. And as I'll show you, those same common alleles actually do not increase risk of infection. So while the complete human knockout increases risk of infection, these partial loss-of-function mutations that protect from autoimmunity have no obvious increased risk of infection, which suggests that there's this therapeutic window that might be valuable for a druggable target. So just to tell you a little bit about the target and the pathway, and actually as it pertains to psoriasis. So there are IL-12 and 23. These ligands can come together and bind to a receptor, and that's shown here, the IL-23 receptor, and then the IL-12 subunit. And they signal then through these intracellular signaling molecules, one of which is TIK2, another is JAK2. What's pretty amazing to me at least is that the genome-wide association studies for psoriasis have identified at least 4 components of this signaling complex as implicated in risk of psoriasis. The gene IL-23A, another gene IL-12B, the receptor IL-23R, and then as I'll describe, TIK2 itself. I think what also is quite interesting is that at least 2 of the extracellular components of this have been shown to be effective drug targets in terms of treating psoriasis, and the IL-23 p19 inhibitors in particular in particular, have shown very remarkable efficacy. And this then leads to the question as, if you were actually to inhibit TIK2, would that be therapeutically beneficial? This is a study that was published just recently by Shoma Rishadri and colleagues in Nature Genetics, and it fine-maps a number of different regions, including TIK2, for both rheumatoid arthritis and type 1 diabetes. And I'll use this a couple of times in the talk because I thought it was a very nice study. But what they actually did was use deep fine mapping to find the underlying causal alleles, and just to highlight one, which is P1104A, as you can see that this is a relatively low frequency variant, about 3% in individuals of European ancestry, but it has a pretty substantial protective effect on risk of, in this case, RA and type 1 diabetes, but as I'll also show you in psoriasis and other autoimmune But there are multiple alleles that also are protective. Another one is also a protein coding variant, I-684S. It's a slightly more frequent variant with a smaller effect in terms of protection from autoimmunity. And there's a third variant which is not shown here, which is even more rare, but also a protein coding variant. All of these are predicted to be loss of function. All of these have been shown to actually protect from autoimmune disease. So now you can actually use this as a tool. This is a phenomenal tool that you can actually now probe in the ways that I've just described, the impact on gene function, but also the impact on pleiotropy, and I'll just walk you through this. This is a paper published a few years ago by Kali Dendroux and colleagues, published in Science Translational Medicine, to actually explore the genotype-phenotype consequences of this one variant, P1104A, and what they actually did was a series of very elegant functional studies. Where they showed that carriers of this particular mutation actually had defects in signaling, shown here for defects in signaling through interferon alpha, but also they showed defects in signaling when the activation was through IL-12 or IL-23. And what's really important to the story is not just that it has a functional effect, but the magnitude of the functional effect. So they showed via these assays that there was about an 80% loss of TIK2 function when measured in patients who carry these different alleles. And that's shown by the figure here for CD4-positive naive T cells. Okay. Then they also took that exact same variant and they went into the UK Biobank. And we've actually done this for other— in other biobanks as well, including Partners HealthCare and with Vanderbilt and working with 23andMe as well. But what they showed here for UK Biobank is that variant is actually not associated with an increased risk of infection. So, yes, it's associated with protecting for the autoimmune diseases that I described, but there's no obvious increased risk for infection. So what this tells you now as a therapeutic hypothesis, that about 80% loss of function of TIK2 activity should protect from autoimmune disease but not increase risk of infection. Okay? Here is showing that it's not just protecting protecting from risk of RA and type 1 diabetes and psoriasis, but a range of autoimmune diseases as shown on the top, including ankylosing spondylitis, Crohn's disease, ulcerative colitis, and MS. But you'll notice that I actually am showing you one of the variants I described at the very beginning, and the reason for that is that a second variant has a slightly more complicated pattern. It looks like it protects from several diseases, including psoriasis and RA, but it may increase risk of other autoimmune diseases including inflammatory bowel disease. Okay. But taking this example, you now have this risk-benefit curve. So this function phenotype map where you can actually say, again, complete loss of function increases risk of infection. So you now know if you develop a drug, you don't want to completely inhibit TIK2's function. But then 80% loss of function looks like it protects from autoimmunity. It has an immunologic phenotype, but is not associated with risk of infection. And we've seen that from the UK Biobank study. Wrong button. OK. So the therapeutic hypothesis is that partial inhibition of about 80% will protect from autoimmunity but not increase risk of infection. OK. So this is another theme to me which is just incredibly important. And it's not necessarily obvious. But as I've done drug discovery more and more since moving from academics into industry, it's to me one of the biggest challenges that we face, which is to take that therapeutic hypothesis and find a modality, a therapeutic modality which matches the mechanism of interest. And the reason this has been so challenging for TIK2 is because a traditional kinase inhibitor, which binds and inactivates the kinase site, so it's called an orthostatic inhibitor, that pocket is highly conserved across a number of different enzymes, including other JAKs. So there have been a number of different companies which have tried to actually develop Tyk2 inhibitors, and they failed because they can't get selectivity. But they failed because they focused on that kinase domain. And so about a couple of years ago, BMS published a study which really unlocked, I think, this area of biology. They took this observation. That TIK2 function in these different immune cell types has a biological effect. We wanted— and they wanted to mimic that. And they actually created an assay which was based upon a phenotypic screen. So rather than try and develop a biochemical inhibitor that was against the kinase function itself, they broadened the search to say, we want to actually look for a phenotypic— we want to use a phenotypic screen to find an inhibitor of TIK2, but we want to be a bit more open-ended as to how that molecule works. It doesn't necessarily have to bind to the kinase domain. It could bind to other sites including allosteric sites. And they were successful at this process. So what they found was that they found a molecule, and I'll tell you in just a minute, it bound to a pseudokinase domain which had an allosteric effect on enzymatic function. And what they could show was that with a molecule that was highly potent, you could get 80% or more inhibition of of the other JAKs, so JAK1, JAK2, and JAK3. The other JAKs are important because they increase risk of infection, and they also have hematologic consequences as well. Okay. This is a model of the— of TIK2 itself. And what I'm showing you here in orange is the site of the kinase domain, which is the orthosteric inhibitor. And this is the site of the mutation. But what they found was that a molecule bound to a different site that may not have been predicted from a straightforward understanding of the gene's function led to dramatic inhibition and also, importantly, selectivity over the JAK inhibitors. So this story is fine, and I've actually been telling the story for a number of years, but what actually to me was quite remarkable was the publication just a few a few weeks ago in the New England Journal of Medicine, which showed that this idea actually works. And so what BMS did is they took this TIK2 inhibitor that I just described. They went into psoriasis. They modeled it to try to get between 50% and 80% inhibition of TIK2. And the drug was remarkably effective. So what I'm showing you here at the top are 3 different doses— 3 milligrams twice daily, 6 milligrams twice daily, or 12 milligrams daily. And what they found was at that level of inhibition that there was a PASI 75, which meant about 75% clearing in about 75% of patients, was effective at 12 weeks. And this is about 2 to 3 times better than any other oral drug that's actually available on the market. But it also achieved PASI 90s in the range of 45%. And these are clinical scores which are approaching that of the biological therapies that I described at the very beginning. So I think this actually is very encouraging that this therapeutic hypothesis will indeed play out and will also benefit patients. Okay. So if that's the model, an allelic series model, I've given you one example, TIK2, trying to say it's important to not only pick the target, match modality and mechanism, and use that to guide the amount of inhibition that you desire. What are the resources that are required to make this kind of more routine in the way that actually drug discovery is done. And I want to describe two to you very quickly. The first is the idea of Mendelian randomization. I think people are very familiar with that. You know, it's nature's randomized control trial, and so that's shown on the left where you have two different groups, an intervention group and a control group. And in Mendelian randomization, you're not randomizing based upon drug, you're actually randomizing based upon alleles. And what you want to find is that allele that actually is present that may have, you know, a higher effect on the biomarker and a higher effect on disease risk. Okay, so what are the tools that you need in order to make Mendelian randomization a reality? You need this causal pathway, genotype to intermediate phenotypes and disease outcome. You need data, genotyping data, sequencing data, functional data as shown in the middle, and ultimately clinical outcome data. You need to actually do the relevant studies. So this first is a quantitative trait study, genotype to biomarker. The second are genotype to clinical outcomes, so the— most of the GWAS that have actually been done to date. And of course human epidemiology is important as well. And you can establish then these causal pathways, and ultimately if it's along the causal pathway, there'll be the genotype to intermediate biomarker to disease phenotype, and that will all line up. But what's important is that, you know, increasingly it's not the genotype to clinical phenotype datasets that are limiting, it's these quantitative trait phenotypes that are limiting, including functional assays that David described, protein quantitative trait loci, and I'll describe those in just a minute, expression QTL levels, metabolomics, et cetera. So this is a study that we did in collaboration with Adam Butterworth and John Dinesh and colleagues. It was published a few months ago in Nature.— and it was a large protein quantitative trait loci study done for about 3,600 plasma proteins in over 3,300 individuals from the interval cohort. A very large number of PQTLS were identified. That wasn't necessarily surprising. But shown on the right is an example of one at a locus, IL-6R. And I think what's interesting about this is that the allele that is protective is associated with higher amounts of soluble IL-6R, but actually associated with decreased amount of membrane-bound IL-6R, which is consistent with the therapeutic hypothesis that preventing IL-6 signaling via an antibody, tocilizumab, as I'll describe in just a moment, should be effective in treating RA symptoms, as an example. And so again, it puts it along this causal pathway, and the therapeutic target therefore should be neutralizing the soluble portion, or IL-6R, directly. And again, that's actually shown here for a drug that's approved for rheumatoid arthritis, tocilizumab, that mimics the effect of the mutation, again, by inhibiting IL-6R signaling. Okay. In the last few minutes, I just want to briefly describe phenome-wide association studies as an emerging resource. And again, this takes advantage of the ability to look at the pleiotropic consequences of alleles as they're identified. A few years ago, I think people probably were not as familiar with this concept of PheWAS. A lot of the work that Dan Roden and Josh Denny and Marilyn Ritchie and others did at Vanderbilt really, I think, pioneered this field, and it's increasingly becoming sort of a commonplace in our vocabulary. But just to describe the approach very quickly, you take a cohort, a large cohort that's a disease-agnostic cohort that has very rich genetic data and also rich clinical data. And you can then merge those two and take variants of interest and now look for the pleiotropic effect of a variant on a wide range of phenotype, as opposed to the reverse, which is often one phenotype and a large number of variants. You can then, again, use this PheWAS approach to look at the pleiotropic consequences, and going back to that allelic series model that I described earlier, you can use those phenotypes as surrogates for efficacy or surrogates of toxicity. And again, this would be one of the tools that would allow you to establish these dose-response curves. So I want to highlight one example very briefly, a gene IFIH1. Again, I'll refer you to Shomari Chaudhary's paper from earlier in the year, fine mapping RA and type 1 diabetes. And again, there are multiple alleles that are protective in terms of risk of type 1 diabetes in particular. But also RA, and one haplotype in particular can be used as a tool in PheWAS. So you can find a tagging SNP for this particular haplotype and take that into a large-scale PheWAS study. The study actually, I say in press here, it came out today in Nature Communications, and it was a PheWAS in about 800,000 individuals from a number of population cohorts, We tested 25 variants that were established as reproducible genetic associations, and we tested these across over 1,600 clinical endpoints, and 10 novel associations were discovered, one of which is shown here for gene IFIH1, also known as MDA5. IFIH1 loss-of-function variants were known to protect against a number of autoimmune diseases, including vitiligo, type 1 diabetes, and psoriasis. But the question then, from a therapeutic hypothesis perspective, is what happens when you inhibit that particular gene, and what adverse events might you expect to— what adverse effects might result from that inhibition? At least this PheWash approach might suggest that there'd be increased risk of asthma and also ulcerative colitis. So it's worth noting that the effect size on that same variant for asthma is much smaller than the effect size, say, for vitiligo. So maybe there is indeed a therapeutic window. Unlike the TIK2 story that I told you about, there are no clinical compounds that inhibit IFIH1, so this is a story that will have to play out over the next several years. But again, at least it highlights this concept of these function phenotype curves, this concept, this model of allelic series, and using this as the starting point for a drug discovery journey. I just want to highlight one emerging resource, that is FinGen, which is recruiting up to 500,000 individuals across the company— country of Finland, and you'll hear more about that over the course of the meeting in the next few days. So I just want to finish up just by summarizing the points that I made. So first and foremost, not only does the pharmaceutical industry use genetics, I would actually argue The pharmaceutical industry needs genetics. We need a way to actually deliver new targets, to drive down costs, and to deliver innovation to people who need it. Human genetics, as you heard from Matt yesterday, increases the probability of success by at least twofold, in some cases even higher than that. I described the Lilac Series model and how you can use genetics to pick targets, think about biomarkers, think about effect size, and also select indications. I gave you an example of TIK2. Which is proving to be an example that's worked in the clinic, and then also described emerging resources for Mendelian randomization in PheWAS. So with that, I'll be happy to take any questions, but I think we'll probably move to the panel discussion. [Speaker:JOSHUA_ELKINGTON] Thank you. Thank you very much for keeping to time. So any burning questions? Before we move to coffee break. We do have a panel discussion later on. Question over here. Okay, thank you.

Thank you so much for the presentation. TIK2 seems like a really good example. You kind of mentioned the disease combinations, but I'm just wondering, I would like to hear your opinion. What if a psoriasis patient develops arthritis or asthma? What is the approach to address that if the mechanism is not the same?

Sorry, ask it one more time. I don't know if I totally followed the question. So for psoriasis— for, for TIK2, multiple different autoimmune diseases including protection from psoriasis, and then IFIH1, a different gene, had increased risk of asthma and protection from other autoimmune diseases. So I may have missed your question.

Yeah, just certain, there are disease combinations. It's not as straightforward as there is one pathway leading so on, because I know that certain percentages of psoriatic patients have psoriatic arthritis and so on and so forth. So things get complicated. So does this mean that you're in square one when you face such a disease?

Yeah, so I think the short answer is that genetics is not It's not going to take that probability of success to 100%. It's still hard. If you can double or triple your probability of success, that will really drive down cost of drug discovery. So it is still hard. It still takes a lot of work. And there's still uncertainty. But I think that probability, it is very clear, at least to me, is doubled if not tripled. And so I think that's the main message is it's still hard. Thank you. Okay, well, thank you very much for your talk.