 So, as I was saying, I'd like to thank Terry, Harrod, Rex, and anyone else who's willing to accept the blame for inviting me to speak. And the topic I was given was integrating model organism data around clinical genomics. I'm going to try and take that fairly literally and give a talk that I actually think Rex had previously heard, showing how we believe that fundamentally just rolling up our sleeves and getting deeply immersed in the middle of this is the only way forward. So this is a fairly typical problem, and in fact, as you've heard earlier in the day, sudden death is perhaps one of the areas where the rubber meets the road in clinical genomics. This is the fundamental genotype-phenotype problem, writ large. So this is one sub-ship of a pedigree that has about 400 affected individuals. I worked on it when I was a post-doc in Cricket's lab almost 25 years ago. So, single gene, sorry, I think you were just a first-year faculty member of that sex. It's a laminopathy. There's perfect segregation for one of two of the phenotypes in this family, and there's a large effect size. If you belong to this kindred and a genotype positive, you have about somewhere between a 510,000 X increased risk of sudden death before the age of 50. But unfortunately, there are multiple phenotypes in this family. Many of them have just asymptomatic phenotypes like this, a first prolongation of the distance between these complexes. Some people will have heart failure as early as the age of 10, and others will die suddenly and unexpectedly with no prior warning. And this is fundamentally the issue is that despite knowing perfectly what the cause is of the underlying disorder, we're unable to actually make any clinically useful predictions. It's worth pointing out that there are 12 different lamin syndromes reported, and actually all but 11 of them are found in this single family when you look deeply enough, although I don't believe that we've ever actually systematically addressed every single one. And the only one that isn't would be the Hutchison-Gilford Progeria, which as Mark mentioned earlier is an example of a single residue. There are multiple modifiers. That's what we would usually say in this setting, either genetic, epigenetic or environmental, but none of these are measured for almost any condition that we would hesitate to tackle. And there's no real empiric support for the strategy to move forward. And importantly, as I mentioned, no real additional information here to change clinical care. And as far as the patients are concerned, that would be improving their symptoms or improving their outcomes. Just as an aside, and we'll come back to this later, these are some of the phenotypes that you might expect outside the heart in a cardiomyopathy kindred, so everything from a Bifid uvula, which actually turns out to be a reasonable index of fusion defects mediated by TGF beta, through palmar plantar carotiderma, abnormalities of the petal arches or the axial skeleton that are based on integrals of long-term muscle tone during development and beyond, and then structural defects such as quartations. Those are all seen in dilated cardiomyopathy kindred. It's not all seen in this particular kindred. Clinical genomics is perhaps the other extreme where we start with a genomic sequence that we're trying then to interpret it. This is an example from the MedSeq study run by Robert Green and Cricket and I are both members of this. It's one of the CSER consortium studies, a randomized controlled trial of whole genome sequencing in a healthy primary care cohort and in a cardiomyopathy cohort. And this was one of the variants that was identified, this upper one here, identified in a primary care patient. It's a likely pathogenic variant in KCNQ1, one of the long QT genes, a potassium channel gene. During the initial patient primary care disclosure, the patient herself had an anxiety attack as a result of immediate concern regarding sudden death risk. And as you might imagine, throughout the rest of her evaluation, the concept of making her feel better or live longer was at the forefront. We could have thought then about what we would do to assess the pathogenicity of this variant. There are numerous assays in in vitro, both in hydrologous and homologous expression systems. There's an example from IPS cells from one of Cricket's papers. But they're also in vivo assays, including measuring along QT, looking at the QT morphology in lots of different ways. Worth pointing out, for example, here that the QT is really a surrogate for the underlying phenotype and it's one that we use simply because you can teach the average person to measure it. And this oversimplification of QT is worth pointing out. If I was to show the ECG to Dan Rodin, for example, he would have a binary yes or no repolarization disorder output, but the average person is left trying to work out where the end of the T-wave finishes or starts. And in fact, one of the most interesting things about this is as we were talking to this, talking about this problem with the patient in the cardiomyopathy clinic, the patient said, but my QT was normal. And so the question, I suppose, we have to ask ourselves at this point, is it always the phenotype? Is segregation, is penetrance, is pleiotropy, always a function of the underlying phenotype? Are there truly biological complexities that move beyond that? And that's worth, I think, thinking about as we go into the discussion. It's worth pointing out that the relationship between all of these metrics and actual objective risk to the patient is obscure. And what we really need are quantitative assays mapped onto people. The one thing I will put in here as a provisor, this is from Cricket's work, nothing to do with me, not that I don't think it's great work, actually quite the opposite. But it's a salutary example of how mutations that are known to cause autosomal dominant, highly penetrant, sarcomeric, hypertrophic cardiomyopathy in individual families in the general population may have no phenotype whatsoever work from the Framingham cohort. We go to the potential clinical studies that we could have done in this individual. We could have looked at her QT. We'd have looked at her corrected QT at the morphology of her electrocardiogram or at subclinical extracardiac phenotypes. None of these actually were particularly useful in this individual. And you can argue that they may not be useful in any individual. It's worth pointing out it was 95 years after the electrocardiogram was first introduced that we recognized the morphologic differences between the different forms of the long QT syndrome, required the genome to be cloned and the genes in each family to be cloned before we recognized it. Again, just an index of how poorly most phenotypes are actually resolved, even the ones that we think are incredibly precise and highly useful. There are a number of provoked phenotypes in the long QT syndrome, including posture, exercise, and recovery. And in fact, actually thanks to Andy Cron, there's actually a test of what the best discriminant might be in somebody who is genotype positive but not known to be phenotype positive. And they are actually the most useful thing, turns out to be the QT, corrected QT at 400, at four minutes in recovery after a standard exercise test. Unfortunately, that also doesn't perfectly apply to this situation. And again, throughout the entire evaluation, we did all of these things throughout the entire evaluation. The only thing the patient was interested in was am I actually at risk of sudden death rather than can you give me a label? And I suppose the other question we have to ask ourselves is, is the risk associated with the genotype or with the phenotype? A detailed family study just shows you how difficult this will be, so this is the actual pedigree when we did a more extended family study. We took a condition-specific family history, so there was actually a sudden death that had been previously labeled as an MI. There was also a congestive heart failure case in a grandparent, but it was complicated by the fact that this individual also had been treated for CLL. When we actually examined this woman, she had an ejection systolic murmur. Again, are you really interested in any of the subjective findings that a physician might record in the electronic health record? I would wager probably not. Certainly not something you'd be willing to spend money on. On the other hand, it is the basis of almost all diagnoses that we use in clinical care today. The QT itself was uncorrected at 466 milliseconds corrected at 461, so right on the margins for a female, but her EKG morphology was normal. But interestingly, our echo revealed focal left ventricular hypertrophy and mitral valve thickening not reported in the long QT syndrome. On the other hand, not systematically studied in any of the large cohorts with long QT. Our MRI was normal, and our provoked phenotype actually was, again, perfectly. This is almost too good to be true, but it was exactly 400 milliseconds of four minutes in recovery. So we have definitive abnormalities observed in the context of a genotype that we're not certain has correlated with this. And with all of these elements, I suppose the most important thing is that we have nothing really useful to show the patient. We've added about $8,000 to the evaluation. And remember, this is a known gene in the typical family. So phenotype, I think, is now limiting in multiple areas, not just in genomics, but also it turns out in clinical care. I think this is the momentum behind the precision medicine initiative, as much as genomics itself. There are fundamental issues with almost every phenotype. They're largely dominated by morphology. They're semi-subjective at best. They're late or even end stage. They are aggregated and have been so for decades for statistical power for clinical trials. We know coronary disease is probably 30 or 40 different disorders, yet we lump everybody together for most of the large trials, which have maybe 10, 20,000 people in them. They're largely legacy phenotypes. We never really thought about them in the biological context, for example, that Mark and Les were talking about this morning. They're often binary. They're cross-sectional, and there are almost no systematic perturbations, with the exception of a few stress tests in cardiology. So where's all the information? Why are these, why might alleles be silent? Well, it's pretty obvious from just standard genetic paradigms. We've looked, we've spent maybe many dollars on looking at Mendelian disorders and at GWAS. We're now in a resequencing phase, but the majority of phenotypes are likely in this middle ground, where there's actually an unmeasured conditioning variable, or there's now a phenotype that we don't have an assay for. And so that when you think about these things, the genetic architecture, I think, is an important piece of how we move forward. We don't know what the genetic architecture is for almost any trait that we think is important in disease. There are huge limitations to the way that we've thought about genetics to date. We've focused on extreme phenotypes out of necessity. There are a few prospective cohorts. There are very few modeling assessments of how many genes might be involved, if, for example, familiarity is detectable. And then the other thing to remember, I suppose, in all of this, is that heterogeneity also scales. So investing many billions of dollars in many much larger GWAS studies, for example, is unlikely to change the underlying thesis. And what we really need is to move away, I think, from the perfection that precision medicine implies to just thinking, perhaps, about more precise medicine as an intermediate step. So how might model organisms actually help here? Well, I think there are a lot of different strategies. And in fact, this is a diagram from Zach Kahane, just thinking about all of the different data sets that you might need in order to build a systems-level understanding of the relationship between genotype and phenotype. Model organisms, if properly tailored and the phenotypes are properly chosen, I think, can infiltrate this space very, very effectively through saturation screens to identify all the known genes for a particular trait. Although, obviously, it's important that we look at all the alleles, not just the F3 leafels. Reverse genetics to take individual alleles and then model what the phenotypic spectrum is. Again, we've tended to focus on areas that have been predetermined by legacy rather than on a universal phenotyping strategy. We need to begin to model environment. We need to think about where the gaps are that we could plug with model organisms. And importantly, I think the final goals will be iterative systems-level modeling and mapping to human genotype and human phenotype. So model organisms, I think, give us really some very important insights. One of the things that we found using the zebrafish, and I'm going to talk simply about the zebrafish simply because that's one of the areas that I've worked in, almost everything we do in the fish is done in 384 or 96 well plates. And I suppose the thing that we've learned is that comprehensive phenotyping is actually a much better way of going about this than taking one gene and a hypothetical phenotype. That you can begin to collect huge amounts of information if you begin to scale the phenotype and make it systematic and orthogonal. So these are just some of the traits that you can measure in fish. I don't have time to go into them in any detail, but it's worth pointing out that simple morphology clusters ion channel blockers as well as patch clamp during the first two or three days of development. We've undertaken strategies to do this in a variety of different disorders. This is a shelf screen for QT that we did in collaboration with Dan many years ago. And it's worth pointing out that in under extreme circumstances mapping exactly the phenotype in fish with a heart that's 100 microns long that we would map in humans, i.e. action potential duration, repolarization abnormalities. We identified this network of 15 genes, three of which have subsequently been found to represent some of the deviations in this QQ plot. We've turned the problem on its head. This is actually something I did almost 20 years ago, which was to look at the genetic architecture of atrial fibrillation, paroxysmal arrhythmia. So when you don't have the symptoms and you don't have any underlying phenotypes, it turns out there's no way of identifying unaffected individuals. And I would hesitate to suggest that that's actually one of the systematic themes in most of genomic medicine is the fact that we don't know who's unaffected because the resolution of our phenotyping is so low. We were able to show, for example, that this has a large Mendelian effect that only about 10% of the heritabilities explained by GWAS loci, even although these are the largest effect GWAS loci bar one in the entire genome. There are large missing intermediate effect sizes, and it's proven incredibly difficult to clone the genes simply because we do not know who's unaffected. So we're going to have to move forward with pedigrees where we basically have to wait for people to declare themselves or we need to identify new phenotypes. We've actually taken an approach in model organisms to try and identify what those phenotypes might be. We used a couple of machine learning strategies to try and identify genes that were at loci that were identified by GWAS that might be shared amongst all these. We found a unifying network that's about an order of magnitude more likely than any other network that is shared by all the 12 loci. And importantly, every gene in this network and either overexpression or loss of function perturbs the cell-cell coupling in the heart. But then when you try to move this to a clinical population, there are no metrics by which you could do that in a clinical population. So the other strategy that we've used is to try and think about what the perturbations might be. And these can be thought of in terms of physiologic tests or in terms of therapies. And as I argued earlier this morning, I think thinking about it in terms of therapies is a quick way to early low-hanging fruit. We modeled this disorder, which is actually a right ventricular disorder in an organism that does not have a right ventricle. It was chosen without a hint of irony. The disorder is characterized by arrhythmia, sudden death, and continuous abnormalities, and has traditional heart failure metrics. We did all the genotype anchoring that you heard about this morning so eloquently from hard. We built a conditional expression model. It turns out this is actually a relevant model of the allylic architecture of the disease. There's codome, co-expression of both the mutant and the wild type. You can see here the fish heart ends up, in fact, you can actually visualize the phenotype of the naked eye. These are siblings. We were able to identify phenotypes that were actually novel for this disorder, showing a dramatic loss of sodium current density at the membrane. There's tremendous heterogeneity of the cell biology in the heart, even in the single chamber fish. We were unable, obviously, to map any of that onto the human, but we were able to identify naturalistic peptide abnormalities. This is a reporter system that we built, and then went on to do a chemical screen for suppressors of this in the genetic mutants, just simply plating the mutants and then looking for suppressors. We found several compounds, including one that was effective at nanomolar potency, which subsequently has been used, and this is published, so I won't go into it, to rescue the phenotype, including the sodium channel defect of a variety of other phenotypes and even ventricular activity in the mouse, now actually also in human IPS cells. So I would hesitate to suggest that actually the fundamental problem with the entire field of genomics is actually this phenotype gap. This was work that we did actually, Teri had asked us to do it as part of the UDN, trying to cost out the scope of phenotyping that might occur in the Undiagnosed Diseases Network. And if you look at this, it's not surprising why we're unable to deconvolute a genome when we have only 10 to the four total phenotypes in the whole spectrum of clinical medicine with 10 to the nine genotypes, 16 plus transcription variants, 10 to the 20 proteomic lipidomic variants. We haven't even gotten to cell-cell connections or exposures. This was chosen by my wife. It's wind as in bag of wind. We need new translatable human phenotypes. The things we've picked are actually almost hysterical when you think about it. Glucose in the urine by taste, 20 year gap between when you develop the metabolic phenotypes of diabetes and when the glucose ends up in your urine. Cholesterol, there are 1,000 things abnormal in the LDL receptor families. We only measure cholesterol because it's visible to the naked eye. There's been deliberate reduction in complexity, limited dimensionality, and no clear organizing stimuli for any of the phenotypes that we look at. I'm gonna have two more slides. We need to reappraise all of our existing phenotypes focusing on resolution and computability. I would wager that there's not a single piece of information in most electronic health records that I care about because it's being collected in such a non-structured and random way. Every MRI done at Brigham and Women's is done using a different protocol. Just even unifying the structure in these would be important. I mentioned the ECG, which is probably the only phenotype for which there are standardizing data. We haven't pushed functional genomics into the clinic, although as you heard from Cricket, that type of pathway analysis could be quite powerful. I think there's a real opportunity to do next generation phenotype. You can imagine almost anything in this space from credit card reports through to facial recognition. I'm not gonna dwell on that. And the last slide, really, we need rigorous probability estimates for every step in this pipeline. I would point out the fact that there's a ton of extant data that we haven't organized in a way that can impact the clinical arena. We need quantitative family histories, population, lifetime risk studies, network structures and their responses, measured exposures, shared phenotypic lexicon that goes all the way from yeast to patients with much more mechanistic and proximate phenotypes is less outlined. Learning information systems and co-clinical modeling I think is an important part of that, all at population scale. And if we've learned anything from the genome, it's the fact that we need comprehensive, multi-scale dynamic phenotyping. This should all be integrated into clinical care with the goal of an next generation computable phenotype. And so in summary, I think it really requires deep knowledge of the conditioning variables for us to be able to interpret even a single genome. And many of the things we talk about here are just not measured any longer. The electronic health record is almost eliminated family history. Exposures, we measure the exposures to five or six things at max. Scalable anal modeling is emerging as a partner for clinical genomics, but I think it has to be embedded at very key points in order to be effective. It's worth pointing out, and I put this paragraph in after this morning's discussion, that at the same time as this is happening in the genomics community, there's a huge effort in the clinical community to redesign clinical care. And I believe that aligning discovery, genomic discovery, and beyond clinical care redesign and cost will actually mean that we have to integrate this stuff immediately to avoid unaffordable duplication. And perhaps one of the most important things we need is, again, just to emphasize this shared lexicon. And then finally, we need a minimal clinical data set, as Dan outlined, some degree of universality, which would have to be cellular, as well as quantitative linear stimulus response pairs, and all embedded in a way that complements existing care and is fundamentally part of how we deliver and educate the information to our providers. Thank you for your attention, and sorry I went a little bit long. Thank you, Callum. So we'll have time for a quick question. Does anybody have a quick question before we move on to Cecilia? Okay, so we'll move on to Cecilia Lowe, and then we'll have follow-up questions during the discussion.