Okay, so I was charged with the task to talk about where we need to go from here, or where do we want to go. And I'll lay out for you a picture of the future that may be a little fanciful, but really incorporates a lot of technologies and knowledge that exists already today. I have many conflicts, but none of the content of this lecture is related to my commercial relationships. So, let's review for a moment what we learned today. What are some of the success stories about precision medicine and hepatology? We learned about the genetic basis for hereditary cholestasis, some unexplained liver disease, GWAS discovery of risk genes, biomarkers, particularly MMP and biliary atresia, lipidomics and NAFLD, predictive clinical algorithms, which now spread across a number of different conditions. I included one that you haven't heard much about, but let's not forget about the important IL-28, a single nucleotide polymorphism that was discovered to predict responsiveness to interferon that could have had a profound effect on treatment efficacy had we not advanced to direct-acting antivirals. Also pharmacogenomics of susceptibility loci and DILI, and then finally you heard about exciting work in the microbiome linked to a bacterial product, cytolysin, and then there's also studies not discussed here that identify bacterial proteins that are really mimics for G-protein-coupled receptors. So those are the success stories. Where are we in terms of our challenges and what we need to confront going forward? These success stories are somewhat sporadic. They're not systematic. It's not as though there's been a global integrated international effort to identify the genetic basis or precision components of liver disease care in an unbiased way. And even with some efforts, while we have some associations, many of the diseases do not really have progress in what I would call providing actionable information. Those that need further study or I would say actionable information include PSC, alcoholic liver disease, PBC, autoimmune hepatitis, DILI, and hepatocellular carcinoma, and we are on the cusp of getting there, particularly with HCC based on the outstanding presentation of Dr. Lovett. The problem is these studies are very data-intensive and expensive. They require very specialized skill sets, and I can tell you if I were advising my children now if they didn't want to do medicine, they should think about going into informatics because there is an extraordinary unmet need for talented people who know how to manage big data sets, not only in medicine but throughout our world. And that includes multiple omics data sets that need integrating. That's genomics, metabolomics, metagenomics, and in general, we have not really incorporated epigenomics, which is yet another important driver or regulator of gene expression. And as I alluded to, we need to distill this into actionable information that changes patient management at the bedside, and that will require software to integrate complex information into the electronic medical record or electronic health record. I'll show you some examples of how that's being done. And then, of course, in our world, we also need to justify this in terms of providing reimbursement, and the climate or landscape for that is very uncertain in the United States, and I assume elsewhere, but ultimately will require evidence of cost savings for justification. So where do we go from here? Well, in many institutions, including our own, we start, as always, with the patient, obtain consent, often blood samples, patient-reported data, other data. We also integrate into our biobank diagnoses, procedures, imaging, laboratory measures, pathology reports, and even more and more information. This ultimately can also be aligned with genomic DNA sequencing or whole exome sequencing, and that becomes a fertile wellspring for research. And one would hope and expect that all of this ultimately returns back in terms of improved outcomes for our patients and improved diagnostics and decision-making based on integration with our different support technologies. Also at Mount Sinai, in addition to establishing this paradigm and a biobank, which I'll talk about in a minute, we have developed a Center for Genomic Health. This is led by Dr. Noor Abulhusen, who provided several of these slides and also performed seminal studies that I'll show you in a moment. And our approach is to slowly but surely develop the infrastructure to actually translate or integrate genomic information into actionable risk items or risk of disease that can lead to genomic counseling, genome-driven surveillance, which certainly we're familiar with, preventive medical and surgical interventions. And the deliverables on this are listed here. They include more knowledgeable patients, and of course they in turn will inform their family, their providers, children. Also early disease detection, which of course intuitively we would hope and expect lead to better outcomes. And finally, reduced cost. So where are we with all that? Well, let me start with a success story. This is actually built on the work of Noor Abulhusen when she was at Regeneron. She's now actually joined the faculty at Mount Sinai. And I'm going to give you an example that you're somewhat familiar with of HSD17B13 as a success story of genomics-driven drug discovery, at least in NASH. And so why NASH and why this effort? I think most of us know that many experimental drugs in NASH are still failing, and preclinical models, while informative, sometimes can be a poor predictor of clinical benefit. So the rationale is if we can provide genetic basis for why a pathway or an outcome is worth addressing, that really establishes a rationale. Or as geneticists like to say, the genetics don't lie. If a genetic polymorphism consistently leads to or predicts a certain biology, then that's telling us something in terms of where we can go and what street lamp we should be looking under. And, of course, there's a high unmet medical need, NASH being a leading cause of morbidity and mortality, need for better treatments, and certainly a need to grow our knowledge base around genomic factors underlying NASH progression. So the work described here was published in the New England Journal last year, and it really represented a modified GWAS study in which investigators at Regeneron were looking for gene variants that correlate with either an elevated or a normal ALT and AST. And this has been an approach that's certainly been used for other genes. Among many, they found and pursued a variant in the gene HST17B13, which, frankly, I had never heard of. Probably most of us hadn't. I think Regeneron hadn't really heard of it either. And therein lies one of the power of these kinds of studies. It's not effectively hypothesis driven. It's hypothesis generating. So this variant popped up, and consistently was associated in patients with decreased transaminase in population studies with a reduced risk of alcoholic and non-alcoholic liver disease and cirrhosis, and apparently protection from NASH in individuals with fatty liver. In other words, the protective variant did not prevent the development of fat, but it seems to correlate with a lower likelihood of that fatty liver converting to the inflammatory ballooning liver that we know is a risk for fibrosis and fibrosis progression. So one obviously asks, well, what does this variant do that protects the patient? And that's effectively illustrated here. The variant encodes an enzyme, but in particular, a truncation mutant of that enzyme that effectively inactivates the enzyme. And so with the presence of this unstable enzyme, it's catalytically defective. And that effectively tells us that the enzyme is generating something that is leading to inflammation and the more advanced forms of NAFLD. Now we actually don't know yet what the native substrate and product of this enzyme is. But we know that it's a druggable enzyme. And that means that even without understanding the entire biology of this enzyme, we can begin to ask, what if we just inhibit the enzyme? And if that sounds a little bit sort of vague, this is an approach that's been wildly successful in uncovering a variant of a gene known as PCSK9 that, when absent, dramatically protected patients from cardiovascular disease. And that led immediately to approaches that are now marketed using either antibodies or small molecules that effectively block PCSK9 and prevent or dramatically lower serum LDL. So one can ask, is this a common variant? And we're very fortunate at Mount Sinai of having the, I take no credit for it, but having the foresight to develop a biobank over 10 years ago. We now have 30,000 participants. These are fully whole exome sequenced individuals. And on the right hand pie graph, what you can see is the ethnic distribution, which we're proud to say reflects our community, a large component of Hispanic, African, as well as European, Asian. And that really gives us a template to begin to ask, well, what is the allele frequency? How many patients? And does it vary across different ethnic groups? And the variability of the allele frequency shown in this cluster here is really striking. You have, at one end of the spectrum, patients with 25% allele frequency, others much lower, shown here. And so we can understand that certain patients have a higher or a lower likelihood of having this protected variant. Overall, the allele frequency was 16.8%. And one of the other benefits of this approach is that it uncovered some new alleles that hadn't really been described before, that were more prevalent in specific subgroups or ethnic subgroups. So this brings us to the point where, as we look into the future, as highlighted by this Washington Post headline, what if your doctor offered genetic testing as a way to keep you healthy? So we can imagine, and in fact, it's already being done more experimentally than routinely, you can imagine for a patient not only obtaining the usual chem panels, but also obtaining a whole genome sequence. And we're really at the point now where the cost of sequencing is so low that it's virtually not an impediment in obtaining this kind of information. So in the case of HSD17B13, how would we use this? Well, it turns out you can combine the risk of this variant with other risk factors, genetic risk factors. And as shown here on the left, you have a graph looking at the impact of whether you have low-risk alleles for both, or high-risk alleles, meaning the disease-promoting variant in PNPILA3 and the lack of the protected variant for HSD17B13. Now this is simplified, and you can do this manually, but imagine now you start to integrate more and more and more genetic information, and maybe a BMI, maybe clinical information. And what this leads us to is a continuum of risk that is simplified in this diagram here, but in its simple terms leads us with these two polymorphisms to identify patients who are at low risk based on those two genes alone, or at high risk based on the variants that predict either a lack of protection and promotion of disease by these two variants. And so this really leads us to needing a multidisciplinary approach to addressing risk factors to prevent and delay onset of liver disease. And then, of course, to tailor or personalize our approach to screening and intervention. But of course, this costs a lot of money, and there is little doubt in my mind that if not now, very soon payers are going to be asking, well, what is this doing to our bottom line? And unfortunately, that's an unavoidable fact of life in medicine. Where are we now? Well, we're really at the very, very early stages of implementing genomic medicine or precision medicine. Currently there are over 65,000 genetic testing products available, but 10 are added every day. The spending currently constitutes 10% of health insurance laboratory costs, but total costs for genetic testing are projected in the United States to reach $15 to $25 billion within two years. The payoff is more theoretical, and as I'll show you, we really have very limited data justifying the investment, even though I think we all agree this is an investment worth making. The payoff, of course, is improved quality of life in many ways, as well as reducing costs by earlier diagnosis and less loss of work, more productive years for our patients. So here's a couple of points, and again, I said the data is still quite limited, and I have little doubt that as we develop these genomic pathways or these pathways of complex integration of data, we should be collecting information about cost saving. It's estimated, at least from this Harvard Business Review white paper, that eliminating unwarranted variations in medical care through precision medicine can reduce the cost of patient management by at least 35%. One hopes that's true. There's very few studies, and in fact, this is one of the only few I could find that project rather than document the potential impact of screening, and this is effectively a study from Australia looking at the relative cost and cost savings if one were to do one cluster of genomic screening, now a second cluster, the cost lowers more, a third cluster that includes other genetic variants, and then finally, if we were to simply combine all of these different tests at one, the prediction is that by doing all of the screening all at once, that ultimately the cost will yield. So welcome to the future. This is really where our medicine will be going, but it also leaves us with some new challenges in terms of data integration. We're living in a world now, a digital data universe. This was highlighted in one of the periodicals, life science periodicals. More broadly, we are in a world where just an extraordinary amount of data is swirling about us and being collected. I thought that Dr. Conwell's example of Waze really fits into this paradigm very well, and within this broad universe of data is the data that surrounds healthcare and personalized medicine, and together, these data sets ultimately require integration, and the promise is this personalized medicine and more effective and personalized therapies. But how are we going to deal with all this data, and this really is where artificial intelligence comes in. This is actually from a diagram we had rendered for a program I just organized at Mount Sinai around artificial intelligence, and I certainly have become much more knowledgeable about the power, the risk, and ultimately the inevitability of artificial intelligence defining many of our principles of care and integration of data. So what is AI? There are many definitions. The one that seems most sensible to me is the simulation of human intelligent behavior by computers with iterative self-learning, and that's really the power of this. It continually integrates more information, refines, refines, refines, and this in turn speaks to the absolutely critical need to not only have high-quality data, but lots of it, and in the case of human disease and biology, as much genetic variation as is possible. The medical applications are already quite evident to many of us, but among all of them, imaging is clearly far and away the most, I would say, mature in the assimilation of artificial intelligence in the decision-making, but this is also going to be true of pathology, creating clinical judgment and diagnosis, streamlining workflow, analyzing relationships between prevention or treatment, and outcomes. And this has not been lost on the commercial or the academic spheres. Tech companies, certainly Google, IBM, Johnson & Johnson, but many, many, many others are integrating AI into whatever their products or platforms are. Many academic centers, and I apologize to those who are omitted from this list, these are just some of the examples that I could easily find, Mayo, Mass General, Stanford, Cleveland Clinic, Mount Sinai, many, many others. Of course, the national health system in the UK has embraced this, and in part have developed an extraordinary biobank, Biobank UK, just as programs in the United States are doing similar efforts through the NIH and the federal government, and of course, many startup companies, some of which are here at this meeting trying to learn more about how to develop predictive models of disease outcomes. So here's an example of how big data can actually change our thinking about disease. This is work from Joel Dudley and his group, actually at Mount Sinai, in which they used a big data integration of not just the genetics, but also our clinical records, our electronic health record, and clinical information of all sorts, lab data, and were able to define what effectively represents three subclasses of type 2 diabetes. So basically, these are the three subclasses shown schematically, and these would not be easily discernible to the clinician in seeing their patient, but they're defined both by three different sectors of phenotypes and genes that cluster together and distinguish one from the other, and this certainly would have critical implications for how you might treat your diabetic patient in terms of what their underlying drivers are and what they're likely to respond to. So the future is that we need to use gene signatures to determine disease classification, risk of progression, disease drivers, and potential targets. We're going to be using genomics more and more for many diseases, but these are three of the most obvious, biomarker discovery, and the use of the electronic health record to translate complex data into clinical decision making. Fortunately, the Obama administration in 2015 made a major commitment to precision medicine to invest in this, and that investment is shown here, $215 million in the year of 2015. And this really brings us to technologies where we can start to think about how all of this complex data actually shows up as a prompt or an advice to a clinician, because after all, none of us have the time or knowledge, if we're examining a patient, to go back and integrate this information. But what if the information is integrated for you, and I think Dr. Cunwell gave a perfect example of that. There are many large-scale efforts. This is one that's most familiar to those in the field in the U.S. It's called the Emerge Network, but suffice to say that it's one of several, here's three other programs around the world, that are also effectively developing tools that distill that genetic or genomic information, integrate it with clinical information, and give you actionable decision making information at the bench or bedside. And so seeing as this is a case-based course, I'm going to close with this case presentation, but it's not from November 2019, it's from November 2029. So here's a patient that, obviously, I made this up, because I can't see the future, but let's see if this resonates with any of us. A 53-year-old female identified by her electronic health record algorithm as high-risk for NASH is referred to her primary MD with abnormal AST on two visits, history of diabetes, hyperlipidemia, elevated BMI. All other indications are entirely consistent with NASH, and an AI-based NASH MRI is performed that's effectively an integration of what may be existing technologies prior to the hepatologist's appointment, and that suggests significant fibrosis. The patient's referred to a hepatologist who has already reviewed the electronic health record. The genomic analysis that's effectively delivered to him or her suggests a disease-associated genotype in that HST gene, but also that the microbiome shows an increase in protobacteria and a decrease in formicides, consistent with what's seen in NASH, and so the patient is prescribed a phage therapy and a small-molecule antagonist of HST17B13. Six months later, the BMI has come down, the AST is normalized, your now imaging test shows moderate fibrosis, the small-molecule antagonist is discontinued, and repeat treatment with phage is implemented. And then finally, 12 months later, the BMI is now 23, AST18, the MRI is consistent with mild to moderate fibrosis, the patient has lost weight, the microbiome analysis now establishes that the improvement is durable and predicts no weight gain and a higher quality of life, and six-month visits are implemented indefinitely with phage therapy as needed, and of course an exercise program at work. And best of all, in my fantasy, the cost to the patient is none, because we will be living in a country where we have a new single-payer healthcare system that assesses it as significantly cost-saving. Thanks for that, that was great. So what are the takeaways? Success stories in precision medicine for hepatology, illustrated by this course, are ample, but there are challenges to overcome. The field is clearly moving towards data-driven diagnosis, classification, and treatment of disease that requires integration of multiple types of data, utilizing artificial intelligence, relying increasingly on genetic counselors, and requiring new skill sets for practitioners. Precision medicine will be most valuable when it leads to earlier detection and actionable information that optimizes prevention strategies and improves treatment, and the best justification for investment in precision medicine will be improved outcomes that reduce costs and enhance patient satisfaction. So finally, I'll leave you with a quote from the founder of Infosys in India, N.R. Narayan Murthy, in which he was quoted as saying, growth is painful, change is painful, but nothing is as painful as staying stuck where you do not belong. So I'm hopeful that we will all belong in the future in this world of precision medicine and hepatology. Thank you.

precision medicine

hepatology

genetic discoveries

predictive algorithms

pharmacogenomics

electronic health records