Modeling Disease: Improving Cancer Diagnosis and Treatment

•Disease Treatment: A patient has symptoms of disease, gets a diagnosis, a treatment, and has an outcome. But this process is actually much more complex. It depends on who the patient is, how engaged they are in their treatment, what they bring to the physician, the environment where the physician works, and the country and culture which surround them. It also depends on who is looking at the process, i.e., the patient, caregiver, physician, pharma, regulator, payer, or researcher. Each stakeholder has their biases. For example, physicians have to be very operational, especially based on economics and time pressures. They are also considering what needs to be done for reimbursement, which may not follow guidelines. Patients want an answer before they leave the doctor’s office. And the treatment process depends on when the clock starts. Usually the clock starts when the patient shows up in the doctor’s office with symptoms, but there is active research for earlier markers that aren't being used.

•Disease Biomarkers: While each of us generates a practically infinite set of measurements continually, what is actually measured is an eentsy subset of that – only the accepted things which can be easily and affordably measured – and even these are only done episodically on say, the next visit to the clinic. It’s like trying to explain a day- long wild melee with a couple of grainy pictures. While they capture something, much is lost. Two patients may present with the same biomarkers or clinical variables, but progress very differently over time. And two patients that are progressing exactly the same may not present with the same value because they were observed in the doctor's office at different points in time. Patients with similar symptoms may have very different illnesses.

•Drug Development: The target for drug development should be a mechanism, something that enhances our understanding of an individual’s path through a disease, but we're frequently dealing with phenotypes (attributes of the patient). For example, a clinical trial showed no effect over six years with a very large population, and analysis of subgroups did not find any groups for which this drug was appropriate. The problem was the definition of the phenotype, and then using that phenotype to enroll patients into that trial. Subsequent analysis identified five completely different diseases. Dr. Liebman challenges us to consider patient testing and treatment decisions as part of an enormously large and complex multidimensional system. This perspective will lead us to consider ways in which the conventional learnings we get from much simpler analyses (like the gold-standard randomized clinical trials, with their own serious biases) can lead us astray. Instead he asks us to build more robust models of disease, using techniques well-known to the modeling profession (e.g., mathematics, Bayesian and other statistical techniques). Then, looking creatively back at existing data, and creatively forward to generate ideas for new measurements, see how these models can be applied to give more granular views of each person’s condition, and therefore more specific ideas for treatment. This is a grand vision. In his talk, which is included below, he discusses the problems and opportunities of models of systems and processes, giving concrete examples of their application and many ways in which the conventional wisdom relies on some dubious foundations. In this, he is encouraging patients to not be afraid to (respectfully) challenge these learnings and perhaps find a better path for their own care. Key Principles to Avoid the Frustrations of a Simplistic Process Model

•Expand the Scope: Think in terms of systems and processes that are bigger than your main focus and link them to everything else. Include temporal relationships that are commonly ignored. For example, how early should we identify a person as entering a given disease path, and what systems, including more peripheral effects, might be relevant? Consider integrated health providers.

•Test the Model: Enable clinicians to evaluate models retrospectively against their patient populations so they can test performance and see if their patient population may exhibit unique characteristics that through feedback could further enhance the model. Individual physician practices may encounter unique patient populations or more diverse populations and for them to gain confidence that a new algorithm might be appropriate, they should evaluate it against previous patients. If it is not effective, then feedback to the developer can aid in refining it, and/or identify unique characteristics of that physician's patient population.

•Humility: The physician likely does not have all the answers and should share their degree of confidence in any specific decision or recommendation in a spirit of collaboration. Transparency and willingness to accept some degree of uncertainty can help form better patient-physician relationships/interactions and lead to better and more informed decision-making. Physicians are focused on finding the right drug for a patient with a diagnosis. Some physicians will admit that they don't trust a diagnosis, but they don't have the option to try to change that. This is an opportunity for empowering the patient and for research. The information and opinions expressed on this website or platform, or during discussions and presentations (both verbal and written) are not intended as health care recommendations or medical advice by Cancer Patient Lab/Prostate Cancer Lab, its principals, presenters, participants, or representatives for any medical treatment, product, or course of action. You should always consult a doctor about your specific situation before pursuing any health care program, treatment, product or other course of action that might affect your health. Meeting Notes Prostate Cancer Lab Meeting - For Precision Medicine, First We Need Accurate Medicine (Michael Liebman) Wednesday, 9/7/2022

•58:05 SUMMARY KEYWORDS patients, disease, physicians, questions, drug, observations, condition, clinicians, problem, understand, diagnostics, groups, guidelines, treatment, diagnosis, process, phenotype, early, complexity, test SPEAKERS Michael Liebman, Jeff Waldron, Brad Power, Brian McCloskey Michael Liebman 00:20 I'm a modeler. I'm trained as a theoretical chemist. I'm not a clinician. But I have experience, both in industry and pharma as Global Head of Genomics and Computational Biology for Roche. I worked on the original HER2 testing with Abbott Diagnostics’ Vysis Group. And I ran a breast center for the Department of Defense, where we did everything from surgery to molecular studies, and I have some academic appointments as well. We've all heard of the central dogma for molecular biology, which is that DNA goes to RNA and then goes to protein. There's also a central dogma for medicine, which is that first a patient has signs and symptoms of disease, and as you all know, they then progress to get a diagnosis, a treatment and have an outcome. But this process of interacting with a patient, with a physician, is actually more complex. Because it depends on what the patient is bringing to the physician. And it also depends on the environment where the physician works. As a point of reference, I'm co-leading an EU effort right now in cardiovascular disease, where we have 30 partners across 15 countries. And you can imagine the complexity of this, across all of those different countries and research groups in just a single disease. The takeaway that I want to leave you with is remembering that biology is about systems and about processes. And it's always much more accurate to describe change than to describe something in an absolute state. That's just a general scientific observation. You may be familiar with the term “phenotype,” and you may have encountered it in your personal healthcare journey. Phenotype refers to an individual's observable traits, such as height, eye color, and blood type. We're going to introduce the concept of a “next generation phenotype,” which refers to how an individual progresses from a non-disease condition into a disease condition over time. That will enable us to bring higher resolution into disease subtyping. Michael Liebman 03:58 An endotype is the subtype of a healthcare condition which is specifically defined by a mechanism. As has been written recently by [FDA Commissioner Robert] Califf, and others, a mechanism really should be the target for drug development. But one of the challenges we have is that we're frequently dealing with phenotypes, not something that enhances the resolution, and understanding, of an individual’s path through the disease. And so that's what I want to try to address. I'll give you an example of what I mean by that. Here is a clinical trial. The outcome showed no effect over six years with a very large population study. Statistical subgroup analysis to try to recover subgroups did not enable them to determine any groups for which this drug was actually appropriate. The problem was the definition of the phenotype, and then using that phenotype to enroll patients into that trial. Because when we applied some novel methods, we found what you can see here: There are actually five different progressions, meaning that there were five different populations, and they never should have been treated the same way. And not understanding that that complexity existed in the population meant that the drug may have worked for some of these subpopulations, and not others. But it also indicates that – in this case it was a heart failure drug – these patients never should have been treated the same way because they don't fundamentally have the same disease. And that should back up not only into the clinical trials, but into drug development. You need to understand that the mechanisms for these five diseases, or five presentations or subtypes of diseases, are actually very different. I'm a modeler with a very coarse model of what a patient journey looks like, and your experiences are a lot more detailed than this. But essentially, you're going from a risk assessment to a diagnosis because of the presentation of symptoms, or clinical signs, with the physicians actually measuring into that diagnosis, treatment, and then a variety of different outcomes. But in reality, you're just cycling back to become a patient again, into possibly another disease or another state. The thing that you need to understand is that, in all of this, You need to understand the question that needs to be addressed. As an economist observed: “If you don't know how to ask the right question, you're not likely to discover anything that's going to be of significant value.” And the question in the patient journey is going to be very different, depending on who's looking at that patient journey. Because each of these groups is emphasizing a different part of this journey, based on what they do, what their daily activities are, and what their priorities are. And none of them are looking at that entire patient journey. What we've focused on is building a model of what this patient journey looks like, the complexity of that, and then allowing different groups to have their individual views, but also being able to expand upon their views, to see how other factors may be actually interrelated to what they need to know. Michael Liebman 08:30 We initially use an ontology, which means we're looking at those concepts, and figuring out all of the factors that are critical for those concepts. This goes many, many layers deeper. You can see what we built is general, and because we can use it across any disease, is something that's disease agnostic. There's no element in here that's specific to any given disease. This can become a model for how we integrate the data to represent and be able to analyze the disease. It also gives you an idea of the kinds of resolution we're looking at. We frequently talk about comorbidities, but comorbidities could be active comorbidities (in other words, another condition that you have, even as you’re being treated for another one, in this case, prostate cancer), or they could be inactive (i.e., previous comorbidities that may have been managed), and they may even be undiagnosed. The complexity of this patient journey is something we try to encapsulate in this form, and we convert to a knowledge graph. There's an ambiguity in diagnosis and staging. If we take a clinical variable, and we take three patients, and we observe them when they come into the clinic (this is for simplicity, it scales up to many many variables), we see two patients may present with the same biomarkers, or clinical variables, but progress very differently over time. And two patients that are progressing exactly the same, may not present with the same value, because they came to be observed in the doctor's office at different points in time. This is a complexity that physicians are always dealing with. Margaret Hamburg from the FDA used to say that clinicians have long observed the patients with similar symptoms may have very different illnesses. It's something that's known, but it impacts a lot of what we need to do to understand the disease. We need to understand that disease is a process that evolves over time. It's not a state, although we tend to treat it as a state in terms of a diagnosis, but a disease process evolves in very, very high dimensional space. Any of your clinical variables, or even your non-clinical variables (environment, lifestyle, social determinants) can impact you as well. What the physician would really like to know is, what's the direction of this? Let’s call it a “vector”. How far are you along that vector and how quickly are you progressing? That is the ideal set of questions a physician would like to ask. But they don't have a way to actually get those details. Mathematically, that means this isn't a vector at all, it’s a tensor [an algebraic object that describes a multilinear relationship between sets of algebraic objects related to a vector space], but we don't have the data to support that. Yet we know that if a patient comes in here, just as I showed in that previous slide, they may get a different diagnosis than if they came in at different numbers along the “disease process” vector, and be managed differently, and have different outcomes.