|Year : 2016 | Volume
| Issue : 4 | Page : 140-143
Delivering precision medicine: Personalization at scale
RowAnalytics Ltd, 3 King's Meadow, Oxford OX2 0DP, United Kingdom
|Date of Web Publication||3-Mar-2017|
RowAnalytics Ltd., Oxford
Source of Support: None, Conflict of Interest: None
|How to cite this article:|
Gardner S. Delivering precision medicine: Personalization at scale. Digit Med 2016;2:140-3
Precision medicine promises to deliver improved patient outcomes and lower health-care costs, potentially benefiting millions of patients and saving global health-care systems tens or even hundreds of billions of dollars per year through improved efficiency, reduced mis-prescription and over-medication and better compliance. Precision medicine's ultimate benefits may, however, be even wider than this, as people learn more about their own disease risks and are given tools that educate and empower them to become active partners in the management of their own health. By helping people to change their own behaviors to remain healthier and independently active for longer we will ultimately reduce the economic and social burden of avoidable chronic lifestyle diseases, and the dependence on social and clinical care, which are continually increasing the cost of care globally, across all developed and developing economies. Delivering the potential of precision medicine will however present a number of challenges to scientists, clinicians, technologists and healthcare systems.
The stakes are large. The societal impact of long-term conditions on patients' and carers' quality of life and their associated economic burden are both huge - 7 conditions (cancer, cardiac disease, hypertension, central nervous system diseases, diabetes, stroke, and chronic obstructive pulmonary disease) cost the US economy over $2.2T p.a. By 2023 this will be $4T p.a. Much of this, for example, £8B of the £10B UK diabetes bill, is caused by avoidable hospitalizations due to inaccurate diagnosis and prescription, complications arising from poor adherence, and adverse drug reactions. About 30% of chronic patient admissions are medication related, 50–70% of which are estimated to be preventable  by more accurately personalizing chronic patients' drug regimens. This could save the NHS £1.35B per year in reduced drug prescriptions alone.
| The Pharmacoeconomic Challenge|| |
The hard way to deliver personalized medicine is always going to be to develop new medicines that suit ever more specific, molecularly determined cohorts of patients. The traditional approach is to discover a novel biomarker, clinically validate it, and then use it to segment patients into corresponding therapy response groups, design a clinical trial and approval process and put a new drug on market. This is hard for a number of reasons– first it takes a minimum of 10–15 years to put a drug on the market and costs upward of two billion dollars. Second, smaller responder populations may improve the efficiency and success rate of approvals, but they will likely mean a reduced economic potential of the drug once on market. Third, having more specific patient cohorts can vastly increase the cost and complexity of patient recruitment.
Beyond that, we are discovering that the closer we look at diseases with a molecular lens, the more complex they appear to be. In oncology, it is often remarked that we are turning one disease into a whole series of rare diseases. While this is undoubtedly the correct view of those diseases, it necessitates a fundamental shift of thinking in terms of the technical and analytical tools we use to investigate disease, and the diagnostic and reimbursement procedures that we use to detect and manage diseases in the future.
| The Analytics Gap|| |
Chronic diseases are caused by complex combinations of factors, which makes delivering precision medicine both challenging and essential, as traditional therapies may not benefit the whole patient population and are more frequently rejected by payers as not providing value for money. Our ability to generate raw data describing individuals and their disease states has increased exponentially in the past 20 years. The most striking example is our ability to use next-generation sequencing or single-nucleotide polymorphism (SNP) chips near-routinely in specialist oncology centers to identify clinical variants in tumors, helping to guide therapy selection based on a patient's genotype.
While cancer treatment may currently be the poster child for personalized medicine, the ability to put the right treatments in a patient at the right time, every time, across a range of diseases is not yet an ambition that has been widely met. There are a number of conditions, for example, pancreatic and ovarian cancers that are very poorly served for effective medications. There are diseases such as some breast cancers where clonal heterogeneity frustrates simple selection of single drug therapies. There are diseases such as bipolar disorder and asthma which, although strongly genetically determined, are very poorly understood and nonspecifically diagnosed. And there are devastating diseases such as Motor Neurone Disease/Amyotrophic Lateral Sclerosis where the disease only has a 5–10% heritability and is typically late onset, suggesting the involvement of multiple environmental, lifestyle, infectious, or other nongenomic factors.
Our ability to identify the factors causing chronic diseases in large populations, even at a purely genetic level, remains relatively basic. The original promise of the Human Genome project, that having access to a large number of sequences of patients and controls would enable us to spot disease associated variants, which would in turn lead to an understanding of disease pathways, identification of intervention points, and discovery of new drugs, has turned out to be overly simplistic. The tools that we use, for example, Genome Wide Association Studies (GWAS), are at the limits of our current computational capacity, and yet have often proved inadequate to represent and unravel the full complexity of disease. Single or two SNP association studies take weeks or months of computation time, and each additional factor considered in combination can add 5 orders of magnitude to the number of possible combinations to be computed.
Even GWAS studies looking for just one or two disease-associated SNPs in large populations with tens of thousands of patients and controls have often failed to find reproducible and clinically useful biomarkers that can be used to accurately segment patient populations. There are a number of reasons underpinning this frustrating lack of success. Diseases are more complex than we originally imagined– rather than the one or two factors (i.e., SNP variants) that we can currently detect, in our current studies, for example in breast cancer, we have evidence that up to 17 factors in combination are driving disease risk and protective effects for individual patients. These signals are highly differentiated between patient and control populations, highly penetrative in patient populations, and reproducible. This level of multi-factor association analysis goes far beyond the existing single gene biomarkers, for eample, BRCA1/2 testing, and may have a profound impact on our ability to make much more personalized predictions for patients.
Even then, disease related factors are rarely entirely genetic, and patients do not present to a clinician with just a single, idealized condition. Patients not only have unique genotypes, but they also have a unique set of other factors such as age, gender and ethnicity, clinical histories, co-morbidities and co-prescriptions, all of which need to be considered as potential disease risk factors. Suggesting that genomics alone can produce personalized advice for all patients and all diseases is obviously wrong. It is the scientific equivalent of making everything into a nail because we have a really good hammer in our hands. We absolutely do need that hammer, but we need to use it along with a range of other tools in a coordinated fashion to achieve the impact that we want.
We can now generate hundreds of assay data points from automated analysis of blood samples in minutes, even while a patient is being admitted to hospital. We can build fully integrated longitudinal electronic health records describing the patient's phenotype, lifestyle, clinical history, co-morbidities, and co-prescriptions. Moreover, we have not stopped in our pursuit of an ever more detailed molecular description of disease, beginning to use new technologies such as epigenetics and phosphoproteomics in the same way as we once unraveled the genome using ever more efficient nucleotide sequencing technologies.
We will need to commit the same resources to build a new generation of analytical tools to enable us to associate all of these multiple heterogeneous factors, at the scale of modern genomics studies with tens or hundreds of thousands of patients and full-genome sequences, for us to make real progress in delivering precision medicine. These new analytical tools will inevitably be based on different mathematics, computational architectures, and assumptions about diseases.
| The Next Generation of Clinical Decision Support Tools|| |
With all of this knowledge about the patient and their disease state available, it is paradoxically harder than ever to be a clinician. The use of these data to deliver precision medicine routinely at the point of care has lagged a long way behind the volumes of data available. The data streams that are potentially at our disposal equate to several fire hoses trained on the clinician, and yet they have almost no tools with which to integrate and interpret these data and make them useful and actionable in the context of a highly time-constrained consultation with a patient.
This is in part due to the inevitably slow development of new therapies and companion diagnostics, but mainly due to a lack of accuracy, scalability, and flexibility in developing decision support tools for clinicians. Even with huge supercomputers, the technical challenge of routinely personalizing decisions at the point of care is daunting due to the complex, combinatorial nature of the hyperdimensional datasets, and the diseases themselves.
This key obstacle to the routine delivery of precision medicine is not however intractable and real progress is being made in several quarters. Various methodologies, from large scale EHR and health analytics platforms such as Kaiser Permanente's HealthConnect  and large-scale deep learning and AI technologies such as IBM Watson  and DeepMind, to a new generation of more nimble point-of-care focused clinical decision support tools, are becoming available.
The ultimate challenge however is likely not to be technical, but cultural, with clinicians' training needing to emphasize that the complexity of disease is too great for them to ever be able to properly diagnose and treat an individual patient without a range of clinical decision support tools to help guide them in selecting their treatments.
| Giving Patients Tools to Manage Their Own Health|| |
Oddly in all of this massive technical, scientific and clinical complexity, the easiest and quickest way to have a meaningful impact on the health of a population may be to develop relatively simple, patient-focused digital health tools, giving them actionable advice that enables them to engage with their health and change behaviors that lead to disease. While the first generation of digital health apps were at best single disease focused, the next generation of digital health apps will help patients understand their personal combination of disease risks and educate them on managing their lifestyle, creating informed and active partners in managing their health.
One such example is the HealthySwaps app (www.healthyswaps.diet), which gives fully personalized advice to patients on the foods that have the potential to interact badly with their individual combination of multiple diseases and prescriptions. HealthySwaps uses a systems pharmacology model to identify all the potential drug-drug, drug-disease, and drug-food interactions that a given patient might encounter. It can suggest alternative products in the same grocery category that do not pose risks to the patient. All of this functionality can be delivered on a patient's own phone/tablet, with their personal data not needing to be transmitted to a server, so that it remains securely in the patient's possession at all times. This is an example of new digital health tools which put sophisticated decision support in the hands of patients, carers and clinicians at the point of care or integrated into their day to day lives.
The cumulative impacts of these new technological developments have the potential to transform the delivery of medicine from a discipline that changed relatively slowly up until the end of the 20th century, to one which has the potential to deliver best clinical practice to every patient every time. The biggest remaining challenges are as always, gaining secure access to high quality, multidimensional patient data, building better analytical tools, and changing the way in which medical knowledge is delivered.
| References|| |
DeVol R, Bedroussian A. Unhealthy America: The Economic Burden of Chronic Disease. Altarum Institute Annual Research Report; 2009.
National Diabetes Audit 2012-2013: Report 2 Health and Social Care Information Centre. Diabetes UK; 2015.
Barnett K, Mercer SW, Norbury M, Watt G, Wyke S, Guthrie B. Epidemiology of multimorbidity and implications for health care, research, and medical education: A cross-sectional study. Lancet 2012;380:37-43.
2016 Profile. Biopharmaceutical Research Industry. PhRMA; 2016.
Al-Chalabi A, Calvo A, Chio A, Colville S, Ellis CM, Hardiman O, et al.
Analysis of amyotrophic lateral sclerosis as a multistep process: A population-based modelling study. Lancet Neurol 2014;13:1108-13.
Moore JH, Asselbergs FW, Williams SM. Bioinformatics challenges for genome-wide association studies. Bioinformatics 2010;26:445-55.
Liang L. Connected for Health: Using Electronic Health Records to Transform Care Delivery. San Francsico: Jossey Bass; 2010.
Chen Y, Elenee Argentinis JD, Weber G. IBM watson: How cognitive computing can be applied to big data challenges in life sciences research. Clin Ther 2016;38:688-701.