An operating system for cancer
By Richard Sallari, Kevin White & Eric Lefkofsky
Introduction
Since the initial draft of the human genome was released in 2000, we have collected an unprecedented amount of genomic data as a community. Comprehensive views of cancer genomes were first produced in 2007. For the past decade, a great deal of our scientific energy, as it relates to mining this data, has been focused on the hunt for driver genes that can cause cancer, known as oncogenes and tumor suppressors, and using this biological knowledge to develop new targeted therapies and therapeutic strategies. But perhaps the most important lesson we have learned so far is that we will need much more data in order to fully understand cancer.
Nixon declared war on cancer in 1971. For the past 45 years, we have invested hundreds of billions of dollars in understanding the drivers behind cancer. The government alone has invested over $90 billion since 1995 [1]. After 2000 our pace of identifying oncogenes and tumor suppressors accelerated, and yet today we still have only identified around 250 cancer driver genes [2], and only about 27 [3] are well established clinically and directly tied to FDA approved drugs. To put that in perspective, there are over 20,000 genes and there is an order of magnitude more gene-controlling DNA sequence in the portion of the genome that does not encode proteins (also referred to as “non-coding”); so after all this time and money, we still have only mapped a small percentage of the genome as it relates to oncology. Although it has been argued that a plateau has been reached with regard to identifying the most common cancer driver genes [4] we have a long road ahead before we have a comprehensive map of all segments of the genome that are relevant to cancer. And we have barely started mapping genes that drive metastases or the associations between genes and therapies in a formal and systematic manner, in large part because the process requires vast amounts of data that up until now have been too expensive to collect. The comprehensive search for driver genes in primary tumors alone is estimated to require 100,000 patients [2]. If we also consider metastasis and therapeutic associations, a mapping of cancer biology and patient response will require the joint analysis of even more patients.
The requirement for large data sets is due in part to the heterogeneity of cancer. Each patient’s tumor is, to a great extent, unique; each cancer a novel manifestation of the relentless force of evolution thrust against the individual. Cancer is also heterogeneous within each patient. A tumor is often made up of many different cell populations, some capable of resisting treatment and others able to metastasize to distant locations; and these mutant cells often cohabitate within the same tumor geometry. And yet today, the largest genomic data sets we have amassed are through The Cancer Genome Atlas (TCGA) and International Cancer Genome Consortium (ICGC) and cover roughly 15 thousand unique exomes (the 1.5% of the genome that codes for protein) [6]. Unfortunately, the difficulty of cancer biology makes this data inadequate if we are to envision an end to cancer in our lifetimes; the volume of patients required is far beyond what any single hospital or institution can collect or analyze. The complexity of the cancer problem requires a highly scalable technology solution.
An Institutional Problem
The comprehensive elucidation of cancer biology amounts to a massive search algorithm executed in parallel across hundreds of institutions globally. Universities, hospitals and companies are all probing the myriad facets of cancer biology in search of insight. Some researchers specialize in a given cancer subtype, like lung or breast, others tackle a specific gene or biomarker, while others put their faith in novel drug discovery. Most of these individual efforts are rational and intelligent, often brilliant, but at the macro scale the current approach represents a brute force implementation, where each institution is looking for a piece of the puzzle largely blind to other efforts.
The biology of cancer is difficult, as the highly redundant system that evolved to keep us alive transforms into the very enemy we face when fighting cancer. Our bodies have adapted to ensure that there are numerous ways to achieve critical functionality. Mutated and dysregulated genes that drive cancer are a part of that complex system; as such, the task of the massive search algorithm embodied in the basic, translation and clinical research of the global cancer community is to not only identify them all, but also to understand their relationship with the circuitry of each cell type, including how they affect and relate to the non-coding portion of the genome. And while identifying driver genes might seem relatively straightforward now, the combinatorial nature of cancer therapies (chemo, radio, targeted or immune) will likely require exponentially larger datasets to discover novel therapeutic associations. This is due, in part, to the curse of dimensionality where, as the degree of nuance in the patterns we search increases, the amount of data available to interrogate each of the patterns becomes sparser.
We are thus confronted with a statistical hurdle that cannot be easily overcome. The inefficiencies of the current macro strategy are frequently referred to in the community as “misaligned incentives” [7]. But the truth is that the cancer community, which is composed of patients, nurses, doctors, researchers, professors, legislators, lobbyists, insurers, advocates, and many other professionals is a complex system in its own right, with emergent behaviors that are not entirely predictable, or always operating in a rational manner. And because the system is riddled with regulation, and highly dependent on academic institutions and governmental agencies, the pace of innovation and adoption of new technologies has lagged behind other sectors.
An Industrial Solution
Aligning the cancer community might be as complex a task as curing cancer itself, but simply stating its flaws is an exercise in futility. Any effort to realign the system will require, at the very least, a platform for everyone involved to orient themselves around data (molecular, phenotypic, and therapeutic), that can be accessed and analyzed in both the clinical and research setting in a frictionless manner. What is needed is an industrial scale systems biology platform for cancer care, with state-of-the-art decision support tools for oncologists, pathologists, and surgeons. A platform to gather large amounts of molecular data, combine it with phenotypic and therapeutic data, analyze it in search of clinically relevant patterns, and ultimately test alternative therapies in patient-derived biological models. An operating system for cancer, if you will.
With the volume of clinical trials and new therapeutic options rising, especially those tied to molecular data, oncologists have the nearly impossible task of keeping up with ever-increasing amounts of data being generated during the course of a patient’s treatment. This problem is aggravated by a lack of technology and infrastructure that physicians should be able to access. Most Electronic Medical Record (EMR) and Electronic Health Record (EHR) systems were designed decades ago and have failed to keep up with modern software and database architecture improvements. Data is highly siloed and often locked, analysis is cumbersome and requires significant time and bandwidth, and feature development and support tools are lacking. On average, physicians spend one to two hours on EHR and desk work for every hour of direct face time with patients. EHR burden is correlated with physician burnout, which is on the rise at a staggering 54% in 2014 [8]. This burden is likely to continue increasing, as most of the relevant patient data is actually in free text fields and is nearly unsearchable and unanalyzable without sophisticated Natural Language Processing (NLP) tools or advanced machine learning. As such, physicians need software and analytic tools that work within a hospital’s existing infrastructure to analyze data and create a roadmap for patients who are not responding to the standard of care. Today, despite advances in genomic sequencing, physicians still lack the basic technology to analyze molecular data and use basic pattern recognition tools that are prevalent in other industries to help them deliver more effective and personalized care.
The technologies needed for such a system already exist. We can now sequence patients and gather rich amounts of molecular data at a fraction of the cost required just a decade ago [9]. We have the tools needed to bring deep learning and big data analytics to cancer, as the core infrastructure necessary to compile and analyze large data sets is within our reach. Additionally, researchers have made incredible strides in the areas of Patient Derived Xenograft (PDX) modeling, organoid development, and microfluidic organs-on-chips such that we can build highly predictive in vitro and in vivo models of a patient’s tumor. All that is needed now is for someone to invest in scaling these technologies and deploying them throughout hospitals worldwide. Once the system is connected and data is aggregated, the problem becomes computational. How do we find patterns, associations, and connections among what appear to be unrelated or loosely related data sets? These problems are outside the area of expertise of biologists and chemists; but they are the everyday work of mathematicians and computer scientists.
As this new and budding operating system is constructed, marrying vast arrays of omic data with patient outcomes, we proffer that patterns will be recognized that will offer clinically relevant avenues for physicians and researchers to explore. Associations that have been historically invisible due to the limits of existing public and private databases will become recognizable with scale. Once discovered, validating these patterns will involve equal effort in the areas biological modeling and clinical trials. We recognize that there are numerous roadblocks and bottlenecks between discovery and clinical implementation. This is why we are deeply committed to both data-driven discovery and rigorous experimental validation.
A Biological Problem
The difficulty of cancer biology
The vast majority of cancer patients die from invasion of their critical tissues by metastatic cells. In contrast to many other diseases, these metastatic cells are not acquired from other patients, nor are they present in the patient at birth. Metastatic cells emerge and evolve from normal cells and do so independently from the cancer cells in any other patient; in each case, their trajectory is unique. With many other diseases we confront the end result of an evolutionary process. With cancer, we confront the evolutionary process itself. This is the source of difficulty for cancer biology. It is a disease of the genome, a product of molecular progression where apoptosis (programmed cell death) and mitosis (cell division) are hijacked, along with other hallmark molecular systems that in health allow us to grow, heal, and remain in balance [10]. As mutations occur in the genome, our molecular equilibrium is compromised, and cancer ensues.
Mutation is the driving force behind biological evolution. The accumulation of mutations in the genome of normal cells can eventually lead to the formation of a primary tumor. This mass of cells can be initially benign, simply expanding inside the body without harm to its host, but it can also create an environment that fosters the exponential increase of mutations and variability in the cells that compose the growing tumor. The primary tumor is thus a springboard for more aggressive cells to evolve, as the environment is now ripe for cells to grow, propagate and survive in an uncontrolled manner. It is therefore important to understand the genes that, when mutated, drive the formation of primary tumors. However, it is even more critical to understand the genetic mutations and cellular states that lead to metastasis. Fortunately, despite every cancer being unique, there are some similarities between tumors. These are likely due to inherent genomic vulnerabilities, inherited variants that increase risk, environmental carcinogens, and convergent evolution. Mapping these similarities is critical to understanding the origins and progression of cancer.
The unfinished map of primary tumor drivers
The discovery of a single driver gene is usually the first step in the development of targeted or immune therapies. To date, despite considerable effort and vast sums of research investment, we have found about 250 statistically rigorous driver genes out of the pool of over 20,000 protein-coding genes. Through great feats of coordination, sequencing and computing, massive international research groups have defined a list of the most common cancer driver genes. However, they have also uncovered a “long tail” of recurrently mutated genes that also appear to bear the hallmarks of cancer drivers. Understanding how the common aberrations in these genes relate and interact with each other, and with the long tail of increasingly rare but recurrently mutated genes, is a major frontier facing cancer researchers today. This challenge requires large sample sizes and rigorous statistical models; methods that separate signal from noise in cancer genomes that can account for a variety of confounders, such as mutational heterogeneity across the genome. However, issues of statistical scale are the main hurdle if we are to identify the majority of cancer genes. The recent estimate of 100,000 patients across all cancer types will allow us to identify approximately 90% of driver genes affecting 2% or more of patients [2]. Current analyses for single tumor types are based on data sets of a few thousand exomes. These data sets are the result of nearly 10 years of cancer genome data accumulation. So how long will it take the community to collect ten times more data? In order to shorten the natural time horizon to break the statistical barrier of cancer genomics, we need to accelerate our sequencing efforts. For example, more than 60% of oncologists have never requested tumor sequencing in breast cancer [11]. Among those that do, the majority do so in less than 5% of their patients; the primary limitations to use being access and funding [11]. Service providers that offer sequencing of even a few hundred genes are still prohibitively expensive and slow, despite significant hardware advancements that should have produced a greater impact on reducing prices. The first major challenge is so large it is beyond the scope of any single institution.
The uncharted landscape of metastasis drivers
While the discovery of the complete set of driver genes in primary tumors will take time, we have already found many of the frequent culprits for major cancer subtypes and at least there is a trajectory to accumulate the necessary data, albeit over a long period of time. The search for metastatic driver genes, however, is less developed. The vast majority of sequenced tumors in public databases are primary tumors that lack the additional transformative mutations that metastases acquire. Metastatic cells are the culmination of a long arms race with the host’s immune system, often over several years. These cells have been shaped by each patient’s unique battle against cancer; the drugs they took, their response, their environment, their diet, etc. They are therefore more heterogeneous and more evolved than primary tumor cells, and capable of adapting to a variety of therapeutic regimens. The mapping of all driver genes for metastasis may require an additional hundred thousand patients, if not many more. But for now, despite being responsible for the deaths of more than 90% of cancer patients, the metastatic space remains unmapped with virtually no drugs available for metastasis-specific gene alterations [12]. Many physicians continue to prescribe the same drugs to patients in an often linear fashion, as if the primary tumor was still the culprit years later. In many cases, it is not. The tumor has mutated in a manner that makes it largely disconnected from the initial disease. That said, just as we continue to find common driver genes in primary tumors, it is probably a matter of time before we identify common genomic culprits and cellular states that are driving metastasis. The same solutions that will speed up the process of data collection on the primary tumor side are necessary to help metastatic patients.
Bottlenecks in drug discovery
Side effects arise frequently when treating cancer because, by its very nature, the cancer we target is a part of us, and by “treating the disease” we often are poisoning ourselves. Survivors understand the hardships of untargeted chemotherapy and radiation treatment. In the cases where a targeted therapy exists, the complications and side effect profiles are lower by orders of magnitude, because what is being treated is the thin margin which differentiates the tumor’s biology from that of the patient. Identifying a new cancer gene is a significant scientific breakthrough, but it has little benefit for cancer patients unless it, or the mayhem it promulgates throughout the cell, can be targeted by a drug. Despite having identified dozens of cancer driver genes, only a small fraction of them are clinically actionable. As we perform systematic, scaled-up studies and identify even more cancer genes, we will have to design drugs for each of them. However, designing a drug is even harder than identifying a cancer gene in the first place, and it is getting harder every year. The cost of developing a new drug doubles approximately every nine years [13]. This problem is known as Eroom’s law, which is a reverse spelling of Moore’s law that described the doubling of components in integrated circuits every two years in the semiconductor industry. The ecosystem in which drugs are developed is routed in regulation, medicinal chemistry (which is as much a practitioner’s art as it is a science), and a statistical methodology developed nearly a hundred years ago that relies on outcome data collected over a long period of time and validated through controlled, randomized trials. It is a system that made sense when we were largely testing highly toxic chemotherapies, but that must be re-examined in today’s world of emerging targeted and immune therapies. When you take into consideration the number of chemical compounds implicated in oncology procedures (over 10,000) and their varying degrees of toxicity, pervasive institutional regulation and patient heterogeneity, it is no wonder that less than 200 cancer drugs have been approved by the FDA over the past 20 years [14].
Roadblocks in treatment data access
Unfortunately, identifying a cancer gene and developing a targeted drug for it are still insufficient to produce an effective and viable therapy. An additional layer of variability still needs to be overcome as every gene can be mutated in thousands of different ways. Additionally, every patient has their own genetic background, and every tumor might find a way to evade the drug’s effect. In other words, even once you find a suitable target and a patient that initially responds, the cancer’s genome can mutate again to produce new tumor cells that are resistant to whatever therapy was developed in the first place.
The mapping of all patient responses to therapy as a function of their genomes is largely uncharted territory. The confounders can be imagined, but remain untested, as we have no rigorous statistical framework with which to estimate the power of current approaches, let alone to determine which new approaches would be necessary to obtain a comprehensive map. Because response is a combination of multiple genes and drugs, the combinatorics are much larger than for identifying single cancer drivers. Hence it is highly likely that the number of patients needed for such an endeavor will be even higher than those needed to identify all driver genes in both primary and metastatic tumors. The exploration of response and resistance is a daunting and ill-posed challenge. It requires something that currently does not exist at scale in any public or private data base, namely a massive data set that combines molecular, phenotypic and therapeutic data, that is also tied to patient derived biological modeling so we can boost power and rapidly analyze disease and response, and so that we can iterate in a time frame that mirrors the patient’s own medical timeline. If a typical stage 4 metastatic patient has two years to live, we cannot take 10 years to test a small handful of therapeutic options; it is just too slow. We must build a system that is capable of real-time learning.
Outdated strategies in clinical trials
Mapping out the entire space of genes, drugs and responses is a massive undertaking that is beyond the scope of any single hospital or research institution. Given the current pace of collaboration, this could take decades to accomplish, but there is hope that disruptive technologies might be able to break through these statistical barriers and bring us to a solution much sooner. Therapies that, unlike a drug, can be tailored dynamically to the specific needs of a patient as their cancer evolves have great potential. T cells engineered to recognize tumor antigens, therapeutic CRISPR aimed directly at leveraging cancer mutations, or anti-metastatic drugs that prevent progression beyond a primary tumor are just a few examples. But again, these solutions require large amounts of data and a modern technology platform in order to be fully tested and implemented in a clinical setting, as they are a hostage to the same hurdles that all other therapeutics face today.
At the heart of pharmacological progress is the clinical trial. Clinical trials provide the basic evidence of efficacy and outcome that regulators, providers, and insurance companies rely on when they determine whether or not to approve a drug, prescribe a drug, or reimburse a patient for taking a drug. Trials are incredibly slow and expensive for a myriad of reasons, but perhaps the most important one that relates to our lack of rapid therapeutic progress is the mismatch between disease, trial design, and patient enrollment. Trials provide the evidence of efficacy of a drug within a specific clinical indication and are currently designed around cancer subtypes, for example: women, pre-menopausal, who have stage I or II breast cancer, are node negative, and hormone positive. Clinical trial criteria regularly do not take into consideration molecular characteristics. In others words, the trial is not further limited to patients that fit the phenotype above and also have a PIK3CA and RB1 mutation. And yet, it is the very presence of these unique genomic elements that often governs which patients will and will not respond in the trial.
Trials are limited by cost, both of enrollment and validation. It is typically so expensive to enroll patients, and so time consuming, that drug companies often prefer to design trials that are broadly defined. The cost of conducting a phase I trial (where a drug is first screened in humans; primarily concerned with safety and tolerability) can range between $40,000 and $60,000 per patient. As the trial advances, the costs rise. A typical phase III trial (where a drug is tested for its value in clinical practice) can cost between $70,000 and $125,000 per patient [15]. As a result, pharmaceutical companies try to avoid small or targeted trials, which are problematic to them for several reasons. First, although a small trial might be more likely to be successful, it implies a small market size and drug companies are sensitive to making large investments if the end market is not big enough to support the up-front cost. Second, the more targeted the trial, the harder to find and enroll patients and validate their results. As a result, some potentially effective therapies might not even reach phase III because it is too slow and costly to conduct the trial.
The perplexing fact, however, is that because of cancer’s heterogeneity, a large collection of very specific drugs might be exactly what we need; and the more specific, the smaller the relevant market becomes. It is likely that to maintain therapeutic effectiveness, drugs will have to become increasingly tailored as they target smaller subsets of the patient population. This is a natural consequence of cancer heterogeneity and the long tail problem, whereby sets of increasingly rare driver genes, in aggregate, are crucial in driving the cancer’s growth in a substantial number of patients. As a result, if we are ever going to combat these “mutations of unknown significance” in genes that are currently not oncogenes but might one day be recognized as true drivers, our clinical trials will have to become narrower, more specific and molecularly driven in order to maintain their effectiveness. Basket and umbrella trials offer a possible solution to this problem, assuming the baskets can be dynamically subdivided according to the latest molecular sub-classifications emerging from the cancer genomics community.
Combinatorics in treatment pathways
Clinical trials determine which new investigational drugs are allowed to enter medical practice, but the unit on which value is measured in a clinical setting is not a drug, but a treatment pathway. Combination therapies hold even greater promise if we can use data and analytics to prescribe custom, genome-guided cocktails that are tailored to each patient’s unique molecular composition. For example, we have already seen progress with the combination of Palbociclib and Letrozole which doubles the progression free survival rate in certain breast cancer patients. Combining Ipilimumab and Nivolumab increases survival by nearly five months for melanoma patients. In multiple myeloma, the combination of Revlimid, Velcade, and Dexamethasone became the standard of care in 2010, after producing such dramatic results that nearly all patients who received the cocktail went into remission [15]. Increasingly, cancer therapies will be administered in combinations and pathways that add a final layer of complexity to the quantification of their effectiveness. In order to accelerate and disrupt cancer care we must integrate the whole process: from the discovery of novel driver genes in primary tumors to tracking the value of treatment pathways in patients. Despite its imposing difficulty, all aspects of the cancer problem hinge on the access and aggregation of molecular, phenotypic and clinical data.
A Data-Driven Solution
Unify siloed data sources
As previously stated, in TCGA and ICGC alone, we have collected about 15,000 exomes to date. Yet this data is devoid of rich therapeutic and phenotypic data collected and observed over long enough patient timelines to be of optimal clinical use. It is akin to collecting 15,000 different locks and trying to unlock them with a handful of keys. We can clearly see, especially in primary tumors, the vast mutation set that is affected at a genomic level when someone develops cancer, yet we cannot follow these patients as they are being treated and see how their tumors respond and adapt to different forms of therapy, allowing us to track the evolution of their disease as they regress or metastasize.
For the first time in our history, we have the ability to generate, store, and analyze genomic data affordably at scale. The cost of sequencing a patient’s genome 15 years ago was roughly $100 million; today, that same sequencing can be performed in days for around $1,000, representing a million-fold decrease in the cost of collecting genomic data [16]. This means that we are able to sequence the 100,000 patients that might be necessary to uncover the majority of primary tumor driver genes for less than the cost of the original human genome. Similar advancements in the underlying cost effectiveness of big data allow us to peer into the vast array of therapeutic data we have collected. Each year, nearly 1.7 million new cases of cancer arise in North America, and roughly 17 million people are treated in totality [5]. With the propagation of EMRs, there is a large volume of patient data that has been digitized and can be mined using NLP techniques and machine learning. By marrying these two large and growing datasets (omics data on the one side collected from sequencing DNA, RNA, protein, etc. and phenotypic and therapeutic data on the other side extracted from EMRs), we should be able to amass a dataset large enough for us to see patterns emerge that were historically invisible to standard analytical techniques. We have the capability to gather unprecedented amounts of data and utilize intense computing to understand cancer at both the macro and micro scale; to prospect the boundaries, peaks and valleys of the cancer landscape.
The primary challenge lies in aggregating the data. First, we are talking about a lot of data. A whole genome can be ~200 Gb of data per patient; so to amass a library of exomes and genomes, we need to inexpensively capture and store genomic data. Second, patient data may be in digital form residing in an EMR system, but the most useful data is generally unstructured and these data sets are growing at a rapid rate. It is estimated that medical information is doubling nearly every five years, and much of the data is in free text fields that are both hard to query and devoid of a common language schema (in other words, doctors do not always use the same words when describing something). That said, the data can be collected, cleaned, stored, and normalized.
Provide deep genome sequencing at scale
As we have discussed, the inherent difficulty of cancer biology requires large volumes of patients and a rich patient characterization, not just at a genomic level but also at a phenotypical level. But, even if we collect a million data points for each patient, if these are too noisy or are missing essential attributes, the size of the data will inhibit rather than enhance our approach. Any data collection effort needs to satisfy the statistical power requirements for identifying the driving genomic factors in a given cancer while ensuring an accurate, rich and reproducible representation of the patient’s disease. This begins with the need for a low-cost sequencing solution that is universally available in both the clinical and research setting. We need to collect large amounts of DNA mutation data (a panel of a few hundred genes is just too small), and of transcriptome data, concurrently. To understand protein dysregulation and interaction, we need to gather data on both DNA and RNA. At the same time, the data we gather has to be of high quality, which in the world of sequencing means high depth of coverage (ideally at least 150-200x for whole exome, and 400-500x for larger gene panels).
Probe unknown drivers beyond gene panels
Gene panels are the current de facto standard in clinical cancer sequencing, as they are effective at revealing mutations in genes that are relatively well understood. However, gene panels limit the opportunity to learn beyond the biology that is already known. There are much richer sources of information such as exome or whole genome sequencing, not to mention the contributions that could be made by measuring the proteome, epigenome, and microbiome. Many more samples are sequenced for genomic DNA than for products of the genome, like RNA transcripts and proteins, because both transcriptomes and proteomes are mutable and noisy, as opposed to DNA whose mutations are permanent and encode the natural history of the tumor; not to mention that adding these other forms of sequencing is more expensive. That said, while panels can be effective tools, especially in the short term for gathering genomic data at a coverage level that is high enough to allow in depth analysis, they are far from sufficient to fully understand the complexity of a tumor. We need to collect broader sets of data, and we need to collect it across a large population of both prospective and retrospective patients. It is imperative that we operationalize and standardize the collection of molecular data, so it can be combined with EMR data for analysis. For this to occur, we need to make sequencing a routine practice in hospitals and cancer centers. It cannot be a “nice to have”, it has to become a “must have”. If we can find a way to sequence a large enough number of cancer patients, then we can build a truly transformative data repository in just a few years.
Characterize patient phenotypes and timelines
This unprecedented volume of molecular data (assuming we can find a way to capture it) will provide the statistical power to map the effectiveness of therapies as a function of patient genomes. Understanding the spatiotemporal dynamics of the tumor genome is especially critical in mapping its response to different therapeutic agents. Here we are not looking for single signals of therapeutic selection across many genes (n to 1) but associations between a myriad of genes and biomarkers with hundreds of therapeutic agents (n to m). Furthermore, both genes and therapies determine response in a combinatorial fashion, requiring a large number of patients, so as to be able to slice the data over thousands of attributes and combinations. By marrying molecular data with phenotypic and therapeutic data, combining both genomic and clinical markers, a new platform could emerge that supersedes both basket and umbrella trials in its ability to identify response in both large and highly specific patient populations, across groupings of any type.
Discover patterns in patient response
As we amass data, we will begin to see patterns emerge. The addition of rich clinical data to phenotypes and outcomes will allow us to identify previously unknown associations. For example, which mutations in circulating tumor cells in blood lead to metastases in the lung? If metastases occur in the brain, what is the clonal structure in the tumor and does it vary based on the the primary tumor? When a primary prostate tumor is treated with drug A, how many types of resistance exist, and which drugs seem to promote or inhibit them? As a breast cancer tumor metastasis produces new clones, what makes them susceptible to a particular drug and do they vary depending on whether or not the patient was a smoker, or had a thyroid condition, or was taking diabetes medication? As a lung cancer tumor spreads to the bone, does its fundamental circuitry and proteomic profile continue to align to a lung cell, or should we consider reclassifying the patient based on the molecular signature of the new tumor? These are just some of the questions that we can begin to answer once we have collected and analyzed the appropriate variants within our new data set.
Rigorously validate novel insight
As treatment hypotheses are generated, we will inevitably need to test them in a biological setting (synthetic or living) in order to validate our findings; biological systems are far too complex to assume that all discoveries made in a computer dry lab can be replicated in an experimental wet lab, let alone in a patient. In order to properly replicate the complexity of a tumor and its microenvironment, we must attempt to mimic its three dimensional structure so as to account for the interactions between cancer cells and the surrounding normal tissues, or find more effective tools for in vivo modeling. One way this can be accomplished is with 3D cell cultures, organoid systems and perhaps eventually with 3D printing of cells. Furthermore, we need to supply the counterforce of an immune system to understand the many evolutionary drivers in the tumor. Injecting cultured tumor cells into mice to create a PDX can replicate some of the aspects of the tumor’s biology within the patient’s body. However, key factors are missing, especially since the mice used for growing PDX tumors typically have no immune system. In vivo models can be humanized to varying degrees of fidelity, but in vitro models, especially microfluidic organs-on-chips and organoids, offer perhaps the greatest hope for testing environments that match a patient at scale. The combination of cost-effective in vivo modeling, and advancements in 3D organoids and microfluidics, will provide a host of new tools to validate therapeutics in both the clinical and research setting.
Accelerate learning of cancer biology
In a world where doctors are making decisions with the assistance of massive datasets that are impossible to interpret without the aid of a computer, a system must be put into place that dynamically learns with each new set of data points. This is the realm of Artificial Intelligence (AI). AI is not a single approach; it can use a combination of probabilistic methods, statistical classifiers and deep learning, to extract patterns over structured and unstructured data. Using this smorgasbord of computationally intensive tools, AI has taken an expanding role in almost every field where there are copious amounts of data. For example, the current resurgence of neural networks (more frequently referred to as “deep learning”) has been partly due to algorithmic improvements but to a great extent to increases in computational power and expanded learning corpora. Image and speech recognition are two research areas that have benefited spectacularly from deep learning approaches. Face and voice recognition software are now pervasive and perform at levels that are uncannily similar to those of a human.
In contrast to images and speech, machine learning in cancer genomics is still nascent. One of the primary requirements in effectively teaching a machine to learn is knowing the right answer beforehand. With an image, we are teaching a machine to learn the patterns that we recognize effortlessly and once the machine output is produced, we can easily validate that the results are correct. With cancer genomics, we cannot be sure we have the information necessary to even begin to solve the problem. However, there are reasons to believe that deep learning is well suited to cancer genomics. Deep learning works well with signals that are compositional hierarchies [17]. As cells integrate stimuli, information flows and converges until it reaches a handful of molecules that determine the state of the cell. The operations guiding this hierarchical integration, and its compositionality, are still to be resolved, but deep learning has the potential to help us organize the vastness of these biological interactions. If we see even the most modest degree of learning, perhaps by identifying small sets of extreme responders across dozens of hospitals, our efforts will quickly translate into hundreds of lives being saved.
Conclusion
In our daily lives, we have the power of technology at our fingertips. Algorithms, in the form of cutting-edge analytics and sophisticated software systems, help us navigate our social communities, explore books, music and movies, and search for local businesses and activities right from our phones. However, technology has not permeated healthcare, and in particular cancer care. Many cancer patients are still treated with a one-size-fits-all approach that is eerily similar to the way patients were treated many decades ago.
If we hope to have an impact on the nearly 1.7 million people who will be newly diagnosed with cancer this year in the United States, we need to disrupt the system; and that disruption begins by assembling all the necessary components of an operating system that unifies the collection and analysis of clinically relevant data.
When the personal computer was first built, it was just a pile of sensors and circuit boards, until someone wrote the first operating system. That system connected the keyboard and the screen, it fired up processors when the power was turned on, and provided an interface that allowed the user to program the machine and create their own algorithms. It unified a bunch of disparate functions into a cohesive experience.
We need a similar system in cancer care today. A system that connects anatomic pathology with molecular pathology, and genomic data to therapies and outcomes. A system that communicates bioinformatics and computational biology outputs to a patient’s physician or tumor board. A system that integrates validation and modeling with its analytics engine.
This is exactly what Tempus has built: a system that unifies disparate technologies and isolated activities and provides the basic technology infrastructure to offer a seamless experience to a varied array of users, especially physicians that need to consider numerous options to dispense personalized care. In other words, an operating system to battle cancer.
We have recruited a team of accomplished geneticists, data scientists and engineers who have developed software and analytic tools that work within a hospital’s existing infrastructure to augment the care that physicians are able to provide – arming healthcare providers with data and insights to help them make real-time, data-driven decisions.
We have built a platform with the capacity to analyze the molecular and clinical data of millions of patients fighting cancer. With this, we can provide physicians with the insight generated from those who have come before.
The first step to personalizing medicine is to gather the necessary data one would need to customize care in units of one, which requires a common platform combining rich molecular, phenotypic, therapeutic, and outcomes data. Only through the universal adoption of a truly ubiquitous learning system can we hope to lay the foundation for precision medicine in cancer care.
Acknowledgements
We thank Casey Frankenberger for his comments on the manuscript.
References
1 Reuben, Suzanne H., Milliken, Erin L., Paradis, Lisa J., “The Future Of Cancer Research: accelerating scientific innovation, President’s Cancer Panel Annual Report 2010-2011”, http://deainfo.nci.nih.gov/advisory/pcp/annualReports/pcp10-11rpt/FullReport.pdf
2 Lawrence, Michael S., et al. “Discovery and saturation analysis of cancer genes across 21 tumour types.” Nature 505.7484 (2014): 495-501.
3 MD Anderson Cancer Center Personalized Cancer Therapy Knowledge Base for Precision Oncology, https://pct.mdanderson.org/#/
4 Vogelstein, Bert, et al. “Cancer genome landscapes.” Science 339.6127 (2013): 1546-1558.
5 American Cancer Society. Cancer Facts & Figures 2016. Atlanta: American Cancer Society; 2016.
6 NCI’s Genomic Data Commons (GDC), https://gdc.cancer.gov/
7 Vice President Joe Biden, “What I Said to the Largest Convening of Cancer Researchers in the Country Yesterday”, https://medium.com/cancer-moonshot/here-s-what-the-vice-president-said-to-the-largest-convening-of-cancer-researchers-in-the-country-3007bb196dbd#.siyj9xn71
8 Sinsky, Christine, et al. “Allocation of Physician Time in Ambulatory Practice: A Time and Motion Study in 4 Specialties” Ann Intern Med. (2016) doi:10.7326/M16-0961
9 Muir, Paul, et al. “The real cost of sequencing: scaling computation to keep pace with data generation.” Genome biology 17.1 (2016): 1.
10 Hanahan, Douglas, and Robert A. Weinberg. “Hallmarks of cancer: the next generation.” Cell 144.5 (2011): 646-674.
11 Gingras I, “The role of precision medicine in “real-life” management of breast cancer patients: A survey assessing the current use and attitudes towards tumor molecular sequencing in clinical practice.” [abstract]. In: Proceedings of the Thirty-Eighth Annual CTRC-AACR San Antonio Breast Cancer Symposium: 2015 Dec 8-12; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2016;76(4 Suppl):Abstract nr P6-04-13.
12 Kiberstis, Paula A., et al. “Metastasis: An evolving story.” Science 162 (2016): 163.
13 Scannell, Jack W., et al. “Diagnosing the decline in pharmaceutical R&D efficiency.” Nature Reviews Drug Discovery 11.3 (2012): 191-200.
14 CenterWatch FDA Approved Drugs for Oncology, https://www.centerwatch.com/drug-information/fda-approved-drugs/therapeutic-area/12/oncology
15 Scarlett, Uciane K., et al. “High-Throughput Testing of Novel–Novel Combination Therapies for Cancer: An Idea Whose Time Has Come.” Cancer Discovery 6.9 (2016): 956-962.
16 The Cost of Sequencing a Human Genome, https://www.genome.gov/sequencingcosts/
17 LeCun, Yann, Bengio, Yoshua, and Hinton, Geoffrey. “Deep learning.” Nature 521.7553 (2015): 436-444.
