Freenome is developing a blood-based early colorectal (CRC) screening test using artificial intelligence. Patients newly diagnosed with colon cancer or advanced adenoma are encouraged to sign up for the Freenome study, AI EMERGE.
Freenome is developing a blood-based early colorectal (CRC) screening test using artificial intelligence. Patients newly diagnosed with colon cancer or advanced adenoma are encouraged to sign up for the Freenome study, AI EMERGE.
Working at Freenome
Sometimes, when you’re up to your neck in multi-analyte data, trying to identify patterns that could help you cure cancer, you just need a little help from your friends. At Freenome, we’re fortunate to find those friends hard at work by our side, and from time to time we make a special excursion just to celebrate that fact.
The annual Freenome company picnic last Friday was just such an occasion, as we gathered both our work family and our family-families together to enjoy a fog-free, golden summer afternoon in Westborough Park, South SF.
Around 3PM Freenomers begin trickling in with families, toddlers and pets (a.k.a. Freenome Mascots and Mascotas), and a smorgasbord of favorite potluck dishes in tow.
Special shout out to Freenome’s office manager and Chief Happiness Officer, Jenn Harrison, for organizing, and to Ben Sutherland, software engineer plenipotentiary and Freenome Grill Master Extraordinaire, for laying out a sumptuous BBQ to satisfy and delight those of all dietary persuasions.
It proved to be the perfect Freenome recruiting event for junior computational researchers and aspiring commercial specialists. Before you judge, this is important work in a highly competitive space!
And, because this is Silicon Valley, after all, we found time to partake in some wholesome team competition, sharpening mind and body in such future Olympic events as Buddy Walker, Giant Slingshot, and Human Targets. (That last one needed special clearance through HR, but we got it.)
All in all, it was a day to relax and reflect on where we were just one year ago. From a core team of 25 biologists, machine learners, clinicians, engineers, and operations staff we’re now at 77 strong and growing fast. (One of the many, many lessons learned? Taking the team photo BEFORE the medal ceremony!) Facing the challenges of marrying two of the most important fields of our time can be equal parts inspiring, challenging, and sometimes, just plain frustrating. But realizing that the people you need to succeed are all around you? That, my friends, is a picnic. And we’re already cooking up ideas for next year. Hope to see you there.
Cell-free DNA Cancer Testing
Much of the recent research in the field of early cancer detection has centered on cell-free DNA, a concept that most people, clinicians included, have never heard of. But its existence makes intuitive sense. Cells, like anything else, don’t simply up and vanish when they die. They break down into the raw materials from which they were comprised. Some of those component parts are fragments of nucleic acids from the cells’ DNA and RNA. Think of them like cellular garbage blowing along before being swept up by the blood’s street sweeping system.
Such real-time information on cellular turnover in the body is, of course, of great interest to anyone studying what’s going on with a disease like cancer. To this point, much of the work around cell-free DNA (cfDNA) in cancer has centered around trying to detect tiny traces of mutated DNA shed into the bloodstream from cancerous cells. The problem with using this technique for early cancer detection is that, during the early stages of a cancer, the fraction of tumor signal as a total of cfDNA in the blood is incredibly, infinitesimally small—less than a tenth to a hundredth of 1%. If you do the math (and we did, in this Perspective Paper), you start to realize some serious limitations on tumor DNA as an early detection tool.
Challenges of detecting tumor DNA
To reliably detect cancer (at around 95% sensitivity) you’d need to draw 15 tubes of blood; not something most patients will be eager to participate in as routine blood work. That’s one problem. The other challenge you run into is that there is only so much you can say with confidence even when you do detect a mutation.
Why? Because not every mutation, even in known cancer-driver genes, represents a malignancy. For example, in the human eyelid, multiple cancer genes are under strong positive selection (found in 18%-32% of cells), including most of the key drivers of cutaneous squamous cell carcinomas.1 That’s almost up to a third of cells showing cancer-associated mutations in otherwise physiologically normal skin. If cells from that epidermal layer (for example) were to shed DNA into the bloodstream, a test based solely on calling mutations might incorrectly identify those somatic mutations as cancerous—leading to poor specificity (false positive results).
A new technique to analyze cell-free DNA
Recognizing that identifying circulating tumor DNA (ctDNA) in the blood was a classic needle-in-the-haystack problem, some of us at Freenome started to wonder about the rest of the haystack—the other 99.99% of the cell-free DNA circulating in the blood. Where does it come from and what can it teach us about the health status of an individual?
Our partners at the Medical University of Graz in Austria are doing some leading-edge work in this area. In 2016 graduate student Peter Ulz and Professors Ellen Heitzer and Michael Speicher published a study in Nature Genetics, “Inferring expressed genes by whole-genome sequencing of plasma DNA,” showing how patterns of gene expression could be inferred from cell-free DNA. To make their results more accessible, we’ve put together a short summary of the paper here.
Their study was based on the knowledge that cfDNA that resists degradation long enough for analysis consists primarily of sequences that were bound within, and protected by, nucleosomes—the DNA-protein complexes within which DNA is organized. Because nucleosomes stick to DNA very tightly, they block other cellular components, like the proteins that transcribe DNA to RNA, from accessing the DNA to which they’re bound. Consequently, nucleosome binding in the cell is a dynamic process: they will bind, unbind, and move around as necessary to allow or restrict access to the underlying DNA.
Because nucleosomes must move around to allow the cell’s transcriptional machinery access to DNA, patterns in this nucleosome positioning, or “nucleosome footprints,” at a particular gene vary depending on whether the gene is actively expressed or not. In particular, the “beginning” of an actively expressed gene (the transcriptional start site, or TSS) tends to be less tightly packaged within nucleosomes to allow transcription to occur more readily. Given their lack of protection by the nucleosome, TSS’s corresponding to actively expressed genes were expected to be under-represented in cfDNA.
To test this hypothesis, the study authors compared differences in cfDNA sequencing coverage between transcriptionally silent and highly-transcribed genes. After establishing that silent and highly-transcribed genes had different coverage patterns, and that larger changes in transcription level lead to larger changes in coverage, they then assessed the sensitivity and accuracy of gene-expression predictions based on cfDNA sequencing–coverage analysis.
Finally, they confirmed the technique of nucleosome footprint analysis by determining whether blood samples from patients with cancer were informative for expressed cancer driver genes. They were, as predicted. Again, you can find a helpful review of their foundational work here.
The most important implication of this work is that nucleosome footprints—inferred from cfDNA sequencing coverage and analyzed through machine-learning techniques—can be used to develop classifiers to sensitively and accurately predict expression of certain genes from cfDNA alone, both in healthy individuals and those with cancer.
Understanding tumor-host interaction through nucleosome footprints
Nucleosome footprints are known to vary by cell type, and prior research has demonstrated that, in healthy individuals (theoretically more representative of early-stage patients with low-tumor fraction), most circulating cfDNA is derived from immune cells.2,3 Freenome is currently investigating whether the techniques outlined above may be used to similarly infer epigenetic changes in immune cells to provide us with valuable insights into cancer’s interaction with the rest of the body.
As Freenome continues to move toward a systems-biology approach to cancer detection, our artificial intelligence platform is the key to realizing the true clinical potential of cell-free DNA. Only through advanced machine-learning techniques can we hope to determine clinical significance from subtle correlations among billions and billions of data sets. Inferring gene-expression patterns in this way will provide important clues to cancer’s underlying biology, leading to new and noninvasive ways to detect and monitor tumor activity over time, and, eventually, helping physicians disrupt tumor formation altogether.
_Imran Haque, PhD, Chief Scientific Officer, Freenome
1. Martinicorena I, Roshen A, Gerstug M, et al. Science. 2015 May 22; 348 (6237): 880-886.
2. Lui yy, et al. Clin Chem. 2002 Mar; 48(3): 421-7.
3. Valouev A, et al. Nature. 2011 May 22; 474 (7352): 516-20. doi: 10.1038/nature10002.
Early Cancer Detection
In 2018, nearly 1.7 million new cases of cancer are expected to be diagnosed in the United States alone, which will result in an estimated 609,640 deaths. With statistics this profound, chances are high that you or someone you know has been—or will be—diagnosed with cancer in your lifetime. In fact, recent facts and figures from the American Cancer Society show that nearly 40% of men and women in the US will, at some point, be diagnosed with cancer.
The most common cancer diagnoses in order of prevalence 1
- Lung and bronchus
- Colon and rectum
- Melanoma of the skin
- Non-Hodgkin lymphoma
- Kidney and renal
One reason cancer mortality remains high is that screening for and detecting all the specific types can be difficult. In addition to being lengthy, expensive, and complex, common screening methods like lab tests and imaging procedures are often only designed to detect a certain type of cancer, such as mammograms for breast cancer and colonoscopy for colorectal cancer. Along with prostate, cervical, and lung cancer, these 5 are the only cancers commonly screened for in the general population.
Cancer blood testing today
Cancer screening using blood testing could make early detection more accessible; however, currently available cancer blood tests are hindered by limited accuracy and only apply to specific cancers. Some common blood tests used in cancer today are:
The complete blood count (CBC) test measures the number of blood cells and/or abnormal cells within the body.
Blood protein testing is effective in identifying abnormal immune system proteins commonly present in multiple myeloma. A blood protein test, also known as electrophoresis, is often the most reliable way to study and examine the health and vitality of proteins in the blood.
Tumor-marker tests such as PSA testing in prostate cancer, detect the chemicals created by cancerous tumor cells. Because tumor markers are also produced by normal cells, this test is typically used in conjunction with other tests to make a firm cancer diagnosis.
Circulating tumor cell tests are used for detecting cells floating in the bloodstream that have broken away from the initial cancer site. This test is often used to monitor patients with breast, colorectal, and prostate cancer.
Even though these tests are scientifically proven, their results are not always accurate and must be closely read and interpreted.Variables like body type and diet can sometimes influence test results. In order to make a definitive diagnosis, most forms of cancer require a biopsy. A biopsy is an invasive surgical procedure in which doctors remove a small piece of tumor tissue in order to diagnose, and in some cases determine how best to treat, a given cancer.
How can blood work detect cancer better?
In many cases, cancer can be present in the body long before any symptoms appear. Unfortunately, many people never consider getting tested for cancer until their symptoms become obvious, at which point the tumor may already have grown and spread. Recently, however, scientists have discovered that, even in the early stages of tumor growth, small amounts of DNA, RNA, and proteins are released into the blood. These fragments can come from both the tumor as well as the body’s own surveillance system, which includes infiltrating immune cells. Patterns among these cell-free molecules could provide the earliest indicators of cancer—if they can be identified.
Next-generation cancer blood tests
Companies like Freenome are now experimenting with new methods to digitize and examine a much broader range of molecules in the blood than ever before possible. While this creates an unprecedented amount of data that can be analyzed for patterns associated with specific types of cancer, it’s so complex that specially-designed artificial-intelligence (AI) is needed to draw meaningful conclusions.
AI pattern recognition among billions of cell-free biomarkers should enable doctors to detect cancer much earlier, and with greater sensitivity and specificity than current blood tests. So thanks to advanced medical technology and ongoing cancer research, it won’t be long before multiple life-threatening cancers can be detected using a simple AI-powered blood test.
Stay connected with the latest news on AI genomics and early cancer detection blood tests by entering your email below.
1. American Cancer Society: Cancer Facts and Figures 2018. Atlanta, Ga: American Cancer Society, 2018. Also available online. Last accessed February 26, 2018.
Artificial Intelligence in Medicine
In the past century, the field of medicine has made miraculous advancements in lifesaving technology. Many diseases and conditions, fatal in previous generations, now have reliable cures, treatments, and preventions. As technology in communications, engineering, chemistry, and physics continues to develop, more possibilities open up for those who work in medicine. Until recently, though, nothing we’ve developed has been able to do the work for doctors.
Artificial intelligence (AI) is quickly changing that.
Advances in the field of “deep learning” have made it possible for machines to begin lightening the workload of doctors by performing certain tasks faster—and in some cases more accurately—than humans. This frees up medical professionals to focus on their true priority: treating patients. Today, we’ll take a look at AI, and several ways it’s making life easier for medical practitioners.
AI and Deep Learning
The idea of artificial beings that are capable of thought has been around since at least ancient Greece. Much like the eponymous doctor in Mary Shelley’s Frankenstein, humans have postulated for years that if divinity has created sentient life capable of cognition, man may likewise be capable of such acts of creation. The result has been the romanticising of the concept in literature and culture, as science simultaneously pursues its realization.
Both approaches have become more sophisticated over the centuries, with writers and poets broadening the scope of what hypothetical synthetic cognition might be capable of. Likewise, as the progression of technology has led to the development of computer systems, it has finally become possible to create systems that can accomplish tasks previously considered too abstract for non-human minds.
The most significant advancement thus far in the field is that of “deep learning,” a subcategory of machine learning that attempts to mimic the way the human brain learns new concepts. A simple definition of deep learning is when a system or algorithm is given a large dataset, and told to look for patterns, without being programmed how to differentiate between the patterns.
The most convenient example of this is image recognition, and how Google built a computer network that taught itself what a “cat” was. Fed enough YouTube videos, it started to recognize patterns among the images it was presented, and it started to group felines together as a category. It’s an impressive feat, considering there’s a non-trivial amount of variety of characteristics among cats.
None of the data Google fed the algorithm was labeled, meaning this was “unsupervised” deep learning, and the computers didn’t have any base examples to work from first. The alternative is “supervised” deep learning, where the computer is given a number of labeled examples to start it off, teaching it what kind of things classify as “category A” or “category B.” In supervised learning, they’re still not instructed what to look for, just given examples of what qualifies; determining qualifying characteristics is up to the system.
This is very similar to how we teach a child what a cat is. We show them pictures and videos of a cat, tell them what sound they make, and maybe even let them pet a cat (if one is available). Most toddlers, even ones who have never seen a cat in person, can identify them by sight or sound. We give them sufficient data, and they can draw the conclusion on their own. That’s one of the foundational principles of intelligence.
Image recognition is where much of the deep learning research has been done, and it’s where a majority of the AI use cases in medicine can be found. That’s because so much of a doctor’s job is looking at images and scans and trying to determine if there’s a problem. It can consume large amounts of valuable time, so employing AI to help process the large number of radiology results can save both time and money on the task.
AI has already been tested (with successful results) to prove it can identify the presence or absence of a number of conditions via reviewing images alone. At Jefferson University Hospital, researchers proved that A.I. could detect tuberculosis with an accuracy of up to 96%. Over in China, Infervision is using the same methodology in detecting lung cancer, helping the mere 80,000 radiologists process the over one billion radiology scans a year.
Lungs aren’t the only things being inspected by computers. In England, several clinical trials have proven that a computer can predict future heart problems in patients simply by looking at a patient’s heart scans. It does this with accuracy levels greater than the doctors themselves. Even diabetic retinopathy can be detected, as can the metastasizing of breast cancer.
As the technology improves, and as the algorithms are fed more data, an increasing number of diseases and conditions should be detectable or predictable by machines, enabling doctors to save more lives.
Artificial intelligence in medicine isn’t limited to image recognition. In some cases, the AI doesn’t even need to see images to make a diagnostic prediction. Take the “Deep Patient” project. Researchers at the Icahn School of Medicine at Mount Sinai gave their AI 700,000 electronic health records (EHRs), hoping that the machine would be able to make the connections between disease or condition predictors, and eventual diagnosis.
Once the AI had been given the chance to parse all the information (patient X developed lung cancer, patient Y developed heart disease), it was tested. It was given data on 76,000 patients whose diagnosis was known, but not given to the computer. Their results were impressive, drastically outperforming “evaluations based only on raw EHR data, doing particularly well at predicting severe diabetes, schizophrenia, and various cancers.”
They weren’t the only ones who had positive results. A team of UK researchers performed a similar test, to see if their learning algorithm could accurately predict heart attacks. It “correctly predicted 7.6% more events than the ACC/AHA method, and it raised 1.6% fewer false alarms.” What’s more, the machine used different guidelines to make its judgement calls, highlighting the fact that doctors may be using the wrong metrics:
Several of the risk factors that the machine-learning algorithms identified as the strongest predictors are not included in the ACC/AHA guidelines, such as severe mental illness and taking oral corticosteroids. Meanwhile, none of the algorithms considered diabetes, which is on the ACC/AHA list, to be among the top 10 predictors.
The corporate world has been benefiting from “big data” for several years, and now, with the help of AI, medicine is too.
The Human Diagnosis Project
The Human Diagnosis Project, also known as Human Dx, is similar to the above-mentioned EHR projects, but on a much larger scale. Using the same methodology of machine learning, but crowdsourcing both the patient data and the solutions, Human Dx aims to aggregate comprehensive diagnostic and prognostic data on a legion of conditions.
The AI then parses the information, and makes it more accessible and user friendly. Then, when a practitioner needs information or a second opinion, they can consult the archive for a wealth of knowledge related to their use case. The project makes answering difficult medical questions easier, and faster.
Mental Health and Cogito
Shifting from physical health to mental health, a new app is actually making it easier for mental health professionals to track the status of their patients. The Cogito Companion app is designed to track a patient’s activity, social connectedness, and mood. It can tell when you’ve left your house, or if you’ve stayed in one place for an extended period. It can tell if you’ve been calling or texting other people, or if you’ve been out of contact. And it allows your counselor to see your progress.
On the surface, it seems like other apps on the market: it tracks your activities, and reports back to your therapist. What makes it different is how you can record audio logs as a sort of diary. While this helps to record your feelings and thoughts, an AI algorithm analyzes the speed, tone, and energy in your voice to assess your mood. It can tell you if you’re having a down day, or if you’re doing better than you expect, and those results can be viewed by the practitioner.
The tool is still in its early stages, but as it learns and becomes more sophisticated it stands to help a great many people who struggle with mental health challenges.
The most direct application of AIs to medical practice is that of surgical robots. Though robots have been assisting surgeons since the 1980s, they’ve been non-autonomous, human-operated tools until very recently. They were just another device for surgeons to use.
Recent developments in machine learning have made it possible for surgical robots to become increasingly autonomous. Again, the technology is still in its early stages, but tests have begun to demonstrate the potential for the robots to be more accurate and precise than humans in the operating room. While it’s unlikely that machines will totally replace experienced surgeons in the conceivable future, there’s the definite possibility that they will begin making procedures easier for them.
AI and Machine Learning Applications in Genomics
Perhaps the most impressive usage of artificial intelligence in medical diagnosis is in the rapidly developing field of genomics, which includes analysis of cell-free DNA (cfDNA) and other genomic biomarkers. Cell-free DNA tests, if you’re unfamiliar with the practice, screen blood plasma for DNA fragments that are left behind when cells in the body die. Currently, the most prominent usage for these tests is as a pregnancy screening tool—by testing fetal DNA (cffDNA) found in the mother’s blood plasma, medical professionals can test for gender and certain genetic conditions.
The new technique already has a wealth of potential, even before artificial intelligence is applied. Incorporate the AI, however, and the possibilities only become more impressive.
Take our ambitious mission at Freenome, for example. Using advanced deep learning algorithms, Freenome’s AI genomics platform is looking at cfDNA and other cell-free biomarkers, such as cell-free RNA (cfRNA) and proteins, to aggregate and decode genetic data left behind by cancer cells, as well as the patient’s immune system. Using the AI to identify patterns and trends in these cell-free biomarkers enables Freenome to achieve cancer detection at much earlier stages, and determine which treatments will be most effective.
Rather than having to wait until later stages when the cancer causes observable symptoms in the patient, or the growth is visible on X-rays and CT scans, Freenome will help doctors identify cancer at the earliest stages of its growth, making treatments more effective and drastically improving survival rates, all from a simple blood test.
Computers have been changing the world since their inception, and with the advent of functional “weak A.I.s,” that progress has only accelerated. With the introduction of artificial intelligence into healthcare and the continual improvement of technology, clinicians in the next several years can expect to achieve medical miracles long thought impossible, ultimately shifting the practice of medicine from its current focus on treatment to one where disease is prevented altogether.
In 2014, my co-founders, Riley Ennis and Charlie Roberts, and I had what we thought was a relatively straightforward, but powerful idea: to create an early-detection diagnostic test for cancer, one that would become smarter over time as it saw more people's samples. The idea was that we would sequence fragments of DNA from blood and detect signals associated with the presence of cancer through software equipped with a type of artificial intelligence (AI) called machine learning. This concept, now known as AI genomics, became the foundation of Freenome's approach in developing next-generation, smart diagnostics.
Little did we know at the time how hard it was going to be to realize our vision. Bringing artificial intelligence to bear on DNA, RNA, and protein data to detect biomarkers associated with early-stage cancer involves merging two incredibly complicated fields of research in a way that's never been done before. In the three years since we began our research, we've made incredible progress, but we’ve also had complications and setbacks that were largely unexpected, and, sometimes, impossible to even identify the cause of. We also underappreciated the giant chasm of work that lies between a research project and a fully-functioning, validated clinical assay.
Sharing lessons learned
One could argue that the three years that Freenome spent learning these lessons the hard way is now a competitive advantage that we have among other companies who are also making the trek to clinical validation; however, these lessons are also applicable to anyone who is trying to make a science project into an actual product. We're relaunching the Freenome blog to give you a front row seat so that, together, we can progress towards the future quickly and responsibly. Through our posts, we’ll provide updates on our progress and our challenges, and help foster meaningful dialogue on how to move discoveries out of the lab and into everyday clinical practice, where they can help people detect and treat diseases like cancer at their most manageable stages.
Building the right team
Once we realized that doing research in a clinical context was far more difficult than we anticipated, we began to hire key players who could fill in our blind spots. Thanks to successful funding and a large network of supporters, over the past two years we’ve been able to recruit some of the top minds in machine learning, molecular biology, and clinical research. Some of our earliest hires included clinical and reimbursement experts like Girish Putcha, MD PhD, from Palmetto; our VP of Operations, Dan Delubac, PhD, formerly of Guardant Health; Alex Rasmussen, PhD, who insures that our software is built in a clinically-compliant way (no easy feat); and some of the most brilliant scientific minds I’ve ever worked with, led by our Chief Scientific Officer, Imran Haque, PhD.
Halfway through 2018, we’re at 70 employees and growing. Our latest key hires include our VP of Regulatory, Abe Tzou, MD, formerly of the FDA; our VP of Marketing, Lena Cheng, MD, formerly of Doctor On Demand; and our new Chief Commercial Officer, Mike Nolan, formerly of Foundation Medicine.
Along the way, we’ve continued to establish partnerships with pharma and academic institutions, led by our VP of Business Development, Blandine Merino, providing valuable data, perspectives, and insights to help accelerate our research.
To ensure we have the right guidance and strategy to scale our business, we recruited a visionary pioneer in genomic medicine, Randy Scott, PhD, (co-founder of Invitae and Genomic Health) and a scientific advisory board to share their expertise in machine learning and biology. That group includes Anshul Kundaje, PhD, from Stanford, and Olivier Elemento, PhD, from Weill Cornell Medical College.
Living our values
Of all our achievements, I’m proudest of the incredible culture we are building. Trust, empathy, integrity and a dedication to service leadership are the core values that set Freenome apart from anywhere else I’ve ever worked. We voted on these values as a team, and, as a team, we strive to live them every day in our work, our hiring, our volunteering—even in how we evaluate our own performance (more on this later). Our courageous, humorous, kind, and brilliant team inspires me daily to push myself as a leader and to be the better and wiser person that they deserve.
Every new addition has raised our company to new heights in a way that I could not have anticipated. Each person has patiently taught me much more about all of the different aspects of what it takes to launch a clinical test than I ever could have learned on my own. I'm thankful to them, because Freenome would not be where it is without their knowledge and their willingness to challenge me and push me to think different.
For the remainder of this year, Freenome will be focused on bringing a blood-based cancer test to market that will finally make early-stage detection and treatment of colorectal cancer a reality for millions of patients. That of course means serious work ahead, but now that we have the right resources, people, and culture in place, I’m more optimistic than ever that we’ll get there together. Thank you to everyone at Freenome for your positivity, vision, and dedication. To the rest of you, keep an eye on future posts where there will be many more learnings and developments to come.
_Gabe Otte, CEO, Freenome