In launching a new project to identify precision medicine approaches for glioblastoma patients, researchers at the New York Genome Center are anticipating "an absolute gusher" of information from the sequencing and RNA analysis they will perform on approximately 20 study participants.
To aid in their effort to synthesize this data and arrive at the best treatment options for glioblastoma patients – who typically have a median survival of around a year following diagnosis – the NYGC has enlisted the help of the IBM supercomputer Watson, which a few years ago famously decimated the best human players that the quiz show Jeopardy! had to offer.
Although the NYCG is starting small and in a very specific group of cancer patients with limited treatment options, the goal is to use a prototype of Watson that's designed to handle genomic data to quickly analyze and interpret complex disease biology information and make precision medicine more broadly accessible to cancer patients.
At an event held last week at NYGC's Lower Manhattan offices, Robert Darnell, neuro-oncologist and scientific director of the NYGC, characterized the project as an effort to "connect the compassion of the physician with the power of technology." With Watson, the NYGC is hoping to speed up the painstaking process of recognizing the salient molecular characteristics of patients' disease based on the published literature and matching patients to drugs that they are most likely to benefit from. Watson's capabilities will be accessible to the nine New York hospitals participating in this project through a cloud-based system.
This type of work is currently ongoing at some of the best cancer centers around the world and is typically performed painstakingly by multi-disciplinary teams of life sciences experts. "It might be a presumption that just getting the sequence done will lead to the answer, but, in fact, the amount of data is rate-limiting for doctors right now in delivering the power of sequencing back to the patient," Darnell said. "And time, frankly, is not your friend when you have glioblastoma."
It has been evident to Darnell in his work across major cancer hospitals around New York that most patients aren't getting their tumors sequenced so they can get molecularly informed therapies. Moreover, when researchers do attempt to sequence a patient's tumor, it takes a team and several weeks of analysis and interpretation to tease out the potential therapeutic targets from the noise. "This [work] is wonderful and powerful … but it is not scalable," Darnell observed. "It can't be brought to thousands of patients."
This is where IBM is hoping Watson will make an impact. John Kelly, director of IBM Research, is hopeful that Watson can reduce this process down to seconds for a single cancer patient. "For those of you who saw the demonstration on Jeopardy!, [Watson] had two-and-a-half seconds to search millions of volumes of information to determine plausible answers and decide if it was going to engage in the game or not," Kelly said.
A number of genomic data interpretation services are available to physicians today, where much of the bioinformatics analysis is performed in an automated fashion using complex algorithms, but human curators are still largely involved in programming and updating these systems. IBM is promoting Watson as faster, requiring less human programming and curation than these other services, and as having the ability to "learn" from how well or poorly patients do on treatment strategies that doctors recommend in the project.
"In information technology we have begun to cross over a barrier. We are entering a completely new generation of computing systems … that no longer require expensive, massive programming by humans," Kelly said. "These are systems that are capable of dealing with immense amounts of data … [and] that learn as they are used. They are not being programmed. They learn as they experience the data and they experience outcome of the use of that data."
Teaching Watson
Following Watson's Jeopardy! success in 2011, IBM has been working to teach the supercomputer the "language of healthcare." IBM worked with researchers at Columbia University and the University of Maryland to brainstorm the types of uses Watson can have in healthcare. The technology giant has partnered with Memorial Sloan-Kettering Cancer Center to try to combine clinical and molecular data from breast cancer patients' electronic medical records, aiming to improve treatment selection. In this project, MSKCC researchers are preparing IBM to understand the medical and cancer terminology.
It's evident from these collaborations that unlike when Watson competed against people on Jeopardy!, the computer and humans will have to work side by side in order to find solutions to the most pressing health burdens. While Watson is designed to answer questions that researchers and doctors ask it in natural language, it has not been used to diagnose a patient's illness. In this project with the NYGC, Watson will be an aid to doctors and the computer's treatment suggestions will be considered and decided upon by a panel of experts.
"Watson is a tool to help physicians, it is not a replacement for physicians," Toby Bloom, deputy scientific director of informatics at the NYGC, told PGx Reporter. "Through its learning capabilities, Watson will incorporate data from patient outcomes to continuously improve the drug selection process."
In getting Watson versed in healthcare, IBM researchers set the computer loose on the machine-readable information available through the NIH so it could learn about biochemical pathways. The system has "digested" 23 million abstracts on PubMed to get up to speed on the published medical literature, and will continue to draw on new information as it comes up. "[Watson] has a voracious appetite," Kelly said. "It will ingest a million abstracts a day."
Some of those studies might contain outdated or wrong findings, or contradict each other. But unlike humans, Watson doesn't harbor any biases in the way it takes in the information. "Watson is not trying to decide which is a better source," Ajay Royyuru, head of the Computational Biology Center at IBM Research, said at the NYGC event. "It takes in all of it at once. What often happens in scientific literature is people build upon other people's work, and consistency or corroboration emerges from that. So, Watson learns from that."
In other words, while Watson is designed to mimic some human cognitive capabilities, it doesn't actually think like a human. If there is inconsistency in the literature, Watson will not try to decide which paper is right, but catalogue that discrepancy and present the different lines of thought to the physician, Royyuru explained. Ultimately, it is still up to the doctor and other experts to determine which option is more relevant in a particular patient's case.
Bloom explained that the researchers at the NYGC will initially identify the molecular markers relevant to a glioblastoma patient's illness and confer with Watson to get a concise report of the information available in the literature. "As the study continues, and the number of patients expands, we will be able to provide back to Watson the information on which combinations of mutations were treated with which combinations of drugs, and what the patient outcomes were," she said. "That information, over time, can be used by Watson to improve its recommendations and learn to rank treatment options for specific patients."
At the NYGC event, IBM's Royyuru demonstrated what a doctor might find if he or she turned to Watson for assistance in developing a treatment plan for a patient. A physician can log into Watson's cloud-based system, upload a patient's clinical and genomic data files, and ask Watson what can be done based on a patient's clinical and genomic data. Through Watson, the doctor will be able to visualize the patient's tumor and normal genome sequence for comparison. Then, Watson can map that information to show the biochemical pathways, proteins, and genetic mutations implicated in aberrant cell replication and the patient's cancer.
Finally, Watson will search the literature and provide information about drugs –investigational or commercially available – that interrogate the pathways and markers involved in the patient's disease. "That then is the insight that goes back to the practicing neurologist and the tumor board ... in the form of a report and [they can] browse this information in detail," Royyuru said.
State of cancer care
Despite the availability of increasingly cheaper sequencing technologies and advances in genomic research, very few cancer patients in the US benefit from these innovations because of the way clinical trials are conducted and logistical challenges to delivering this type of care. While a few molecular markers, such as HER2, BCR-ABL, and EGFR, are broadly factored into patient care for certain types of cancer (breast cancer, chronic myeloid leukemia, and lung cancer, respectively), for many indications the search for predictive markers hasn't yielded straightforward answers.
A recent study published in Lancet Oncology showed that after recruiting more than 400 metastatic breast cancer patients and analyzing their tumors with CGH array and Sanger sequencing, researchers at 18 French institutions were able to pick out drug-targetable genomic alterations in 195 women. Of these participants, only 43 received the precision therapies recommended by a team of oncologists and life sciences experts, with only four experiencing an objective response and nine having stable disease for more than four months (see related story, in this issue)
Researchers led by Fabrice Andre, director of the INSERM Unit U981, a French lab focused on developing personalized treatments, noted in the Lancet Oncology paper that their ability to identify markers driving the disease was hindered by the limited research on the molecular underpinnings of metastatic breast cancer. And although researchers recommended several patients for drugs in Phase I trials based on a marker they identified to be driving their tumor, 17 women weren't accepted into the studies probably because their disease was too advanced. The study also highlighted the fact that despite the identification of markers associated with disease progression, there are often not available drugs targeting those mutations or pathways.
Like metastatic breast cancer, glioblastoma is certainly and rapidly fatal for patients with the disease. The standard of care for glioblastoma is radiation therapy and the chemotherapeutic temozolomide, which can extend patients' lives for a few months. For researchers and drugmakers working to discover disease and treatment response-linked molecular markers, the biology has proven highly complex. Among positive advances in this regard, researchers have shown that MGMT methylation status of glioblastoma patients is a prognostic marker that can asses which patients will have better outcomes, and is a predictive marker for temozolomide response. MGMT is being studied in most glioblastoma drug development trials, according to MDxHealth, which markets a test that gauges methylated MGMT status in glioblastoma patients and is collaborating with a number of pharma companies.
However, drug developers in their search to advance molecularly-targeted therapies for glioblastoma have also experienced a number of disappointments. Merck KGaA, for example, embarked on a large Phase III study for the drug cilengitide, in which it looked to see if newly diagnosed glioblastoma patients with MGMT-methylated tumors responded better to the drug than those with non-methylated tumors, but this hypothesis did not bear out in clinical trials. Similarly, Roche subsidiary Genentech has also looked for biomarkers of response to Avastin (bevacizumab) in newly diagnosed glioblastoma patients in a Phase III trial, including VEGF-A and VEGFR-2, but these weren't found to predict improved patient survival.
Given these examples, advancing personalized strategies for glioblastoma patients will be a formidable challenge for NYGC researchers, even with the help of Watson. "There are thousands of mutations in each tumor and each tumor is different from the next one," NYGC's Darnell said. "All glioblastomas are not the same."
Genomic analysis necessitates biological samples from patients, and particularly in advanced cancer patients, procuring tumor tissue can be challenging and involve invasive procedures. In the NYGC glioblastoma project, researchers will enroll patients for whom tumor DNA and blood samples are available. The study participants will already be receiving the standard of care for glioblastoma. So, the information from Watson will be provided to treating oncologists in the hopes that they can potentially improve upon that standard care. In the consent process for enrolling in the study, patients are made aware that no personalized treatment options may be available for them or that the treatments administered may not provide any benefit.
"If the tumor board decides that a drug in Phase III for some disease other than GBM is the best course of treatment, we will work with the pharmaceutical company and seek approval to use that drug for this case," Bloom said.
Betting on Watson
Russ Altman, director of the biomedical informatics program at Stanford University, has been working at his own institution to build expert computing systems to advance personalized medicine. While the majority of genomics research is occurring in cancer, it is still early days, he reflected, and papers like the one published by Andre et al. in Lancet Oncology demonstrate that.
"But what keeps everyone going is that these few patients sometimes have amazing things learned from the genome that takes their treatments into entirely new directions," said Altman, who is one of the founders of Personalis, a company that performs whole-genome and exome sequencing and interprets the data for doctors looking to administer precision medicine to patients. "It is still rare, but if I had cancer I would want my genome in case I'm one of the rare ones."
Given that it's still early days for figuring out how genomics data might be best used to inform healthcare, one of the major challenges for a system like Watson, according to Altman, is the limited availability of data sets that it can learn from. "It will be very difficult for these systems to do as well as [human] curators because the curators are mostly doing these things for the first time now, and so there are few 'training sets' for Watson to use, and not even a clear procedure for doing this kind of inference," he told PGx Reporter over e-mail. He said, however, that it is plausible that in a limited disease setting like glioblastoma or when given a specific set of markers to tackle, Watson may be able to come up with some good clues.
"I would never bet against the miracle of informatics and computing, and I certainly think Watson-type technology will be useful, but I would be surprised if it replaces human curation any time soon," he said.
IBM's Kelly is understandably more optimistic about Watson's potential role in advancing personalized care, but he acknowledged that using Watson in healthcare is a big bet. "Normally big bets in our terms take decades," he said. "But I think this one is going to occur very, very rapidly," he said.