Skip to main content

At ASCO, FDA Discusses Use of Real-World Data Drawn From Flatiron Health's Clinico-Genomics Database

Premium

CHICAGO – At the American Society of Clinical Oncology's annual meeting this week, the US Food and Drug Administration highlighted the ways in which it is using real-world patient data to learn more about the safety and efficacy of precision oncology drugs and the extent to which diagnostics are used to guide treatment strategy.

Two years ago, the FDA struck up a research alliance with healthcare technology firm Flatiron Health to explore how real-world data derived from sources outside of the traditional clinical trial setting could be used to gather evidence on the safety and efficacy of cancer treatments. 

Data from clinicals trials create idealized conditions for testing treatment modalities. "In studies, often the way we guide care and assess patients according to the protocols [in clinical trials] has little resemblance to how patients are being treated out in the real world," Sean Khozin, acting associate director of the FDA's Oncology Center of Excellence, said at the ASCO meeting.

In partnering with Flatiron, the FDA hopes to gain a better view of the differences in research and real-world practice, and investigate the impact on the safety and efficacy of drugs after they enter the market. The agency is drawing on data within Flatiron’s clinico-genomics database, in which the company has linked deidentified electronic health records from 30,000 patients treated at more than 200 US cancer centers, with results from genomic tests performed by Foundation Medicine.

Previously, the FDA and Flatiron have discussed the use of real-world data to track the adoption of immunotherapy in non-small cell lung cancer and the utilization of PD-L1 testing. "There are differences in practices, and different velocities of adoption," Khozin said. But generally, since the approvals of Opdivo (nivolumab) and Keytruda (pembrolizumab) in NSCLC in 2015, there has been rapid adoption of these immunotherapies.

However, Keytruda's labeling required PD-L1 testing to identify high-PD-L1 expressors who should receive the drug, while the labeling of Opdivo didn't require companion testing. The difference in labeling indicating Keytruda for a smaller subset of patients was reflected in real-world data, which showed a slower uptake of Keytruda compared to Opdivo.

A year after the approval of these agents, real-world data also revealed that around 94 percent of more than 1,300 patients who received Opdivo didn't get PD-L1 testing first, but 38 percent of 48 patients treated with Keytruda did not get testing. "To our surprise [PD-L1 testing] was very low," Khozin said. "The majority of patients were not being tested in the first year following the approval of these therapies."

More broadly, Khozin noted how the agency captured in real-world data the limited use of companion diagnostics to guide personalized treatment strategies. "The rate of testing is very low, and what we've noticed is that the majority of clinicians are not using [FDA approved] companion diagnostics and laboratory-developed test use is very high in the real world," Khozin said.

In a recently published Health Affairsarticle, Khozin and experts from Flatiron and Foundation, described the development of Flatiron's database and provided some examples of the types of learnings that can be gleaned from the repository. For example, experts were able to query the database to explore the rate of high microsatellite instability across tumor types.

Last year, the FDA for the first time approved a tissue-agnostic indication for a drug when it approved Keytruda for patients with advanced solid tumors characterized by high microsatellite instability (MSI) or mismatch repair deficiency. The accelerated approval was based on data from five single-arm studies involving 149 patients with 15 types of solid tumors characterized by high MSI or mismatch repair deficiency.

Khozin and colleagues wrote in Health Affairs that the data in the clinico-genomics database showed, as expected, that endometrial cancer patients commonly had high MSI status. But the data also revealed that tumors of unknown origin tended to have MSI-high status more often other tumor types, "which suggests that perhaps the management of this disease entity should include MSI-H testing," Khozin and colleagues wrote. They further noted that given how rarely this biomarker shows up in most tumor types in the database, very large sample sizes will be needed to establish the efficacy of immunotherapy in the MSI-high subpopulation.

However, since FDA's tissue-agnostic approval for Keytruda in May 2017, the proportion of MSI-high, advanced cancer patients receiving checkpoint inhibitors has increased in Flatiron's database, from 10 percent before approval to 15 percent in September of 2017. As the number of treated patients get bigger in the clinico-genomics database, it could spur more in-depth studies on the efficacy of these drugs in this subset of patients, Khozin and colleagues wrote.

At ASCO, Khozin also highlighted the case of a single patient, referred to as JD, who was diagnosed with non-small cell lung cancer at a local community cancer center in 2011. From information in this database, one is able to piece together the timeline of this cancer patient.  This woman with early-stage cancer was initially treated with chemotherapy. Approximately four years later, a CT scan detected lung cancer recurrence, at which point, her oncologist sent her tumor biopsy for next-generation sequencing.

The test report from Foundation Medicine identified potentially actionable mutations in KRAS and STK11 genes, as well as around a dozen variants of unknown significance. The lab recommended targeting alterations in KRAS with a MEK inhibitor Mekinist (trametinib), and the published literature suggested that mTOR inhibitors could be potentially efficacious in STK11 mutated cancers. The doctor ended up giving JD Mekinist, even though the published literature contained examples where patients with KRAS alterations responded to the drug and where they didn’t.

JD fared well on Mekinist for eight months, after which her disease progressed, but her doctor continued to give her the drug for another eight months. When her cancer progression accelerated, the doctor switched her to carboplatin and paclitaxel. However, JD passed away two months later from disease complications in December 2016.

"Ultimately, what we want is that n-of-one experience," said Khozin. "Personalized oncology becomes an oxymoron if you cannot understand the n-of-one experience."

Looking back on the information in the clinico-genomic database, one is able to discern the sequence of this patient's treatments according to the molecular features of her tumor, though it's still not clear which specific alteration was responsible for her treatment response: the KRAS short variant or the amplification? Foundation had also reported a KRAS rearrangement that it deemed to be a variant of unknown significance, and this marker may also may have played a role.

"To answer questions like these, data from large cohorts of similar patients must be shared and aggregated, and nuances of both clincial and genomic findings need to be captured," wrote Khozin and others in the Health Affairs paper, where this case study was also highlighted.

"There is no such thing as a typical journey, and that’s one of the things we were hoping to highlight with this [Health Affairs] article." Vineeta Agarwala, director of product management, told GenomeWeb. Because each of the 30,000 cases are unique in their own way in the database, "it’s then really challenging to figure out how to make sense of that many stories," she said.

The first data from the National Cancer Institute's Molecular Analysis for Therapy Choice, presented at this same meeting, demonstrated how genetically complex tumors can get, and how difficult it can be to parse that data for signals into which patients are responding to a drug, and why others aren't.

Real-world data offers another avenue for tracking the treatment outcomes in genetically complex and rare tumor types. "We're saying, 'Let's include every patient who walks in the door and undergoes sequencing.' No one is looking at their sequencing data. No one is looking at their electronic health records," Agarwala continued. "No one is looking back to say, 'Show me all the patients like JD,' until you construct a real-world data set."

However, FDA's interest in real-world data suggests that drugmakers will likely bring such datasets to the agency as part of broader drug approval submission packages. "These datasets will undergo review by the FDA as part of real drug development scenarios," Agarwala said. "If I had to predict, I'd say there is a phase in which the FDA is actively interrogating and evaluating such data."

At the ASCO meeting, researchers led by Gracy Crane from Roche have a study in which they evaluated more than 2,100 NSCLC patients within the clinico-genomics database to investigate time-to-treatment failure as a surrogate endpoint for overall survival. (Earlier this year, Roche acquired Flatiron for $1.9 billion; and it also holds a majority stake in Foundation.)

Researchers grouped patients according to whether they received immunotherapy, targeted treatment, or chemotherapy. They established that a time-to-treatment failure of less than two months — as determined by when patients moved onto their next treatment or recorded progression in their electronic health records — as indication of rapid disease progression, and compared overall survival and genetic profiles of this group against non-rapid progressors.

Crane and colleagues reported that median overall survival was four months among rapid progressors versus 17 months in non-rapid progressors. Non-rapid progressors were more likely to respond to immunotherapy when they had higher tumor mutational burden compared to rapid progressors, but this was not observed for patients treated with targeted drugs or chemotherapy. Additionally, researchers saw associations with certain tumor mutations and rate of progression depending on whether patients were treated with immunotherapy, targeted drugs, or chemotherapy.

"The research team at Roche is very interested in understanding the utility of this type of real-world data," Agarwala said. Although drugmakers usually study therapies in controlled clinical trials, real-world data "provides an opportunity to learn a bit more about the patient population who might be eligible not only for their drug but other therapies coming down the line in the same class."

Interest in using real-world data in drug development has been bolstered by the 21st Century Cures Act of 2016, which requires the FDA to develop guidance for scenarios in which this type of information can be used to approve drugs, expand labeling, and identify new treatment indications.

Flatiron is hopeful that industry and regulators' experience using its database will guide national policy discussions on the collection and sharing of clinical and genomic information. "We're still at a nascent stage in creating such policy but hopefully this case example informs it," she said.