September 12, 2024 | Multi-cancer early detection (MCED) technology is going to be the best way to achieve population-scale cancer control, and Grail hopes its Galleri blood test that screens for more than 50 types of cancer is among the first to be approved by the U.S. Food and Drug Administration (FDA). Galleri has been on the market as a laboratory-developed test for three years with ordering physicians voluntarily reporting “cancer signal detection” cases to the company as it separately conducts an outcome tracking study on Medicare beneficiaries under an FDA-approved Investigational Device Exemption (IDE).
These were among the updates provided on this new front in the 53-year War on Cancer by Megan P. Hall, Ph.D., a distinguished scientist and vice president of corporate affairs at Grail, at the recent Next Generation Dx Summit. She started with the stats: 86% of cancers are diagnosed outside recommended screening, more than 80% of cancer deaths are from types without recommended screening, and the survival rate is about four times higher when cancer is diagnosed in the early, pre-metastatic stage.
In the United States alone, cancer’s deadly toll is 600,000 people every year, Hall reported. Only five cancers even have recommended screening—breast, cervical, colorectal, prostate, and lung.
The case for MCED was made in a 2018 paper published in the npj Precision Oncology (DOI: 10.1038/s41698-018-0066-x), which talks about how positive predictive value (PPV)—meaning the likelihood that a positive test result is actually positive—is driven by both the prevalence of cancer in a population and the specificity or the false-positive rate. Evidence suggests that the PPV is higher with pan-cancer tests, and even more so as specificity goes up, said Hall.
Since people don’t get to choose the cancer they get, getting screened for colorectal cancer raises by tenfold the chance of having a non-colorectal cancer, she continued, “and that’s consistent across the cancer types that we screen for.” This was also reflected in the National Lung Screening Trial, which found “the risk on non-lung cancer was much higher across all of the quintiles of pack years” than the risk of having lung cancer.
Detecting as many cancer types as possible in a population involves a tradeoff between yield (number of cancers detected in the population) and aggregate sensitivity, Hall pointed out. While sensitivity is important from an analytical perspective, for multi-cancer tests it is “extraordinarily hard to interpret and compare across platforms and even across studies.”
For a test looking only at lung cancer where the aggregate sensitivity is 75%, only 110 cancers per 100,000 people would be detected, she offered as an example. Conversely, MCED tests will have their aggregate sensitivity driven down by cancers in the mix that don’t shed DNA and thus have lower detection rates. However, they’d be detecting five times as many cancers in the population.
Grail developed Galleri to push the transition from “screening for individual cancers one at a time to screening an individual for multiple cancers simultaneously,” said Hall. MCED tests should be designed and evaluated differently than single-cancer tests since they represent a “fundamentally different paradigm with a fundamentally different analyte shared across a number of different cancer types.”
The Galleri test meets the mark on several fronts, she said. Via a shared cancer signal, it detects a broad range of deadly cancers with high specificity. Importantly, it also predicts the cancer signal origin (CSO) and therefore doesn’t contribute to overdiagnosis. As with other MCEDs under development, it is designed to complement rather than replace existing single-cancer screens.
Measuring the benefits based on those performance characteristics gets a bit complicated. Aggregate performance of the tests, as measured by PPV and cancer yield, is going to be critical since making comparisons cancer by cancer creates a “denominator problem,” explained Hall.
For longer-term clinical utility studies, both “observed reduction in late-stage cancers” and “modeled mortality” have emerged as potentially appropriate clinical utility measure of MCED tests, she added.
The Galleri test recognizes the signal pattern of “cancer or no cancer” in the blood as well as where in the body the signal is coming from, said Hall. It works by isolating cell-free DNA fragments that tumors have shed into the bloodstream, using a targeted methylation approach looking at over a million cytosine-guanine dinucleotides (CpG) sites across the genome. “These are methylated regions identified as the most informative for both cancer signal detection and for cancer signal origin prediction,” she noted.
Two machine learning classifiers are utilized, the first to detect whether there is a general pattern of cancer and the second, if cancer is detected, to predict where the cancer originated. “That’s really important for directing that downstream diagnostic workup,” said Hall.
Getting to this point with the Galleri test is due to having “one of the largest clinical evidence programs in genomic medicine to date,” Hall said, inclusive of nine clinical trials. Among these are two published foundational studies—the CCGA (Circulating Cell-free Genome Atlas) study covering development and validation of a cell-free, DNA-based MCED test, and the PATHFINDER study to evaluate clinical implementation and perceptions of the test.
Grail is also conducting the first and largest randomized controlled trial of a MCED test in collaboration with the National Health Service (NHS) that will assess the clinical utility of Galleri for population screening of 140,000 participants in the UK, she shared. “We’re through the third year of screening and now in the follow-up period. Results are expected in 2026, and we are very excited about that.”
The company is also now enrolling participants into PATHFINDER 2, which will evaluate Galleri in an eligible screening population. The focus of this study and the one with the NHS is “clinical utility and evaluating test performance and follow-up diagnostic workups in intended use populations” and will help support its premarket approval application with the FDA.
The REACH study, announced in November 2023, seeks to understand the health equity impact of Galleri in a Medicare population, she continued. The real-world study is being done in a unique collaboration with the Centers for Medicare and Medicaid Services, which is covering the cost of 50,000 tests for the more than 385,000 expected participants. More than 300,000 individuals on Medicare have already enrolled in the study.
For both the REACH and NHS-UK studies, the primary endpoint is a reduction in late-stage incidence of cancer, said Hall. “That is also what we are starting to see traction for in the literature, so there are a number of places where late-stage diagnosis and modeled mortality are coming into play as potential clinical utility endpoints.”
The MCED Consortium—a group of industry, academic, and government partners—is also “taking a hard look at clinical utility,” she added. The consortium wants to ensure that “trials and guidelines and approvals keep pace with innovation, which as we all know is moving very fast.”
CCGA was the company’s foundational case control study that supported Galleri’s launch, said Hall, and identified the analytical performance characteristics of the test. “We were able to design the test with a very low false-positive rate of 0.5%,” meaning it wouldn’t be shunting a lot of people into the healthcare system for workups they don’t need.
Additionally, Galleri was found to have about a 90% CSO accuracy, she said, and a modeled PPV—based on the distribution in the National Cancer Institute’s SEER (Surveillance, Epidemiology, and End Results) program, to avoid overestimating—of 44%. To ensure the results would translate to the intended use population, researchers looked at the screening population in the Pathfinder study where cancer status was unknown at the time of blood draws, and tests results were returned to patients.
“Performance characteristics translated almost exactly between case control and intended use, which was really an exciting moment for us as a company,” said Hall. “It really supported the rigor and robustness of the assay and the classifiers the teams had developed.”
Pathfinder was designed to be an interventional, intended use study, ultimately enrolling about 6,700 asymptomatic people who were undergoing regular screening with unknown cancer status at the time of blood draw, she continued. “What we found was that adding the Galleri test more than doubled the number of cancers that were identified by screening [and] about 50% of those were early stage [1 or 2].”
Notably, only four adverse events were reported and “all of them were from the blood draw and all of them resolved and none of them were serious,” she said. “We didn’t observe any adverse events from any of the follow-up diagnostic workups,” she added, which can be a real concern when they are prompted by a false-positive result.
Among the cancers detected in the study, 71% do not have a recommended screening test to enable early detection, Hall reported. The PPV came in at 43% with 88% localization accuracy.
Single-cancer screening tests are designed to be highly sensitive and tolerate a higher false-positive rate, she noted. That’s why the PPV on the Galleri test is roughly 10 times higher than that for mammography (4.4%).
In a “snapshot” of the cancers Galleri detected in the Pathfinder study, Hall homed in on 36 detected in a subset of 35 participants. Among this group, there was one multiple primary cancer and 27 cancers without a recommended screening test. Several stage 1 and stage 2 cancers were found, in addition to the later stage ones. “Significantly, we saw a stage 2 pancreatic cancer and a stage 1 liver cancer,” both of which are particularly aggressive.
Cancer signal origin prediction was also found to aid physicians in efficiently getting to a diagnostic resolution, she said. The median time to confirm cancer or no cancer in the Pathfinder study was 57 days and, among the true positives, almost three-quarters were resolved in under three months. This compares to 156 days seen in a retrospective claims analysis of almost 460,000 participants, agnostic to MCED.
Almost 80% of the time, added Hall, “the very first CSO-directed evaluation got patients to a diagnosis.” This provides further evidence that the CSO capability of Galleri is directive for physicians.
In terms of the potentially harm-causing procedures done on the false-positive group, they represented a much smaller percentage than the true positives, Hall said, and more of them were nonsurgical procedures. Together with the fact that there were no reported adverse events from any of these diagnostic workups suggests the technology can be implemented safely.
In fact, at the end of 2023, an FDA advisory committee highlighted the importance of having what it terms a “signal of origin” prediction capability as an inherent molecular feature of tests specifically so they can direct the next diagnostic workup, she noted.
Analyses to date offer “strong support” for a shared cancer signal across several cancers, reported Hall. One internal experiment found “equivalent performance” when comparing a full machine learning classifier trained on all the different cancer types and an occluded algorithm where lung cancer was dropped out. “We are still able to detect the lung cancers even though the algorithm was not trained on those lung cancers.”
The same experiment was repeated cancer type by cancer type and “the solid cancers did really well overall,” she continued. But there was a slight drop in performance with the hematologic cancers, suggesting systemic differences in their cancer signal. This was also observed in early study results when the classifier underpinning Galleri began using different specificity thresholds for the hematologic and solid cancers to account for the higher number of false positives seen with blood cancers. The Grail team is now examining this hematologic signal difference in greater detail.
Although blood can be drawn at any time during the preclinical diagnosis phase, to be of practical value it needs to happen during cancer’s “shedding window” above the lower limit of detection, said Hall. In respect to study design and test performance measurement during this preclinical detectable window, when testing in an asymptomatic population, the window for defining whether a patient has cancer is arbitrarily set at 12 months.
In the Pathfinder study, blood was drawn at “time zero” and patients with a positive result at that point who went on to develop cancer, clinically defined at 10 months, were deemed true positives, she explained. Similarly, patients with a negative test at time zero who went on to develop cancer at 10 months were deemed false negatives.
But it has become evident that many cancers develop outside of that 12-month window, she said, in some cases much later than one year. The established episode sensitivity definition thus complicates the measurement of sensitivity when looking at asymptomatic or screening individuals. “It’s a very complicated measure to interpret not just platform to platform or technology to technology but even between studies.”
Grail collaborated with the Americana Cancer Society to gain access to its Cancer Prevention Study-3 (CPS-3) cohort of patients—all cancer-free when they enrolled between 2006 to 2013 and whose long-term health outcomes continue to be studied—to learn how fast cancers were moving in the population and the rates for different cancer types, said Hall. Not surprisingly, the time between blood draw and cancer diagnosis was relatively short (about six months) for aggressive cancers such as pancreatic cancer and a bit longer for indolent cancers such as those that present in the prostate.
Some of the lymphomas “can kind of smolder for a while and have a much longer preclinical detection window,” she said. On average, most cancers do in fact land at around one year—323 days, based on this analysis—making it a good interval for a MCED test to start with on the general population age 50 and above.
Within that one-year window, 60% of cancers in the CPS-3 cohort were detected but 40% happened at other time points, Hall reported. “When designing studies in asymptomatic populations or screening populations you have to think about when you are going to find a cancer and how is that going to impact your potential results.” In longer-term follow-up with the Pathfinder cohort, she noted, a colorectal cancer was detected three years after blood draw and staged as regional at the time of diagnosis.
It is well appreciated that the amount of tumor DNA in the blood is driving cancer detection, with tumor fraction and detection going up with cancer stage, said Hall, in sharing figures from the CCGA study. Cancer by cancer and subtype by subtype, the more aggressive and lethal cancers consistently shed more DNA into the bloodstream even at early stages, she noted.
Looking specifically at stage 2 breast cancer, for instance, the hormone receptor-negative (HR-negative) variety that includes the more aggressive triple-negative breast cancers has a nice detection rate, said Hall. The issue here is that the HR-positives account for most of the breast cancers detected in the population, so aggregate sensitivity is likely going to be driven down by the lower signals associated with the less aggressive breast cancers. The same is true for prostate cancer, she added.
At least one study from another group is now “down-sampling” for both breast and prostate cancer, likely due to those findings, Hall said. A similar phenomenon is seen with lung cancer, where (consistent with findings by others) adenocarcinoma has lower tumor fraction than other more aggressive subtypes such as small cell and squamous cell lung cancers.
Prostate cancer is subject to a lot of overdiagnosis and overtreatment, Hall continued, and too much test sensitivity in early stages of the disease which might exacerbate the problem. Galleri is “not terrible” at detecting higher-grade prostate cancers, but that is when they are inherently more detectable by any cell-free DNA technology.
The case-controlled CCGA study included five years of longitudinal follow-up and Kaplan-Meier survival curves broken, revealing a “clear separation” between detected versus non-detected cancer types even at the early stages, she reported. “Cancers that are not detected do better than cancers that are detected.”
These are not “born-to-be-bad cancers that are going to kill you anyway,” as some have suggested, said Hall. Patients with undetected cancers in stage 4 were doing “way better than would otherwise be expected,” based on SEER data. “There is something biological in the cancers, even at late stages, that is defining how these patients do.” The cancers that are detected do “just about as we would expect, if not a little bit better,” she added.
When modeled with a small sample set, the expected population-scale impact of implementing Galleri in addition to the single-cancer screens would be a shift from late- to early-stage stage diagnosis potentially translating into a roughly 40% reduction in deaths in the cancers detected, said Hall. The calculation is based on real-world evidence coming through Grail’s commercial channel on patients who have had two years of testing. The plan is to do the same analysis in larger cohorts.
As voluntarily reported to GRAIL by ordering physicians, and admittedly “super biased,” Galleri has had 462 confirmed cancer diagnoses to date based on short-term follow-up out of 809 cancer signal detection cases, Hall said. The list notably includes 18 stage one cancers and 12 stage two cancers.