Journal Club: CCGA (GRAIL)

Study Design

The Circulating Cell-free Genome Atlas (CCGA) was a large, prospective, case-control, observational study that enrolled 15,254 participants across 142 sites in North America between August 2016 and February 2019. 8,584 participants had a recent cancer diagnosis, and 6,670 did not.

CCGA was designed as three pre-specified sub-studies.

1. The first compared different genomic approaches and found that DNA methylation patterns outperformed the alternatives (whole-genome sequencing and targeted mutation detection).
2. The second refined the methylation assay and trained the machine learning classifiers.
3. This paper reports the third and final sub-study: clinical validation.

The validation set included 4,077 participants (2,823 with cancer and 1,254 without), completely independent from the training data. Among the group of participants with cancer, 55% had stage I or II disease.

This study utilized a case-control design and was not a screening study. Participants were enrolled because they either had a known cancer diagnosis or did not, which is a critical distinction from testing the general population, who would be using this test to detect cancer.

Results

Specificity was 99.5% (95% CI: 99.0--99.8%). Of 1,254 people without cancer, only 6 received a false positive result, yielding a false positive rate of 0.5%. Overall sensitivity was 51.5% (95% CI: 49.6--53.3%) across all cancer types and stages, meaning the test detected a cancer signal in roughly half of known cancer cases. Sensitivity in a group of 12 pre-specified cancers (anal, bladder, colon/rectal, esophageal, head and neck, liver/bile duct, lung, lymphoma, ovarian, pancreatic, plasma cell neoplasm, and stomach), which were chosen reportedly based on being responsible for roughly two-thirds of annual cancer deaths in the US, was 67.6% at stages I-III and 76.3% at stages I-IV. Cancer signal origin accuracy was 88.7% (95% CI: 87.0--90.2%) in true positive cases. Modeled population estimates for adults aged 50-79 yielded a PPV of 44.4% and a negative predictive value of 99.4%.

What I Think

The good news:

A 99.5% specificity (a 0.5% false positive rate) is remarkable for any screening test. The authors note that existing recommended single-cancer screening tests have false positive rates ranging from 9 to 14.5%.

For the 12 pre-specified high-mortality cancers without approved screening tests in people without risk factors, such as pancreatic, ovarian, liver, and esophageal, sensitivity was 67.6% for stage I-III disease. A significant number of patients with these types of cancers present at an advanced stage where treatment is not curative. Detecting these cancers at an earlier stage, when cure is still the goal, has the potential to change outcomes for patients who, upon diagnosis, have no good options.

Several caveats exist.

This is a case-control study, not a screening study. Over 60% of the participants had cancer and were enrolled to calculate sensitivity. Presumably, a significant proportion of their cancers were symptomatic, which led to their cancers being clinically diagnosed in the first place. It's possible the sensitivity in this study likely overestimates what you'd see in a true asymptomatic screening population.

The 88.7% accuracy in predicting cancer signal origin is important. A positive MCED result without a signal of where the cancer is coming from would be a major problem. It would be difficult and expensive to tell a patient they have cancer somewhere and then figure out where to look. An 11.3% misdirection rate means roughly 1 in 9 true positives will initially point clinicians toward the wrong organ.

Detecting only 16.8% of stage I cancers seems to undercut the entire promise of "early detection." But this is across all cancer types, and many early-stage cancers simply don't shed enough cfDNA into the bloodstream to be detected with current technology.