Detection and localization of surgically resectable cancers with a multi-analyte blood test
Comprehensive Overview
Published in Science in 2018, this study asked whether a single blood test could detect multiple common cancers before they spread, while also hinting at which organ a cancer came from, information that normally requires extensive additional workup. Researchers developed a test called CancerSEEK that combines two types of signals: mutations in tiny fragments of tumor DNA that leak into the bloodstream, and the levels of eight proteins previously linked to cancer. They evaluated the test on blood samples from 1,005 patients already diagnosed, mostly through symptoms, with stage I to III cancers, meaning cancers that had not yet spread to distant organs, of the ovary, liver, stomach, pancreas, esophagus, colon or rectum, lung, or breast, and compared results with blood from 812 healthy people with no known history of cancer. A statistical scoring model combined the DNA and protein results into a single positive or negative call, and a separate machine-learning step used the same data to predict which organ a positive result likely came from. As expected from earlier, narrower studies of blood-based cancer detection, the test correctly identified a majority of cancers overall while very rarely flagging healthy people as positive.
Outcomes and Implications
The most important finding is that CancerSEEK detected a median of 70% of cancers across the eight types tested while incorrectly flagging fewer than 1% of healthy individuals, just 7 out of 812. Performance varied enormously by cancer type: detection reached 98% for ovarian cancer but fell to just 33% for breast cancer, and for five cancer types with no recommended screening test at all, ovary, liver, stomach, pancreas, and esophagus, the test caught 69% to 98% of cases, arguably the most valuable result given these cancers are usually found only after symptoms appear. Detection also improved with more advanced local disease, catching only 43% of the earliest-stage cancers versus 78% of stage III cancers, meaning the test is weakest exactly where early intervention matters most. When the test did flag a cancer, machine learning using the same blood data narrowed the likely organ of origin to one of two sites in 83% of cases and to a single organ in 63%, which could meaningfully speed up follow-up testing. Because all patients here already had known, symptomatic cancers rather than being screened before any suspicion arose, and healthy controls excluded anyone with other illnesses that might cause false alarms, real-world screening performance is likely lower, an estimated 55% sensitivity once adjusted for how common each cancer actually is. The authors, estimating the test could cost under $500, conclude that large prospective studies are needed to determine whether CancerSEEK can actually reduce cancer deaths.