Most popular DNA screenings for cancer risk test only a single gene site, like BRCA1. But Grail’s chief medical officer Josh Ofman tells me, “cancer may show up as thousands of methylation changes, a much richer signal to teach machine learning algorithms to find cancer” vs. a single site. “There are 30 million methylation sites in the entire human genome on 100,000 DNA fragments. Grail looks at a million of them.” It takes industrial-grade artificial intelligence to find patterns in all this data, something a human eye would never see.
Mind you, this is not a consumer 23andMe test of your genome that says you might have, say, a 68% chance of getting cancer. Grail is detecting the signature of actual cancer cells in your blood. According to validation data published in the Annals of Oncology, the test can find 50 different types, more than half of all known cancers.
And it can find cancer well before symptoms show up, in Stage I or sooner, when therapies are cheap and effective. If Grail can scale, it will be a massive game-changer. Five-year survival rates can approach 9 in 10 if cancer is detected early, compared with about 1 in 5 in Stage IV.
And while worrywarts brood over artificial intelligence and robot overlords, early detection of cancer is really what machine learning is meant for. This is the Holy Grail.