People with cancer have different amounts of a type of repetitive DNA — called Alu elements — than people without cancer. Now, machine learning can measure that from a blood draw. Researchers at the Johns Hopkins Kimmel Cancer Center have used this finding to improve a test that detects cancer early, validating and reproducing the results by starting with a sample size tenfold larger than typical of such types of studies.
The research was published Jan. 24 in the journal Science Translational Medicine.
Alu elements are small: around 300 base pairs long out of 2 billion steps in a DNA ladder. But, changes in the proportion of Alu elements in people's blood plasma occur regardless of where cancer originates, explains lead study author Christopher Douville, Ph.D., an assistant professor of oncology at Johns Hopkins.
"Blood testing holds great promise for the earlier detection of cancers before people exhibit any symptoms," Douville says. "However, analyzing results with machine learning "has not necessarily translated into long-term success for patients when minor fluctuations produce widely different predictions in these complex models. To have a long-term impact on patient care, physicians and patients must have confidence that models consistently and reproducibly classify cancer status. In our manuscript, we evaluated 1,686 individuals multiple times to assess whether our machine learning model consistently delivers the same answer."
Douville and colleagues developed a test to detect aneuploidy, chromosome copy number alterations found in cancers. The test measured aneuploidy through a blood test called liquid biopsy, which detects fragments of cancer cell DNA circulating in the bloodstream.
However, Douville observed an unexplained signal that distinguished cancer from noncancer but could not be explained by chromosomes being gained or lost.
The team decided to combine their previous test — able to check 350,000 repetitive locations in DNA — with an unbiased machine learning approach.
Douville and colleagues collected samples from 3,105 people with solid cancers and 2,073 without. The study covered 11 cancer types and 7,615 blood samples. The repeats were used as replicates to see how well the model worked. They reached 98.9% specificity, which meant they could minimize false-positive test results. "This is crucial when screening asymptomatic patients, so people aren't told incorrectly that they have cancer," says Douville.
In an independent validation cohort, adding Alu elements to the machine learning model caught 41% of cancer cases missed by eight existing biomarkers and the group's previous test, making "a greater contribution," authors wrote in the paper, "than aneuploidy or proteins." The type of repetitive DNA contributing most to cancer detection was the largest subfamily of Alu elements, called AluS; the blood plasma of people with cancer had less of it than usual.
