Dana-Farber AI traces unknown cancers back to their source: study

Researchers at Dana-Farber Cancer Institute have developed an artificial intelligence tool that could help trace a cancer back to its site of origin in particularly tricky cases.

While a standard diagnostic work-up can include digital imaging scans, tumor biopsies and pathology reports, clinicians are still unable to find the original source of the malignancy in about 3% to 5% of patients.

Known as cancers of unknown primary origin, they are cases of metastatic disease that has spread from a separate and hidden location in the body. These patients may have few treatment options, as most therapies are approved for a specific type of tumor, according to Dana-Farber.

“This patient group has dismal outcomes,” said institute researcher Alexander Gusev, Ph.D., who served as senior author of a study published in the journal Nature Medicine. “This could open the door to more precision treatment for these patients.”

The AI program, dubbed OncoNPC, uses sequencing data taken from tumor DNA to help predict where it came from. The computer model was built using medical records and genetic testing results gathered from more than 36,400 patients. 

According to a retrospective analysis in the study, the machine learning-powered algorithm’s success in classifying patients with these hard-to-diagnose cancers could have more than doubled the number of people eligible for approved, precision treatments.

OncoNPC was able to accurately predict the origin of about 80% of tumors, including metastatic disease, by linking them with known cancer types, according to Dana-Farber. 

And out of the findings that the model described as high-confidence predictions—which amounted to 65% of a group of samples used for validation—researchers showed it was correct about 95% of the time.

“We see the OncoNPC prediction as a nudge, a way to provide a possible explanation for the cancer that helps point to appropriate treatment, including precision medicine,” Gusev said.

Going forward, researchers plan to improve the model’s performance by feeding it additional types of information, such as pathology reports. They also aim to explore how it can complement other diagnostic techniques in the community cancer treatment setting.