Cancer treatment often depends on where a tumor began. That answer can be difficult when cancer has spread or when a biopsy does not fit one clear tumor type. This study examined whether molecular data already collected during tumor profiling could give doctors another clue. ^[1]

The system did not make a diagnosis on its own. When its prediction disagreed with the submitted diagnosis, a molecular pathologist reviewed the case and looked for supporting evidence such as imaging, tissue staining, characteristic mutations, gene fusions or viral reads. ^[1]

The model combines two kinds of molecular information from a tumor sample: changes in DNA and patterns of gene activity measured through RNA. It learned associations between those measurements and 90 reportable cancer labels from 201,612 previously profiled cases. ^[1]

The researchers tested the model retrospectively and prospectively, then examined eight months of routine use. Among all cases profiled during that period, the review process led to a changed diagnosis for 704 patients, or 0.88 percent. These were unusual cases where the model's disagreement triggered a closer look, not a change applied to every patient. ^[1]

A physician survey received 97 responses. Fifty-two respondents said the result prompted a treatment-plan change. Because the response rate was about 10 percent and respondents were selected by convenience, the survey cannot show how often this would happen across all clinicians or hospitals. ^{[1] [1]}

The study did not have enough follow-up to determine whether using the system improved survival or other patient outcomes. It also excluded samples that could not undergo molecular profiling, including some biopsies with too little usable tissue. ^[1]

The work came from Caris Life Sciences, the company that developed and operates the test. The authors reported employment and stock interests, and two authors reported a pending patent. External evaluation in other laboratories would help show how well the approach travels beyond this setting. ^[1]