AI Helps Untangle Cancer Mysteries

Patients with cancer of unknown primary (CUP) face treatment challenges because many therapies are indicated only for diseases tied to a specific anatomic site. Researchers have begun tackling this problem using artificial intelligence (AI) to predict the source of CUP based on a cancer's histologic features (Nature 2021;594:106–10) or patients’ gene expression sequencing (Cancer Discov 2022;12:2566–85). A new strategy that determines mutations in tumor DNA, dubbed Onco­NPC, demonstrates the potential for clinical benefit by identifying cancer type and increasing patients’ eligibility for targeted therapies (Nat Med 2023;29:2057–67).

“We observed that a large number of CUP patients have somatic alterations that may be eligible for existing targeted therapies and the retrospective finding that those CUP patients who receive therapies that are concordant with their molecular classification have significantly better outcomes than those who do not,” says senior author Alexander Gusev, PhD, of Dana-Farber Cancer Institute (DFCI) in Boston, MA.

OncoNPC predicted primary cancer types with high confidence in 41.2% of 971 CUP tumors collected at DFCI. In a retrospective analysis, study participants with CUP who received treatment concordant with their AI-predicted cancers were estimated to be about 65% less likely to die of their disease than patients who did not. Based on the researchers’ model, if OncoNPC was clinically implemented it could have enabled a more than twofold increase in patients with CUP who receive genomically guided therapies. However, researchers cautioned that prospective randomized trials need to confirm the model's suggested benefits.

In the nearly 60% of cases in which OncoNPC could not identify the source of a CUP tumor, the researchers lacked sufficient data to train the algorithm, or disease states were too poorly differ­entiated for AI to distinguish between them, highlighting OncoNPC's limits. They believe that their algorithm is best used in conjunction with conventional diagnostic tools.

To train the tool, records from more than 36,000 patients with known primary solid tumors and next-generation sequencing (NGS) data were used. This large dataset and the fact that it was drawn from three institutions distinguish this study from its predecessors. “We also think the use of next-generation panel sequencing has particular practical relevance, as many cancer hospitals are now using such panels routinely,” notes Gusev.

Salil Garg, MD, PhD, of the Yale School of Medicine in New Haven, CT, who was not associated with this research, agrees. “It is significant and different that this group is primarily using molecular sequencing features to look at mutations in patients’ DNA. Creating an integrated multimodal diagnostic tool that includes looking at mutation signatures, tumors under a slide, and gene expression is where the field is going—this research is a key step in that direction,” he says.

Before a more robust AI tool that can help diagnose CUP is widely available, Garg adds, machine-learning algorithms and the infrastructure for training them will need to be standardized across wide swaths of the health care system.

Despite the limitations of Onco­NPC—that it identifies cancer solely based on NGS, for example—the tool and interpretations for each cancer classification included within it are publicly available at https://github.com/itmoon7/onconpc. “The algorithm produces interpretable predictions, so we hope this is a resource that even oncologists without much algorithmic expertise could derive value from,” concludes Gusev. –Myles Starr

For more news on cancer research, visit Cancer Discovery online at http://cancerdiscovery.aacrjournals.org/CDNews.