OncoNPC – the AI tool that can identify a cancer at its source
Researchers at the Dana-Farber Cancer Institute (MA, USA) have developed an AI-based prediction tool, termed OncoNPC, that could be used to identify the primary source of a patient’s tumor in a clinical setting.
The primary source of a cancer is traditionally identified though radiological and pathological assessments. These methods fail to identify the source of a cancer in 3–5% of patients, these patients are therefore instead diagnosed with cancers of unknown primary (CUP).
As most treatments are cancer-specific, this group of patients have fewer treatment options available to them. Unfortunately, this means CUP patients also have a poorer prognosis.
A paper published in Nature Medicine describes the Oncology NGS-based primary Cancer type Classifier (OncoNPC), an AI predictive tool that has shown promise in predicting the primary origin of a cancer.
To develop and validate the tool, scientists utilized the medical records of 36,445 patients spanning 22 different cancer types from three different institutions. Specifically, the researchers used the tumor genetic sequencing data from these records, as well as clinical patient information.
The tool was tested on many different types of cancers of a known origin, including those that had metastasized, to tests its validity. The OncoNPC accurately estimated the origin of approximately 80% of the tumors with a known origin.
When OncoNPC was applied to database of CUP tumors that clinicians had made significant efforts to identify the primary source and was found to be effective in 41.2% of cases.
It is difficult to fully confirm the effectiveness of OncoNPC at predicting the origin of CUP cancers as there can be no validation from traditional methods. This was highlighted by one of the authors of the paper Alexander Gusev (Dana-Farber Cancer Institute) “Validation is a challenge because there is no ground truth. Existing methods failed to identify the origin”.
“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.”
However, the research team have reviewed the genetic risks of cancer of patients, and these line up with the predictions made by OncoNPC.
The team of researchers responsible for the paper opted to design OncoNPC with interpretability and transparency as priorities, aiming to instil confidence in clinicians and encourage their trust in the tool.
As a next step, this tool would need to be tested in a clinical trial. Additionally, the authors of the paper hope to collaborate with a community cancer center to gain deeper insights into how OncoNPC predictions could enhance and work alongside current diagnostic methods.