Radiology and pathology assessments are the gold standard for diagnosing cancer. But for a small percentage of cancer cases these techniques fail to locate the primary site of a metastatic tumour, which is then classified as a cancer of unknown primary (CUP).
Such CUPs, which represent 3–5% of all cancers, pose unique challenges, such as difficulties in selecting an appropriate treatment plan. The lack of knowledge about the primary site hinders the prescription of precision drugs that are approved for specific cancer types. Such targeted treatments have been shown to be more effective and less invasive than broad-spectrum treatments. But patients with CUP often find themselves without such targeted therapies.
Now, a research collaboration from MIT and Dana-Farber Cancer Institute has come up with a potential solution to this long-standing problem. The researchers have harnessed the power of machine learning to develop a computational model that can predict the site of origin of CUPs.
In their study, published in Nature Medicine, Alexander Gusev and his team used machine learning to predict cancer type based on genetic data. By training their machine learning model on data from almost 30,000 patients diagnosed with 22 known cancer types, the researchers created a tool called OncoNPC. This tool successfully predicted the origins of about 80% of 7289 known tumour samples, and this accuracy rose to nearly 95% for tumours with high-confidence predictions (about 65% of the total). By analysing the genetic sequence of around 400 genes, OncoNPC can accurately predict the origin of tumours and, as such, could significantly improve treatment options for cancer patients.
Building on this success, the researchers applied the model to a dataset of 971 tumours from patients with CUP. The model accurately predicted the origin of at least 40% of these tumours, representing a significant improvement in treatment accuracy for this historically challenging group.
Moreover, the researchers correlated the model’s predictions with germline mutations, inherited genetic changes that can indicate a predisposition to certain cancers. The model’s predictions were notably aligned with the type of cancer suggested by the germline mutations, further validating its accuracy.
“That was the most important finding in our paper, that this model could be potentially used to aid treatment decisions, guiding doctors toward personalized treatments for patients with cancers of unknown primary origin,” explains lead author Intae Moon, an MIT graduate student.
The practical implications of this breakthrough are substantial. Survival data analysis demonstrated that CUP patients predicted by the model to have cancer with a poor prognosis indeed had shorter survival times, while those predicted to have cancer types with better prognoses showed longer survival times. Additionally, the model identified a group of patients who could have benefited from existing targeted treatments had their cancer type been known, potentially sparing them from broad-spectrum chemotherapy drugs.
Novel AI tool predicts colorectal cancer survival from pathology images
Next, the researchers plan to enhance their model by integrating additional data, such as pathology and radiological images. This holistic approach could offer comprehensive insights into tumours, facilitating predictions not only about the cancer type and patient outcomes, but even potentially guiding optimal treatment decisions.
With the convergence of machine learning and medical science, this advanced research shines a light on the future of personalized cancer treatment for patients whose cancers have long puzzled the medical community.