Dana-Farber AI Tool Shows Promise in Predicting Cancer's Primary Source

Researchers at Dana-Farber Cancer Institute have developed an AI-based tool, OncoNPC, that utilizes tumor gene sequencing data to predict the primary source of a patient's cancer. The study, published in Nature Medicine, suggests that OncoNPC could help guide treatment of cancer and improve outcomes in difficult to diagnose cases.

Currently, cancer's primary source is typically diagnosed through a standardized diagnostic work-up involving radiology and pathology assessments based on tumor biopsy slides. However, in 3-5% of cancer cases, the primary source remains unknown, leading to diagnoses of cancers of unknown primary (CUP). Patients in this category often face limited treatment options due to the lack of specific approved therapies for their cancer type, resulting in dismal outcomes.

The researchers discovered that OncoNPC's predictions could significantly benefit these patients. By providing additional diagnostic information about the tumor's primary source, the tool could assist physicians in selecting more effective treatments that improve survival rates.

To develop OncoNPC, the team trained and validated a machine learning classifier using the medical records of 36,445 patients with known primary tumors from major cancer centers, including Dana-Farber. This comprehensive dataset contained both tumor genetic sequencing data and clinical information for each patient.

To enhance clinical trust and transparency, the researchers selected an interpretable machine learning model for OncoNPC, enabling clinicians to understand the reasoning behind the predictions. By analyzing genetic factors that contributed to the model's predictions, especially in the enigmatic nature of CUP tumors, the researchers aimed to offer more actionable insights.

OncoNPC demonstrated high accuracy, predicting the origin of approximately 80% of tumors with known types, including metastatic tumors. For 65% of these cases, the model made high confidence predictions with 95% accuracy. Furthermore, when applied to a separate database of 971 CUP tumors from patients seen at Dana-Farber, OncoNPC accurately predicted the tumor's origin with high confidence in 41.2% of the cases.

The researchers validated these predictions by aligning them with inherited germline risks of cancer and examining the patients' pathology results, medical history, and genetic mutations. The results demonstrated that OncoNPC's predictions held strong, making it a promising tool for guiding precision treatments.