Cancer cartographer navigating tumor origins with AI precision

An AI model, OncoNPC, has shown great accuracy when predicting the origins and type of cancer in patients. In a subset of cancer patients (~3–5% ) the challenge arises, particularly in cases where tumors have spread extensively, where oncologists encounter difficulty in identifying the primary source of a cancer. As such, malignant tumors fall into the category of cancers of unknown primary (CUP). This informational gap frequently limits medical practitioners from prescribing "precision" medications to patients. These drugs are typically used for well-defined cancer types with established efficacy. Unlike treatments that are commonly administered to CUP patients and often lead to broader side effects, these targeted medications demonstrate higher effectiveness and fewer adverse reactions. Professor Alexander Gusev, senior author of the paper stated, “A sizeable number of individuals develop these cancers of unknown primary every year, and because most therapies are approved in a site-specific way, where you have to know the primary site to deploy them, they have very limited treatment options.” Researchers from MIT (MA, USA) and the Dana-Farber Cancer Institute (MA, USA) have pioneered an innovative method that could simplify the process of pinpointing the sources of elusive cancers. Through the utilization of machine learning, the team devised a computational framework that is capable of analyzing approximately 400 gene sequences. This model then harnesses this genetic data to determine the specific location within the body from which a particular tumor likely originated. Utilizing this framework, the team demonstrated their ability to precisely categorize a minimum of 40% of unidentified origin tumors with high confidence, within a dataset covering approximately 900 patients. This method led to a notable 2.2-fold increase in the number of patients who could potentially be suitable for personalized genomically guided targeted therapies, dependent on the cancer's suspected point of origin. “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,” explained Intae Moon, an MIT graduate and lead author of the new study. Moon analyzed routinely collected genetic data at Dana-Farber, seeking its potential for predicting cancer types which encompassed genetic sequences of around 400 frequently cancer-mutated genes. With a dataset of almost 30,000 patients diagnosed with 22 distinct cancer types, including data from Memorial Sloan Kettering Cancer Center (NY, USA), Vanderbilt-Ingram Cancer Center (NA, USA), and Dana-Farber, researchers trained a machine-learning model, named OncoNPC. Testing it on 7,000 previously unseen tumors with known origins, the model achieved approximately 80 percent accuracy, rising to 95% for high-confidence predictions, constituting about 65% of cases. Encouraged by these outcomes, the team analyzed 900 tumors from Dana-Farber's CUP patients using the OncoNPC model which yielded high-confidence predictions for 40% of these cases. By comparing its predictions with germline mutation analyses in available tumor subsets, the researchers found a strong alignment between the model's predictions and the cancer type most indicated by the germline mutations. Validating the model, researchers correlated CUP patients' survival with predicted cancer types. Those predicted with poor prognosis cancers, like pancreatic cancer, exhibited shorter survival, while predictions aligned with better prognosis, like neuroendocrine tumors, saw extended survival. Analysis of treatment types received by CUP patients revealed promising results and among those receiving targeted treatments based on oncologists' estimates, alignment with the model's predictions led to better outcomes. This is when compared to treatments that would usually be given for a different type of cancer than what the model predicted. The model also identified 15% more patients eligible for targeted treatments, highlighting the potential for improved therapy decisions, and avoiding general chemotherapy drugs when specific options were available. Gusev stated, “That potentially makes these findings more clinically actionable because we’re not requiring a new drug to be approved. What we’re saying is that this population can now be eligible for precision treatments that already exist.” The researchers are aiming to broaden their model's scope by incorporating diverse data forms like pathology images and radiology images. This expansion aims to enhance predictions by integrating multiple data modalities. By doing so, the model would gain a holistic understanding of tumors, enabling predictions not only of tumor type and patient prognosis but potentially also guiding optimal treatment choices.