Nature Medicine: AI model can help determine where a patient's cancer arose

Source: Biocom

For a small percentage of cancer patients, doctors cannot determine where their cancer originated. This makes it more difficult to choose a treatment for these patients, since many cancer drugs are often developed for specific cancer types.

A new method developed by researchers at MIT and Dana-Farber Cancer Institute may make it easier to pinpoint where these mysterious cancers originate. Using machine learning, the researchers created a computational model that can analyze the sequence of about 400 genes and use this information to predict where in the body a given tumor originates.

Using this model, the researchers showed that they could accurately classify with high confidence at least 40 percent of tumors of unknown origin in a dataset of approximately 900 patients. This approach resulted in a 2.2-fold increase in the number of patients eligible for genome-guided targeted therapy based on the origin of their cancer.

“This is the most important finding in our paper, and the model could potentially be used to aid treatment decisions and guide physicians in personalizing treatment for cancer patients of unknown origin,” said Intae Moon, an MIT graduate student in electrical engineering and computer science. , who is the lead author of the new study.

Alexander Gusev, an associate professor of medicine at Harvard Medical School and Dana-Farber Cancer Institute, is the senior author of the paper published in the journal Nature Medicine.

Mysterious Origin

In the 3 to 5 percent of cancer patients, especially those whose tumors have metastasized throughout the body, oncologists don’t have an easy way to determine the cancer’s origin. These tumors were classified as carcinoma of unknown primary (CUP).

This lack of knowledge often prevents doctors from giving patients “precise” medicines, which are often approved for specific cancer types that are known to be effective. These targeted therapies tend to be more effective, with fewer side effects, than treatments used for a broad range of cancers, and are often used in patients with CUP.

“A considerable number of people get these cancers of unknown primary each year, and because most treatments are approved in a site-specific way, you have to know the site of origin to use them, so their treatments Choices are very limited.”

Moon, part of the Computer Science and Artificial Intelligence Laboratory, co-advises Gusev. Moon decided to analyze genetic data routinely collected at Dana-Farber to see if it could be used to predict cancer type. The data included genetic sequences for about 400 genes that are frequently mutated in cancer. The researchers trained a machine learning model on data from nearly 30,000 patients diagnosed with one of 22 known cancer types. The data set included patients from Memorial Sloan Kettering and Vanderbilt-Ingram Cancer Centers, as well as Dana-Farber.

The researchers then tested the model on about 7,000 never-before-seen tumors whose location of origin was known. The model, which the researchers named OncoNPC, was able to predict their origin with about 80 percent accuracy. For tumors predicted with high confidence (approximately 65% of the total), its accuracy rose to approximately 95%.

Following these encouraging results, the researchers used the model to analyze approximately 900 tumors from CUP patients, all from Dana-Farber. They found that for 40 percent of these tumors, the model was able to make predictions with high confidence.

The researchers then compared the model’s predictions with existing data analyzes of subsets of tumors for germline or genetic mutations, which can reveal whether a patient has a genetic predisposition to develop a particular type of cancer. The researchers found that the model’s predictions were more likely to match the type of cancer most strongly predicted by germline mutations than any other type of cancer.

Guiding Medication Decisions

To further validate the model’s predictions, the researchers compared survival time data for CUP patients with typical prognosis for the cancer type predicted by the model. They found that patients with CUP who were predicted to have a cancer with a poorer prognosis, such as pancreatic cancer, had a correspondingly shorter survival time. At the same time, CUP patients with cancers that typically have a better prognosis, such as neuroendocrine tumors, were predicted to live longer.

Another indication that the model’s predictions might be useful came from the types of treatments the CUP patients analyzed in the study received. About 10 percent of these patients received targeted therapy, based on oncologists’ best guesses about the cancer’s origin. Among these patients, those who received a treatment consistent with the type of cancer predicted by the model fared better than those who received a typical treatment that differed from the type of cancer predicted by the model.

Using this model, the researchers also identified an additional 15 percent of patients (a 2.2-fold increase) who would have received existing targeted therapies if their cancer type had been known. Instead, these patients ended up receiving more common chemotherapy drugs.

“This could make these findings more clinically actionable, because we don’t need new drugs to be approved. What we’re saying is that these people can now receive precision treatments that already exist,” Gusev said.

The researchers now hope to expand their model to include other types of data, such as pathology and radiology images, to provide more comprehensive predictions using multiple data modalities. This will also give the model a comprehensive view of the tumor, enabling it to predict not only the type of tumor and patient prognosis, but possibly even the best treatment options.