A new AI model helps identify the source of cancer in patients. This enables targeted treatments to be applied more often in cancer patients, including tumours that so far cannot be properly classified.
Increasingly, cancer patients are receiving targeted treatments. Here, medication is precisely tailored to the specific cancer form patients have. In practice, the method is more effective and has fewer side effects than generic treatments. However, doctors cannot identify the source of their cancer cells in a three to five per cent of cancer patients. This complicates selecting the right treatment for these patients.
AI model classifies unknown tumours
An AI model developed by researchers at the Massachussets Institute of Technology (MIT) may help. The OncoNPC AI model should make it easier to identify the origin of such cancer types and classify the species. The model uses machine learning and analyses the sequence of 400 genes. The model uses this information to predict the location in the body where the tumour originated.
"A significant number of people each year develop these cancers of unknown primary nature, 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," says Alexander Gusev, assistant professor of medical sciences at Harvard Medical School and Dana-Farber Cancer Institute. Gusev is lead author of the study.
Making more patients eligible for targeted treatment
The researchers report that the model is able to classify with high certainty at least 40% of tumours whose origin is not clear. Although this also means that in 60% of cases it fails to do so, the model increases the number of patients eligible for targeted cancer treatment by a factor of 2.2.
"That was the key finding in our paper, that this model could potentially be used to support treatment decisions, guiding doctors towards personalised treatments for patients with cancer of unknown primary origin," said Intae Moon, a graduate student in electrical engineering and computer science at MIT and lead author of the study.
Trained on data from nearly 30,000 patients
The model was trained on data from nearly 30,000 patients diagnosed with one of 22 types of cancer. The dataset includes patients from several US hospitals, including Memorial Sloan Kettering Cancer Center, Vanderbilt-Ingram Cancer Center and Dana-Farber Cancer Institute. The resulting model was tested on about 7,000 tumours that the model had not seen before. The origin of these 7,000 tumours is known. In doing so, the model managed to predict with an 80% accuracy which cancer type patients have.
In 65% of the tumours on which the model was tested, doctors could identify the type of cancer with high certainty. When the model analyses these tumours, it can identify the correct cancer type with 95% accuracy.
Classifying unknown tumours
In a follow-up test, the model was unleashed on data relating to 900 tumours that doctors could not classify. In 40% of cases, the model managed to classify these tumours with high certainty. The researchers checked these findings of the model in several ways. Among other things, they looked at the survival time of patients with these tumours. In many cases, this survival time was found to match the cancer type identified by OncoNPC.
In addition, the researchers looked at inherited mutations in genes, which reveal a genetic predisposition to developing a particular type of cancer. The model's predictions were found to match patients' genetic predispositions in many cases. Finally, researchers looked at patients with an unknown form of cancer who nevertheless received targeted treatment from their doctors. Treatment in this case was based on a guess. Patients who received targeted treatment in line with the type of cancer predicted by the AI model were found to have better outcomes than patients where this differed.
Applying targeted treatment more often
Finally, the researchers report that an estimated 15% more patients could have received targeted treatment if the form of cancer they suffer from was known. Since this was not the case, these patients were now receiving more generic treatment.
"That may make these findings more clinically useful because we are not requiring a new drug to be approved. What we are saying is that this population can now qualify for precision treatments that already exist," Gusev says.
The researchers have shared their findings in a paper published in *Nature Medicine*.
Author: Wouter Hoeffnagel
Image: Gordon Johnson via Pixabay
