Form of Artificial Intelligence Proves Superior in Identifying Inherited Cancer-Related DNA Variants

February 1, 2021

A new “deep learning” form of artificial intelligence outperformed standard methods in identifying cancer patients with inherited DNA alterations that could increase their risk of developing cancer or improve their response to certain targeted cancer drugs.

Dana-Farber researchers led by Saud H. AlDubayan, MD, and Eliezer Van Allen, MD, report in JAMA that the analytical method known as “Deep Variant” — a special application of deep learning — identified more inherited disease-causing (also known as pathogenic) variants in DNA from melanoma and prostate cancer patients than were detected by standard techniques. These variants — mutations and other changes — occurred in the germline DNA inherited by patients and were present since conception.

A test of machine learning

Genetic testing of germline DNA is increasingly performed to detect genetic variants that might confer a higher than average risk of developing cancer. Some germline DNA variants make individuals more likely to respond to targeted cancer drugs. Less than 10% of patients with cancer have pathogenic variations that are detectable with current methods.

The aim of the study was to “assess the performance of the standard method to detect germline genetic variants in cancer patients and whether we can use recent advances in machine learning techniques to further improve the detection rate of these genetic alterations,” says AlDubayan. While the standard method and “deep learning” are both forms of machine learning, or artificial intelligence, deep learning has greater potential to detect rare genetic variants.

The investigators compared the two methods in analyzing the genetic sequencing data of 1,072 men with prostate cancer and 1,295 patients with melanoma. The primary objective was to search for DNA variants in 118 genes associated with a predisposition to cancer. The deep learning method identified more patients with pathogenic variants in those genes than the standard method: 198 vs 182 for prostate cancer and 93 vs 74 for melanoma. Not only was the deep learning method better at identifying pathogenic variants, it was also faster and less expensive to run compared with the standard method.

While they focused on prostate cancer and melanoma, the researchers say the deep learning method identified risk variants for other types of cancer. For example, several patients in the study were found to have pathogenic germline variants that are linked to a higher risk of ovarian cancer, and for which removal of the ovaries at a particular age is strongly recommended. The standard method did not detect those variants.

“Further research is needed to understand the relevance of these findings with regard to clinical outcomes and whether they are generalizable to other conditions and patient populations,” the researchers say.