AI behind potential blood test for ovarian cancer

The power of artificial intelligence (AI) is behind the development of a new accurate method for detecting early-stage ovarian cancer. The findings from researchers at Brigham and Women’s Hospital (BWH) and Dana-Farber Cancer Institute (DFCI), both in Boston, MA, were recently published in eLife.

The team identified a network of circulating microRNAs—small, non-coding pieces of genetic material—that are associated with ovarian cancer risk. MicroRNA can be detected from a blood sample.

“MicroRNAs are the copy editors of the genome,” according to lead author Kevin Elias, MD, of BWH’s Department of Obstetrics and Gynecology. “Before a gene gets transcribed into a protein, they modify the message, adding proofreading notes to the genome.”

Most women are diagnosed with ovarian cancer when the disease is at an advanced stage, at which point only about one-quarter of patients will survive for at least 5 years.

For women whose cancer is identified at an early stage, however, survival rates are much higher. Currently, there is no FDA-approved screening technique for ovarian cancer, which presents a diagnostic challenge in women with a genetic predisposition for the disease and in the general population.

Ovarian cancer is relatively rare compared to benign gynecological conditions such as ovarian cysts, but early detection tests have a high false-positive rate for ovarian cancer. In addition, clinical trials have found that these tests do not have a meaningful impact on survival rates when they are used to detect early-stage ovarian cancer.

The Dana-Farber and BWH team sought a tool that would be more sensitive and specific in detecting early-stage disease, which led them to the set of molecules called microRNAs.

“This project exemplifies the synergy of the two institutes DFCI and BWH and the power of clinicians working closely with lab-based scientists,” said senior author Dipanjan Chowdhury, PhD, chief of the Division of Radiation and Genomic Stability in the Department of Radiation Oncology at DFCI. “My lab has been working on microRNAs for a decade. When Kevin [Elias] came to us with the patient samples, it was a no-brainer to initiate this project.”

In the laboratory, Drs. Elias and Chowdhury and their colleagues determined that ovarian cancer cells have different microRNA profiles than normal cells. Unlike other parts of the genetic code, microRNAs circulate in the blood, making it possible to measure their levels from a serum sample.

MicroRNAs were sequenced in blood samples from 135 women before surgery or chemotherapy. This created a “training set” to help a computer program identify microRNA differences between cases of ovarian cancer and cases of benign tumors, non-invasive tumors, and healthy tissue.

Using this machine-learning approach, the team could leverage large amounts of microRNA data and develop different predictive models. The model that most accurately distinguished ovarian cancer from benign tissue is known as a neural network model, which reflects the complex interactions between microRNAs.

“When we train a computer to find the best microRNA model, it’s a bit like identifying constellations in the night sky,” said Dr. Elias. “At first, there are just lots of bright dots, but once you find a pattern, wherever you are in the world, you can pick it out.”

The team then tested this sequencing model in an independent group of 44 women to determine the test’s accuracy. After confirmation, they deployed the model across multiple patient sample sets, using a total of 859 samples to measure the sensitivity and specificity of the model.

The new technique was far better at predicting ovarian cancer than ultrasonography. With ultrasound, fewer than 5% of abnormal test results would be ovarian cancer. With the microRNA test, almost 100% percent of abnormal results represented ovarian cancer.

The group used their final model to predict the diagnoses of 51 patients presenting for surgical care in Lodz, Poland. In this population, 91.3% of the abnormal test results were ovarian cancer cases. Negative test results reliably predicted absence of cancer in about 80% of cases, which is comparable to the accuracy of a Pap smear.

“The key is that this test is very unlikely to misdiagnose ovarian cancer and give a positive signal when there is no malignant tumor,” said Dr. Chowdhury. “This is the hallmark of an effective diagnostic test.”

The team also looked for evidence of biological relevance for distinguishing microRNAs. The researchers found changes in the quantity of these microRNAs in blood samples collected before and after surgery, which suggests that the microRNA signal decreases after cancerous tissue is removed. They also imaged microRNAs from patient samples in cancerous cells, demonstrating that the serum signal was coming from cancerous tissues.

To move the microRNA tool into clinical practice, the research team will need to verify how the microRNA signature changes over time as risk of ovarian cancer increases. This effort will involve the use of prospectively collected, longitudinal samples and following women over time. The team is particularly interested in determining whether the tool will be useful for both women at high risk of ovarian cancer and the general population.

Funding for this work was provided by the Robert and Deborah First Family Fund, NICHD K12HD13015, the Ruth  N. White Research Fellowship in Gynecologic Oncology, the Saltonstall Research Fund, Potter Research Fund, the Sperling Family Fund Fellowship, the Bach Underwood Fund, First TEAM grant of the Foundation for Polish Science and the Smart Growth Operational Programme of the European Union, NIH P50CA105009, Department of Defense grants OC093426 and OC140632, and the Honorable Tina Brozman Foundation.

To read more about this study, click here.