Skip to main content
Premium Trial:

Request an Annual Quote

Johns Hopkins Researchers Use Machine Learning for Early Detection of Ovarian Cancer

Premium
AdobeStock_486657193.jpeg

NEW YORK – Researchers from Johns Hopkins University's Sidney Kimmel Comprehensive Cancer Center have developed a classifier that utilizes artificial intelligence in combination with cell-free DNA fragmentation profiles and two protein biomarkers to detect ovarian cancer in earlier stages than is currently being done with conventional technology.

The approach leverages machine learning to analyze cfDNA fragmentation profiles across the genome, as well as established protein biomarkers cancer antigen 125 (CA-125) and human epididymis protein 4 (HE4). The researchers previously presented results from the study at the annual meeting of the American Association for Cancer Research in April, but in a paper published in Cancer Discovery, the researchers described the validation of their classifier in more detail. They found that it was able to detect ovarian cancer with more than 99 percent specificity and sensitivity between 69 percent and 100 percent, depending on the stage of ovarian cancer. In contrast, using the CA-125 biomarker alone detected ovarian cancer with sensitivity between 34 percent and 100 percent depending on the stage.

The classifier also differentiated between benign masses and ovarian cancers with an area under the curve of 0.88.

The researchers used next-generation whole-genome sequencing to sequence the collection of DNA from blood samples at low coverage, according to Victor Velculescu, codirector of the Cancer Genetics and Epigenetics program at the Kimmel Cancer Center and one of the authors on the paper. By sequencing across the entire genome, they were able to "generate a lot of information" that was then used in an artificial intelligence-based analysis to determine a patient's fragmentation profile. The fragmentation data was combined with the values of CA-125 and HE4 gathered from immunoassays to determine whether a patient was above a certain threshold potentially indicating cancer.

The classifier is a prototype that could ultimately lead to a clinically validated test for screening. In a clinical setting, a patient with a result above the threshold would then be further evaluated for ovarian cancer, he said.

Velculescu noted that both CA-125 and HE4 have been investigated for use as ovarian cancer detection biomarkers but "are not thought to be good enough on their own." CA-125, in particular, is an established biomarker that is higher in women who have ovarian cancer, but when used by itself, the biomarker was able to detect only 34 percent of stage I cancers, according to Velculescu.

Fragmentation profile changes, meantime, are a "hallmark of cancer because of the way the cancer genome is packaged aberrantly," he said. Adding fragmentation profiles to the two protein biomarkers lifted stage I sensitivity to 72 percent, according to the Cancer Discovery paper. The researchers found that fragmentation profiles were homogenous among people in the study without cancer and showed limited changes in people with benign adnexal masses.

However, fragmentation profiles from patients with ovarian cancer showed "marked heterogeneity both between patients and across different regions of the genome for the same individual," the research team wrote.

Combining fragmentation profiles with the protein biomarkers was able to boost the sensitivity of the classifier particularly in early-stage disease, which Velculescu said is essential because "that's the only stage at which early intervention has been shown to make a difference in some cancer types."

"Many tests historically just measure one thing," Velculescu said. But machine learning can help researchers measure a variety of factors, such as the size and other characteristics of DNA fragments and create a stronger method of detection.

"Each [piece of information] alone may not be very strong, but together they might find a pattern that is very powerful and that would not be easy for the human mind to see," he said.

The new classifier, called the DELFI Protein score (DELFI-Pro), was designed to have very high specificity to reduce false positives to avoid sending people on unnecessary "diagnostic odysseys," Velculescu said.

Because ovarian cancer is relatively rare and the next step after an elevated test result is often exploratory surgery, a clinician must balance the benefits of detecting and eventually removing a cancerous ovary with the potential harm that may come from performing unnecessary surgeries, he noted.

One of the benefits of the DELFI-Pro approach is that the test can be calibrated to different levels of specificity depending on what the end goal is, he said. For example, if there are settings where the next step would be an imaging test rather than surgery, lower specificity may be permissible.

The fragmentation technology was first described in a 2019 paper published in Nature that showed it was applicable to seven different cancer types, including lung and liver cancer. Delfi Diagnostics is currently using the technology for its clinically validated FirstLook Lung test. Velculescu is a founder and board director at Delfi.

Steven Skates, an associate professor of medicine at Massachusetts General Hospital and Harvard Medical School who was not involved in the development of the test, said that the ability of the classifier to differentiate between benign and cancerous pelvic masses was particularly useful to provide more accurate referrals to specialists. For example, in a community hospital setting where a gynecologic oncologist isn't available, the classifier could be used to refer patients with cancer to a tertiary care facility with gynecologic oncologists for next steps. But if a mass is benign, the operation could be performed at the community hospital, which would save time and money.

This application for the classifier is "immediately clinically relevant," he said.

However, for the early detection application of the classifier, Skates said that he would like to see the researchers validate the test in samples from ovarian cancer patients taken prior to diagnosis to see how much earlier cancer could be detected while maintaining a low false positive rate. To get those samples, a large screening trial of currently healthy patients with no symptoms would have to be conducted with routine follow-up or the researchers would need access to samples saved from a previously conducted screening trial, he noted.

The performance of the classifier would also likely be lower in a true screening setting, and you "can't make a claim for validated early detection until you examine the classifier in samples that were taken prior to diagnosis of ovarian cancer," Skates said, adding the study is a "great first step" in the validation of the test, and that it is "showing promise."

Skates also noted that there are other ovarian cancer tests for pelvic masses, such as the risk of malignancy index, a tool that combines imaging results and CA-125 measurements, and the risk of malignancy algorithm, which combines CA-125 and HE4. There is also the risk of ovarian cancer algorithm (ROCA) for earlier detection, which looks for patients with significantly higher CA-125 values than the patient's own baseline and which he developed. However, Skates noted that the ROCA results, while showing a significant increase in early stage detection, didn't show a significant mortality difference, which is why it's not available commercially.

Aspira Women's Health has also developed a machine learning-based algorithm, the MIA3G multivariate index assay, for ovarian cancer risk assessment using seven protein biomarkers, age, and menopausal status.

Skates said he would have liked to see the Johns Hopkins researchers provide a quantitative assessment of the added value of cfDNA-based fragmentation profiles by comparing the full classifier to a classifier based only on the standard biomarkers for ovarian cancer.

Velculescu noted that the researchers aim to further validate DELFI-Pro in additional populations and will initially be testing it in other screening populations where the classifier would be particularly useful and for individuals with different racial and ethnic backgrounds. They are "looking to confirm that this technology works on a larger scale."