Cancer cells leave a trace. The AI learns to read it.
For decades, ovarian cancer has claimed lives not because it is untreatable, but because it is found too late — a silence the disease keeps until it has already spread. Researchers at Johns Hopkins have now developed a blood test called DELFI-Pro that teaches artificial intelligence to read the disordered genetic traces cancer leaves in the bloodstream, detecting early-stage ovarian cancer at twice the rate of existing protein markers. If larger clinical trials confirm what these initial results suggest, a routine blood draw could one day interrupt a pattern of late diagnosis that has haunted oncology for generations.
- Ovarian cancer kills roughly half of those it touches within five years, largely because it hides until it has already spread — a diagnostic silence that has resisted every prior attempt to break it.
- DELFI-Pro combines two established protein markers with an AI analysis of fragmented DNA in the blood, exploiting the fact that cancer cells die chaotically, leaving a genetic signature distinct from healthy tissue.
- In validation studies across European and American patient groups, the test detected 72% of stage I cancers compared to 34% for protein markers alone, and 100% of stage IV cases, with almost no false positives.
- Beyond detection, the test can distinguish benign ovarian growths from malignant ones — a capability that could spare thousands of women from exploratory surgery currently triggered by inconclusive ultrasound findings.
- Researchers are now planning larger randomized clinical trials, the necessary bridge between a promising laboratory result and a test that could appear on a standard annual checkup.
Ovarian cancer is the fifth leading cause of cancer death among American women, and roughly half of those diagnosed do not survive five years. The disease is so lethal in part because it is so quiet — by the time symptoms appear, the tumor has usually spread far beyond the ovaries. This timing problem has defined the disease for decades.
Researchers at Johns Hopkins Kimmel Cancer Center believe they have found a way to interrupt that silence. Their test, DELFI-Pro, draws on three sources of information: two protein markers, CA-125 and HE4, that have long been associated with ovarian cancer but are unreliable on their own, and a newer AI-powered analysis of cell-free DNA — the genetic fragments that drift through the bloodstream after cells die. Healthy cells leave behind orderly fragments. Cancer cells do not. Their DNA breaks apart in irregular, chaotic patterns, and the AI learns to recognize that signature.
Tested on blood samples from women in the Netherlands and Denmark, DELFI-Pro detected 72 percent of stage I ovarian cancers, compared to just 34 percent for protein markers alone. At stage IV, it found every case. Across all stages, it identified 87 percent of cancers while generating almost no false alarms. A separate validation on American women produced similar results.
Perhaps as important as what the test detects is what it can distinguish. When ultrasound finds an ovarian growth, the standard response is exploratory surgery — even when the growth turns out to be benign. DELFI-Pro showed it could tell the difference, potentially sparing women from unnecessary procedures with their attendant risks and recovery.
The study, published September 30 in Cancer Discovery, builds on earlier work by the same team showing that DNA fragmentomics could detect lung cancer. Senior author Victor Velculescu described the results as encouraging while stressing that larger randomized trials are needed before the test could enter routine clinical use. If those trials succeed, a disease long defined by late discovery may finally meet its match at the annual checkup.
Ovarian cancer kills more American women than any cancer except lung, breast, colorectal, and pancreatic cancer. Half of those diagnosed do not survive five years. The disease often announces itself too late—by the time a woman feels something wrong, the tumor has usually spread beyond the ovaries, where it is far harder to treat. This timing problem has haunted oncology for decades. Now researchers at Johns Hopkins Kimmel Cancer Center, working with collaborators across the United States and Europe, say they have found a way to catch it earlier: a blood test that uses artificial intelligence to read the fingerprints cancer leaves in the bloodstream.
The test, called DELFI-Pro, combines three pieces of information. Two of them—proteins called CA-125 and HE4—have been known markers of ovarian cancer for years, but neither one alone is reliable enough to screen for the disease. The third piece is newer: an AI-powered analysis of cell-free DNA, the fragments of genetic material that float in the blood after cells die. Healthy cells, when they break apart, leave behind DNA fragments in neat, orderly patterns. Cancer cells are chaotic. Their DNA fragments are irregular, disorganized, distinctive. The AI learns to recognize that signature.
The researchers tested this combined approach on blood samples from three groups of women: 94 with ovarian cancer, 203 with benign ovarian tumors, and 182 with no ovarian growths at all. The samples came from hospitals in the Netherlands and Denmark. When they ran DELFI-Pro on these samples, the results were striking. For stage one cancers—the earliest and most treatable form—the test caught 72 percent of cases. The old protein test alone caught 34 percent. For stage four cancers, DELFI-Pro found 100 percent. Across all stages, it detected 87 percent of ovarian cancers while producing almost no false alarms in women without cancer.
The researchers then tested the approach again on a separate group of American women: 40 with ovarian cancer, 50 with benign growths, and 22 with no ovarian lesions. The results held. The test detected 73 percent of all cancers and 81 percent of the most aggressive subtype, high-grade serous carcinoma. Again, almost no false positives. What made this especially valuable was something ultrasound cannot do: the test could tell the difference between a benign tumor and a cancerous one. This matters because when ultrasound finds an ovarian growth, the standard next step is exploratory surgery. A blood test that could spare women with benign growths from that surgery—from anesthesia, from recovery time, from the small but real risks of surgery itself—would change practice.
The study, published September 30 in Cancer Discovery, a journal of the American Association for Cancer Research, represents years of work building on earlier successes. The same research team had previously shown that this DNA fragmentomics approach could detect lung cancer. The principle is simple but powerful: cancer cells leave a trace. The AI learns to read it.
Victor Velculescu, the senior author and a professor of oncology at Johns Hopkins, called the results encouraging but cautious. The next step is larger randomized clinical trials to confirm that the test works in real-world screening—that it actually saves lives, that it actually reduces unnecessary surgery, that the benefits outweigh any harms. If those trials succeed, DELFI-Pro could become a routine blood test, something a woman could have drawn at her annual checkup. It would be affordable, accessible, and potentially transformative for a disease that has long been caught too late.
Citas Notables
Early detection of ovarian cancer may save lives but most women are diagnosed late in the course of the disease when survival rates are much lower.— Jamie Medina, postdoctoral fellow at Johns Hopkins Kimmel Cancer Center
This study provides further evidence demonstrating the benefit of genome-wide, cell-free DNA fragmentation and artificial intelligence to detect cancers with high accuracy.— Victor E. Velculescu, senior author and professor of oncology at Johns Hopkins
La Conversación del Hearth Otra perspectiva de la historia
Why does ovarian cancer stay hidden so long?
Because it doesn't announce itself early. There are no specific symptoms in stage one or two. By the time a woman notices something—pain, bloating, changes in her body—the cancer has usually spread. That's when survival rates drop sharply.
And the old protein tests—CA-125 and HE4—they don't work?
They work sometimes, but not reliably. CA-125 alone catches only 34 percent of stage one cancers. It's not sensitive enough to screen millions of women. You'd miss too many cases.
So what does the AI add?
It reads the DNA fragments in the blood. Cancer cells die and break apart differently than healthy cells do. The fragments are chaotic, irregular. The AI learns that pattern and recognizes it. When you combine that with the protein markers, you get something much stronger.
The test distinguished benign from cancerous tumors. Why is that so important?
Because right now, if ultrasound finds an ovarian growth, the next step is surgery. You go under anesthesia, they open you up, they look around. But many of those growths are benign—they'll never hurt you. A blood test that says "this one is benign, you don't need surgery" spares women from unnecessary risk and recovery.
What happens next?
Larger trials. They need to prove this works in real screening, in real populations, that it actually saves lives. If it does, it becomes routine—a blood test at your annual checkup.