Pancreatic Cancer Kills 78% of Everyone It Touches. Two Labs Just Attacked It From Opposite Ends.

Seventy-eight percent. That is the mortality-to-incidence ratio for pancreatic cancer in the United States. Of the 67,530 Americans who will be diagnosed this year, 52,740 will die of it. No other major cancer comes close. Breast cancer kills 8.3% of those it touches; prostate, 2.1%. Pancreatic cancer is the third leading cause of cancer death in the country and, on current trajectories, will become the second by 2030, overtaking colorectal cancer despite being far less common.

Why? Because more than 85% of patients receive their diagnosis after the disease has already metastasized, spread so far beyond the pancreas that surgery is no longer an option, chemotherapy becomes palliative rather than curative, and five-year survival drops to 3.2%. Average life expectancy at that point: one year. But catch it before it spreads and five-year survival jumps to 44%. Catch it early enough for resection, follow that with six months of adjuvant chemotherapy, and half of patients are alive five years later. Not uniquely lethal. Uniquely invisible. That distinction matters.

In 2026, two independent research groups published results that attack that invisibility from opposite directions. One built an AI that sees what radiologists cannot. Its counterpart built a vaccine that destroys what neither can yet see. Both worked. And the math of what happens when you combine them is worth running.

The AI That Reads the Invisible

In April, a team led by Dr. Ajit Goenka at Mayo Clinic published a validation study in Gut of an AI model called REDMOD, the Radiomics-based Early Detection Model. REDMOD analyzes routine abdominal CT scans, the kind ordered for unrelated complaints, and looks for radiomic patterns: quantitative features of tissue texture and structure too faint for the human eye to resolve.

Goenka's team fed the model nearly 2,000 CT scans, including images from patients who were later diagnosed with pancreatic cancer but whose scans had been read as perfectly normal at the time. REDMOD flagged 73% of those invisible cancers at a median of 16 months before clinical diagnosis, nearly double what specialists achieved reviewing the same scans without AI assistance.

At earlier time points, the advantage widened. Dramatically. On scans obtained more than two years before diagnosis, REDMOD caught nearly three times as many early cancers as unaided specialists, and in patients with multiple scans over time, the model produced 90% to 92% concordance across serial images, confirming it was tracking a real biological signal rather than randomly firing on incidental noise that happened to correlate with eventual disease.

Specificity, the ability to correctly identify patients who did not develop cancer, landed at 81% across a multi-institutional validation cohort and 87.5% on an independent NIH dataset. Not perfect. A 19% false-positive rate on routine CTs would generate substantial follow-up imaging and anxiety. But REDMOD is not designed for mass screening. It targets high-risk patients, particularly those with new-onset diabetes, a known early marker of pancreatic pathology, who are already getting abdominal CTs for other reasons and whose prior probability of disease is high enough to make an 81% specificity clinically actionable rather than noise-generating.

Mayo Clinic has already advanced the work into a prospective study called AI-PACED, Artificial Intelligence for Pancreatic Cancer Early Detection, which will evaluate how clinicians integrate REDMOD into actual care workflows.

The Vaccine That Prevents the Cancer

In July, a team at Johns Hopkins led by oncologists Elizabeth Jaffee and Neeha Zaidi published Phase 1 results in Cancer Discovery for a peptide-based vaccine designed not to treat pancreatic cancer but to intercept it before it forms.

Pancreatic ductal adenocarcinoma does not appear overnight. Tumors develop slowly, over decades, from benign cysts and lesions. Roughly 10% of people who eventually develop PDAC carry a genetic predisposition that can be identified in advance. Jaffee's team recruited 20 such individuals, all of whom had confirmed pancreatic lesions that had not yet progressed to cancer, and vaccinated them against the six most common mutations in the KRAS gene, a cell-growth regulator mutated in approximately 90% of all PDAC tumors.

Immune response was immediate and strong: a median 18.2-fold increase in T cell responses against the mutant KRAS antigens. Half of the patients mounted strong responses against all six peptides. Side effects were mild: fatigue, chills, flu-like symptoms, all self-resolving.

Then came the follow-up data. Over a median observation period of 16.5 months, zero vaccinated patients progressed to cancer. Zero. Not one. Three patients saw their cysts completely resolve. Three more saw their lesions regress. Everyone else held stable. When researchers compared these outcomes to an unvaccinated cohort monitored over a similar period, they observed more progression in the unvaccinated group.

In patients who returned for optional follow-ups one to two years after their final vaccination, Jaffee and Zaidi confirmed that the vaccine had stimulated both CD4+ and CD8+ T cells, including effector and central memory phenotypes, the cellular machinery of durable immune protection. Memory populations declined over time, as they do with most vaccines, but showed robust expansion potential when re-exposed to mKRAS antigen in laboratory testing.

"This study represents the first proof of concept for the use of vaccines for interception of pancreatic cancer in human patients," Zaidi said. Twenty patients is far too small a sample to draw clinical conclusions. But zero progressions in 16.5 months, with three complete resolutions, is not ambiguous.

The Math Nobody Has Run

Neither team designed their work with the other's in mind. But combine them on paper and the implications demand calculation.

Consider the population of Americans with new-onset diabetes over age 50, a group at elevated pancreatic cancer risk whose standard diagnostic workup already includes the abdominal CT scans that REDMOD is designed to analyze, scans that are currently being read, filed, and forgotten. If REDMOD were deployed on those scans, and its published detection rate held, it would flag suspicious radiomic patterns in patients whose scans currently read as clean.

Here is a rough model. Approximately 10% of the 67,530 annual PDAC diagnoses occur in patients whose cancer was detectable in retrospect on prior abdominal imaging, roughly 6,750 patients per year. Apply REDMOD's 73% detection rate and you catch approximately 4,928 of them earlier. If half of those patients shift from a late-stage diagnosis (Stage III or IV) to an early-stage one (Stage I or II) because the cancer was caught before visible tumor formation, that gives you approximately 2,464 patients per year who gain access to surgical resection.

What Stage-Shifting Buys: 5-Year Survival by Stage at Diagnosis
Stage at Diagnosis	5-Year Survival	Typical Treatment	Surgical Eligibility
Stage I	42.9%	Surgery + adjuvant chemo	Yes
Stage II	16.8%	Surgery + adjuvant chemo (sometimes neo-adjuvant)	Usually
Stage III	7.4%	Chemo/radiation (rarely surgery)	Rarely ("borderline")
Stage IV	1.5–3.2%	Palliative chemo	No

Moving 2,464 patients from a 3.2% survival tier to a 44% survival tier produces a net survival gain of approximately 40.8 percentage points. Multiply that by the population and you get roughly 1,005 additional patients alive at five years who would otherwise be dead, and that estimate covers only one high-risk subgroup. The total addressable population for AI-assisted early detection on routine abdominal imaging is considerably larger.

Now add the vaccine layer. KRAS mutations are present in about 90% of PDAC tumors, and the six variants targeted by the Hopkins vaccine cover the most common mutations in that set. Conservatively, assume these six variants account for 80% of KRAS-mutated PDAC. That means roughly 72% of all pancreatic cancers arise from mutations the vaccine is designed to intercept: about 48,600 of the 67,530 annual US cases.

If REDMOD identifies patients with radiomic anomalies before visible tumors form, and the vaccine prevents those anomalies from progressing, you have a detection-interception pipeline: find the precursor, vaccinate against the mutation driving it, and prevent the cancer from ever arriving. That pipeline does not exist yet. Neither technology is approved. One is in prospective clinical validation and the other has completed Phase 1 with 20 patients. But the logic is concrete, the mechanisms are independently validated, and the data points in the same direction.

The Funding Paradox

One number makes the urgency plain. The National Institutes of Health allocated $440 million to pancreatic cancer research in fiscal year 2025, which works out to $8,945 per death. Compare that. Breast cancer received $1.58 billion, or $69,800 per death. Prostate cancer received $663 million, or $126,992 per death. Pancreatic cancer gets 7.8 times less funding per death than breast cancer and 14.2 times less than prostate cancer, despite having the highest mortality-to-incidence ratio of any major cancer.

NIH Research Funding Per Cancer Death (FY 2025)
Cancer Type	5-Year Survival	Annual Deaths	NIH Funding	Funding Per Death
Prostate	97.9%	5,219	$663M	$126,992
Breast (female)	91.7%	22,606	$1,578M	$69,800
Colorectal	65.4%	49,576	$495M	$9,979
Pancreatic	13.3%	49,211	$440M	$8,945
Liver	22.0%	27,816	$291M	$10,447
Small-cell lung	9.1%	22,240	$63M	$2,818

This disparity is not scandalous on its face. Breast cancer's high funding reflects decades of advocacy, a large patient population, and genuine scientific opportunity. Prostate cancer's extreme per-death funding reflects a tiny denominator created by that cancer's extraordinary survival rates. But the pancreatic cancer column tells a story about prioritization: a disease that kills almost as many Americans as colorectal cancer, with a mortality rate six times worse, and receives slightly less funding.

Both the REDMOD and mKRAS vaccine studies were supported in part by NIH grants. The Mayo team's work was also funded by the Hoveida Family Foundation and the Funk-Zitiello Foundation's Champions for Hope program. The Hopkins vaccine research led to a licensing deal with Adventris Pharmaceuticals, co-founded by Jaffee and Zaidi, the lead researchers. Neither breakthrough required a massive budget. They required focused questions, clever use of existing infrastructure (routine CT scans in one case, established peptide vaccine platforms in the other), and years of patient work on a disease that most of the scientific funding establishment treats as a lower priority.

The Strongest Counter

Here is the strongest case against reading these two papers as a pipeline is that they are separated by enormous practical gaps. REDMOD has been validated retrospectively; its prospective performance in clinical practice, where radiologists must act on AI recommendations and patients must consent to follow-up procedures based on invisible risk signals, is entirely unknown, and the history of AI in radiology is littered with models that performed brilliantly in validation studies and then degraded catastrophically in deployment because of dataset shift, clinician distrust, or the mundane friction of integrating a new tool into overloaded clinical workflows where radiologists are already reading 50 to 100 cases a day.

Vaccine development is further away still. Twenty patients is a safety and immunogenicity study, not an efficacy trial. The zero-progression result is encouraging but statistically fragile: with only 20 participants over 16.5 months, the confidence intervals around that zero are wide enough to accommodate meaningful rates of progression. Phase 2 and Phase 3 trials, each requiring hundreds or thousands of patients observed over years, stand between this result and clinical use. Jaffee and Zaidi's financial ties to Adventris Pharmaceuticals, which has licensed the vaccine technology, introduce an additional layer of scrutiny that independent replication will need to resolve.

And the combination, detect with REDMOD, then vaccinate with mKRAS peptides, has never been tested or even proposed by either team. It is an analytical extrapolation, not a clinical program. Whether radiomically flagged patients actually carry the specific mKRAS mutations targeted by the vaccine, whether vaccination at that point in disease progression would have the same effect as vaccination in patients with known genetic predisposition, whether the immune response would be sufficient to eliminate established radiomic anomalies rather than merely prevent their initial formation: none of this is known.

Limitations

Our stage-shift calculation assumes that earlier detection translates directly to earlier staging, which oversimplifies the biology. A patient flagged by REDMOD at 16 months before clinical diagnosis may still present with locally advanced disease that is technically Stage III at the time of clinical workup. The 73% detection rate was measured on a retrospective cohort; prospective sensitivity may differ. The 10% subgroup estimate (cancer patients whose disease was detectable on prior imaging) is derived from published estimates of incidental pancreatic findings and may not precisely correspond to REDMOD's target population. Treatment cost comparisons were not included because reliable US-specific per-stage cost data for PDAC are fragmented across literature, insurance types, and treatment regimens, and the available estimates range too widely to produce a responsible number.

The Bottom Line

Pancreatic cancer is not a hard cancer because the disease is especially cunning. It is a hard cancer because it is invisible until Stage IV, and at Stage IV, nothing works well. For the first time, two independent technologies address that specific problem from opposite ends: one renders the invisible visible, and the other destroys what the invisible would become. Neither is ready for clinical deployment. Both have cleared the hardest conceptual barrier, which is demonstrating that their approach works at all. The AI sees real signal on scans that specialists read as clean, and the vaccine provokes real immune memory against the mutations that drive the cancer.

If you are over 50 and your doctor orders an abdominal CT for any reason, ask whether your institution uses AI-assisted pancreatic screening. Most do not yet, but the question registers demand. If you have a first-degree relative who died of pancreatic cancer, ask about genetic testing for PDAC predisposition; the Hopkins vaccine trial specifically recruited patients with known genetic risk. If you are a physician, read the REDMOD paper in Gut and the mKRAS vaccine paper in Cancer Discovery; both are open-access and both are methodologically rigorous enough to be worth your time. If you are a researcher or funder, notice that the deadliest major cancer in America receives $8,945 in NIH funding per death. That number has consequences, and they are measured in years of life lost.