Pancreatic cancer is usually pancreatic ductal adenocarcinoma — a cancer of the cells lining the ducts that carry digestive enzymes out of the pancreas. The pancreas sits deep in the abdomen, behind the stomach, wrapped around major blood vessels. Nothing presses on it early. Nothing shows on the outside. A tumour can grow there for a long time without producing a symptom anyone would take to a doctor.
That anatomy, more than anything about the cancer's biology, is why pancreatic cancer has one of the worst survival records of any common cancer. Overall five-year survival is around 9 percent [1], reported variously as 10 percent [2] and under 5 percent in older series [3] [4]. It is projected to become one of the leading causes of cancer death in high-income countries, not because it is becoming more aggressive but because other cancers are being caught and treated better [5] [6].
But that 9 percent is an average over stages, and it conceals the single most important fact in this review. A SEER analysis of stage IA pancreatic cancer found five-year overall survival of 83.7 percent for patients diagnosed at that stage in 2012, up from 44.7 percent in 2004 — while the age-adjusted incidence of stage IA diagnoses rose by 14.5 percent per year across 2004 to 2016 [7]. Caught early enough, this is a survivable cancer. The problem is that no more than 20 percent of patients are candidates for surgery at the time of diagnosis [4].
This review is deliberately non-sensational. Pancreatic cancer is genuinely serious and the honest numbers are grim. They are also more structured than the headline suggests, and the structure points at where the gains are.
Start here: why late diagnosis is not inevitable
There is a widespread assumption that pancreatic cancer spreads so early that finding it sooner would not help. The genomics say otherwise. Sequencing of matched primary and metastatic tumours found that distant metastasis occurs late during the genetic evolution of pancreatic cancer [8] — there is a window, measured in years, between a tumour becoming malignant and it seeding elsewhere. The disease is not diagnosed late because it metastasises immediately. It is diagnosed late because nobody is looking, and because when they do look the symptoms are indistinguishable from ordinary complaints.
That reframes the whole problem from a biological one to a detection one, and it is why the figure below is about stage rather than about drugs.
What the tumour is made of
The mutational landscape is unusually stereotyped. Activating KRAS mutations and CDKN2A/p16 inactivation are near-universal events in pancreatic ductal adenocarcinoma [9]; integrated genomic profiling of 150 specimens found recurrent somatic mutations in KRAS, TP53, CDKN2A, SMAD4, RNF43, ARID1A, TGFβR2, GNAS and others, with the rare KRAS wild-type tumours instead carrying alterations in GNAS, BRAF, CTNNB1 or other RAS-pathway genes [10] [11]. Mouse work showed Kras G12D initiates premalignant ductal lesions and cooperates with Trp53 or Ink4a/Arf loss to produce metastatic disease [12] [13], and oncogenic KRAS also rewires the tumour's metabolism [14] [15]. PDAC has been called the most RAS-addicted of all cancers [16].
Transcriptional subtyping has identified distinct molecular subtypes, including tumour- and stroma-specific ones separable by virtual microdissection [17] [18].
The stroma is the other half of the tumour. PDAC is characteristically dense, fibrous and poorly vascularised, with distinct populations of inflammatory fibroblasts and myofibroblasts [19]. That stroma is a physical barrier to drug delivery — enzymatically ablating it improved treatment in models [20] — and an immunological one, which is why PDAC has resisted the checkpoint blockade that transformed melanoma and lung cancer. Targeting CXCL12 from FAP-expressing fibroblasts [21], CSF1/CSF1R blockade to reprogramme macrophages [22], and focal adhesion kinase inhibition [23] all rendered pancreatic tumours responsive to checkpoint immunotherapy in models. In patients, vaccine and checkpoint combinations have so far produced modest results [24] [25] [26].
Pillar 1: measurement and diagnosis
The symptoms are the problem
The early symptoms are weight loss, vague upper abdominal or back pain, and loss of appetite — each of which has a hundred benign explanations. Painless jaundice, from a tumour in the head of the pancreas obstructing the bile duct, is the one classical presentation that reliably brings people in quickly, and it is a matter of tumour location rather than tumour stage.
Two other presentations deserve specific attention because they are the closest thing to an early warning.
New-onset diabetes. A meta-analysis of 36 studies covering 9,220 people with pancreatic cancer found type 2 diabetes associated with pancreatic cancer at an odds ratio of 1.82 (95% CI 1.66–1.89) — and, crucially, the risk was higher in people whose diabetes had only recently been diagnosed (OR 2.1) than in those who had had it five years or more (OR 1.5, P = 0.005) [27]. That gradient runs the wrong way for a purely causal story. If diabetes caused pancreatic cancer, longer exposure should mean more risk. The fact that recent diagnosis carries the higher risk is the signature of reverse causation: in some people the diabetes is an early consequence of the tumour, not a cause of it. New diabetes in an older adult with weight loss and no obvious reason is therefore a signal worth taking seriously — while remembering that the overwhelming majority of new diabetes is just diabetes.
Unexplained pancreatitis. Chronic pancreatitis is an established risk factor and has its own genetic contributors [28] [29] [30] [31], and an episode of pancreatitis with no gallstones, no alcohol and no other cause in an older adult can be the tumour announcing itself.
The established risk factors — smoking, obesity, diabetes, alcohol, age, family history — are set out in the epidemiological reviews [1] [6].
Imaging and tissue
The workup is imaging-led — dedicated pancreas-protocol CT to define the tumour and its relationship to the surrounding arteries and veins, MRI as a complement, and endoscopic ultrasound, which can both visualise small lesions and take a needle biopsy in the same sitting. This substrate, assembled by citation crawl during an OpenAlex outage, returned the acute- and chronic-pancreatitis imaging guidelines but no dedicated diagnostic-accuracy study of pancreas-protocol CT or endoscopic ultrasound in cancer, so no performance figures are quoted here. Cystic lesions found incidentally — intraductal papillary mucinous neoplasms in particular — have their own international management guidelines, because some are precursors and most are not [32]; molecularly, GNAS mutations are present in 66 percent of IPMNs and either KRAS or GNAS in 96 percent [33].
CA 19-9: what it is and is not
CA 19-9 is a carbohydrate antigen measurable in blood. It carries genuine prognostic information — in the NAPOLI-1 analysis, lower CA 19-9 was among the characteristics associated with long-term survival [34] — and CA 19-9 responses are tracked as a secondary readout in trials [25] [26].
What it is not is a screening test, and there is a specific piece of evidence here worth quoting: in a study developing a two-gene methylation biomarker panel that reached 97.3 percent sensitivity and 91.6 percent specificity, adding pre-operative CA 19-9 values to the panel did not improve sensitivity [35]. A marker that adds nothing to a diagnostic panel in people who already have the disease is not going to find it in people who do not.
An honest note: this substrate was assembled by citation crawl during an OpenAlex outage and did not return a dedicated study of CA 19-9's screening performance. The claim above is what the recovered sources support; the broader case against CA 19-9 as a population screening test rests on literature not available here.
Staging is the decision
Everything that follows depends on one split, made on imaging: resectable, borderline resectable, locally advanced (unresectable but not spread), and metastatic. The distinction is anatomical — how far the tumour has grown into the major vessels behind the pancreas — and it determines whether the patient is on a potentially curative path or a life-extending one. Those categories are not bookkeeping: trials are designed around them, and the borderline-resectable group in particular is defined by being the one neoadjuvant therapy might move [36] [37] [38] [39].
Genetic testing
Germline testing matters for two reasons. It identifies the small group with BRCA1, BRCA2, PALB2 or CDKN2A mutations who have a targeted option [40] [41] [42], and it identifies families who should be under surveillance — including Lynch syndrome families, whose pancreatic cancer risk has been quantified separately [43]. Somatic testing looks for the same, plus mismatch-repair deficiency, which is rare in PDAC but opens the door to immunotherapy when present.
High-risk family surveillance rests on a real risk gradient. In a prospective registry study of 5,179 individuals from 838 kindreds, members of familial pancreatic cancer kindreds — defined by at least one pair of affected first-degree relatives — had an observed-to-expected pancreatic cancer rate of 9.0 (95% CI 4.5–16.1), while members of kindreds without such a pair had a rate of 1.8 (95% CI 0.22–6.4), which was not significant [44]. That nine-fold excess is what justifies screening programmes in these families, and prospective controlled screening for early pancreatic neoplasia in high-risk individuals has been done [45].
Centerpiece: a simple simulatable model of the two levers
Two randomised trials, eight years apart, tested the same regimen family at two different stages. Together they form a two-by-two table that prices both of the levers available in this disease.
PRODIGE 4/ACCORD 11 randomised 342 patients with metastatic disease to FOLFIRINOX or gemcitabine: median overall survival 11.1 versus 6.8 months (hazard ratio for death 0.57, 95% CI 0.45–0.73, P < 0.001), with an objective response rate of 31.6 percent [46].
PRODIGE 24 randomised 493 patients with resected disease to modified FOLFIRINOX or gemcitabine as adjuvant therapy: median overall survival 54.4 versus 35.0 months, median disease-free survival 21.6 versus 12.8 months (stratified hazard ratio 0.58, 95% CI 0.46–0.73, P < 0.001), three-year disease-free survival 39.7 versus 21.4 percent [47].
Stage moves survival about five-fold. The drug moves it about 1.6-fold. The stage ratio is 35.0/6.8 = 5.15 on gemcitabine and 54.4/11.1 = 4.90 on FOLFIRINOX; the drug ratio is 11.1/6.8 = 1.63 in metastatic disease and 54.4/35.0 = 1.55 in resected disease. Whichever way you slice the table, stage is worth roughly three times what the drug is worth.
The check the arithmetic was never given. If stage and treatment acted independently and multiplicatively, that table would be rank one — and any three cells would predict the fourth. Taking the gemcitabine resected median and applying the drug effect measured in an entirely different trial in metastatic patients: 35.0 × 1.63 = 57.1 months, against an actual 54.4 — an overshoot of 5.0 percent. Going the other way, 6.8 × 1.55 = 10.57 against an actual 11.1, an undershoot of 4.8 percent. Two trials run by overlapping investigators but analysed separately, with different patients, different stages and different primary endpoints, produce a table that is multiplicative to about five percent. Nobody computed this in either paper.
A second, independent agreement. The two trials' hazard ratios were derived from full survival curves rather than from medians, by separate analyses: 0.57 and 0.58. The same regimen delivers essentially the same proportional benefit at two very different stages. Note honestly that the hazard ratios and the median ratios do not coincide — 1/0.57 is about 1.75 while the median ratio is 1.63 — because those two quantities only agree if survival is exponential, and it is not.
The right-hand panel is illustrative and labelled as such on the figure, because it does something the data do not license: it weights two median survivals as if they were the means of a two-group mixture, and assumes every patient receives the regimen named. Neither is true. It is there to compare the size of two levers, not to predict population survival. On those terms: moving from gemcitabine to FOLFIRINOX for everybody, at today's stage distribution, is worth about 7.3 months. Raising the resectable share from 20 percent to 40 percent, with no new drug at all, is worth about 8.7 months. Earlier detection outperforms the single largest chemotherapy advance in this disease's history.
The teaching point. Survival in pancreatic cancer is dominated by stage at diagnosis, metastasis happens late in the tumour's genetic evolution so there is a window to find it in [8], and stage IA disease already has five-year survival near 84 percent [7]. The biggest available lever is not a better drug. It is finding the disease earlier — and because population screening for a cancer this uncommon would produce far more false alarms than cancers, the realistic version of that lever is surveillance in defined higher-risk groups: familial kindreds [44] [45], germline mutation carriers [41], and possibly people with new-onset diabetes [27].
Three honest limits. Both trials enrolled patients fit enough for multi-agent chemotherapy — ECOG 0 or 1 in the metastatic trial — so these medians are not what an unselected or frail population achieves. The resected cohort is defined by having survived surgery to receive adjuvant therapy, which selects favourably twice over, so part of the five-fold stage ratio is selection rather than biology. And the multiplicative property is an empirical observation on four numbers, not a law; a third stage or a third drug could break it.
Pillar 2: treatment
Surgery is the only potentially curative option
Pancreaticoduodenectomy — the Whipple procedure — removes the head of the pancreas, the duodenum, the gallbladder and bile duct, and sometimes part of the stomach, then reconstructs the whole assembly. It is one of the largest operations in general surgery. Perioperative care has been formalised in Enhanced Recovery After Surgery guidelines [48] [49], and nutritional status before the operation predicts outcome after it [50].
The hard constraint remains: no more than 20 percent are candidates at diagnosis [4], and surgery alone is usually not enough. In CONKO-001, which randomised patients after gross complete resection, recurrent disease developed in 92 percent of the observation group and 74 percent of those given adjuvant gemcitabine [51]. Resection is necessary for cure and nowhere near sufficient for it.
Chemotherapy
Gemcitabine became the standard in 1997 on the basis of improved survival and clinical benefit over fluorouracil [52], and it defined a era in which the effect sizes were very small.
FOLFIRINOX replaced it for fit patients with metastatic disease in 2011 [46]. Gemcitabine plus nab-paclitaxel is the main alternative: in 861 patients with metastatic disease, median overall survival was 8.5 versus 6.7 months (HR 0.72), with one-year survival 35 versus 22 percent and two-year survival 9 versus 4 percent [53]. Those two-year figures are worth stating plainly — a doubling, from 4 percent to 9 percent. Liposomal irinotecan with fluorouracil is a second-line option [34].
In the adjuvant setting, gemcitabine after complete resection improved disease-free survival against observation in 368 patients [51], and modified FOLFIRINOX then beat gemcitabine [47]. Earlier work established chemotherapy rather than chemoradiotherapy as the adjuvant standard in Europe [54], against a US trial of adjuvant radiotherapy with fluorouracil [55].
Neoadjuvant therapy — chemotherapy before surgery — is the most active area. A systematic review and meta-analysis assessed preoperative therapy [36]; total neoadjuvant FOLFIRINOX followed by individualised chemoradiotherapy has been reported [37]; and the PREOPANC approach of neoadjuvant chemotherapy with preoperative chemoradiotherapy has been trialled in borderline resectable and resectable disease [38]. The rationale is partly biological selection: patients whose disease progresses during two months of chemotherapy were never going to be cured by an operation, and giving the chemotherapy first spares them a Whipple they cannot benefit from.
Precision therapy, honestly scoped
PARP inhibitors for germline BRCA. The POLO trial randomised patients with a germline BRCA mutation and metastatic pancreatic cancer whose disease had not progressed on first-line platinum chemotherapy to maintenance olaparib or placebo. Median progression-free survival was 7.4 versus 3.8 months (HR 0.53, 95% CI 0.35–0.82, P = 0.004) — but an interim overall survival analysis showed no difference (18.9 versus 18.1 months, HR 0.91, P = 0.68), and no difference in quality of life [56].
Two things about that trial should be said out loud. First, progression-free survival improved and overall survival did not, which is a real limitation and not a technicality. Second: 3,315 patients were screened to randomise 154 — 4.6 percent. That is the honest scale of this precision option in pancreatic cancer today.
Mismatch-repair deficiency and microsatellite instability are uncommon in pancreatic cancer but real, and have been characterised in terms of histology, molecular pathology and clinical implications [57] [58] [59] [60]; where present they identify patients who may respond to checkpoint blockade. For everyone else immunotherapy has so far not worked in this disease — durvalumab with or without tremelimumab was well tolerated and ineffective [61], and vaccine and checkpoint combinations have produced modest results at best [24] [25] [26] — for the stromal and immunological reasons above.
KRAS-targeted agents are the change most likely to matter, because KRAS is mutated in nearly every one of these tumours [9] [16]. The substrate here predates the current generation of G12D and pan-RAS inhibitors and this review will not characterise their clinical results, which are not in it.
Supportive care is not optional
Biliary stenting relieves obstructive jaundice. Pancreatic enzyme replacement treats the exocrine insufficiency that follows tumour or surgery and is a major determinant of whether a patient can maintain weight; nutritional guidelines for pancreatic disease exist [62], and nutritional status predicts surgical outcome [50]. Pain management and early palliative-care involvement belong in the plan from diagnosis rather than at the end of it: patients with unresectable disease benefit from palliative interventions including biliary decompression and coeliac plexus block [4].
Pillar 3: what is unresolved
Early detection is the field's central problem and the one with the largest payoff. Approaches under investigation include circulating tumour DNA combined with protein markers — one such liquid biopsy detected nearly two-thirds of pancreatic cancers that had no evidence of distant metastasis at resection [63] — glypican-1-positive exosomes [64], DNA methylation panels [35], ctDNA in localised disease, detectable in 43 percent of cases at diagnosis and predictive of relapse after resection [65], and artificial intelligence applied to imaging and records [2]. Whole-genome sequencing has been returned inside the clinical decision window in advanced disease [66], and personalised RNA neoantigen vaccines with atezolizumab have been shown to stimulate neoantigen-specific T cells in resected pancreatic cancer, with 18-month recurrence-free survival reported in a 16-patient study [67]. The fact that stage IA incidence is rising and its survival improving suggests something is already working [7].
Who to screen. Population screening is not viable for a cancer this uncommon. Defining the enriched groups — familial kindreds [44], germline carriers [41], new-onset diabetes with weight loss [27], suspicious cysts [32] — is where the practical progress is.
Drugging KRAS, and what happens when you do. The pathway is near-universal [9] and metabolically load-bearing [14] [15]; resistance is the obvious next question.
The stroma. Every model-based attempt to break the fibrotic and immunosuppressive barrier has looked promising [20] [21] [22] [23] and none has yet translated into a practice-changing result. Newer single-cell and spatial work is redrawing the map of what is actually in there [68] [69] [19], and organoid models allow the question to be asked per patient [70].
Dig deeper in lmmol
The contrast with the other oncology reviews is the point of reading them together. Colorectal cancer is the disease pancreatic cancer would like to be: an accessible organ, a detectable precursor lesion, and organised screening that shifts stage at diagnosis — which is precisely the lever the figure above prices. Lung cancer shows the same stage arithmetic with a screening programme that works in a defined high-risk group rather than in the whole population, which is the realistic template here. Melanoma is the sharpest contrast on treatment: checkpoint blockade transformed it and has so far failed in pancreatic cancer, and the reason is the stroma and the immunologically cold microenvironment [23] [21]. Breast cancer and prostate cancer share the germline BRCA and PARP-inhibitor thread [56] [40]. And type 2 diabetes and glycemic control is the closest non-cancer link, because new-onset diabetes can be the tumour's first sign rather than an independent illness [27]. The full collection is at health.