After surgery to remove a pancreatic tumor, patients typically receive chemotherapy with gemcitabine to reduce recurrence risk. But not all patients benefit equally, and there is currently no reliable way to tell in advance who will respond well.
The lack of actionable biomarkers for precision therapy in pancreatic cancer means most patients receive the same treatment regardless of whether it is likely to help them, missing opportunities for personalized care.
Researchers hypothesized that patterns visible in tumor tissue slides — detectable by AI but not the human eye — might predict which patients benefit from gemcitabine after surgery.
The team digitized surgical tumor tissue slides from 93 patients in The Cancer Genome Atlas (TCGA) dataset, all of whom received adjuvant gemcitabine. They used 46 patients for training and 47 for testing the AI model.
The AI was trained to extract subtle visual features from slides and link them to disease-specific survival (DSS). The resulting signature — called VPG (Valar Pathology Gemcitabine) — assigned each patient a score based on patterns in their tumor tissue.
To check whether the VPG signature predicted gemcitabine response specifically (rather than just overall prognosis), the team also tested it on patients who received no adjuvant treatment — a treatment-predictive biomarker should only work in treated patients.
In the internal test set of 47 TCGA patients, VPG-positive patients had a median disease-specific survival of 67.9 months, compared to just 16 months for VPG-negative patients. The hazard ratio for death in VPG-negative patients was 2.94 — nearly three times the risk.
The signature was validated in an external cohort from the University of Pittsburgh Medical Center (UPMC), where VPG-positive patients again showed significantly better survival after gemcitabine treatment (median DSS 43.1 vs. 16 months, p=0.02).
When tested in patients who received no treatment after surgery, VPG status did not predict survival outcomes — confirming the signature identifies gemcitabine responders specifically, not just patients with generally better-prognosis tumors.
Pancreatic cancer has been previously classified into molecular subtypes based on gene expression including the Moffitt, Collisson, and Bailey systems. The VPG histologic signature did not correlate with any of these existing systems.
This suggests the AI has identified a new layer of biological information visible in the physical appearance of tumor tissue that is not captured by gene expression profiling.
This independence means VPG could add predictive power even for patients already genetically profiled, opening the door to multi-modal precision medicine.
If validated in prospective trials, the VPG signature could enable oncologists to identify patients likely to benefit from gemcitabine before treatment starts, sparing non-responders from unnecessary side effects while ensuring responders receive this effective therapy.
The approach demonstrates that AI applied to standard tumor slides — a type of sample already routinely collected during surgery — can generate clinically meaningful predictions without additional tests or costs.
The authors call for prospective trials to validate the VPG signature and explore whether VPG-negative patients might benefit more from alternative chemotherapy regimens such as FOLFIRINOX.