AI-Selected Predictive Biomarker Guides First-Line Treatment Selection in Advanced Pancreatic Cancer

A computational histology–based artificial intelligence (AI) platform was able to identify a biomarker that could predict treatment benefit between two chemotherapy options for patients with advanced pancreatic cancer, according to the results of a study presented in a poster at the 2026 ASCO Gastrointestinal Cancers Symposium (Abstract 764). The performance of the predictive biomarker was validated in two prospective clinical studies for determining first-line chemotherapy treatment preferences.

“[This] tool could be used to optimize first-line treatment selection in pancreatic cancer,” the study authors, led by Andrew E. Hendifar, MD, Associate Professor of Medicine, Cedars-Sinai Medical Center, and Medical Director of Pancreatic Cancer, Samuel Oschin Comprehensive Cancer Center, Los Angeles, concluded in their poster.

A report on the performance of the predictive biomarker has also been accepted for future publication in the Journal of Clinical Oncology.

Study Methods

The researchers used the Computational Histology Artificial Intelligence (CHAI) biomarker platform—which pulls out quantitative histomorphologic features from whole-slide images—to identify predictive biomarkers from pretreatment biopsy specimens of patients with locally advanced or metastatic pancreatic ductal adenocarcinoma who were to receive either fluoropyrimidine- or gemcitabine-based chemotherapy regimens in the first-line setting. 

First, the platform was assessed in a retrospective development cohort of 178 patients from Cedars-Sinai and the University of Pittsburgh Medical Center. All patients had a digitized whole-slide image of a biopsy sample from before first-line treatment and had a known time to next treatment or death.

The platform was initially pretrained on hematoxylin-and-eosin whole slides of solid tumors, and the base model was subsequently focused on pancreatic cancer using tissue- and cell-level segmentation and classification. Features were then extracted from the slides of the development cohort, quantified, and weighted to develop a relevant biomarker. The platform created a continuous histologic signature that was associated with outcomes from the two treatment options in terms of time to next treatment or death. Then, the patients were separated into treatment preference groups according to their expected outcome.

In the validation cohort, the locked biomarker was tested on 299 patients from two prospective clinical studies: COMPASS and the Pancreatic Cancer Action Network’s Know Your Tumor initiative.

Findings

Patients in the validation cohort who were treated with gemcitabine- vs fluoropyrimidine-based treatment were found to have similar time to next treatment or death (P = .9) and overall survival (P = .13) outcomes.

A total of 173 patients were classified based on biomarker status as preferring treatment with a fluoropyrimidine (58%) or gemcitabine (42%). In those with a preference for a gemcitabine backbone, the time to next treatment or death appeared to be significantly improved when they were treated with gemcitabine vs a fluoropyrimidine (P = .038), whereas overall survival showed no significant difference (P = .52). The median time to next treatment or death was 9.6 months in patients treated with gemcitabine vs 7.2 months in those who received a fluoropyrimidine (hazard ratio [HR] = 0.6; 95% confidence interval [CI] = 0.4–1.0). The median overall survival was 14.3 months with gemcitabine and 12.4 months with a fluoropyrimidine (HR = 0.9; 95% CI = 0.5–1.4).  

Patients with a preference for a fluoropyrimidine were found to have both a longer time to next treatment or death (P = .035) and improved overall survival (P = .003) when treated with a fluoropyrimidine vs gemcitabine. The median time to next treatment or death was 8.6 months with a fluoropyrimidine vs 7.5 months with gemcitabine (HR = 1.5; 95% CI = 1.0–2.1). The median durations of overall survival were 14.4 and 11.7 months with a fluoropyrimidine and gemcitabine, respectively (HR = 1.7; 95% CI = 1.2–2.5).

By propensity score–weighted analysis, the differential treatment effect of the predictive biomarker showed significant interaction value for both time to next treatment or death (P < .001) and overall survival (P = .005).

DISCLOSURE: Dr. Hendifar has served as a consultant or advisor for Alcresta Therapeutics, Amgen, Exelixis, Faraday Pharmaceuticals, Ipsen, Novartis, Pancreatic Cancer Action Network, Pfizer, Regeneron, and Valar Labs; has received research funding from Ipsen and NGM Biopharmaceuticals; has received reimbursement for travel expenses from Halozyme; and has other financial relationships with RayzeBio.

Expert Point of View

William A. Hall, MD, Professor and Chair, Radiation Oncology, and Bob Uecker Endowed Chair, Department of Surgery, School of Graduate Studies, Medical College of Wisconsin, Milwaukee, spoke during an artificial intelligence (AI)–focused session of the 2026 ASCO Gastrointestinal Cancers Symposium and highlighted a poster (Abstract 764) showing that a computational histology–based AI platform could identify a biomarker predicting treatment benefit between two chemotherapy options for patients with advanced pancreatic cancer. He suggested that such AI-generated biomarkers will increasingly guide treatment decisions going forward.

“These are the types of biomarkers that we as oncologists have to be grabbing onto and looking for, because our management of patients cannot, and should not, be [based on] personal opinions such as, ‘I don't think neoadjuvant radiation [therapy] has a role in pancreatic cancer’ or ‘I don't think FOLFIRINOX [leucovorin, fluorouracil, irinotecan, and oxaliplatin] is the best type of chemotherapy.’ In reality, much of the randomized data show complete equipoise between many different strategies of treatment. We must be doing better for our patients. The future is going to hold readily available and novel biomarkers that will bring us closer to more robust objective answers on which of these treatments are best for our patients.”

DISCLOSURE: Dr. Hall reported owning stock or other financial interests in Sonoptima; has served as a consultant or advisor for Aktis Oncology and Sonoptima; has received research funding and expense support from Elekta; and has a patent pending for a wearable device for radiation therapy treatment planning.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.