Developing a Histology-Based Artificial Intelligence Biomarker to Predict Adjuvant Chemotherapy Benefit in Pancreatic Cancer Authors: Audrey Beaufils, Julien De Martino, Xiaofeng Jiang, Théau Blanchard, Nicolas Fraunhoffer, Diana Mendes, Camille Pignolet, Taib Bourega, Asier Rabasco Meneghetti, Srividhya Sainath, Miguel Albuquerque, Nathalie Colnot, Matthieu Tihy, Anthony Turpin, Meher Ben Abdelghani, Alice Wei, Emmanuel Mitry, Thierry Lecomte, James Biagi, Pascal Artru, Ludovic Evesque, Aurélien Lambert, Daniel J Renouf, Marjorie Mauduit, Nelson J Dusetti, Pascal Hammel, Thierry Conroy, Jean-Baptiste Bachet, Louis de Mestier, Vinciane Rebours, Jérôme Cros, Jakob Nikolas Kather, Rémy Nicolle Published: 2026 ABSTRACT AbstractPURPOSEThe choice of adjuvant chemotherapy in pancreatic ductal adenocarcinoma (PDAC) is mainly guided by patients' general condition. We hypothesized that tumor morphology may predict differential treatment benefit and tested whether deep learning applied to histology images could derive a biomarker of relative benefit from gemcitabine (GEM) versus modified FOLFIRINOX (mFOLFIRINOX) in resected PDAC.PATIENTS AND METHODSStandard whole-slide images from a retrospective multicentric series of 231 patients who underwent curative-intent pancreatectomy and received adjuvant mFOLFIRINOX (n = 54) or GEM (n = 177) were used to train regimen-specific histology models on disease-free survival (DFS), which were then combined into PANCprAId, a biomarker estimating personalized relative benefit from adjuvant GEM versus mFOLFIRINOX. External validation was performed in the randomized PRODIGE-24/CCTG PA6 trial (n = 313).RESULTSIn PRODIGE-24/CCTG PA6, the treatment-specific histology scores used to construct PANCprAId stratified outcomes among patients treated with GEM (hazard ratio [HR], 1.69 [95% CI, 1.04 to 2.73]; P = .03) and mFOLFIRINOX (HR, 2.02 [95% CI, 1.4 to 3.0]; P < .001). When combined into PANCprAId, the biomarker identified subgroups with differential relative benefit from adjuvant GEM versus mFOLFIRINOX, with significant treatment interactions for DFS (interaction P = .003) and cancer-specific survival (interaction P = .001). Predicted sensitivity to each regimen was associated with distinct epithelial and stromal features.CONCLUSIONHistology-based deep learning can derive a predictive biomarker of relative benefit from adjuvant GEM versus mFOLFIRINOX in resected PDAC. INTRODUCTION Pancreatic ductal adenocarcinoma (PDAC) remains one of the deadliest cancers, with only 8% of 5-year survival rate estimated. Despite significant therapeutic advancements, it is projected to become the second leading cause of cancer-related deaths in Western countries by 2030.1,2 For more than two decades, gemcitabine (GEM) was the standard adjuvant chemotherapy treatment for patients with resected PDAC.3-6 This paradigm changed with the PRODIGE-24/CCTG PA66,7 trial, which demonstrated the interest of a multidrug adjuvant regimen with folinic acid, fluorouracil, irinotecan, and oxaliplatin, titled modified FOLFIRINOX (mFOLFIRINOX; FFX). As a result, the choice of adjuvant chemotherapy is almost exclusively based on the patient's performance status rather than on tumor biology. Approximately one third of patients are unfit for adjuvant mFOLFIRINOX and therefore can receive either a GEM-based or fluorouracil-based regimen.8 Predictive tools relying on the tumor biology to predict which regimen each patient would most benefit from are greatly needed. CONTEXT Key Objective Key Objective Can histology identify which patients are more likely to benefit from adjuvant gemcitabine (GEM) versus mFOLFIRINOX? Can histology identify which patients are more likely to benefit from adjuvant gemcitabine (GEM) versus mFOLFIRINOX? Knowledge Generated Knowledge Generated In the randomized PRODIGE-24 trial, histology models predicted sensitivity to mFOLFIRINOX and GEM. A combined PANCprAId biomarker enabled direct patient-specific comparisons of the expected benefits of adjuvant GEM versus mFOLFIRINOX, showing improved disease-free survival and cancer-specific survival. In the randomized PRODIGE-24 trial, histology models predicted sensitivity to mFOLFIRINOX and GEM. A combined PANCprAId biomarker enabled direct patient-specific comparisons of the expected benefits of adjuvant GEM versus mFOLFIRINOX, showing improved disease-free survival and cancer-specific survival. Relevance (R.B. Parikh) Relevance (R.B. Parikh) AI-based pathology biomarkers can predict benefit from adjuvant GEM- versus fluoropyrimidine-based chemotherapy in pancreatic cancer and should be incorporated as prespecified correlatives or stratification factors in future phase III studies.* AI-based pathology biomarkers can predict benefit from adjuvant GEM- versus fluoropyrimidine-based chemotherapy in pancreatic cancer and should be incorporated as prespecified correlatives or stratification factors in future phase III studies.* *Relevance section written by JCO Associate Editor Ravi B. Parikh, MD, MPP, FACP. *Relevance section written by JCO Associate Editor Ravi B. Parikh, MD, MPP, FACP. Several transcriptomic signatures, such as GemPred9-11 and Pancreas-View,12 were developed to quantify chemotherapy sensitivity from bulk transcriptomics on tumor biopsies. These demonstrated that patients predicted to be sensitive to GEM and treated as such achieve similar progression-free survival to FFX-treated patients with less grade 3 to 4 toxicity.11 Digital pathology, defined by the computational analysis of digitized whole-slide histology images (WSI), has emerged as an alternative and highly scalable source of biological information.13 WSI are routinely acquired in standard clinical workflows and contain rich, yet largely unexplored, morphologic information reflecting both tumor cells and their microenvironment. Applied to PDAC, histology-based models have shown that WSI can capture clinically relevant tumor and microenvironmental features.14 In addition, direct prediction of treatment response from WSI has the potential to outperform genomic biomarkers.15 Recent studies have demonstrated the effectiveness of deep learning models in predicting general treatment response across different types of cancer.16-18 The flexibility of these models to capture complex morphologies and relate them to survival contributed to their success.19,20 In this study, we investigated whether deep learning computer vision approaches applied to histologies of chemotherapy-naïve surgical specimens can be used to develop and validate a histology-based biomarker to estimate relative benefit from adjuvant GEM versus mFOLFIRINOX in resected PDAC. As a biomarker comparing two regimens, we first derived treatment-specific histology scores for GEM and mFOLFIRINOX, then integrated them into a combined biomarker. This biomarker is intended to support oncologists in selecting adjuvant chemotherapy after surgical resection of PDAC and is positioned within the postoperative decision-making pathway. Finally, taking advantage of the tumor areas highlighted by the model as most associated with chemotherapy benefit, we investigated their molecular features and shed light on resistance mechanisms. PATIENTS AND METHODS Training and Validation Cohorts Resection specimens from upfront surgeries of patients with resectable PDAC from two retrospective multicentric cohorts were processed to generate routine hematoxylin, eosin, and saffron (HES) tissue stains. Digitized WSI from two cohorts were used: GEM adjuvant cohort (cGEM) and FFX adjuvant cohort (cFFX). The first training cohort cGEM consisted of 204 consecutive patients who underwent curative-intent surgery for PDAC across multiple French university centers and who received adjuvant GEM treatment.9,21 The tissue and WSI of 177 patients were available. There were no significant clinical differences between the whole cGEM cohort and the assessable subgroup, except for resection margins (P = .02; Data Supplement, Table S1, online only). Aiming to maximize robustness, the cGEM cohort data were split by clinical center: all patients from one center were reserved for validation (n = 34; 19%), while the remaining were used to train the model (n = 143; 81%). Training and validation sets showed no clinical differences (Data Supplement, Table S2). The second training cohort (cFFX) included 61 consecutive patients who underwent curative-intent surgery for PDAC and received adjuvant FFX between March 2017 and May 2022 from Beaujon (n = 34) and Pitié Salpêtrière Hospitals (n = 25). Tissue was unavailable for seven patients, leaving 54 assessable patients. Assessable patients (n = 54) showed no clinical differences, except for a slightly lower proportion of T1 tumors (P = .02; Data Supplement, Table S3). The smaller size of the bicentric cFFX cohort entailed a random split at a 3:1 ratio to generate a training set (n = 37; 69%) and a validation set (n = 17; 31%). This split was homogeneous in all clinical criteria, except for disease-free survival (DFS), which was higher in the train set (P = .04; Data Supplement, Table S4). The randomized phase III PRODIGE-24/CCTG PA6 trial was used as an external validation cohort. Among the 493 included patients, surgical specimens were available for 366. Eleven patients were excluded because of lack of tumor cells, 14 because of low scan quality, and 28 because of nonanalyzable tissue, leaving 313 assessable patients included in the final analysis, who received FFX (n = 176) or GEM (n = 137), showing no clinical differences, except for slightly fewer patients from the GEM arm (Data Supplement, Table S5) and no clinical differences between assessable patients in both arms (Data Supplement, Table S6). The trial protocol was approved by an independent ethics committee in France (Comité de Protection des Personnes Est III) and by ethics committees at participating centers in Canada. All the patients provided written informed consent. This study was approved by the PRODIGE ethical review board (EudraCT: 2011-002026-52). Statistical Analysis and Experimental Design Hazard ratios (HR) with 95% CI were estimated using Cox proportional hazards models, stratified by resection margin, lymph node status (pN0 v pN1), and CA 19-9 levels in the PRODIGE-24/CCTG PA6 cohort, as in the original study. Survival distributions were compared using log-rank tests. Predictive value was assessed using stratified Cox models including treatment, biomarker subgroup, and their interaction; significance was determined from the interaction term. Kaplan-Meier curves were used to compare DFS and cancer-specific survival (CSS) between groups. Categorical and continuous variables were compared using Fisher's exact test and Wilcoxon or Kruskal-Wallis tests, respectively. All tests were two-sided with α = .05. Reporting follows all the relevant criteria of the TRIPOD + AI statement.22 Image Preprocessing and Deep Learning Networks WSI from HES-stained tissue slides were tiled into 224 × 224-pixel patches. Tissue regions were classified using the PACpAInt14 framework, separating tumor and nontumor areas and distinguishing stromal from tumoral cells. Features were extracted from tumor tiles using the pretrained foundation model UNI.23 For patient-level prediction, tile features from all patient-associated slides were aggregated into bags and processed using attention-based multiple instance learning. Models were trained to predict DFS, with checkpoint selection based on validation Cox loss, using an open-source pipeline.24 Binary stratification was done on the median predicted risk from the training set. The PANCprAId biomarker was defined as the difference between treatment-specific linear predictors (log HRs), after calibration of each model (binarization for histology model for gemcitabine [hGEM] and centering for histology model for FOLFIRINOX [hFFX]). PANCprAId is defined on the log-hazard scale; therefore, the difference corresponds to the log ratio of predicted hazards between regimens. Additional methodological details are provided in the Supplementary Methods. Transcriptome Profile Generation and Analysis Total RNA was extracted from tumor-centered microdissections corresponding to the 25 most resistant and the 25 most sensitive regions highlighted by each model, selected from a cohort of 97 patients with PDAC from Beaujon Hospital. Microdissections were performed using the Millisect system (Roche, Tuscon, AZ). RNA was then isolated using the Quick-RNA FFPE by ZYMO according to the manufacturer's instructions. Subsequently, the RNA was used to generate an RNAseq library using the Smarter stranded pico V3 kit, and the library was sequenced on a NovaSeq X Plus (2 × 100 cycles), with a target of at least 30 million paired reads. The raw RNAseq reads were then mapped to the human genome (Ensembl GRCH38 version 107) and Ensembl's reference transcriptome using the STAR25 aligner. Gene counts were obtained using FeatureCount.26 Transcriptome data for the PRODIGE24/CCTG PA6 trial were previously generated.11 Differential gene expression analysis was performed using DESeq227 using univariate models (sensitive v resistant) with Wald and multi-test correction using the Benjamini-Hochberg method.28 Gene Set Enrichment Analysis29 was performed using genes ranked by the Wald test statistics from DESeq2, comparing them against pancreatic cancer subtypes signatures. Details on immunohistochemistry and quantifications using Qupath30 are presented in Supplementary Methods. RESULTS Derivation of Treatment-Specific Histology Models To build the components of a predictive biomarker, a first model for GEM (hGEM) was trained and tested on WSI of the chemotherapy-naïve specimen histologies of 177 patients who had received adjuvant GEM (cGEM; Fig 1A). The deployment of hGEM in the cGEM training set identified hGEMres and hGEMsens patients as, respectively, resistant and sensitive to adjuvant GEM. hGEMres patients (n = 72; 50%) had a significantly lower DFS (median 11.3 months [95% CI, 9.4 to 14.7]) than hGEMsens patients (n = 71; 50%; median DFS 42.6 months [95% CI, 26.0 to not reached]) and a HR of 3.48 ([95% CI, 2.28 to 5.31]; log-rank P < .001; Fig 1C). In the cGEM validation set (n = 34; 19%), 15 patients (44%) were predicted as hGEMsens and 19 (56%) as hGEMres (Fig 1D) with a significantly shorter survival (HR, 3.44 [95% CI, 1.43 to 8.22]; P = .003). Chemotherapy sensitivity in the training and internal validation sets: (A) cGEM flowchart (train/validation set of the hGEM model) and (B) cFFX flowchart (train/validation set of the hFFX model). Kaplan-Meier curves for DFS in the cGEM cohort (n = 177): (C) training set cGEM (n = 143) and (D) validation set (n = 34) stratified by chemotherapy sensitivity prediction. Kaplan-Meier curves for DFS in the cFFX cohort (n = 54): (E) training set cFFX (n = 37) and (F) validation set (n = 17) stratified by chemotherapy sensitivity prediction. cFFX, FFX adjuvant cohort; cGEM, GEM adjuvant cohort; DFS, disease-free survival; FFX, folinic acid, fluorouracil, irinotecan, and oxaliplatin; HR, hazard ratio; mFOLFIRINOX, modified FOLFIRINOX. To develop the FFX predictive model (hFFX), we relied on a multicentric cohort of 61 patients who had received adjuvant FFX (cFFX; Fig 1B). The hFFX model also showed a significant association with DFS in the cFFX training split (n = 37; P = .005) with a median DFS of 25.9 months for hFFXres patients (predicted resistant to adjuvant mFOLFIRINOX, n = 19; 51%) and median DFS not reached for hFFXsens (predicted sensitive to adjuvant mFOLFIRINOX, n = 18; 49%; HR, 4.52 [95% CI, 1.42 to 14.4]; Fig 1E). The model remained significant (P = .03) in the small validation set, with hFFXres patients (n = 10; 59%) showing a median DFS of 19.9 months versus 31.2 months for hFFXsens patients (n = 7; 41%; HR, 4.14 [95% CI, 1.03 to 16.6]; Fig 1F). Finally, both models were applied in the other cohort, showing neither prognostic value for the hGEM model in the cFFX cohort (HR, 1.25 [95% CI, 0.38 to 4.17]; P = .71) nor for the hFFX model in the cGEM cohort (HR, 1.03 [95% CI, 0.72 to 1.47]; P = .88; Data Supplement, Fig S1). External Validation of the Treatment-Specific Histology Scores in the Randomized Phase III PRODIGE-24/CCTG PA6 Cohort To assess the performance and the generalization potential of the models, the multicentric randomized phase III PRODIGE-24/CCTG PA6 cohort was used as an external validation cohort (Fig 2). Of the 313 assessable patients, 137 (44%) had been randomly assigned to the adjuvant GEM arm and 176 patients (56%) to the adjuvant FFX arm. No significant clinical or survival differences were observed between the whole cohort of patients (N = 493) and the subset included in this histologic study (n = 313), except for a lower proportion of patients in the GEM arm (P < .001; Data Supplement, Table S5). External validation of treatment-specific histology scores in PRODIGE-24/CCTG PA6. (A) The multicentric randomized phase III PRODIGE-24/CCTG PA6 flowchart. Kaplan-Meier curves for DFS in the PRODIGE-24/CCTG PA6 cohort (n = 313): Treatment groups stratified by chemotherapy sensitivity prediction and arm for (B) the hGEM model and (C) the hFFX model. DFS, disease-free survival; GEM, gemcitabine; FFX, folinic acid, fluorouracil, irinotecan, and oxaliplatin; hFFX, histology model for FOLFIRINOX; hGEM, histology model for gemcitabine; mFOLFIRINOX, modified FOLFIRINOX. Applied to WSI of these patients (n = 313), the hGEM model classified 248 patients (79%) resistant to GEM (hGEMres) versus 65 patients (21%) sensitive to GEM (hGEMsens), and the hFFX model classified 163 patients (52%) as resistant to FFX (hFFXres) and 150 patients (48%) as sensitive to (hFFXsens). In the GEM arm (n = 137; 44%), hGEMres patients (n = 107; 78%) had a significantly shorter DFS compared with hGEMsens patients (n = 30; 22%), with median DFS of 11.0 versus 23.1 months, respectively (HR, 1.69 [95% CI, 1.04 to 2.73]; P = .03; Fig 2B). The survival difference between hGEMres and hGEMsens patients was specific to adjuvant GEM and was not observed in the FFX arm (HR, 1.15 [95% CI, 0.74 to 1.81]; P = .5). Likewise, in the FFX arm (n = 176), hFFXres (n = 95; 54%) patients had significantly shorter survival compared with hFFXsens (n = 81; 46%) patients (HR, 2.0 [95% CI, 1.4 to 3.0]; P < .001), while no significant differences were found in the GEM arm (HR, 1.02 [95% CI, 0.68 to 1.51]; P = .9; Fig 2C). PANCprAId: An Algorithm for Tailoring Adjuvant Chemotherapy The clinical utility for tailoring treatment depends on the combined predictive performance of the models. In the PRODIGE-24/CCTG PA6 cohort (n = 313), a direct comparison of both models revealed minimal overlap between patients predicted as sensitive to GEM and FFX (Table 1) and an absence of correlation of the predicted log HRs (Fig 3A). Few patients (n = 25; 8%) were predicted to benefit from both regimens. This limited overlap suggests combining both models could refine adjuvant treatment tailoring. The difference between the predicted log HR was used as a single biomarker, termed PANCprAId, which estimates the personalized difference in sensitivity to one adjuvant regimen versus the other. This results in two groups of patients, those for which the log HR is predicted to be lower under adjuvant GEM (favGEM, n = 46; 15%) versus mFOLFIRINOX (favFFX, n = 267; 85%). Table 1 shows the PANCprAId distribution across patients according to the hGEM and the hFFX predicted log HRs. TABLE 1.Association Between Histology-Based hGEM and hFFX Models in the PRODIGE-24/CCTG PA6 Cohort (n = 313)Model Predicted Category hFFXres hFFXsens Total hGEMres 123125248 hGEMsens 402565Total163150313Abbreviations: hFFX, histology model for FOLFIRINOX; hGEM, histology model for gemcitabine. Association Between Histology-Based hGEM and hFFX Models in the PRODIGE-24/CCTG PA6 Cohort (n = 313) Abbreviations: hFFX, histology model for FOLFIRINOX; hGEM, histology model for gemcitabine. PANCprAId algorithm: combination of hGEM and hFFX models. (A) Scatter plot of the PANCprAId distribution across patients according to the hGEM and the hFFX predicted log hazard ratios. Kaplan-Meier curves and forest plot for (B, C) DFS and (D, E) CSS in the PRODIGE-24/CCTG PA6 cohort (n = 313) stratified by the PANCprAId algorithm. CSS, cancer-specific survival; DFS, disease-free survival; hFFX, histology model for FOLFIRINOX; hGEM, histology model for gemcitabine; nr, not reached; sHR, stratified hazard ratio. Patients with lower predicted log HR under GEM (favGEM) who were randomly assigned to adjuvant GEM had the longest DFS (median 33.5 months [95% CI, 21.8 to not reached]), higher than those randomly assigned to FFX (median 23.6 [95% CI, 12.1 to 36.3], HR: 0.48 [95% CI, 0.23 to 1.02]; P = .09; Fig 3B). However, in the favFFX patient group, patients randomly assigned to adjuvant FFX had a higher DFS (median 21.4 [95% CI, 17.1 to 31.5]) than those randomly assigned to GEM (median 11.1 [95% CI, 9.26 to 13.6]; HR, 2.0 [95% CI, 1.48 to 2.67]; P < .001). Subgroup analyses demonstrated a significant interaction between PANCprAId and adjuvant treatment in DFS (P = .003; Fig 3C). The analysis of the CSS (Figs 3D and 3E) further corroborated these findings (interaction term: P = .001). Histology Signatures Reveal Underlying Biologies of Chemoresistance To characterize the biologic correlates of histology-predicted chemotherapy sensitivity, we analyzed tissue regions associated with treatment benefit identified by the models. Tumors from the PRODIGE-24/CCTG PA6 cohort with RNA and HES available (n = 309) were classified using the RNA-based GemPred (rGEM) and PancreasView-FFX signatures and compared with our histology-based models (hGEM and hFFX). The two modalities identified different subsets of patients, suggesting they measure different mechanisms of resistance or sensitivity to chemotherapies (Data Supplement, Fig S2). Using attention-weighted scores to identify tissue subregions most influential in each treatment-specific model (Fig 4), we selected, for each model, 25 sensitive and 25 resistant regions from the training cohorts (cFFX and cGEM; n = 231) for RNA sequencing, followed by immunohistochemistry (IHC) validation in selected hFFX cases. Chemoresistance biological mechanisms. (A) The six top tiles (max two per patient) weighted attention associated with the according label, squared by a red line with the adjacent tiles, along with the heatmap of the patient with the highest score for hGEMsens, hFFXsens, hGEMres, and hFFXres. (B) Differential expression analysis of gene in microdissection data set between hGEMsens/hGEMres (n = 50) and hFFXsens/hFFXres (n = 50) classes. (C) Gene Set Enrichment Analysis on differential expression analysis results for hGEMmodel and hFFXmodel. hFFX, histology model for FOLFIRINOX; hGEM, histology model for gemcitabine; NES, normalized enrichment score. For the hGEM model, sensitive regions (hGEMsens; Fig 4) were characterized by tumor-adjacent inactive/deserted stroma, consistent with mature, paucicellular fibrosis31 and supported by inactive stromal RNA signatures. By contrast, hGEMres regions showed less differentiated epithelium with basal-like features and increased neutrophil infiltration, with overexpression of proliferative and basal-like signatures21 and neutrophil-associated genes (CXCR2 and FCGR3B) confirmed by MCP-counter.32 hFFX-sensitive regions (hFFXsens) predominantly displayed well-differentiated glandular epithelium, consistent with a classical phenotype (Fig 4), supported by RNA overexpression of classical markers and a high proliferative index, confirmed by Ki-67 IHC (P = .02). RNA deconvolution further revealed increased lymphoid immune infiltration, including B and T cells (Data Supplement, Fig S3). By contrast, hFFX-resistant regions (hFFXres) were characterized by activated tumor-adjacent stroma, with enrichment of TGFB-driven CAF signatures and extracellular matrix components (ACTG2, POSTN, SPARC, and collagens), confirmed by both RNA signatures and IHC (ACTG2 P = .005; POSTN P < .001). DISCUSSION In this study, we developed AI-driven histology models to predict response to adjuvant chemotherapy in resected PDAC. Current treatment selection relies primarily on patient performance status, with limited consideration of tumor biology. Our results suggest that histology-based models can identify patients who may benefit from GEM or mFOLFIRINOX, supporting a more biology-driven treatment strategy. RNA-based predictors such as GemPred and Pancreas-View have demonstrated predictive value but remain constrained by technical requirements, cost, and sensitivity to tumor content. By contrast, histology-based approach provides a faster and cost-effective complementary layer to RNA-based predictive signatures. Leveraging routinely available HES slides enables rapid phenotyping of tumor architecture and stromal organization. This allowed us to assess of tumor-stromal interactions, a key determinant of PDAC progression and chemotherapy response. Comparison with RNA-based signatures revealed largely independent patient stratification, with improved predictive performance when both modalities are combined. The treatment-specific histology models were developed as independent views relating tumor histology to the sensitivity-versus-resistance balance for each treatment. These models can be considered as building blocks for subsequent approaches that may integrate additional treatment regimens or combinations. PANCprAId is a direct application of this idea, comparing the expected benefit of GEM versus mFOLFIRINOX in the adjuvant setting. The PANCprAId biomarker integrates these treatment-specific predictions into a single patient-level estimate of relative benefit, providing a patient-tailored estimate of the proportional change in hazard associated with each regimen. This enables direct comparison between therapeutic options and supports clinical decision making in patients eligible for multiple regimens. In practice, the histology of a patient is used to predict the difference in hazard between the two regimens, thereby providing an estimate of the survival benefit expected from favoring one treatment over the other (favGEM and favFFX). Clinical implementation would likely require calibration steps related to tissue preparation and image acquisition, while the underlying biological signal appears robust across cohorts. Prospective validation will be required before routine clinical use. For instance, a clinical trial randomly assigning patients between a histology-informed adjuvant treatment selection arm and a reference arm in which treatment choice is left to the investigator would prospectively validate the clinical value of the proposed biomarker. Further research should include extension of the proposed biomarker to other settings, including neoadjuvant and advanced PDAC, as well as to other treatment regimens such as GEM plus nab-paclitaxel. It is important to note some limitations. First, the algorithm was not tested on biopsy samples, which could expand the clinical applicability. The limited representativity of small biopsy specimens, typical with endoscopic ultrasound-guided fine-needle biopsies, may challenge accurate characterization of tumor architecture. Furthermore, the predictive capacity of our algorithm was not designed for individual components of mFOLFIRINOX. Despite these limitations, our findings provide a proof-of-concept for using AI-driven histologic analysis to predict chemotherapy benefit in resectable PDAC. This approach may complement existing clinical decision making by enabling personalized treatment selection based on tumor and microenvironmental features. ACKNOWLEDGMENT The authors thank the patients and their families, as well as the investigators, nurses, site staff, and the laboratory technicians for their technical assistance. The authors also thank the Unicancer Data center labeled “Center for Data Handling” by the National Cancer Institute (INCa) for providing clinical data from PRODIGE-24/CCTG PA6 patients. Théau Blanchard Employment: GE Healthcare Stock and Other Ownership Interests: GE Healthcare Research Funding: GE Healthcare Nicolas Fraunhoffer Patents, Royalties, Other Intellectual Property: Coinventor of the patent PCT/EP2022/065222: Simple transcriptomic signatures to determine chemosensitivity for pancreatic ductal adenocarcinoma. SATT Sud-Est Anthony Turpin Honoraria: SERVIER, Daiichi Sankyo/AstraZeneca, AstraZeneca, BMS, Incyte, MSD Oncology, Merck, Norgine, MSD Oncology Research Funding: BeiGene (Inst), Oncoscience (Inst) Travel, Accommodations, Expenses: Pierre Fabre, AstraZeneca, Merck, Takeda Meher Ben Abdelghani Consulting or Advisory Role: Incyte, Pierre Fabre, BMS, Merck Serono, SERVIER, Bayer, Takeda (Inst), AstraZeneca, Daiichi Sankyo/Lilly (Inst) Travel, Accommodations, Expenses: Amgen, SERVIER, Viatris, Merck Serono, Roche, Pierre Fabre Alice Wei Honoraria: Medtronic, AstraZeneca Consulting or Advisory Role: Histosonics Research Funding: Ipsen (Inst) Other Relationship: BioNTech SE (Inst), Epistem (Inst), Clarity Pharmaceuticals (Inst) Emmanuel Mitry Honoraria: Pierre Fabre, MSD, Sandoz Consulting or Advisory Role: Tarian, Novocure, MSD, GlaxoSmithKline Travel, Accommodations, Expenses: MSD, Pierre Fabre Thierry Lecomte Consulting or Advisory Role: Merck Serono, SERVIER, Amgen, IPSEN, Pierre Fabre, AstraZeneca, BMS, Deciphera, Chugai Pharma Europe, AstraZeneca, Incyte, Overseed, Natera, Astellas Pharma Research Funding: Leo Pharma (Inst), Pierre Fabre (Inst) Travel, Accommodations, Expenses: Ipsen, Servier/Pfizer, Takeda, Amgen, Pierre Fabre, IPSEN Pascal Artru Consulting or Advisory Role: SERVIER, Pierre Fabre, AstraZeneca, MSD Oncology, Viatris, Merck Serono, Takeda, Astellas Pharma Travel, Accommodations, Expenses: Servier, Pierre Fabre Ludovic Evesque Honoraria: SERVIER, AstraZeneca, Merck, BMS GmbH & Co. KG, MSD Oncology Aurélien Lambert Consulting or Advisory Role: Janssen, Merck KGaA, Bayer, Servier/Pfizer, Advanced Accelerator Applications, AstraZeneca/MedImmune, GlaxoSmithKline, Ipsen, Merck Serono, Pierre Fabre, Novartis Travel, Accommodations, Expenses: Merck KGaA, Pfizer Daniel J. Renouf Honoraria: Roche, Bayer, Boehringer Ingelheim Consulting or Advisory Role: Roche, Bayer, Pfizer Research Funding: Bayer, Roche (Inst) Nelson J. Dusetti Consulting or Advisory Role: Cure51 Research Funding: Servier (Inst) Patents, Royalties, Other Intellectual Property: PCT/EP2022/065222. Simple transcriptomic signatures to determine chemosensitivity for pancreatic ductal adenocarcinoma. SATT Sud-Est. Inventors: N. Dusetti, N. Fraunhoffer, J. Iovanna, 2021. Licensing to Predicting Med Other Relationship: Predicting Med Pascal Hammel Honoraria: SERVIER, Astellas Pharma, Pfizer Consulting or Advisory Role: VECT-HORUS, AstraZeneca (Inst), Servier/Pfizer, Revolution Medicines Speakers' Bureau: AstraZeneca, SERVIER, Mylan Research Funding: ERYTECH Pharma (Inst), AstraZeneca (Inst) Travel, Accommodations, Expenses: SERVIER, Pfizer, Revolution Medicines Jean-Baptiste Bachet Honoraria: Amgen, Merck Serono, SERVIER, MSD Oncology, Takeda, Incyte, AbbVie, BMS Consulting or Advisory Role: Amgen, Merck Serono, Servier, GlaxoSmithKline, BMS, MSD, Incyte, AbbVie, Takeda, Nutricia, Novocure Travel, Accommodations, Expenses: Merck Serono, Amgen, Servier, Sanofi Louis de Mestier Consulting or Advisory Role: Ipsen, Novartis, Boehringer Ingelheim, Mayoly-Spindler, Sirtex Medical, Esteve, BMS, Viatris, Servier Research Funding: Boehringer Ingelheim (Inst), ITM Solucin (Inst), Novartis (Inst), Crinetics Pharmaceuticals (Inst), Ipsen (Inst), RayzeBio (Inst) Travel, Accommodations, Expenses: Novartis, Ipsen, Mayoly-Spindler Jérôme Cros Honoraria: Novartis Consulting or Advisory Role: HISTALIM Research Funding: ERYTECH Pharma (Inst) Travel, Accommodations, Expenses: IPSEN, Novartis Jakob Nikolas Kather Stock and Other Ownership Interests: StratifAI, Synagen, Spira Labs, Tremont AI, Saterra AI Honoraria: Roche, Pfizer, AstraZeneca, Daiichi Sankyo Europe GmbH, Merck Consulting or Advisory Role: Owkin, DoMore Diagnostics, Panakeia, AstraZeneca, Bioptimus, Mindpeak, MultiplexDx Research Funding: GlaxoSmithKline (Inst), AstraZeneca (Inst) Rémy Nicolle Stock and Other Ownership Interests: Cure51 Consulting or Advisory Role: Cure51 No other potential conflicts of interest were reported. See accompanying Editorial, p. 2467 SUPPORT Supported by the Ligue Contre le Cancer and the Institute National du Cancer (INCa-DGOS_14504, project No. PRT-K19-131), through “SiRIC InsiTu: Insights into cancer: From inflammation to tumor” (grant No. INCa-DGOS-INSERM-ITMO Cancer_18008), and the FRAP network for pancreatic cancer. This study has been performed in the context of the IdEx Université de Paris Cité (ANR-18-IDEX-0001, France 2030). PRODIGE 24 was supported by R&D Unicancer, Clinical Research Hospital Program grant (PHRC11-006) from the French Ministry of Health and the Institute National du Cancer, and the French National League against Cancer. The Canadian Cancer Trials Group Pancreatic Adenocarcinoma (CCTG PA.6) part of the trial was supported by a Program Grant (704970) from the Canadian Cancer Society and by grants from 7 Days in May. J.C., J.N.K., and R.N. contributed equally to this work. Supplementary Materials DATA SHARING STATEMENT A data sharing statement provided by the authors is available with this article at DOI https://doi.org/10.1200/JCO-26-00327. Model weights, preprocessing steps, and inference pipelines are available in the associated repository (https://github.com/GeNeHetX/PANCprAId), enabling full reproducibility of the predictions. AUTHOR CONTRIBUTIONS Conception and design: Audrey Beaufils, Jérôme Cros, Jakob Nikolas Kather, Rémy Nicolle Financial support: Jérôme Cros, Rémy Nicolle Administrative support: Marjorie Mauduit, Jérôme Cros, Rémy Nicolle Provision of study materials or patients: Diana Mendes, Nathalie Colnot, Meher Ben Abdelghani, Alice Wei, Emmanuel Mitry, Thierry Lecomte, James Biagi, Ludovic Evesque, Daniel J. Renouf, Marjorie Mauduit, Pascal Hammel, Thierry Conroy, Jean-Baptiste Bachet, Louis de Mestier, Jérôme Cros Collection and assembly of data: Audrey Beaufils, Julien De Martino, Diana Mendes, Camille Pignolet, Taib Bourega, Miguel Albuquerque, Nathalie Colnot, Meher Ben Abdelghani, Thierry Lecomte, Pascal Artru, James Biagi, Ludovic Evesque, Aurélien Lambert, Daniel J. Renouf, Marjorie Mauduit, Nelson J. Dusetti, Pascal Hammel, Thierry Conroy, Jean-Baptiste Bachet, Louis de Mestier, Vinciane Rebours, Jérôme Cros, Anthony Turpin, Alice Wei, Emmanuel Mitry Data analysis and interpretation: Audrey Beaufils, Xiaofeng Jiang, Théau Blanchard, Nicolas Fraunhoffer, Camille Pignolet, Taib Bourega, Asier Rabasco Meneghetti, Srividhya Sainath, Matthieu Tihy, Nelson J. Dusetti, Jérôme Cros, Jakob Nikolas Kather, Rémy Nicolle Manuscript writing: All authors Final approval of manuscript: All authors Accountable for all aspects of the work: All authors AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST Developing a Histology-Based Artificial Intelligence Biomarker to Predict Adjuvant Chemotherapy Benefit in Pancreatic Cancer The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/jco/authors/author-center. Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments). Théau Blanchard Employment: GE Healthcare Stock and Other Ownership Interests: GE Healthcare Research Funding: GE Healthcare Nicolas Fraunhoffer Patents, Royalties, Other Intellectual Property: Coinventor of the patent PCT/EP2022/065222: Simple transcriptomic signatures to determine chemosensitivity for pancreatic ductal adenocarcinoma. SATT Sud-Est Anthony Turpin Honoraria: SERVIER, Daiichi Sankyo/AstraZeneca, AstraZeneca, BMS, Incyte, MSD Oncology, Merck, Norgine, MSD Oncology Research Funding: BeiGene (Inst), Oncoscience (Inst) Travel, Accommodations, Expenses: Pierre Fabre, AstraZeneca, Merck, Takeda Meher Ben Abdelghani Consulting or Advisory Role: Incyte, Pierre Fabre, BMS, Merck Serono, SERVIER, Bayer, Takeda (Inst), AstraZeneca, Daiichi Sankyo/Lilly (Inst) Travel, Accommodations, Expenses: Amgen, SERVIER, Viatris, Merck Serono, Roche, Pierre Fabre Alice Wei Honoraria: Medtronic, AstraZeneca Consulting or Advisory Role: Histosonics Research Funding: Ipsen (Inst) Other Relationship: BioNTech SE (Inst), Epistem (Inst), Clarity Pharmaceuticals (Inst) Emmanuel Mitry Honoraria: Pierre Fabre, MSD, Sandoz Consulting or Advisory Role: Tarian, Novocure, MSD, GlaxoSmithKline Travel, Accommodations, Expenses: MSD, Pierre Fabre Thierry Lecomte Consulting or Advisory Role: Merck Serono, SERVIER, Amgen, IPSEN, Pierre Fabre, AstraZeneca, BMS, Deciphera, Chugai Pharma Europe, AstraZeneca, Incyte, Overseed, Natera, Astellas Pharma Research Funding: Leo Pharma (Inst), Pierre Fabre (Inst) Travel, Accommodations, Expenses: Ipsen, Servier/Pfizer, Takeda, Amgen, Pierre Fabre, IPSEN Pascal Artru Consulting or Advisory Role: SERVIER, Pierre Fabre, AstraZeneca, MSD Oncology, Viatris, Merck Serono, Takeda, Astellas Pharma Travel, Accommodations, Expenses: Servier, Pierre Fabre Ludovic Evesque Honoraria: SERVIER, AstraZeneca, Merck, BMS GmbH & Co. KG, MSD Oncology Aurélien Lambert Consulting or Advisory Role: Janssen, Merck KGaA, Bayer, Servier/Pfizer, Advanced Accelerator Applications, AstraZeneca/MedImmune, GlaxoSmithKline, Ipsen, Merck Serono, Pierre Fabre, Novartis Travel, Accommodations, Expenses: Merck KGaA, Pfizer Daniel J. Renouf Honoraria: Roche, Bayer, Boehringer Ingelheim Consulting or Advisory Role: Roche, Bayer, Pfizer Research Funding: Bayer, Roche (Inst) Nelson J. Dusetti Consulting or Advisory Role: Cure51 Research Funding: Servier (Inst) Patents, Royalties, Other Intellectual Property: PCT/EP2022/065222. Simple transcriptomic signatures to determine chemosensitivity for pancreatic ductal adenocarcinoma. SATT Sud-Est. Inventors: N. Dusetti, N. Fraunhoffer, J. Iovanna, 2021. Licensing to Predicting Med Other Relationship: Predicting Med Pascal Hammel Honoraria: SERVIER, Astellas Pharma, Pfizer Consulting or Advisory Role: VECT-HORUS, AstraZeneca (Inst), Servier/Pfizer, Revolution Medicines Speakers' Bureau: AstraZeneca, SERVIER, Mylan Research Funding: ERYTECH Pharma (Inst), AstraZeneca (Inst) Travel, Accommodations, Expenses: SERVIER, Pfizer, Revolution Medicines Jean-Baptiste Bachet Honoraria: Amgen, Merck Serono, SERVIER, MSD Oncology, Takeda, Incyte, AbbVie, BMS Consulting or Advisory Role: Amgen, Merck Serono, Servier, GlaxoSmithKline, BMS, MSD, Incyte, AbbVie, Takeda, Nutricia, Novocure Travel, Accommodations, Expenses: Merck Serono, Amgen, Servier, Sanofi Louis de Mestier Consulting or Advisory Role: Ipsen, Novartis, Boehringer Ingelheim, Mayoly-Spindler, Sirtex Medical, Esteve, BMS, Viatris, Servier Research Funding: Boehringer Ingelheim (Inst), ITM Solucin (Inst), Novartis (Inst), Crinetics Pharmaceuticals (Inst), Ipsen (Inst), RayzeBio (Inst) Travel, Accommodations, Expenses: Novartis, Ipsen, Mayoly-Spindler Jérôme Cros Honoraria: Novartis Consulting or Advisory Role: HISTALIM Research Funding: ERYTECH Pharma (Inst) Travel, Accommodations, Expenses: IPSEN, Novartis Jakob Nikolas Kather Stock and Other Ownership Interests: StratifAI, Synagen, Spira Labs, Tremont AI, Saterra AI Honoraria: Roche, Pfizer, AstraZeneca, Daiichi Sankyo Europe GmbH, Merck Consulting or Advisory Role: Owkin, DoMore Diagnostics, Panakeia, AstraZeneca, Bioptimus, Mindpeak, MultiplexDx Research Funding: GlaxoSmithKline (Inst), AstraZeneca (Inst) Rémy Nicolle Stock and Other Ownership Interests: Cure51 Consulting or Advisory Role: Cure51 No other potential conflicts of interest were reported. REFERENCES Associated Data Supplementary Materials Data Availability Statement A data sharing statement provided by the authors is available with this article at DOI https://doi.org/10.1200/JCO-26-00327. Model weights, preprocessing steps, and inference pipelines are available in the associated repository (https://github.com/GeNeHetX/PANCprAId), enabling full reproducibility of the predictions.