Purpose Modified FOLFIRINOX (FFX) and gemcitabine plus nab-paclitaxel (GNP) are standard first-line treatments for metastatic pancreatic ductal adenocarcinoma (PDAC), but no validated biomarker guides treatment selection. We developed MULTIPL, a multimodal machine learning system, and established the PASS-01 Challenge to benchmark prognostic and predictive biomarkers. Patients and Methods MULTIPL was trained in the COMPASS study (N=268), integrating clinical, digitized histopathology, whole-genome, and RNA-seq data. MULTIPL, PurIST, hENT1 expression, and HRDetect were evaluated in the PASS-01 trial, a randomized phase II trial of FFX versus GNP (N=160), within the Challenge. The primary endpoint was differential treatment benefit measured by concordance-for-benefit for progression-free survival. Results MULTIPL had the highest concordance index for OS among individually evaluated biomarkers (0.595; 95% confidence interval [CI], 0.55-0.65) and separated high- versus low-risk patients (hazard ratio, 1.62; 95% CI, 1.13-2.33; P=0.009). Patients recommended for GNP by MULTIPL had significantly longer OS with GNP than with FFX (hazard ratio, 0.47; 95% CI, 0.28-0.82; P=0.007), whereas patients recommended for FFX had similar OS between treatments. Interpretability analysis of MULTIPL in COMPASS identified KDM6A alterations and SSTR1 expression as prognostic biomarkers, which were validated in PASS-01. However, none of the tested biomarkers significantly predicted differential treatment benefit in the PASS-01 Challenge. Conclusion MULTIPL demonstrated robust prognostic performance in external validation, identified a subgroup enriched for benefit from GNP, and enabled discovery and validation of prognostic biomarkers in metastatic PDAC. However, no biomarker met the primary endpoint for differential treatment benefit, underscoring the value of the PASS-01 Challenge. ### Competing Interest Statement R.C.G. discloses paid consulting or advisory roles for AstraZeneca, Eisai, Incyte, Knight Therapeutics, Guardant Health, and Ipsen, all unrelated to this work. The other authors declare no competing interests. ### Clinical Trial NCT04469556 ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: Ethics committee/IRB of University Health Network gave ethical approval for this work. I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes All data produced in the present study are available upon reasonable request to the authors Princess Margaret Cancer Foundation, https://ror.org/00ctsa544 Ontario Institute for Cancer Research, https://ror.org/043q8yx54 University of Toronto, https://ror.org/03dbr7087 Terry Fox Research Institute, https://ror.org/00mtf6c40

Background: Histomorphology is a strong prognostic biomarker correlated with basal-like and classical programs in surgically resected pancreatic ductal adenocarcinoma (PDAC). However, the spectrum of morphology and its biological associations remain poorly defined in advanced disease. Objectives: We explored the transcriptomic and genomic underpinnings and clinical relevance of morphological classes across localized and metastatic PDAC. Design: We unified morphological classifications into four classes: glandular, cribriform, solid, and squamous. We integrated transcriptome and whole–genome sequencing following laser–capture microdissection with morphological classifications in 348 PDAC patients, where half of the cohort included locally advance and metastatic stages to uncover molecular associations. Results: Non–glandular morphologies comprised three distinct classes that were enriched in metastatic disease. Transcriptomic profiling exhibited that glandular tumours predominantly expressed classical epithelial programs, although a subset displayed partial or full epithelial–mesenchymal transition signatures. In contrast, non–glandular morphologies showed basal–like transcriptional programs with subtype–specific pathways, including ciliogenesis in cribriform tumours, extracellular matrix remodelling and immune evasion in solid tumours, and keratinisation programs in squamous tumours. The solid class was significantly enriched in liver metastatic lesions and was associated with increased intra– tumoural morphological heterogeneity, whole-genome doubling, KRAS major allelic imbalance, and elevated KRAS–ERK signalling. Conclusion: Non–glandular morphologies identify biologically distinct PDAC tumor states that are enriched in liver metastases and associated with subtype–specific transcriptional programs and KRAS–driven genomic alterations.

754 Background: Pancreatic ductal adenocarcinoma (PDAC) is the third deadliest malignancy with most patients (pts) presenting with advanced disease. In those with localized disease at diagnosis who undergo curative intent surgery followed by adjuvant chemotherapy, the relapse rate remains high (70%). While there are known clinicopathological risk factors to predict relapse, little is known about genomic and transcriptomic features differentiating recurrent PDAC. Methods: We identified PDAC pts who had curative surgery at Princess Margaret Cancer Centre in Toronto, Canada between August 2008-2022 who had whole genome (WGS) and transcriptomic (RNA-Seq) data available. Patients were classified by timing of relapse after surgery: early (≤6 months), intermediate (6–12 months), late (>1 year), or no recurrence. Clinicopathologic and survival data were collected via chart review, and groups were analyzed using Fisher’s exact/Wilcoxon rank sum tests. Kaplan-Meier survival analyses were conducted. Results: Of 715 surgical resections, 170 had clinical and WGS/RNA-Seq data available. Ninety-three pts (55%) were female and 126 (74%) had at least N1 disease. Timing of relapse was classified as early (n=37, 22%), intermediate (n=46; 27%), late (n=61; 36%), and no recurrence (n=26; 15%) at time of data cut-off. There were no statistically significant differences in sex, Moffitt subtype, HRD/MMR status, or driver mutations such as KRAS , CDKN2A , and SMAD4 . Clinicopathological analysis revealed that vascular involvement (p=0.0231), nodal (N) status (p=0.0384), non-adenocarcinoma histology (p=0.0393) and G3 tumors (p<0.001) were significant predictors of early recurrence. Hepatic metastasis accounted for most early recurrences (51%), whereas late recurrences were often locoregional (36%) or pulmonary (20%). Genomic analysis revealed a higher incidence of ARID1A mutations in early vs. late recurrence (20% vs. 3.5%, p=0.0179), with 71% demonstrating monoallelic loss, while low ploidy (p=0.0371), PTEN mutation (p=0.0166) and TP53 wild-type status (p=0.0237) were significantly associated with no recurrence. Higher structural variant counts were also observed in the early recurrence group (p=0.0163). Differential expression analysis demonstrated significant enrichment in gene sets associated with MYC activation, glycolysis and hypoxia response (FDR q-value < 0.001) in early recurrence pts, while the late-recurrence pts showed enrichment in pancreatic beta cell-related gene sets (FDR q-value < 0.01). Conclusions: While clinicopathologic factors have been confirmed as prognostic for recurrence, we identified a higher incidence of ARID1A mutations in the early recurrence group and significant enrichment in gene sets associated with MYC/hypoxia. Multi-omics analysis could further improve recurrence prediction, guiding future therapeutic strategies.

774 Background: The optimal approach in r-PDAC remains unclear. Perioperative Modified FOLFIRINOX (mFFX) is increasingly used yet we lack biomarkers to guide this approach. NeoPancONE demonstrated GATA6 high expression at baseline and classical RNA subtype on resection tissue in pts receiving perioperative mFFX was associated with a trend toward improved EFS and OS. Low GATA6/basal-like predicts early disease progression and poor outcomes 1 . We now report on additional transcriptome and genomic subtypes identified in the NeoPancONE resection specimens. Methods: Patients with r-PDAC were enrolled following central review and underwent an EUS FNB for central analysis of baseline GATA6 by in-situ hybridization (ISH). Surgical resection specimens were processed in each centre and fresh frozen tumor tissue was transferred for central whole genome and transcriptome sequencing (WGTS) following laser capture microdissection. WGTS reports annotated mutations, copy number alterations, mutational signatures, and structural variant patterns. KRAS allelic states were reported with major imbalances in mutant KRAS (KRAS maj ) defined as mutant: Wildtype (WT) copy number ≥3. RNA transcriptomic subtypes (classical vs. basal-like) were classified by PurIST. We report a descriptive and survival analysis of the genomic subgroups. Results: Of the 84 enrolled baseline GATA6 ISH was analysed in 74 (88%); 81% high, 19% low and 73 (87%) proceeded to surgery. PurIST subtyping was available in 51 (70%) of resections; 86% classic, 14% basal. Spearman correlation coefficient between GATA6 ISH and RNA subtype = 0.51 (p<0.001). In those who recur/progress there was no association between GATA6 and RNA subtype and site of recurrence p=0.57 and p=0.18 respectively. WGS reports are available in 36 (49%) resections. KRASm were evident in 94% and KRAS G12D represented 50%. MTAP deletion was present in 22%, mostly correlated with CDKN2A deletions. Tumor suppressors TP53, CDKN2A and SMAD4 were lost in 86%, 69% and 50% respectively. KRAS maj was present in 8% of cases and polyploid tumors in 28%. Polyploid and KRAS major were enriched in the basal-like, and associated with significantly worse OS (p=0.046 and 0.0081 respectively). There was also enrichment of SMAD4 wt in basal-like tumors. ARID1A wt tumors demonstrated improved OS (P=0.032). Unstable genotypes were present in 6% with somatic-HRD evident in 3%. Conclusions: NeoPancONE represents a heterogeneous group of genotypes and molecular subgroups in r-PDAC. GATA6 low and basal-like subtype, associated with inferior outcomes, demonstrate less genomic stability. These results substantiate consideration of an alternative to mFFX based peri-operative therapy in GATA6 low/basal-like and reinforces the value of incorporating upfront genomic subtyping. Future r-PDAC trials should now incorporate RNA subtyping and KRAS directed therapies. Ref: 1. McLaughlin et al. ASCO 2025. Clinical trial information: NCT04472910 .

696 Background: Patients (pts) with advanced pancreatic ductal adenocarcinoma (PDAC) experience significantly worse outcomes when nutritional support is inadequate; characterized by weight loss, GI symptoms, and metabolic derangements. These correlate with increased toxicity, reduced survival, and diminished quality of life. Prior prospective cohorts 1 highlight that most pts deteriorate nutritionally from baseline with a significant reduction in weight (p<0.001), albumin (p<0.001) and symptoms using the Edmonton Symptom Assessment Scale (ESAS) tool in the first months on chemotherapy. The NUANCE study was designed to evaluate a prospective interprofessional psychoeducational program by measuring routine clinical endpoints. Methods: Pts with metastatic and locally advanced (LA) unresectable PDAC were consented if undergoing systemic chemotherapy. Pts undergo comprehensive initial assessments and are then pre-emptively supported by a registered dietitian, nurse specialist and social worker to identify and optimize GI/nutritional symptoms. We report assessments of nutrition, ESAS scores and ECOG PS from baseline to 4 months. Results: Since Oct 2022, a planned 100 pts (78 metastatic, 22 LA) have consented. We report on baseline measurable nutritional parameters and symptoms and at 4-mos in pts reaching 4-mos. With a med F/U of 8 mos; med age 67.5 yrs (40-84), 44% female, the med OS on study was 12 mos (9-16). Diabetes was present in 39% at baseline, 54% were prescribed insulin. Pancreatic enzyme replacement was offered in 88%. Baseline and 4-mos clinical measures and patient reported symptoms by ESAS are described in table 1. Pain, appetite and diarrhea symptom ESAS scores improve (NS) and energy level ESAS scores are maintained. No significant worsening is observed in weight and ECOG PS, and 68% of patients at 4 mos maintain albumin levels > 36g/L. ESAS completion rate was low (50%). Conclusions: In these early NUANCE results, we observe no significant deterioration in weight, ECOG PS and in maintaining albumin, a marked improvement over the previous cohort suggesting this program helps maintain pt’s nutrition and wellbeing during the critical first 4 months on chemotherapy. This supports the integration of nutritional screening and interprofessional support as a critical component of PDAC routine care to mitigate chemotherapy toxicity and improve patient-centered outcomes. Next steps include building the NUANCE design into a smart-app, providing more accessible support to pts with advanced PDAC. Ref: 1. Knox et al. COMPASS unpublished data. Characteristic/Symptom (ESAS score) N=100(baseline) Baseline value med (range) n(4 mons) 4 mos med (range) P value Weight (Kg) 100 67.0(37.0 – 105.9) 47 63.2(40.6 – 109.2) 0.33 Pain scores 52 3 (0 – 8) 45 1.0 (0 – 8) 0.18 Energy 51 5 (1 – 10) 46 5 (0 – 9) 0.95 Appetite 51 5 (0 – 10) 45 4 (0 – 9) 0.27 Diarrhea 50 1 (0 – 8) 43 0 (0 – 7) 0.67 Baseline n 4 mos n ECOG PS 0/1/2/3 100 13 / 73 / 12 / 2 39 4 / 31 / 3 / 1 0.95

Abstract Cancers acquire alternate transcriptional states as they evolve, but the origins, timing and determinants of this plasticity are poorly understood in many tumours. We investigated the transcriptional states of pancreatic cancer by integrating ∼1000 tumour-enriched genomes and transcriptomes from 464 patients combined with scRNA-seq, multiome profiling, and spatial proteomics. Four epithelial states covering the spectrum of lineage plasticity were identified (Classical-1, Classical-2, Basal-1, Basal-2). Comparing these states to normal and pan-cancer human single cell atlases showed each state reflects distinct tissue programs found in other malignancies. Single cell analysis uncovered that the main transcription state of this disease (Classical-1) emerges before KRAS mutations. Spatial proteomics from patients and cancer-free donors showed that the Classical-1 program emerges during acinar-to-ductal metaplasia, and also unexpectedly, in normal ducts without disrupting their morphology. Overall, these findings link the extensive lineage plasticity potential of this organ to the origins of the transcriptional states.

PURPOSE:Biliary tract cancer (BTC) is the leading cause of death in patients with primary sclerosing cholangitis (PSC). PSC-related BTC is poorly understood, and the risks and benefits of conventional and immunotherapy treatments are unknown. We aimed to characterize clinical outcomes and genomes of PSC-related BTCs. EXPERIMENTAL DESIGN:This was a retrospective cohort study of patients with BTC with underlying PSC treated at MD Anderson Cancer Center (N = 46) and Princess Margaret Cancer Centre (N = 16), which were contrasted to patients with non-PSC-related BTC (N = 146). We compared outcomes between PSC and non-PSC, and PSC treated with and without immunotherapy. A combination of targeted sequencing (N = 139), whole-genome sequencing (WGS; N = 27), and WGS with paired RNA sequencing (N = 33) delineated the genomic and transcriptomic landscape of PSC-associated BTCs. RESULTS:In PSC-related BTC, the addition of immunotherapy to chemotherapy was associated with improved first-line progression-free survival (PFS; N = 22 vs. 11; median PFS, 12.2 vs. 4.7 months; P = 0.01). Immune-related adverse events were rare (N = 2, 12.5%) and improved after treatment discontinuation. Classic actionable genomic alterations, including IDH1 mutations and FGFR2 fusions, were absent in PSC-related BTCs. PSC tumors had a 2.6-fold higher tumor mutational burden (P = 3.28e-05) compared with non-PSC tumors. Transcriptomic profiling revealed a subset of PSC tumors displaying RNA signatures of immunotherapy response. CONCLUSIONS:Immunotherapy in PSC-associated BTCs seemed safe, with a potential signal of effectiveness. Given the sample size and retrospective design, these results are hypothesis-generating. Together, these results demonstrate the unique biology underlying PSC-associated BTCs, highlighting the need for prospective trials and the development of specialized treatment strategies.

Abstract Background & Objectives Biliary tract cancer (BTC) groups a family of rare cancers with high recurrence rates after surgery and poor survival, especially in advanced disease. Several molecular classification systems based on transcriptomics demonstrate prognostic relevance, however clinical implementation has been limited by a lack of consensus or elucidation of their biological underpinnings. Methods We performed whole genome and transcriptome sequencing of 169 tumors enriched with laser capture microdissection and annotated with clinical records. Results Network integration and clustering on existing classifiers revealed that BTC transcriptomes can be summarized into two consensus molecular subtypes (BTC-CMS). BTC-CMS-A tumors almost exclusively carry the clinically actionable alterations characteristic of BTC (IDH1 mutations and FGFR2 fusion). These tumors are restricted to the liver and confer superior survival after surgery (P = 9.5e-05) and in advanced disease (P = 0.02) in cohorts without precision therapy treatments. BTC-CMS-B tumors span the biliary tree and are associated with chronic inflammation risk factors including primary sclerosing cholangitis (P = 0.016), fluke infection (P = 1.8e-10) and cholecystitis (P = 0.0029). These tumors exhibit mutations characteristic of pancreatic cancer (TP53, SMAD4, CDKN2A and KRAS), a connection further supported by similarities in their transcriptomic profiles. BTC-CMS can also be reliably differentiated by tumour proteomes and methylomes, offering multiple modalities for tumor classification. de novo expression signature extraction by non-negative matrix factorization and subsequent projection of these signatures onto single cell RNAseq data supports that individual cells from a given tumour are exclusively one subtype, with divergent copy number and expression profiles. Furthermore, we show patient tumors are restricted to one subtype across metastases and through progression or recurrence, suggesting these are immutable programmes. Conclusion The BTC-CMS robustly defines two distinct molecular lineages of BTC that improve clinical and biological stratification over anatomy atone.

Better patient selection for treatment is critical to improving both cancer care and therapeutic development in oncology. The ability to predict individual patient responses to cancer treatment ahead of time would transform cancer care with substantial impact on outcomes, quality of life and cost. We generated molecular digital twins of individual participants using a Bayesian foundation model of cancer (FarrSight®) in the COMPASS clinical trial (a non-randomized study of mFOLFIRINOX and gemcitabine + nab-paclitaxel in first-line advanced pancreatic cancer). We compared these individual digital twin predictions to the existing Moffitt classification of pancreatic ductal adenocarcinoma. Individual digital twin predictions of response to mFOLFIRINOX outperformed the Moffitt classification in the basal-like subtype with an AUC of 72.3% compared to the conventional biomarker AUC of 44.8%; overall accuracy of 65.8% vs. 47.4%; PPV of 60% vs. 40%; and NPV of 72.2% vs. 52.2%. Individual patient response predictions using models such as FarrSight® have the potential to better select patients for treatment with established therapeutics and in therapeutic development compared to biomarkers based on population averages.

PURPOSE:Predicting recurrence of pancreatic cancer after surgery could inform clinical decision making, including adjuvant therapies and follow-up. This study aimed to develop and validate a deep learning model using digitized whole-slide images (WSI) of histopathology. METHODS:Publicly available WSI of pancreatic ductal adenocarcinoma resections from three cohorts were used for training. The model consisted of a pan-cancer foundation model to generate embeddings, mean-pooling across tissue patches, and then a fully connected neural network. Model predictions were compared with human-labeled histopathologic features and genomic alterations. The model was externally validated in a meta-analysis of a single-center cohort from Princess Margaret Cancer Centre, a multicenter cohort from France, and the PRODIGE 24 trial of adjuvant chemotherapy. RESULTS:The deep learning model was trained on 12,594 tissue patches from 257 patients. High-risk classifications were associated with squamous morphology, reactive stroma, tumor cellularity, and necrosis, whereas low-risk classifications were associated with tubulopapillary and conventional morphologies, as well as deserted stroma. High-risk cancers were enriched for basal-like gene expression profiles and distinct oncogenic pathways. In a meta-analysis of the external cohorts, the hazard ratio (HR) for death comparing high-versus low-risk cancers was 1.49 (95% CI, 1.25 to 1.79, P < .001), whereas the HR for recurrence or death was 1.41 (95% CI, 1.19 to 1.68, P < .001). The classifications remained prognostic among moderately differentiated cancers. CONCLUSION:An open-source deep learning model using WSI from pancreatic cancer resections generated risk classifications that correlated with histopathologic and genomic features. Classifications were externally validated in a meta-analysis of three cohorts. This model could be applied to WSI to provide individualized prognostic information for patients.

772 Background: The increasing use of neoadjuvant systemic therapy for resectable pancreatic adenocarcinoma (rPDAC) highlights the need for biomarkers for patient selection in this setting. High GATA6 expression has been shown to be useful in identifying the classical subtype, compared to the GATA6 low basal-like subtype associated with chemoresistance. GATA6 expression and CA19-9 levels have been evaluated independently as potential biomarkers, with their respective limitations. We therefore sought to assess whether the use of both, in tandem, may enhance their prognostic utility in the phase 2 NeoPancONE trial where CA19-9 levels were not part of exclusion. Methods: Patients with rPDAC were analyzed within four subgroups based on GATA6 ISH (high vs low) and baseline CA19-9 level (≥200 vs <200 U/mL), following the cut-off used in the PRODIGE48 study. CA19-9 non-secretors (≤2 U/mL) were excluded. Overall survival (OS) and event-free survival (EFS) were assessed using the Kaplan-Meier method and log-rank test. The prognostic effect of GATA6 expression and CA19-9 levels were assessed by uni- and multivariate analyses. Results: Out of 84 patients enrolled between 09/2020 to 09/2023, 74 (88%) had GATA6 ISH baseline testing and 78 (93%) had CA19-9 secreting disease, while 68 (81%) had both and were included in this analysis. Median OS was significantly longer in subgroup A compared to subgroups C and D, but not B (p<0.001, p<0.001, p=0.149). Median EFS was shorter in subgroup D compared to the other subgroups but not significant with B (p<0.001, p=0.014, p=0.104). The remaining pairwise comparisons did not reach statistical significance, however the subgroups were small. In univariate analysis, GATA6 high was associated with improved OS (HR 0.47; 95% CI: 0.22-0.98; p=0.043), while CA19-9 ≥200 U/mL was associated with worse OS (HR 3.52; 95% CI: 1.74-7.12; p<0.001). Both remained significant in multivariate analysis (p=0.044, p<0.001), and no significant interaction was observed. The findings on univariate analysis for EFS were not significant for GATA6. Conclusions: This study explores potential biomarkers at baseline in rPDAC treated with neoadjuvant FOLFIRINOX. GATA6 and CA19-9 are independent prognostic markers and so their prognostic utility is improved when used in tandem. High GATA6 expression with low CA19-9 was associated with best OS, while low GATA6 expression with high CA19-9 was associated with worse EFS, suggesting a strong rationale for stratification at baseline. Further studies are warranted to establish biomarkers in rPDAC to guide patient selection for neoadjuvant therapy. Clinical trial information: NCT04472910 . Survival outcomes by subgroup. Subgroup Median OS (months) Median EFS (months) A: CA19-9 <200 U/mL, GATA6 high (n=31) 40.9 27.1 B: CA19-9 <200 U/mL, GATA6 low (n=7) 36.9 21.8 C: CA19-9 ≥200 U/mL, GATA6 high (n=24) 26.9 15.8 D: CA19-9 ≥200 U/mL, GATA6 low (n=6) 15.0 6.4

Abstract Pancreatic ductal adenocarcinoma (PDAC) has consistent genomic drivers and established classical and basal transcriptional subtypes, but despite this, there exist many intermediate state tumors, a multiplicity of copy number aberrations and low frequency mutations, and a complex, spatially heterogeneous tumor microenvironment. These elements are not random; they are evidence of coordinated biological programs and evolving signaling networks that underlie tumor progression. Merging data from these varied molecular and spatial layers is essential to clinically define and therapeutically target a) tumors that fall into the ‘grey zone’ of pancreatic cancer subtyping and b) specific collusive epithelial-stromal niches.To identify the in situ interdependencies between distinct tumor cell states, fibroblast phenotypes, deposited extracellular matrix, immune infiltrate, and vasculature, we performed imaging mass cytometry on three serial sections of a PDAC tissue microarray (221 resected tumors, ∼4 cores each), generating >800 multiplexed images (40-43 channels) each focused on deeply profiling a different cell lineage. We captured 76 immune and stromal cell types and states, as well as six cancer cell types that recapitulated the classical and basal PDAC signatures, plus four discrete “intermediate” states with distinct associations to RNA subtype (n = 92), tumor ploidy (n = 192), and patient outcome. We clustered our immune and stromal cell populations to define 8 recurrent microenvironments, and found the microenvironment dominated by CD105+ CAFs was significantly spatially associated to classical tumor cells with strong epithelial differentiation transcription factor expression (pairwise Fisher’s exact test, odds ratio = 3.7), ECM-rich microenvironments were proximal to basal tumor cells (odds ratio = 4.3), and pMLC2+ CAFs were enriched near the poor-prognosis, low-ploidy S100A4+ tumor phenotype (odds ratio = 4.2). We additionally denote a specific fibroblast-centric microenvironment associated with neoadjuvant treated tumors (n = 26). Using matched 30X whole-genome sequencing (n = 192), we found specific mutations and copy number changes that were associated changes in tumor and microenvironment composition, and performed Lasso-based machine learning to determine the most important cross-omic features for overall survival prediction. Together, our findings define a phenotypic and molecular framework of PDAC from genome to tumor-microenvironment, provide insight into the connection between tumor phenotype and stromal niches, and offer a refined basis for patient stratification. Citation Format: Ferris Nowlan, Noor Shakfa, Sibyl Drissler, Tan Tiak Ju, Elizabeth Sunnucks, Edward LY Chen, Cassandra J. Wong, Brendon Seale, Zhen Yuan Lin, Michelle Chan-Seng-Yue, Amy Zhang, Sabiq Chaudhary, Chengxin Yu, Golnaz Abazari, Michael Geuenich, Matthew Watson, Jiaxi Peng, Somaieh Afiuni-Zadeh, Ayelet Borgida, Ricardo Gonzalez, Sheng-Ben Liang, Klaudia Nowak, Miralem Mrkonjic, Anna Dodd, Julie M. Wilson, Kieran Campbell, Jenn Gorman, Barbara Grünwald, Robert C. Grant, Jennifer J. Knox, Anne-Claude Gringas, Faiyaz Notta, Steven Gallinger, Grainne O'Kane, Hartland Jackson. Single cell imaging uncovers a coordinated tumor-immune-stroma spectrum with genomic associations in pancreatic ductal adenocarcinoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6213.

Abstract Pancreatic ductal adenocarcinoma (PDAC) remains one of the most lethal malignancies, with most patients diagnosed at advanced stages and limited therapeutic options. The aggressive biology and substantial molecular heterogeneity of PDAC underscore the need for approaches that enable more personalized treatment selection. Patient-derived organoids (PDOs) have emerged as powerful in vitro tumor model that retains key genetic, transcriptional, and phenotypic features of the tumors from which they are derived. As such, PDOs provide an opportunity to bridge fundamental cancer biology with translational and clinical applications, including functional precision medicine. Our work focuses on integrating PDO models into prospective clinical studies of PDAC, most prominently through the PASS-01 clinical trial, a multi-institutional study of patients with stage IV disease. Within this trial, biopsies were collected for molecular correlatives including PDO establishment across six institutions in the United States and Canada. A total of 186 biopsies from 183 enrolled patients were shipped to Cold Spring Harbor Laboratory for PDO generation and characterization. Biopsy samples originated primarily from liver metastases but also included primary pancreatic tumors and other metastatic sites such as peritoneal, omental, lymph node, lung, and brain lesions. KRAS mutation analysis using droplet digital PCR was used to confirm neoplastic organoid cultures and identify pseudonormal epithelial outgrowth, which occurred in approximately 25–40% of samples.Overall, malignant PDOs were successfully established from approximately 48% of patient biopsies. Interestingly, successful organoid establishment was associated with patients who experienced more rapid clinical progression, and organoid morphological features also correlated with patient outcomes. Molecular characterization revealed that PDOs were more frequently generated from tumors belonging to the classical transcriptional subtype and from tumors harboring KRAS and TP53 mutations. Established PDO lines were expanded, biobanked, and subjected to high-throughput pharmacotyping against a library of more than 120 standard-of-care and investigational compounds. The median time from tissue receipt to drug screening data was 70 days, with several lines yielding results within two months, enabling pharmacotyping data to be presented during monthly molecular tumor boards. In several cases, PDO drug response profiles were considered alongside genomic and clinical data to help inform second-line therapy decisions following disease progression. Ongoing work integrates PDO pharmacologic response data with genomic and transcriptomic profiling to identify predictive biomarkers of therapy response and resistance. Our results thus far demonstrated some concordance with patient responses to first-line chemotherapy, particularly gemcitabine/nab-paclitaxel. In parallel, PDO models are being used to evaluate emerging targeted therapies, including RAS inhibitors and combination strategies aimed at PDAC vulnerabilities. Together, these studies highlight the potential for patient-derived cancer models to accelerate translational discovery and support clinically actionable precision medicine approaches in pancreatic cancer. Citation Format: Amber N Habowski, Fatim Kouassi, Hardik Patel, James Rouse, Caitlin Tsang, Luce Kelly, Dennis Plenker, Gun Ho Jang, Deepthi Budagavi, Grainne O'Kane, Raditya Utama, Julie Wilson, Anna Dodd, Stephanie Ramotar, Julien Hohenleitner, Kimberly Perez, Robert C Grant, Eileen M O’Reilly, Steven Gallinger, Dan A Laheru, Brian M Wolpin, Andrew J Aguirre, Kenneth H Yu, Elizabeth M Jaffee, Jennifer Knox, Daniel A King, Faiyaz Notta, David A Tuveson. Human cancer models: From patient-derived systems to clinical translation [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(8_Suppl):Abstract nr SY36-02.

600 Background: Biliary tract cancers (BTC), including cholangiocarcinomas and gallbladder adenocarcinomas, remain poorly understood. Stratifying this heterogeneous group of diseases based on molecular features may lead to novel precision treatments and yield urgently needed improvements in outcomes. Methods: We performed whole-genome and transcriptome sequencing of 165 laser-capture microdissected BTC tumors with paired normal tissues, representing 165 patients. First, we assigned all samples to 28 gene expression-based BTC signatures from nine previously published classifiers to create a signature similarity network. Using Markov clustering on the optimized network, we identified broad consensus molecular subtypes. Next, we used non-negative matrix factorization to decompose the consensus molecular subtype signatures into underlying components to identify associated cell types and processes. Finally, we integrated subtypes with de novo derived mutational signatures and drivers as well as electronic health records to search for prognostic and outcome implications. Results: Published subtyping classifications reflected only two underlying consensus molecular subtypes, which we named BTC-CMS-A and BTC-CMS-B. BTC-CMS-A tumors represented two thirds of intrahepatic cholangiocarcinomas and all mixed hepatocellular carcinomas, but no extrahepatic BTCs or gallbladder cancers, while BTC-CMS-B tumors were found along the whole biliary tract. BTC-CMS-B components exhibited markers of common bile duct epithelium but also shared expression profiles with intestinal Goblet cells, suggestive of an intestinal metaplasia histopathology. In contrast, BTC-CMS-A expressed hepatoblast markers while also exhibiting upregulation of pathways associated with kidney cell lineage. These results were validated by projecting all CMS-associated expression components onto two public single-cell RNAseq datasets. Of 16 drivers identified de novo from WGS, BAP1 and TP53 mutations were most frequent, mutually exclusive, and each was restricted to BTC-CMS-A and BTC-CMS-B, respectively. Mutational signature extraction revealed BTC-CMS-B tumors had increased tumor mutational burden and APOBEC-associated mutational signatures, whereas BTC-CMS-A was associated with increased non-homologous end-joining. Clinically, hepatitis B was more frequent in BTC-CMS-A cases, while patients with primary sclerosing cholangitis exclusively had BTC-CMS-B tumors. BTC-CMS-A was associated with significantly longer overall survival compared to CMS-B, after adjusting for clinically relevant covariates. This finding was validated in three external cohorts. Conclusions: BTC-CMS unify previous BTC classifications and display distinct clinical and genomic profiles and therefore should guide the development of targeted therapies and biomarker-driven trials.

Abstract In pancreatic cancer, 50% of the patients are diagnosed at the metastatic stage due to a lack of symptoms. The overall survival rate dramatically reduces from 44% in early-stage resectable patients to 3% in metastatic cases. However, even in early-stage patients, >75% of them still recur after resection and adjuvant therapies. Therefore, there is an urgent need to develop a sensitive strategy to detect this disease earlier before it spreads. Characterizing circulating tumor DNA (ctDNA) in plasma is an effective approach to detect and monitor cancer. Whole-genome sequencing (WGS) tracks all the tumor mutations simultaneously, and has a higher sensitivity than targeted approaches, especially in samples with extremely low tumor burden. In this study, to investigate the ctDNA dynamics and early dissemination in pancreatic cancer, we established a cohort of 1,013 samples from 277 donors, including plasma WGS at 20-60x, alongside germline DNA WGS, tissue WGS and transcriptomic sequencing. Using a tumor-guided approach, we found that the early-stage resectable cases shed very little ctDNA (<1% tumor fraction, TFx). Plasma TFx levels were elevated at the metastatic stage, but were also dependent on metastatic tissue site, where patients with liver metastases had higher TFx than the ones with only non-hepatic lesions. We further discovered the tumor-intrinsic features that were related to increasing ctDNA shedding, such as whole-genome duplication (WGD), high cell cycle activity and non-glandular morphology, as well as extrinsic features related to reduced shedding, including a reactive microenvironment and B cell immunity. Our analysis on tumor clonal architecture revealed that subclonal mutations were more frequently detected than the clonal ones in plasma samples with low ctDNA levels or from early-stage patients, which strongly suggests that dissemination from early-stage primary tumors mostly derived from subclones. In the longitudinal plasma, we observed that subclones seeded metastasis years before imaging diagnosis. This study with a unique large cohort of paired tumor and plasma sequencing data provided a comprehensive insight on ctDNA dynamics, disease monitoring and early detection in pancreatic cancer. Citation Format: Yuanchang Fang, Michelle Chan-Seng-Yue, Karen Ng, Amy Zhang, Tuan Hoang, Gun Ho Jang, Sabiq Chaudhary, Catia Gaspar, Eugenia Flores-Figueroa, Daniela Bevacqua, Stephanie Ramotar, Ayelet Borgida, Shawn Hutchinson, Anna Dodd, Barbara Grünwald, Julie Wilson, Robert Grant, Erica Tsang, George Zogopoulos, Masoom Haider, Jennifer Knox, Steven Gallinger, Faiyaz Notta. Pervasive early dissemination in pancreatic cancer uncovered by tissue-paired plasma whole-genomes [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1126.

765 Background: In resected pancreatic cancer (PDA), morphology from routine histopathology slides provides a rapid inexpensive biomarker that predicts overall survival and correlates with transcriptomic subtypes. In this study, we evaluated clinical and genomic associations of morphological subtypes in resected and advanced disease and validated the consistency of subtypes in patient-derived organoids (PDO) and mouse xenografts (PDXs) during in vitro and in vivo modeling. Methods: Our cohort included PDA tumor tissues from 152 resectable (stage I/II) and 228 advanced cases (Stage III/IV). Hematoxylin and eosin-stained slides were blindly reviewed by two pathologists and classified into subtypes based on Kalimuthu (1). Morphological subtypes were correlated with clinical and genomic data from whole-genome and transcriptome sequencing. Histological preparations from PDOs and PDXs obtained from pancreatic resections and metastases were reviewed using the same criteria. Results: Morphological subtypes were significantly associated with clinical patterns. Locally advanced PDA exhibited the highest proportion of glandular tumors. Metastatic tumors were enriched for non-glandular morphologies. The non-glandular morphologies were significantly associated with lower survival rates in both resected and advanced tumors. Furthermore, morphological subtypes were significantly associated with unique genomic alterations. Compared to glandular tumors, non-glandular tumors were associated with increased KRAS copy number, KRAS imbalances, and polyploid genomes. In advanced settings, glandular and non-glandular tumors mostly exhibited classical and basal-like transcriptional subtypes, respectively. Squamous tumors had the highest mutation burden and the highest proportion of Basal A signature (2). Using differential gene expression, we identified transcriptional signatures of the morphological subtypes. PDOs and PDXs maintained the morphological subtypes, although non-glandular tumors had a lower success rate for PDO establishment. Conclusions: Morphological subtypes have distinct clinical and genomic associations across all stages of pancreatic cancer, which can be consistently modeled both in vitro and in vivo . These results demonstrate that morphological subtyping offers a rapid and biologically-relevant classification of PDA that could be used for drug development and stratification to predict therapy selection. 1. Gut; 69:317-328 (2020). 2. Nat Gen; 52:231-240 (2020).

Accurate predictions of future cancer risk can increase early detection through selecting high risk individuals for screening. Existing risk prediction tools have limited predictive performance and are underutilized due to workflow disruption and reliance on structured data. We investigated whether large language models (LLMs) can predict cancer risk directly from routine free-text clinical notes recorded by primary care physicians. The dataset used for this study was from individuals aged over 18 years living in Ontario (2010–2016), obtained from ICES. ICES is an independent, non-profit research institute whose legal status under Ontario’s health information privacy law allows it to collect and analyze health care and demographic data, without consent, for health system evaluation and improvement. The development dataset consisted of 1,080 individuals selected from a cohort of 109,378 patients in Southern Ontario using stratified sampling. The testing dataset included 1,080 individuals from 135,894 patients in Toronto. A pipeline was developed using source-available LLMs to estimate cancer risk. Prompt and sampling methods were optimized on the development dataset. Prompts were designed to elicit probabilistic cancer risk estimates from progress notes. The study employed a fixed exclusion window of 365 days prior to cancer diagnosis and evaluated performance across lookahead windows of up to ten years. Lung cancer diagnosis was used as a case example to optimize the system, which was tested against the current screening eligibility guidelines based on age and pack-years of smoking. We evaluated generalizability to other geographic regions (1,080 individuals from 16,130 patients in Northern Ontario) and adapted the prompting pipeline for other cancer types (200 individuals from Southern Ontario). The best-performing model was a Mistral Small model using a multiple prompting strategy (chain-of-thought, emotion, persona). The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.787 (95% CI 0.749 - 0.824) in the testing dataset in Toronto for lung cancer prediction over a 10 year lookahead period. At the equivalent specificity to current lung cancer screening eligibility (0.941, 95% CI 0.920-0.961), predictions from the system had increased sensitivity (0.421 [95% CI 0.362-0.479] compared to 0.207 [95% CI 0.158-0.254], p<0.05). The system generalized to Northern Ontario (AUROC 0.807, 95% CI 0.771-0.844). When extended to other cancer types, the system achieved AUROC of 0.780 (95% CI 0.702–0.841) for pancreatic cancer and 0.729 (95% CI: 0.728–0.730) for prediction of any cancer diagnosis. LLMs can predict future cancer risk directly from routine clinical progress notes, offering a scalable, privacy-preserving, and generalizable alternative to structured-data-based models. This approach could be readily integrated into clinical workflows to improve cancer screening and early detection. Daniel Mau, Karl Everett, Ning Liu, Jason Chai-Onn, Liisa Jaakkimainen, Anna Dodd, Spring Holter, Steven Gallinger, Rahul G. Krishnan, Kelvin Chan, Robert Grant. Large language models to predict cancer risk from free-text clinical notes [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B015.

Figure S1. hENT1 expression by modified Moffitt subtypes Figure S2. OS in GnP cohort stratified by hENT1 and modified Moffitt Classifier