Abstract IA004: Toward a Predictive Science of Pancreatic Cancer: AI-Guided Spatial Multi-Omics for Detection, Prognosis, and Target Discovery

Abstract

Pancreatic ductal adenocarcinoma (PDAC) remains among the most lethal solid tumors, and it is defined less by any single driver mutation than by its architecture: a dense desmoplastic stroma, a spatially organized immunosuppressive microenvironment, and profound intratumoral heterogeneity. Bulk profiling and manual histologic assessment average away exactly the information that governs therapeutic failure. My talk emphasizes how AI, applied across routine histology, spatially resolved multi-omics, and blood-based assays, can convert this architecture into clinically actionable readouts.AI-powered spatial pathology of the TME. The cancer–immune set point framework distinguishes inflamed, immune-excluded, and immune-desert tumors, but has been difficult to apply reproducibly at scale. Deep learning–based segmentation of H&E whole-slide images separates cancer epithelium from cancer-associated stroma, quantifies intratumoral and stromal lymphocyte densities, and assigns immune phenotypes automatically; in non–small-cell lung cancer these phenotypes predicted response to checkpoint blockade and complemented PD-L1 scoring. In 304 resected PDACs, the inflamed phenotype showed the longest overall and recurrence-free survival within every pathologic stage and higher CD8+ T-cell fractions and cytolytic scores — prognostic information beyond stage, from a slide already made for every patient.From cell catalogs to ecosystem maps. Transcriptional subtypes carry prognostic weight, but their behavior is dictated by the microenvironment they sit in. Integrating a PDAC single-cell atlas with imaging-based spatial transcriptomics resolves recurrent cellular neighborhoods, or spatial ecotypes, shared across patients: stromal and basal-like neighborhoods correlate with shorter survival, whereas a pre-myCAF neighborhood is associated with favorable outcome. An ANO1/LRRC15+ CAF axis marks ECM-remodeling, T-cell–excluded niches and nominates ecotype-restricted targets.Multimodal cfDNA analysis for early detection. The same logic applies upstream of treatment. Because CA19-9 lacks specificity and is uninformative in Lewis antigen–negative individuals, most patients still present with unresectable disease. Genome-wide cell-free DNA profiling combining methylation, copy-number, and fragmentation features in a machine-learning ensemble detected eight cancer types at 93.2% sensitivity and 95% specificity, with 92.3% sensitivity for stage I disease and 91.5% for pancreatic cancer. Applied to a PDAC-enriched cohort spanning stages I–IV, it detected disease at high specificity; performance in stage I–II PDAC and in surveillance of high-risk groups remains the decisive question. These efforts point toward an AI-guided spatial oncology in which detection, prognostication, and target selection are grounded in tissue and plasma architecture rather than averaged molecular summaries. Prospective validation and standardized reporting of AI-derived spatial metrics will determine whether these readouts change practice.