Abstract B025: A spatial proximity-weighted algorithm (SPAN) defines reproducible tumor niches and a persister ecosystem underlying acquired resistance to pan-RAS inhibition in PDAC

Abstract

Spatial transcriptomics extends single-cell analysis of the tumor microenvironment into a mappable ecosystem by resolving gene expression in situ. Within it, cells assemble into recurrent multicellular niches whose spatial organization shapes tumor progression, immune evasion, and therapy response and resistance. Yet these niches are typically defined by arbitrary fixed-radius neighborhoods and unstable clustering, limiting reproducibility and obscuring how ecosystems remodel under therapy. To address this, we developed SPAN (Spatial Proximity-weighted Analysis of Niches), a generalizable framework that defines robust, reproducible niches from any spatial dataset. We focus on pancreatic ductal adenocarcinoma (PDAC), a stroma-rich, treatment-refractory tumor where niche architecture is thought to shape therapeutic failure. This is now especially pressing for pan-RAS/KRAS inhibitors, which are changing how PDAC is treated but face acquired resistance; how the ecosystem remodels under them remains unclear, with consequences for rational combinations and second-line therapy. We therefore applied SPAN to the autochthonous KPC model of PDAC, profiled by 10x Xenium across three states: untreated, one week of RMC-7977, and end-stage tumors that regrew despite continued therapy. For each cell, SPAN performs a multi-radius neighborhood search and builds a distance-weighted composition vector across radii, so proximal cells receive more weight. Composition matrices are partitioned by silhouette-optimized k-means, and spatial co-localization of the resulting niches is validated by neighborhood-enrichment testing. SPAN recovered reproducible niches whose composition tracked treatment state, revealing discrete, interconvertible states, not a static map. Vehicle tumors comprised epithelial and mixed epithelial/fibroblast/myeloid niches. Response abolished five baseline niches that never re-formed, including the epithelial-dominant ones, and the residual bed became fibroblast-dominated. One niche, epithelial-enriched fibroblasts, persisted across all three states, marking a treatment-persistent reservoir at the epithelial-stromal interface. Resistance was not a reversion to baseline but rather was associated with the development of a distinct ecosystem: a fibroblast/endothelial niche shared with treated tumors plus three resistance-restricted niches, including a fibroblast/myeloid/B-cell niche unique to regrowing tumors, implicating stromal-immune reorganization in relapse. A TGF-β/Notch persistence signature from the persistent niche stratified overall survival in human PDAC, linking the mouse reservoir to clinical outcome in untreated patients. Together, SPAN converts spatial data into reproducible, permutation-validated niches and reframes pan-RAS-inhibitor response and resistance as ecosystem-level state transitions. It nominates a treatment-persistent epithelial-fibroblast niche as a candidate residual-disease reservoir and a resistance-specific stromal-immune niche as spatially defined targets for combinations to deepen response and delay relapse.