Abstract A106: Virtual Stromal Staining of Routine Histology as an Exploratory, Stroma-Aware Prognostic Readout in Pancreatic Cancer

Abstract

Introduction. Pancreatic ductal adenocarcinoma (PDAC) is defined by a dense desmoplastic stroma central to its biology, yet clinical staging discriminates survival weakly. For over a decade the stroma was pursued as a therapeutic target; yet stroma-directed strategies, have shown mixed results. These challenges recast the stroma as heterogeneous, both tumor-restraining and tumor-promoting, and exposed a gap: no reliable way to read prognostic stromal signals from routine material. We asked whether stromal information in H&E whole-slide images, made explicit by virtual staining, could serve as an interpretable prognostic readout. Methods. From each digitized H&E whole-slide image we generated virtual stains spanning stromal (trichrome), epithelial (pan-cytokeratin), immune (CD45), and molecular (p53) compartments. Each virtual stain was synthesized from the H&E at the same spatial coordinates, producing a pixel-aligned representation of stromal, epithelial, immune, or molecular-associated morphology without additional tissue sections or chemical staining. Tiles were encoded with a pathology foundation model adapted per stain, with auxiliary molecular supervision, and aggregated into patient-level risk by cross-attention multiple-instance learning under a Cox objective. Treating H&E as one channel among the virtual stains, we compared stain subsets and evaluated image-derived risk against clinical variables on prespecified endpoints, using patient-level 5-fold cross-validation and a 15-seed ensemble in TCGA-PAAD (181 patients; 99 OSS and 79 DSS events). Results. Adding the virtual stromal (trichrome) channel to H&E produced the best discrimination (out-of-fold C-index 0.674; H&E alone ∼0.63); the stromal channel, rather than the epithelial, immune, or molecular channels, provided the strongest incremental value, consistent with stroma as PDAC's dominant prognostic compartment. Image-derived risk exceeded clinical staging (0.673 vs 0.565) and provided largely complementary information. Image-only performance was at or above published TCGA-PAAD benchmarks (staging ∼0.60; H&E-only ∼0.58–0.62; histology-plus-genomics 0.653) without sequencing. A median risk split separated OSS and DSS (OS hazard ratio 2.42, log-rank p=1.6×10−5; DSS 2.34), with signal strongest in resectable subgroups (node-negative OS C-index ∼0.76; T1–2 OS C-index ∼0.77). Conclusions. In this exploratory single-cohort retrospective analysis, a virtual stromal stain from routine H&E behaved as an interpretable, stroma-aware prognostic readout, adding information beyond clinical staging and reaching multimodal-range discrimination without sequencing, raising the hypothesis that virtual staining surfaces molecular-associated signal latent in H&E. Findings are prognostic associations, not treatment benefit, from one public cohort with normal-tissue admixture, and require external validation with tumor-region curation. If validated, it could provide the patient-level stromal characterization that stroma-directed programs lacked.Generative AI was used to assist in drafting this abstract. Citation Format: Megan Rothney, David Sriker, Yair Rivenson. Virtual Stromal Staining of Routine Histology as an Exploratory, Stroma-Aware Prognostic Readout in Pancreatic Cancer [abstract]. In: Proceedings of the AACR Conference on Pancreatic Cancer: New Frontiers in Biology and Therapeutic Development; 2026 Sep 25-28; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(18_Suppl_2): nr A106.