Histology-derived prediction of a stromal remodeling program in pancreatic ductal adenocarcinoma with transcriptomic and spatial validation.

OBJECTIVE: To develop and validate a histology-based model for predicting a stromal stiffness-related molecular surrogate in pancreatic ductal adenocarcinoma (PDAC) and to assess its spatial biological plausibility. METHODS: A stromal stiffness score was defined from COL12A1, LOX, and LOXL2 expression. Candidate genes were derived from two independent PDAC chemotherapy-response studies, refined by Cox-LASSO regularization, and filtered using survival analysis in the TCGA-PAAD cohort. StRoMA, an encoder-decoder Transformer framework with three feature encoders, was trained on 206 H&E whole-slide images from 178 TCGA cases and externally validated on 492 slides from 153 CPTAC cases. Spatial concordance was evaluated using independent Visium HD data, with Tangram-based fibroblast mapping used for sensitivity analysis. RESULTS: In internal validation, UNI achieved the best performance, with a Spearman correlation of 0.655 and an AUC of 0.811. In CPTAC external validation, CONCH showed the strongest generalization for predefined labels, with a Spearman correlation of 0.458 and an AUC of 0.744. In spatial validation, CONCH attention was most concordant with the three-gene transcriptomic score (Spearman correlation = 0.211; permutation P = 0.002), exceeding myofibroblast and inflammatory CAF programs. Tangram-derived fibroblast mapping did not reproduce this spatial signal. CONCLUSION: Routine H&E histology encodes morphologic information related to ECM remodeling and stromal stiffness in PDAC. This three-gene histology-derived phenotype represents a candidate biomarker for future translational studies of stromal remodeling and therapeutic relevance in PDAC.