Evaluating LongNet Spatial Modeling of Lung Adenocarcinoma Mutations from Prov-GigaPath Foundation-Model Slide Embeddings

In lung adenocarcinoma (LUAD), identifying actionable driver mutations guides therapy, but next-generation sequencing is costly, slow, and unevenly available. Predicting mutation status from routine H\&E whole-slide images (WSIs) is a proposed low-cost complement, and pathology foundation models with spatially aware slide encoders can use tissue organization. We test this by pairing frozen Prov-GigaPath embeddings with lightweight classical classifiers to predict TP53, PIK3CA, and NTRK3 status from TCGA-LUAD WSIs on a unified 335-patient cohort. We isolate spatial context with a coordinate-shuffle ablation that permutes inter-tile positions while preserving tile morphology. Only the common driver TP53 scores above a blank size-and-stain control (LongNet AUC 0.757 [0.700--0.807] vs.\ 0.520, $p < 0.001$); the rare actionable oncogenes PIK3CA (0.512) and NTRK3 (0.469) score below their controls, so no predictive information for them is present at this cohort size. Shuffling tile coordinates does not change any gene's AUC (median $|\Delta\mathrm{AUC}| \le 0.05$, all adjusted $p \ge 0.9$), and permutation-invariant MIL baselines exceed LongNet on TP53, so the predictive information is tile-local rather than spatial. We report where foundation-model embeddings work, why, and where they do not, as a calibration point for image-based mutation prediction.