Stress-Testing Histology-Based KRAS Prediction in Colorectal Cancer

Gene-level discrimination from histology does not establish transfer to new specimen settings, finer genotype resolution, or faithful score reconstruction from tissue composition. We stress-test colorectal $\textit{KRAS}$ prediction in 1,239 source patients and five target settings using frozen-feature multiple-instance learning (MIL). Source UNI-v1 AUROC was 0.693; after restricting fixed scores to MSS/pMMR, $\textit{BRAF}$-WT tumors, common-support standardization retained an AUROC gain of 0.028. Source-only models reached AUROCs of 0.706 in CPTAC and 0.747 in RIH primaries, whereas metastatic discrimination remained uncertain. For molecular resolution, shared-positive controls held positive patients fixed while replacing mutant negatives with matched wild-type patients. Codon-12-versus-other-mutant and G12D-versus-other-mutant AUROCs were 0.483 and 0.495, while repeated wild-type controls ranged 0.623-0.667 and 0.632-0.679, supporting two source-population, protocol-specific limits; three within-codon tasks remained unresolved. A 32-concept abundance representation achieved $\textit{KRAS}$ AUROC 0.601 but reconstructed only part of MIL-score variance ($R^2=0.195$). These results show that validation must match the intended population, molecular endpoint, and explanatory claim.