PanRareWSI: Systematic UNI2 Foundation-Model Benchmark for Biomarker Prediction Across Rare TCGA Cohorts

Pre-registration, analysis code, frozen cross-validation splits, and supplementary materials for a systematic benchmark of UNI2 (ViT-H/14) frozen slide-level embeddings for molecular biomarker prediction across seven rare TCGA cohorts (ACC, UVM, MESO, CHOL, THYM, KICH, DLBC; 596 patients total). These cohorts are absent from all existing pathology foundation-model benchmarks including Patho-Bench (95 tasks, 33 datasets). The study evaluates 22 primary and 9 exploratory biomarker × cohort prediction tasks using L2-regularised linear probes and gated-attention ABMIL, with pre-registered calibration, permutation testing, and a tripartite outcome classification (recoverable / inconclusive / not recoverable).