Computational nomination of candidate repurposing drugs for five rare cancers via GDSC-to-TCGA pharmacogenomic transfer learning

Rare cancers account for ~24% of diagnoses yet attract little therapeutic investigation. We trained ridge regression on 700 Genomics of Drug Sensitivity in Cancer (GDSC) cell lines and imputed sensitivity to 278 drugs for 401 TCGA patients across five rare cancers (mesothelioma MESO, adrenocortical carcinoma ACC, uveal melanoma UVM, cholangiocarcinoma CHOL, thymoma THYM; n=87/79/80/35/120), linking imputed sensitivity to overall survival in 369 patients with evaluable follow-up. The analysis nominated drug-survival associations in ACC (59), MESO (39), and UVM (16) ; none in CHOL or THYM. Because the imputed profiles are strongly correlated (mean pairwise r=0. 78), these counts index a ranking rather than independent discoveries; label-permutation testing confirmed each per-cancer aggregate exceeded its own correlation-preserving null (p=0. 01-0. 02), and each headline drug survived conditioning on the dominant prognostic axis. Firth-penalised per-standard-deviation hazard ratios were bounded (MESO Remodelin 1. 72, ACC Methotrexate 0. 40, UVM Serdemetan 0. 48). Repeated split-half reproducibility was strongest in ACC (mean between-half r=0. 24), modest in MESO (0. 19), and weakest in UVM (0. 14). Candidates were pathway-concordant: bromodomain/BET inhibitors in BAP1-mutant MESO, antiproliferatives in ACC, MDM2 inhibition in p53-intact UVM; SHAP recovered BCL2L1. Critically, an independent CTRPv2 cell-line screen did NOT corroborate the leading mesothelioma BET theme (one-sided Mann-Whitney p=0. 76) even though it recovered the canonical haematopoietic BET dependency (p=9x10^-51), underscoring that mechanistic plausibility does not guarantee replication. The reported associations are prognostic, not predictive — partly entangled with a generic poor-prognosis transcriptional signature — and should be read as a prioritised, mechanistically interpretable hypothesis set for experimental follow-up rather than evidence of clinical efficacy. The complete analysis pipeline is available as a self-contained Jupyter notebook with extension scripts.