Abstract
Lung adenocarcinoma (LUAD) remains molecularly heterogeneous, and many tumors lack clearly tractable vulnerabilities. We developed DepPrior, a computational framework that ranks candidate LUAD therapeutic targets by requiring concordant evidence of CRISPR dependency separability, molecular predictability, and cross-cohort expression/protein reproducibility. DepMap dependency scores were modeled from matched expression and copy-number features using linear and non-linear learners, and gene-level AUROC and R
2
were combined into a heuristic DepScore. The final candidate set included FERMT2, CRKL, MYC, CHMP4B and related genes. The set formed a coherent tumor expression module in TCGA-LUAD, was strongly associated with proliferation-linked features, and showed rank-based concordance across GEO transcriptomic cohorts and CPTAC transcriptomic/proteomic resources. Five-fold cross-validation supported the ranking of non-linear models, although performance gains were moderate and should be interpreted as model-ranking evidence rather than as large effect-size proof. Orthogonal experiments in HCC827 cells showed modest but reproducible protein-level reductions after FERMT2 and CRKL knockdown, accompanied by a directionally stronger apoptosis-associated protein shift after combined suppression than after single perturbation. These findings support DepPrior as a reproducibility-oriented, hypothesis-generating approach for target nomination. Because cross-cohort expression concordance does not prove patient-tumor dependency conservation, and experimental validation was restricted to selected genes and cell-line systems without rescue or proliferation/clonogenic assays, the prioritized genes should be considered candidates for further perturbation, rescue, patient-derived model, and therapeutic tractability studies.