Cancer dependency maps are pivotal for identifying genes crucial to cancer cell proliferation, laying the groundwork for targeted treatment strategies. Despite the preservation of core biological processes, significant distribution discrepancies between cancer cell line (CCL) models and patient-derived data challenge the direct application of CCL findings to clinical practices. To address this, we introduce a machine learning algorithm utilizing deep unsupervised domain adaptation with CORAL loss, statistically align feature distributions between distinct data domains. Trained on labeled CCL data and validated against unseen CCL and unlabeled patient data from The Cancer Genome Atlas (TCGA), our model predicts patient dependency maps with better accuracy than compared baselines. This unsupervised approach not only precisely predicts cancer dependency for patient-derived tumors but also informs the promising advancement of out-of-distribution generalization in therapeutic interventions.