BACKGROUND: Cancer treatments often target specific molecular pathways but are typically approved for anatomically defined indications. Large observational studies are urgently needed to better understand response patterns in clinical practice. When collected systematically at a large scale, these data hold promise to identify exceptional responders, accelerate learning about relevant pathways, and support subsequent approvals for new indications. Sharing of such datasets for real-world evidence (RWE) analyses continues to face serious barriers because of privacy and security issues. PROBLEM STATEMENT: Data privacy concerns hamper efforts to share real-world data on a large scale. Targeted analyses of smaller data collections by individual organizations may, however, have insufficient power to detect heterogeneous responses reliably. PROPOSED SOLUTION: We show in this proof-of-concept work that homomorphic encryption can balance data privacy concerns with utility when implemented in a practical version that focuses on simple arithmetic operations. This approach allows researchers to encrypt all identifiers and quasi identifiers and perform analyses on encrypted data. We use the PACE Continuous Innovation Indicators (freely available at http://scoringprogress.com) and published literature to construct a simulated population of patients receiving one or more hypothetical cancer treatments for several cancer diagnoses. We demonstrate that analysis of all prescribing in the US for these treatments in common cancers between 2004-2016 would have been feasible on a standard desktop computer. Calculating or updating a vector of 100 simulated encrypted patient-associated values, for example, takes approximately 16 seconds for one individual and scales linearly with increasing sample size. PRACTICAL IMPLICATIONS: The system hosting the homomorphically encrypted data would perform the analyses and control the number of performed tests while not having any knowledge about the raw data except its structure. A systematic implementation of such a system could identify exceptional responders to treatments in very large datasets and accelerate learning without jeopardizing patient privacy.