Abstract Thousands of patients have benefited from the growing use of cancer immunotherapies. However, the success of these therapies can be highly variable. Strikingly, some murine tumor models show large variability in the outcome of cancer immunotherapies, even when the mice, tumor cells and anti-tumor immune cells injected into mice are all genetically identical. Here, we sought to analyze this variability in adoptive cell therapies, in order to identify the immune population driving this variability. To search for the occurrence of large variability ex vivo, we extracted mouse TCR-transgenic CD8+ T cells, and co-cultured them with antigen-expressing tumor cells. By using multiplexed in vitro assays and single-cell analysis of thousands of samples, we identified conditions (e.g. cell numbers, antigen quality, tumor cell types) where large variations in the immune activation against cancer cells is observed even between technical replicates. We then developed a quantitative framework that uses statistical modeling and machine learning to extract useful information from this immunotherapeutic variability. Our framework allowed the prediction and identification of a rare population of naïve CD8+ T cells (“Spark T cells”) that is necessary and sufficient to spark massive anti-tumor immune reactions. We are currently performing experiments to test the efficacy and functional significance of the identified immune population in vivo. We are also applying this framework in human TCR-engineered T cell blasts to identify the equivalent of the Spark T cells in humans. We envision this framework being applied to identify other relevant immune cell types that act as catalysts for successful cancer immunotherapies. Supported by the Intramural Research Program of the National Institutes of Health, the National Cancer Institute