Immunotherapies are among the most promising avenues for personalized and effective cancer medicines. Cellular crosstalk and appropriate signaling between the relevant immune cells and the targeted cancer cells is critical for the effectiveness of the therapy. Despite extensive research, routine clinical application of these strategies is met with hurdles in regards to immune cell efficacy, immune exhaustion, safety and cancer resistance. With the development of single-cell RNA sequencing technologies, we are now able to investigate the transcriptional changes in engineered T cells interacting with cancer cells at much higher resolution, offering new insights into the mechanisms of action of these therapies. We hypothesize that the productive interaction between cancer cell and T cell is critical for killing efficiency and could lead to a distinct transcriptional program in both the immune cell (and the target cell). Having access to such a transcriptional signature could in turn be exploited to better predict CAR or TCR sensitivity and specificity, thus improving the efficacy of cancer immunotherapies. The same concept could also be applied to study T cell exhaustion or resistance mechanisms, where this transcriptional program could likely be disturbed. In this project, we aim to measure transcriptional responses of physically interacting T cell and cancer cell pairs. Based on this data we derive a gene expression program that predicts the most effective T killer cells. The identity of this program is trained using a large language model (LLM). We then utilize our novel model and identify naturally occurring killer cells in public datasets of primary tumor-infiltrating T cells.