T-cells are central to adaptive immunity, yet predicting which antigens they recognize and how strongly they respond remains difficult. This difficulty stems from the complexity of molecular recognition—governed by the 3D complementarity of immune receptors and antigens—and from the signaling programs that translate recognition into robust and sustained responses. In this talk, I will first introduce physically motivated machine learning methods that learn protein and immune receptor representations. The implicit physical reasoning in these models improves generalization to unseen data, enabling accurate prediction of antigen-induced T-cell activity, rational design of immunogenic antigens, and quantitative assessment of recognition specificity. Next, I will introduce a theoretical framework that views T-cell responses as feedback-controlled programs, integrating antigenic, pro-inflammatory, and anti-inflammatory signals to regulate activation, proliferation, differentiation, and death. Exploring alternative controller designs reveals an inherent trade-off between efficient pathogen clearance and immunopathology. I will discuss how these principles extend to cancer immunotherapy, suggesting signal-processing modifications that could improve efficacy while limiting toxicity. Together, these approaches reveal design principles of T-cell response and identify new levers of control for engineering safer and more potent T-cell therapies.