Cytokines are signalling proteins produced during infection and control immune responses central to antimicrobial immunity via binding and activating cell surface receptors. However, their current therapeutic development faces substantial limitations including incomplete efficacy, off-target effects, and unpredictable immunogenicity. Most research focuses exclusively on human and mouse systems, overlooking hundreds of millions of years of vertebrate immune evolution that may reveal functional constraints and alternative signalling mechanisms critical for drug design.    We present a multidisciplinary framework integrating evolutionary biology and artificial intelligence (AI), exploiting deep learning models informed by phylogeny to uncover novel biology and druggability of cytokine systems. This involves compiling cytokine-receptor sequences across diverse vertebrate taxa, providing unprecedented training datasets for deep learning architectures including protein language models, graph neural networks, and generative models. Unlike general-purpose models, this approach explicitly incorporates evolutionary constraints like selection pressure analysis and coevolutionary patterns between cytokines and their cognate receptors to improve prediction accuracy for therapeutically relevant properties.    We provide a proof-of-principle by focusing on one cytokine system, that of type II interferon (IFN), IFN-gamma (g). IFNg is an important antimicrobial and immunoregulatory cytokine involved in infection, cancer and inflammation. Already IFNg-targeting interventions exist, although with limiting effectiveness. By reconstructing the co-evolution of IFNg and its two receptors, and integrating with AI models, we identify novel functional changes in IFNg sequence and structure, validated experimentally in a virus infection model. As such, we propose this methodology may offer a generalisable strategy for AI-enabled understanding of complex biomolecular systems relevant for diverse therapeutic development.