Large language model-guided CAR-T in silico platform for cytokine optimization in liver cancer with low antigen density.

Abstract

CAR-T cell therapy has shown remarkable success in hematologic malignancies but remains limited in solid tumors such as liver cancer due to antigen heterogeneity, low target antigen density, and an immunosuppressive tumor microenvironment (TME). Cytokine engineering can enhance CAR-T persistence and effector function; however, the optimal cytokine payload may vary depending on tumor type, target antigen expression level, and microenvironmental context, making systematic experimental comparison time-consuming and labor-intensive. Here, we applied a large language model (LLM)-based CAR-T in silico platform to systematically evaluate cytokine engineering strategies, including IL-2, IL-7, IL-12, IL-15, and IL-18, in glypican-3 (GPC3)-targeted CAR-T cells for liver cancer. We used cytokine selection as a biologically grounded benchmark to test whether the platform could recover known CAR-T cell-relevant cytokine biology and support future novel predictions. Computational predictions identified IL-15 as the most effective enhancer, particularly against tumor cells with low GPC3 expression. Guided by these results, we generated cytokine-armored GPC3 CAR-T cells and performed in vitro and in vivo validation. IL-15-engineered CAR-T cells exhibited superior proliferation, persistence, and serial cytotoxicity against GPC3-low liver cancer cells. In human liver cancer xenograft models, IL-15-enhanced CAR-T cells achieved improved tumor control compared with conventional and other cytokine-engineered CAR-T cells. The recovery of IL-15 served as a positive benchmark supporting the validity of the LLM-guided CAR-T in silico workflow. Collectively, this study establishes an LLM-guided framework, schema-constrained for rational cytokine selection in CAR-T engineering and identifies IL-15 as a potent enhancer for targeting antigen-low liver cancers.