P14.16.A REINFORCEMENT LEARNING IDENTIFIES TIM-3 BLOCKADE AS A PROMISING IMMUNOTHERAPEUTIC STRATEGY IN CHORDOMA

Abstract BACKGROUND Chordoma, a rare notochord-derived bone malignancy, remains largely resistant to chemotherapy and radiation, making en-bloc surgical resection the therapeutic cornerstone. However, due to anatomically challenging locations at the skull base and sacrum, achieving complete (R0) resection is frequently impossible, resulting in local recurrence rates of 40-60% after incomplete (R1) resection. Effective systemic treatment strategies remain urgently needed. MATERIAL AND METHODS We performed comprehensive multimodal immune profiling on six newly diagnosed sacral chordomas, integrating CD45-enriched single-cell RNA sequencing (total of 85,921 cells: 39,938 from tumour tissue and 45,983 from matched peripheral blood), coupled with paired T/B-cell receptor sequencing. Spatial context was defined via 10X Visium spatial transcriptomics performed on four tumours (total 20,164 spatial spots, median 5,041 per tumour). To identify promising therapeutic targets, we employed a reinforcement-learning-driven, agent-based computational model (CellAgentChat) to simulate the effects of immune-checkpoint inhibitor perturbations. RESULTS All analysed chordomas consistently exhibited a distinct “nest-septa” architecture: physaliphorous tumour nests were largely devoid of immune cell infiltration, while surrounding fibrous septa were densely populated with immunosuppressive M2-like macrophages, DC2-type dendritic cells, and enriched in TGFB/IL-10 signalling. Despite this exclusionary environment, we identified T-cell receptor (TCR) clonotypes shared between peripheral blood and tumour-infiltrating lymphocytes. Convergent CDR3 sequence motifs suggested the presence of chordoma-specific antigens; notably, these T-cell clones adopted an exhausted phenotype (PD-1⁺TOX⁺) upon entering the tumour microenvironment. Reinforcement-learning simulations predicted the downstream transcriptional effects of immune-checkpoint inhibition, identifying TIM-3 blockade as the most promising strategy. In silico TIM-3 inhibition robustly enhanced antigen presentation, cytokine signalling, and immune-regulatory gene expression specifically within the immunosuppressive septal compartment. CONCLUSION This multimodal characterization of six sacral chordomas reveals a spatially structured immunosuppressive barrier, penetrated—but not overcome—by circulating tumour-reactive T cells. Although generalisability is constrained by cohort size and anatomical specificity, our reinforcement-learning models strongly support clinical evaluation of TIM-3 blockade, either as monotherapy or in combination with PD-1/CTLA-4 inhibitors, as a promising therapeutic avenue for sacral chordoma patients.