30Meta-analysis of single-cell RNA sequencing of tumor-infiltrating T-cells in clear cell renal cell carcinoma reveals distinct phenotypes and ­prediction markers of expanded T-cell clonotypes

Background The recent clinical success of tumor-infiltrating lymphocyte (TIL) therapy in immune checkpoint inhibitor-refractory melanoma has renewed interest in applying this approach to other malignancies. Renal cell carcinoma (RCC) is characterized by prominent CD8+ T-cell infiltration and elevated expression of cytolytic genes, suggesting the presence of tumor-reactive T-cells that could be harnessed for personalized cell therapy. TIL products are polyclonal and may recognize a diverse repertoire of tumor antigens, which could reduce the likelihood of immune escape through antigen loss. Despite this potential, past efforts to implement TIL therapy in RCC have been largely unsuccessful, likely due to the dominance of TILs exhibiting an exhausted phenotype, and the inability to effectively rejuvenate these cells during ex vivo expansion. Thus, the development of strategies that can selectively isolate and expand tumor-specific T-cells with minimal off-target toxicity and durable cytotoxic function will likely be key to the successful development of TIL therapy for RCC. We recently demonstrated that highly expanded clonal CD8+ T-cell populations in RCC tumors exhibit an antigen-experienced phenotype, indicative of chronic stimulation and consistent with tumor antigen recognition. These clonotypes represent a key feature of RCC TIL populations. Based on this, we hypothesize that the selection of non-exhausted, clonally expanded T-cells using specific surface markers will enhance the therapeutic potency and specificity of RCC-derived TIL products. Methods To identify surface markers that distinguish non-exhausted, clonally expanded tumor-reactive T-cells, we compiled a comprehensive single-cell RNA sequencing (scRNAseq) and paired TCR sequencing dataset. This includes tumor, normal adjacent tissue (NAT), peripheral blood mononuclear cells (PBMC), and lymph nodes (LN) from 23 clear cell RCC (ccRCC) patients, combining seven published and in-house studies. We are applying both differential gene expression (DGE) analysis and multi-instance-based machine learning (ML) to identify candidate surface markers linked to non-exhausted, clonally expanded phenotypes. These models are designed to accommodate the inherent noise and sparsity of single-cell data and to uncover combinatorial marker sets. Results In RCC tumors, we observed significant clonal expansion of T-cells with a memory-like, antigen-experienced phenotype. Tumor samples showed significantly higher T-cell clonality compared to matched normal tissues. Cells sharing the same TCR clonotype displayed heterogeneous transcriptional states, including both exhausted and non-exhausted phenotypes. Markers of tumor reactivity identified for other cancers (eg, CD39, CD103) labeled mainly exhausted subsets of expanded clonotypes, suggesting a need for improved marker combinations for RCC. Clonality tended to decline with increasing tumor stage, possibly indicating a link between exhaustion and tumor progression. DGE analysis revealed several surface markers associated with expanded TCR clonotypes, for which commercial antibodies are available for flow sorting. To overcome limitations of DGE alone, we are applying ML models that integrate gene expression, TCR identity, and clonal frequency to predict tumor reactivity. We successfully trained and validated multiple machine learning models, including scFormer (transformer-based), attention_MIL (attention-based deep multi-instance learning), miBoost (tree-based), and miSVM (support vector classifier), on sparse and heterogeneous scRNAseq data using a local GPU platform. All models showed consistent improvement in validation accuracy during training. Feature analysis revealed gene sets predictive of large versus small clonotypes, offering candidate markers for isolating tumor-reactive, non-exhausted T cells for downstream applications. Conclusions T-cells sharing the same T-cell receptor can display both terminally exhausted and progenitor exhausted states. Selecting clonally expanded, non-exhausted T cells may enhance the tumor reactivity of TIL products. Our ongoing analyses, including model hyperparameter optimization and gene pathway comparison, focus on identifying surface markers that are antibody-compatible, broadly expressed across patients, and applicable to CD8+ and CD4+ T-cells while excluding Tregs. These insights will support a refined TIL manufacturing strategy for improved RCC therapy.