Decoding immunotherapy responders in hepatocellular carcinoma patients with stable disease: Multi-omics insights and machine learning model development.

e16188 Background: Over 40% of hepatocellular carcinoma (HCC) patients treated with immune checkpoint inhibitors (ICIs) manifest stable disease (SD). This proportion is much higher than other cancers. Heterogeneous treatment efficacy, existing both in the duration of response and depth of response, was observed in SD. Hence the definition and features of SD with long-term survival need to be urgently explored. Tumor microenvironment (TME) plays a crucial role in ICIs therapy response and can be effectively decoded with multi-omics analysis. Machine learning (ML) can effective mine high-throughput data for precise prediction. In this study, we aim to recognize the responder in SD (SDR), characterize TME features, and establish a ML model to early identify SDR, leaving time-window for the therapeutic regime adjustment of non-responder in SD (SDNR). Methods: 264 HCC patients treated with ICIs were retrospectively selected in 5 centers. 139 SD patients were identified with RECIST1.1. A subgroup of SD patients, defined by progress-free survival (PFS) and the best percentage change of tumor burden (%BOR) were identified as SDR, exhibiting a comparable overall survival (OS) with partial response (PR) patients. The remaining SD patients were classified as SDNR. Clinical (n = 139), transcriptional (n = 20), pathological (n = 22), and radiological (n = 139) data of SD were collected to decode the structural and functional features of TME. Integrating radiomic and clinical features, ML models (Logistic regression, Random forest, Support vector machine, and XGBoost) were trained and externally validated. The best-performing model was selected and interpreted biologically and mathematically by transcriptional correlation and SHAP method. Results: We identified SD with PFS > 180 days and %BOR ≤ 0 as SDR, who exhibited similar OS with PR and significantly better OS than SDNR, along with early tumor regression. Multi-omics revealed that SDR exhibited a hot TME and elevated intra-tumor tertiary lymphoid structure (iTLS) abundance. The iTLS was associated with higher CD8 T cell and cytotoxicity function, as well as higher CD20 B cell, improved B cell maturation, and higher tumor reactive IgG production in SDR. Additionally, the XGboost model with the highest Area under the Curve (AUC) of 0.96 in training set and 0.89 in external validation set was selected as the final model for SDR prediction. Conclusions: In HCC with SD to ICIs, the ML model, incorporating radiomic and clinical features, can effectively recognize the responder in SD. The SDR can retain long-term benefits from ICIs and exhibit elevated iTLS in TME, which is associated with enhanced anti-tumor cellular and humoral immunity. This study brings more attention to SD and may contribute to drawing more confident conclusion in translation research, guiding precision medicine and avoiding unnecessary intervention.