Abstract
BACKGROUND: Cancer progression involves distinct stages, with a critical tipping point marking the transition from early to advanced phases, driven by complex tumor-immune dynamics. While immunotherapy has significantly improved outcomes, current biomarker models lack integration of cancer-immune interactions and progression dynamics. Leveraging advances in machine learning, there is an urgent need for a comprehensive framework to systematically analyze these dynamics, predict immunotherapy responses, and improve patient outcomes. METHODS: We developed ImmuProgML framework by integrating multi-omics data and dynamic network biomarker (DNB) analysis to identify key pathways and critical stages in cancer progression, tested in melanoma and non-small cell lung cancer (NSCLC). We introduced the DNEX score, which combines expression changes with immunotherapy-driven network topologies, and employed machine learning algorithms for prognostic and immunotherapy response predictions. We utilized molecular docking to identify potential therapeutic targets and drug candidates. RESULTS: ImmuProgML pinpointed tipping points at stage III for melanoma and stage II for NSCLC, characterized by accelerated disease progression, significant survival differences, heightened DNA damage repair mechanisms, and enhanced immune responses, with lymph nodes as pivotal hubs. By introducing the DNEX score, an integrative metric combining differential expression and network analysis, ImmuProgML evaluated gene immunomodulation activity during tumor progression and identified immunotherapy targets. High DNEX score correlated with immune-related pathways, including T cell activation and PD1 signaling, in melanoma and NSCLC. Using DNEX score, 62 machine learning models were integrated to create DNEX-SM, which predicted immunotherapy prognosis in melanoma with a C-index of 0.69, a perfect 3-year survival AUC of 1.0 in the GSE78220 dataset, and an AUC of 0.94 in the VanAllen_Science_2015 dataset, outperforming 35 published signatures. DNEX-RM, another immunotherapy response classifier within ImmuProgML, achieved an F1 score of 81.91% and AUCs of 0.912 in training, 0.877 in cross-validation, and 0.749 in testing, with an average AUC improvement of 0.053 across three datasets compared to other methods. Furthermore, DNEX ranking and molecular docking analysis identified four potent protein-drug pairs with strong binding affinities and unique binding pockets: CXCR4 with PIK-93, LCK with PAC-1, PRKCB with SNX-2112, and PRKCB with PIK-93. CONCLUSIONS: ImmuProgML offers a promising avenue for understanding the intricate relationship between tumors and the immune system, providing a machine learning framework for personalized cancer immunotherapy selections.