PAN-SYSTEM CANCER INTELLIGENCE: INTEGRATING BLOOD, IMMUNE, MICROBIOME, AND TUMOR MICROENVIRONMENT DATA USING FOUNDATION MODELS

Abstract

Cancer is no longer seen as just a disease of mutated cells. Over the last decade, researchers have come to realize that tumors live inside a much larger ecosystem that includes circulating blood biomarkers, immune cell populations, the gut and tumor microbiome, and the local tumor microenvironment. Each of these layers carries useful information, but they are often studied in isolation. This paper proposes a pan-system approach to cancer intelligence that uses foundation models to integrate data across all four of these biological layers. We worked with a curated multi-omics dataset combining publicly available cohorts and synthetic clinical records, covering nearly 4,200 patient samples across five cancer types. A transformer-based foundation model was pretrained on the combined modalities and then fine-tuned for tasks like cancer subtype classification, survival risk prediction, and treatment response forecasting. Comparisons were made against single-modality baselines and traditional multi-modal fusion models. Results show that the foundation model approach improved classification accuracy by about 11 percent and lifted the concordance index in survival prediction by nearly 9 percent compared to standard methods. The microbiome layer, often overlooked, turned out to add meaningful predictive value, especially in colorectal and lung cancers. We discuss the practical implications of moving toward integrated cancer intelligence and the challenges that come with it, including data harmonization, interpretability, and clinical validation. The work points toward a future where foundation models could serve as a backbone for personalized oncology, helping clinicians look at cancer as a whole-body phenomenon rather than a localized event.