Review of Spatial AI combining multi-omics with deep learning for cancer biomarkers

Immune checkpoint blockade has transformed cancer therapy, achieving lasting responses in some patients, yet most still encounter primary or acquired resistance. Recent evidence demonstrates that this resistance is driven not only by intrinsic cellular features but also by the spatial organization of the tumor microenvironment (TME), including physical barriers, localized immunosuppressive niches, and organized immune cell aggregates that collectively regulate anti-tumor immunity. This review synthesizes advances in Spatial AI, combining high-resolution spatial multi-omics with deep learning approaches, particularly graph neural networks (GNNs), to elucidate the topological mechanisms of immune evasion and inform therapeutic development. Technological platforms enabling spatial molecular mapping, tools for multi-modal alignment and normalization, and computational frameworks for graph-based TME representation are covered. We define spatial phenotypes associated with immune resistance, such as immune exclusion, dysfunctional inflamed regions, and maturation states of tertiary lymphoid structures, and demonstrate how Spatial AI generates interpretable topological biomarkers that surpass conventional assays. The discussion addresses translational pathways for spatial biomarker validation and highlights key obstacles, including data standardization, computational scalability, explainability, and regulatory approval. Ultimately, immune evasion is a topological challenge, and Spatial AI offers a robust computational solution to translate complex spatial data into actionable clinical strategies to overcome architectural resistance in cancer immunotherapy.