Spatial multi-omics and AI for decoding tumor heterogeneity: implications for target discovery and therapy resistance

Tumor heterogeneity remains a major bottleneck for effective cancer therapy because many determinants of target expression, biomarker performance, and treatment resistance are spatially organized within the tumor ecosystem. Bulk molecular profiling averages these signals across cell populations, whereas dissociation-based single-cell approaches disrupt the tissue architecture that shapes cell state, signaling, and drug access. Spatial multi-omics addresses these limitations by mapping transcriptomic, proteomic, metabolomic, and epigenomic features in situ, while AI enables integration of these modalities with histology, spatial coordinates, and clinical outcomes. Current evidence suggests that graph-based and neighborhood-aware models are most useful for spatial domain detection and niche mapping, whereas multimodal and histology-based models require stronger external validation before clinical deployment. In this review, we examine how spatial multi-omics and AI are redefining tumor heterogeneity as a translational variable in oncology. We focus on target discovery, patient stratification, and resistance biology. We also discuss major barriers to clinical translation, including platform bias, FFPE compatibility, assay cost, molecular coverage, segmentation and registration errors, limited benchmarking, domain shift, and the need for independent multicenter validation.