Abstract
Recent breakthroughs in spatially resolved multi-omics have unlocked the ability to simultaneously profile multiple molecular layers within tissues, offering unprecedented insights into their coordinated roles in development and disease. Despite these advancements, integrative analysis of multi-omics data remains a formidable challenge due to inherent biological and technical discrepancies across assays, often leading to gradient conflicts during joint learning. These conflicts arise as optimization trajectories from different omics compete or contradict, thereby constraining integration performance. To overcome this challenge, SpaBalance, a unified computational framework designed to harmonize cross-omics learning via gradient coordination and adaptive feature decomposition, is proposed. SpaBalance introduces a novel gradient equilibrium mechanism that dynamically balances inter-omics contributions during backpropagation, resolving conflicts through task-specific prioritization without requiring manual weighting. Concurrently, SpaBalance leverages a dual-stream architecture to simultaneously learn shared representations and preserve omics-specific features. Extensive evaluations across a variety of spatial omics datasets, including paired epigenome-transcriptome and proteome-transcriptome data from human tumors and brain tissues, demonstrate SpaBalance's superior ability to delineate complex spatial domains and uncover previously hidden multi-omics regulatory hubs, significantly improving clustering accuracy and biological interpretability. Moreover, SpaBalance flexibly scales to integrate multiple omics, bridging data integration with biological discovery and advancing spatially resolved systems biology.