MELON: Multimodal Learning Framework for Spatial Multimodal Omics Data Integration

Spatial multi-omics technologies measure multiple molecular modalities on the same tissue section, but existing integration methods optimize for shared structure and rarely preserve the cross-modal synergistic signal that emerges only from joint observation across modalities. We present MELON, a representation-learning framework built around a partial-information-decomposition (PID)-guided contrastive objective that explicitly preserves redundant, unique, and synergistic cross-modal information, com- bined with a learned neighborhood-aware spatial bias that respects local tissue structure. On a controlled simulation isolating cross-modal synergy, MELON recovers a synergy-only label well above chance while seven multi-omics baselines remain at chance. On three real-data benchmarks spanning RNA–ATAC and RNA–protein settings, MELON achieves consistently higher agreement with anatomical reference labels than seven established baselines and produces more spatially contiguous domains with sharper boundaries; a trimodal tonsil extension confirms that these gains transfer beyond the bi-modal setting.