SchedOmics:Adaptive Modality Scheduling for Single-Cell Multi-Omics Integration Across Pairing Topologies

Single-cell multi-omics integration combines complementary molecular assays to characterize cellular heterogeneity beyond any single modality. However, existing methods often rely on predefined alignment and fusion policies that overlook modality-dependent learning dynamics. Under incomplete cell-level pairing and without cell-type supervision, such policies may prematurely incorporate under-optimized or unstable representations, compromising integration. To address this challenge, we propose SchedOmics, a self-supervised framework for adaptive modality scheduling across pairing topologies. First, mask-aware representation learning encodes observed assays and provides comparable representations for state estimation. Second, state-aware modality scheduling estimates remaining difficulty, stable learning progress, and temporal volatility to allocate bounded additional optimization, prioritizing reducible learning demand while discounting instability. Third, topology-aware cross-modal interaction uses these states to weight directional guidance according to target learning demand and source stability. Together, these components coordinate modality learning to exploit complementary information while limiting interference from unstable representations, enabling effective integration across pairing topologies. Across six datasets spanning three pairing settings, SchedOmics achieves consistently strong performance in label discrimination, clustering, and manifold preservation, with clear advantages in cross-assay transfer under incomplete pairing.