ATLAS: Adaptive Temporal Learning for Single-Cell Multi-Omics Alignment and Dynamics

Single-cell multi-omics technologies provide a powerful basis for characterizing cell-state transitions and dynamic regulatory processes. However, existing methods for multi-omics dynamic modeling still face two major limitations. From the data perspective, many methods require high-quality paired multi-omics measurements; from the modeling perspective, dynamic inference often relies on predefined regulatory structures or kinetic assumptions. These limitations restrict their applicability to partially paired, unpaired, and broader cross-omics settings. To address these challenges, we propose ATLAS, a unified framework for single-cell multi-omics alignment and dynamic modeling that jointly learns cross-omics consistency and temporal dynamics from partially paired or unpaired data. ATLAS adaptively models temporal lag effects between omics layers to characterize asynchronous cross-omics regulation, and further introduces a reliability-guided temporal distillation strategy to improve model-based temporal ordering. We systematically evaluate ATLAS on five datasets across four tasks, showing strong overall performance in multi-omics alignment, cross-omics prediction, trajectory inference, and future-state prediction. Our code is available at https://github.com/anomity/ATLAS.