OmiDos

Integrating single-cell multi-omics data is crucial for understanding the complex regulatory mechanisms governing cellular functions. However, a significant challenge arises from the concurrent existence of shared mechanisms across diverse cellular states and condition-specific features, particularly within dynamic and heterogeneous microenvironments. Most existing integration approaches struggle to effectively distinguish between shared and modality-specific features, leading to misinterpretation of regulatory dynamics. Here, we introduce OmiDos, a novel dynamic orthogonal deep generative model, specifically designed to disentangle shared and unique molecular signatures across multi-omics layers in single-cell multi-omics data. Unlike prior methods that rely on simplistic assumptions about shared features, OmiDos leverages deep transfer learning to extract invariant shared signals from paired data while enforcing orthogonality constraints to separate modality-specific signals into distinct latent spaces. To address unpaired data, OmiDos incorporates an adversarial discriminator combined with noise contrastive estimation. Furthermore, OmiDos features a decoder layer enhanced by a maximum mean discrepancy regularization, enabling robust batch correction across diverse experimental conditions. We benchmark OmiDos against a wide range of multi-omics datasets, demonstrating superior accuracy and robustness in integrating complex, high-dimensional data. In a mouse secondary palate development dataset, OmiDos precisely identified a critical enhancer at the genomic locus chr16: 32697085-32697585, elucidating its essential role in regulating \emph{Muc4} expression. Its activity strongly correlates with the regulatory dynamics observed in E13.5 and E14.5 epithelial cell differentiation and migration. When applied to a medulloblastoma dataset, OmiDos revealed a dynamic regulatory transition, characterized by a shift from predominant reliance on distal regulatory elements during early developmental stages to a marked preference for proximal regulatory elements in advanced tumor stages.单细胞多组学（single-cell multi-omics）数据的整合，对于解析调控细胞功能的复杂分子机制至关重要。然而，当前该领域面临一项重大挑战：不同细胞状态间共享的调控机制与条件特异性特征同时存在，在动态且异质性的微环境中这一问题尤为突出。多数现有整合方法难以有效区分共享特征与组学模态特异性特征，进而导致对调控动态过程的误判。本研究提出OmiDos——一种全新的动态正交深度生成模型，专为解耦单细胞多组学数据中不同组学层级间的共享与独有分子特征而构建。与此前依赖共享特征简化假设的方法不同，OmiDos借助深度迁移学习从配对数据中提取不变的共享信号，并通过正交性约束将组学模态特异性信号分离至独立的隐空间中。为处理非配对数据场景，OmiDos集成了对抗判别器与噪声对比估计（noise contrastive estimation）模块。此外，OmiDos采用经最大均值差异（maximum mean discrepancy，MMD）正则化增强的解码器层，可在多样实验条件下实现稳健的批次校正。本研究通过多组学数据集基准测试对OmiDos进行评估，结果显示其在整合复杂高维数据时具备更优异的准确性与稳健性。在小鼠继发性腭发育数据集的分析中，OmiDos精准定位到基因组位点chr16: 32697085-32697585处的关键增强子，并阐明其在调控Muc4基因表达中的核心作用。该增强子的活性与E13.5及E14.5时期上皮细胞分化和迁移过程中的调控动态变化显著相关。将OmiDos应用于髓母细胞瘤数据集时，其揭示了一种动态调控转变过程：发育早期主要依赖远端调控元件，而在肿瘤进展晚期则显著偏好近端调控元件。