UDEC-MO: an uncertainty-guided deep embedded clustering framework for bulk and single-cell multi-omics data

Bulk and single-cell multi-omics technologies provide complementary molecular views for characterizing disease heterogeneity and cellular diversity. However, robust multi-omics clustering remains challenging due to high dimensionality, pervasive noise, modality heterogeneity, and substantial reliability variation across features, modalities, and samples or cells. Existing clustering methods often insufficiently account for such multilevel data quality variation. Here, we present UDEC-MO, an Uncertainty-guided Deep Embedded Clustering framework for robust Multi-Omics clustering. UDEC-MO first estimates feature-wise heteroscedastic uncertainty through uncertainty-aware reconstruction and summarizes it into modality-level and instance-level uncertainty scores. The instance-level uncertainty is further transformed into reliability weights to modulate the Kullback-Leibler-divergence loss in deep embedded clustering, allowing reliable samples or cells to guide cluster refinement while reducing the influence of highly uncertain instances. We evaluated UDEC-MO on both bulk cancer multi-omics datasets and single-cell multi-omics datasets generated by different sequencing technologies. The results demonstrate that UDEC-MO achieves competitive or superior clustering performance across multiple metrics and provides uncertainty-derived reliability indicators that may offer auxiliary information for characterizing potentially unreliable features, less reliable modalities, and ambiguous instances.