scURL: Uncertainty-Sharpened Representation Learning for Single-cell Multi-omics Clustering

Single-cell multi-omics clustering requires integrating complementary omics layers while accounting for their unequal reliability across individual cells. However, cell-wise reliability can vary substantially within each omics layer, causing different omics to provide reliable signals for some cells but ambiguous or noisy signals for others. This challenge is further complicated by omics-level heterogeneity, where differences in sparsity, signal distribution, and biological resolution may introduce cross-omics conflicts and bias the fused representation.

We present scURL, an uncertainty-sharpened representation learning framework for single-cell multi-omics clustering. scURL introduces representation uncertainty (RU) and establishes its connection to clustering generalization risk (GR), enabling the derivation of cell-wise omics weights for uncertainty-aware fusion (UAF). To support stable uncertainty estimation, scURL further introduces a multi-granular calibration module (MCM) that calibrates omics-specific representations from both cluster-level semantic and local-level structural perspectives before fusion. Experiments on ten datasets demonstrate that scURL outperforms existing clustering methods, remains robust under dropout noise, and identifies biologically meaningful cell populations, as supported by marker gene analysis.

Source code is available at https://github.com/Yaolab-fantastic/scURL.

Supplementary data are available at Bioinformatics online.