CIPR: Class Imbalance-Aware Pseudo-Label Regeneration for Single-Cell Multi-Omics Clustering

Single-cell multi-omics clustering has emerged as a powerful paradigm for dissecting cellular heterogeneity and advancing precision medicine. Despite recent progress, most existing methods prioritize the strict alignment of heterogeneous modalities. However, this alignment process is often dominated by majority cell populations, which inadvertently suppresses the representation of rare cell types and exacerbates the long-tail effect inherent in biological data. This imbalance distorts the latent feature space, resulting in suboptimal clustering performance. To address this, we propose CIPR, a Class Imbalance-aware Pseudo-label Regeneration framework. First, CIPR constructs a unified embedding by integrating cellular attributes with local and global structural relationships. Crucially, we formulate the multi-omics pseudo-labeling process as an imbalanced Optimal Transport problem. Instead of enforcing rigid equal-sized clusters that distort true biological proportions, we introduce a soft uniform prior to guarantee a baseline representation for all cell groups. This explicitly prevents rare cell populations from being absorbed by dominant cell types, while allowing the model to naturally adapt to the inherent imbalance of single-cell data. Furthermore, CIPR incorporates a curriculum learning strategy that incrementally learns from high-confidence samples, thereby reducing noise propagation during pseudo-label generation. Finally, a cross-omics interaction mechanism mutually guides prediction distributions, minimizing inter-modality conflicts and maximizing information complementarity. Extensive experiments on various multi-omics datasets demonstrate that CIPR significantly outperforms state-of-the-art methods, particularly in identifying rare cell populations and handling severe class imbalances.