scPhoenix: A Contrastive Learning-based Framework with Aux-Core Feature Disentanglement Enhances Sparse Cellular Multi-omics Translation

Abstract

Recent advances in single-cell multi-omics co-assays and spatiotemporal sequencing technologies have provided unprecedented opportunities for systematically characterizing cellular heterogeneity. However, severe sparsity and pronounced spatial heterogeneity--hallmark features of complex diseases and tumor microenvironment--remain major obstacles in deciphering cellular multi-omics data. Here, we present scPhoenix, a contrastive learning-based framework for single-cell cross-modality translation. scPhoenix adopts a two-stage Aux-Core strategy to disentangle modality-specific feature extraction from cross-modality feature interaction. Across diverse datasets, it preserves cellular heterogeneity during translation and demonstrates significant advantages for data with high sparsity. In addition, the framework integrates a contrastive learning framework with five effective data augmentation methods tailored to single-cell data. Moreover, scPhoenixs design supports extensions to unpaired data training and spatial multi-omics translation, enabling robust performance in scenarios with high spatial heterogeneity. scPhoenix is freely available at https://github.com/liwz-lab/scPhoenix.