Topology-guided dual-masked contrastive learning for weakly supervised cell-type annotation in single-cell multi-omics.

Single-cell multi-omics technologies provide complementary views of cellular states, but are challenged by pervasive feature sparsity, pronounced cross-modality heterogeneity, and a critical shortage of labeled cells, hindering accurate cell-type annotation. We propose dual-masked contrastive learning (DMCL), a topology-guided weakly supervised cross-modal representation learning framework that addresses these challenges within a unified architecture. DMCL first constructs modality-specific K-nearest neighbor graphs to capture cell-cell topological relationships within each omics layer. A dual masking strategy combining global column-wise masking and local random masking is then applied to simulate realistic feature missingness and regularize the model against technical noise. Topology-aware latent representations are subsequently derived through a graph autoencoder that faithfully preserves the local cellular neighborhood structure of each modality. Cross-modal integration is achieved via a multi-head self-attention module that dynamically weights modality-specific features according to per-cell informativeness, enabling context-adaptive modality fusion. To further improve representation quality, DMCL incorporates a dual contrastive objective that jointly enforces intra-cell view consistency and inter-cell neighborhood structure, thereby enhancing both cross-modality alignment and structural preservation in the latent space. Extensive experiments on diverse real and simulated datasets demonstrate that DMCL achieves accurate cell-type annotation and superior clustering performance under minimal label supervision, establishing a principled and scalable foundation for weakly supervised single-cell multimodal analysis.