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

Topology-guided dual-masked contrastive learning achieves accurate cell-type annotation with minimal labels in single-cell multi-omics.

• Why it matters: Accurate cell-type annotation is hindered by feature sparsity, cross-modality heterogeneity, and limited labeled cells, which restrict current methods' effectiveness.
• What they did: The authors developed DMCL, a framework that constructs modality-specific graphs, applies dual masking, and uses a graph autoencoder with self-attention and contrastive objectives to learn robust, integrated representations.
• The result: DMCL demonstrates superior clustering and annotation accuracy on diverse datasets with minimal supervision, providing a scalable approach for weakly supervised multimodal single-cell analysis.