Identifying multi-omics spatial domains via masked autoencoder and diffusion-based latent regularization with SpaMEDM.

SpaMEDM identifies spatial multi-omics domains by learning graph embeddings that integrate modalities despite sparsity and noise, using masked reconstruction plus diffusion-based latent noise prediction during training.

• Why it matters: Spatial multi-omics integration is hard because spot features are sparse, technical noise is present, and each modality has distinct distributions.
• What they did: SpaMEDM trains a dual-attention graph backbone with a masked track that reconstructs randomly masked spot features from modality- and graph-specific neighborhoods, alongside a diffusion-style objective that predicts injected Gaussian latent noise at sampled noise levels.
• The result: Across diverse spatial multi-omics datasets, SpaMEDM achieves competitive spatial-domain identification and yields integrated embeddings that support downstream exploratory analyses.

AI summary of the abstract · may contain errors · not medical advice · read the paper.