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

Spatial multi-omics data provide complementary molecular views of tissue organization. However, their integration remains challenging due to data sparsity, technical noise, and modality-specific distributions. Here, we present Spatial Masked autoEncoder and Diffusion-based Model (SpaMEDM), an end-to-end graph representation learning framework that augments a dual-attention graph-integration backbone with masked reconstruction and diffusion-based latent noise-prediction regularization. In the masked track, randomly masked spot features are recovered by leveraging modality- and graph-specific neighborhood information. In parallel, Gaussian perturbations are applied to modality-specific latent representations at randomly sampled noise levels, and a spatial-graph-informed network is trained to predict the injected noise. This diffusion-based objective is used exclusively during training to encourage robustness to latent perturbations, rather than to perform iterative reverse diffusion or sampling during inference. Furthermore, cosine-based paired-spot directional alignment and adjacency reconstruction encourage compatible cross-modality representations and preservation of the selected graph structure. SpaMEDM flexibly supports purely spatial, dual feature-spatial, and modality-specific feature-pruned spatial graph configurations. Evaluations across diverse spatial multi-omics datasets demonstrate competitive spatial-domain identification performance and produce integrated embeddings that facilitate downstream exploratory analyses.