MGCF-Net: a multi-view granular-ball contrastive fusion network for spatial transcriptomics clustering

Spatial domain identification from spatial transcriptomics (ST) data is essential for characterizing the spatial organization of complex tissues. Existing contrastive learning methods face a persistent trade-off: instance-level approaches are susceptible to false negatives, whereas cluster-level methods may overlook fine-grained local structures. Moreover, direct multi-view fusion may introduce redundant or noisy correlations, thereby degrading representation quality.

To address these limitations, we propose MGCF-Net, a multi-view granular-ball contrastive fusion network for spatial domain identification. MGCF-Net constructs a spatial graph and a feature graph to represent physical proximity and transcriptomic similarity, respectively. A redundancy-reduction objective preserves spot-level correspondence across the two views while suppressing spurious similarities between different spots. Multi-view granular-ball contrastive learning groups spot-level representations into local structural units and aligns them across views according to their membership overlap.

On eight dorsolateral prefrontal cortex (DLPFC) tissue sections, MGCF-Net attains the highest mean adjusted Rand index (ARI) among the evaluated methods. Its normalized mutual information (NMI) remains competitive with the comparison methods.

By contrasting local structural units rather than individual spots, MGCF-Net reduces the influence of false negatives arising from instance-level contrastive learning. This group-level formulation preserves coherent spatial structures without requiring a generative decoder.