Learning Shared and Modality-Specific Structure in Spatial Multi-Omics
SpaMV disentangles shared biological structure from signals that are private to each spatial omics modality.
Interpretable Spatial Multi-Omics Integration
Different spatial omics modalities observe complementary aspects of the same tissue. Mapping all measurements into a single latent space can conflate what is shared across modalities with signals that are meaningful only within one modality.
SpaMV
SpaMV uses a multi-view representation-learning framework to separate shared latent structure from omics-specific private structure. This decomposition supports spatial domain analysis while retaining interpretable molecular programs that would otherwise be lost in a single shared embedding.
Research Role
This work forms an early foundation for my interest in multimodal biological representation learning: how models can connect modalities without erasing what makes each modality scientifically distinct.
Resources
Published in Nature Communications (2026).