Unified Representation Learning for Spatial Multi-Omics

Integrating spatial multi-omics data is crucial for decoding tissue function, yet existing methods struggle with noise and cross-modal alignment.

We introduce SpaAlign, a novel two-stage framework for the robust integration of multiple omics and histological data. At its core, a contrastive learning stage creates a unified semantic space, followed by a self-supervised clustering stage that enforces structural coherence. Our evaluations demonstrate its effectiveness in accurately delineating the capsule-cortex-medulla architecture in human lymph nodes and precisely identifying germinal centers in human tonsil data, outperforming baseline methods.

The source code is available at https://github.com/VitaIntelli-CQU/SpaAlign. The source code is archived at https://doi.org/10.5281/zenodo.20250508.