The paper introduces Local Pooling (LP), a revolutionary computational framework designed to process ultra-large-scale spatial omics data with unprecedented speed and efficiency. By utilizing a neighbor-indexing strategy instead of traditional adjacency matrices, the system significantly lowers GPU memory consumption, allowing it to analyze millions of cells on a single device. The researchers implemented this framework through SpaLP, a tool that excels at tasks such as niche identification, 3D organ atlas construction, and multi-omics integration across various biological platforms. Experimental results demonstrate that SpaLP is up to 300 times faster than existing graph neural network methods while maintaining superior accuracy in identifying complex tissue structures. Furthermore, the model exhibits strong generalization capabilities, suggesting it could serve as a foundation for future large-scale biological models. Ultimately, this lightweight approach removes significant hardware barriers, enabling researchers to decode intricate cellular microenvironments using minimal computational resources.

References:

• Dai B, Liang Y, Yi L, et al. A lightweight, ultrafast and general embedding framework for large-scale spatial omics data[J]. bioRxiv, 2026: 2026.02. 04.703814.