Life is composed of cells arranged in an orderly manner. Once detached from the organism, a single cell is unable to function independently. Therefore, a comprehensive understanding of cellular function requires considering the microenvironment and spatial location of cells. With the rapid advancement of spatial omics technologies in recent years, vast amounts of spatial omics data from different tissues and organs have been generated using different technical platforms. Accurately representing and analyzing these rapidly growing, large-scale, and heterogeneous data is not only the foundation for the development of corresponding computational models, but also a prerequisite for effectively deciphering the rich biomedical information embedded within them.

Unlike current methods that primarily focus on spot/cell-level features in single or a few spatial slices, PASSAGE proposed a multi-level attention-based large-scale heterogeneous spatial omics representation learning approach across spatial slice, cellular and genes. Specifically, PASSAGE designed an attention pooling layer based on a graph attention autoencoder for spot-level representations, adaptively weighting and aggregating all spots within the same slice to generate slice-level representations. These slice-level representations are then optimized using a contrastive learning strategy guided by phenotypic information. Furthermore, non-negative matrix factorization (NMF) was employed to obtain gene-level attention scores, thereby enabling effective analysis of spatial omics features associated with specific phenotypes (Figure 1).

Figure 1 Framework of PASSAGE

For example, in the case of 103 breast tissue and breast cancer slices from 42 different patients across 2 different spatial transcriptomics platforms (ST and Visium). PASSAGE effectively addressed batch effects between different samples, accurately identifying spatial regions within breast cancer tissues (Figure 2a). Meanwhile, the molecular-level attention representation introduced by PASSAGE enables the extraction of gene sets highly correlated with the phenotype from the learned weights of attention pooling layer. Specifically, genes identified by PASSAGE have been previously associated with disease progression. Such approach highly enhanced the biological interpretability of PASSAGE (Figure 2b).

Figure 2 A) PASSAGE successfully learned informative slice-level representation and identified phenotype-associated spatial signatures in breast cancer slices. B) PASSAGE successfully learned gene sets that are highly associated with phenotype-associated spatial signatures in breast cancer slices.

It is worth noting that such multi-scale representation learning architecture enables PASSAGE to effectively perform systematic analysis of phenotype-associated spatial signatures in large-scale heterogeneous spatial transcriptomics data. A typical consumer-grade RTX 4090 GPU could process at an average speed of 600 spots/cells per minute.

This research was conducted by Chen-Kai Guo, a graduate student from the Guangdun Peng’s lab at the Guangzhou Institute of Biomedicine and Health, Chinese Academy of Sciences, during his visiting at the Ge Gao’s lab, and the study was completed in collaboration with PhD student Chen-Rui Xia, with Prof. Ge Gao, Dr. Zhi-Jie Cao, and Prof. Guangdun Peng as co-corresponding authors. The research was supported by the National Natural Science Foundation of China, the China Postdoctoral Science Foundation, the National Key Laboratory for Gene Function Research and Control, the Beijing Future Gene Diagnosis and Innovation Center, and the Changping Laboratory. This paper was invited to be published online in Small Methods (Special Issue on Single-cell and Spatial Transcriptomics) on February 5, 2025. All code has been open-sourced and is available at https://github.com/gao-lab/PASSAGE.

Paper URL：https://doi.org/10.1002/smtd.202401451 

Code URL：https://github.com/gao-lab/PASSAGE