Integrating cross-sample and cross-modal data for spatial transcriptomics and metabolomics with SpatialMETA
Ruonan Tian, Ziwei Xue, Yiru Chen, Yicheng Qi, Jian Zhang, Jie Yuan et al. · 2025 · 30 citationsRead the paper
Simultaneous profiling of spatial transcriptomics (ST) and spatial metabolomics (SM) on the same or adjacent tissue sections offers a revolutionary approach to decode tissue microenvironment and identify potential therapeutic targets for cancer immunotherapy. Unlike other spatial omics, cross-modal integration of ST and SM data is challenging due to differences in feature distributions of transcript counts and metabolite intensities, and inherent disparities in spatial morphology and resolution. Furthermore, cross-sample integration is essential for capturing spatial consensus and heterogeneous patterns but is often complicated by batch effects. Here, we introduce SpatialMETA, a conditional variational autoencoder (CVAE)-based framework for cross-modal and cross-sample integration of ST and SM data. SpatialMETA employs tailored decoders and loss functions to enhance modality fusion, batch effect correction and biological conservation, enabling interpretable integration of spatially correlated ST-SM patterns and downstream analysis. SpatialMETA identifies immune spatial clusters with distinct metabolic features in cancer, revealing insights that extend beyond the original study. Compared to existing tools, SpatialMETA demonstrates superior reconstruction capability and fused modality representation, accurately capturing ST and SM feature distributions. In summary, SpatialMETA offers a powerful platform for advancing spatial multi-omics research and refining the understanding of metabolic heterogeneity within the tissue microenvironment.
What the full text shows
Working prototype- SpatialMETA outperformed existing non-spatial single-cell and spatial multi-omics tools in both vertical (cross-modal) and horizontal (cross-sample) integration of ST and SM data. — Highest overall score across 11 samples (Tested on ccRCC, mouse brain, and GBM datasets)
- The framework demonstrates high computational efficiency in terms of memory and processing time. — < 2 GB memory and < 100 s processing time (Standard processing of ST and SM datasets)
- Identified a specific immune sub-cluster (Imm_4) in ccRCC with upregulated lipid metabolism at both transcriptional and metabolic levels. (ccRCC Y7_T sample)
- The model scales approximately linearly with increasing data size. — 10,000 to 100,000 spots (Simulated ST and SM datasets)
- Successfully aligned and reassigned SM data to ST coordinates to resolve resolution discrepancies. — Reduced 10,145 SM spots to 2,018 ST spots (ccRCC Y7_T sample)
- Setup
- Publicly available ST (10x Genomics Visium) and SM (MALDI, DESI) datasets from human ccRCC, GBM, and mouse brain samples.
- How
- A CVAE-based framework utilizing tailored decoders and MMD loss for modality fusion and batch effect correction. It includes a preprocessing pipeline for spatial alignment (rotation, translation, non-linear distortion) and resolution reassignment via K-nearest neighbors (KNN).
- Limitations
- Alignment and reassignment are currently limited to two adjacent sections within a single sample. · Metabolite identities based on m/z values require additional experimental validation (e.g., LC-MS).
Summarized by Gemma 4 from the open-access full text · 12 Oct 2026. Read it