JADE: Joint Alignment and Deep Embedding for Multi-Slice Spatial Transcriptomics

Y. P. Guo, Jun S. Liu, Huimin Cheng, YingLiang MabioRxiv 2025 · DOI

TL;DR

This paper introduces JADE, a novel computational framework that simultaneously aligns and integrates multiple spatial transcriptomics (SRT) datasets. By jointly learning spatial correspondences and shared gene expression patterns, JADE overcomes limitations of existing methods that tackle these tasks in isolation, leading to more accurate reconstruction of tissue structures and identification of biological signals across different tissue slices.

What Problem Does This Paper Address?

Spatial transcriptomics (SRT) technologies allow us to map gene expression within the 3D context of tissues. As more studies generate multiple SRT datasets from adjacent or replicated tissue slices, the need to integrate this data becomes crucial for tasks like reconstructing 3D tissue structures, tracing developmental trajectories, or comparing biological conditions. However, integrating these datasets is challenging due to:

  • Physical Distortions: Tissues can deform during sectioning and mounting, making it difficult to align corresponding regions across slices.
  • Technical Variability: Differences in sample preparation, sequencing, and other technical factors introduce "batch effects" that obscure true biological signals.
  • Isolation of Tasks: Existing methods typically address spatial alignment (finding correspondences) and representation learning (creating meaningful gene expression summaries) separately. This separation limits their ability to achieve coherent and biologically accurate integration.

Methods

JADE tackles these challenges with a unified, "roundtrip" framework that iteratively refines both alignment and embeddings:

  • Joint Alignment and Deep Embedding: JADE simultaneously learns two key components:
  • Probabilistic Alignment: It infers a probabilistic mapping between spatial locations in different slices.
  • Shared Low-Dimensional Embeddings: It generates compressed representations of gene expression that capture biological information and are comparable across slices.
  • Attention-Based Alignment: To infer alignment, JADE uses attention mechanisms. These mechanisms dynamically weigh the importance of different gene expression dimensions, allowing the model to focus on features most relevant for alignment and suppress noise.
  • Roundtrip Learning: A core innovation is the alternating refinement of alignment and embeddings within each training iteration. This ensures that the learned embeddings guide better alignment, and vice versa, leading to a more cohesive integration.
  • Graph Autoencoder: A graph autoencoder is used to learn initial low-dimensional embeddings by leveraging the spatial proximity of cells within each slice.
  • Self-supervised Graph Contrastive Learning: To enhance the biological relevance of embeddings, a contrastive learning approach is employed. This method encourages similar spatial locations to have similar embeddings while pushing dissimilar ones apart.
  • Fast Computation (Hyperspots): To address the computational cost of aligning high-resolution datasets, JADE offers a "Fast JADE" version that uses "hyperspots" (aggregated spatial units) for initial alignment, significantly reducing computation time while maintaining performance.

Datasets

JADE was evaluated on two publicly available multi-slice SRT datasets:

  • 10x Visium Human Dorsolateral Prefrontal Cortex (DLPFC): This dataset consists of 12 serial tissue sections from human brains, allowing for the evaluation of alignment and domain detection across cortical layers and different donors.
  • Stereo-seq Axolotl Brain: This dataset captures gene expression across five developmental stages of the axolotl brain, enabling the assessment of JADE's ability to integrate data across different developmental time points.

Key Insights and Ideas

  • The Power of Joint Optimization: The paper strongly suggests that treating spatial alignment and representation learning as coupled problems, rather than independent ones, is crucial for achieving robust multi-slice integration.
  • Attention for Alignment: Using attention mechanisms to dynamically focus on relevant gene expression features for alignment is a novel approach that improves the model's ability to handle complex spatial variations.
  • Roundtrip Refinement: The iterative, alternating refinement of alignment and embeddings is a key mechanism that allows JADE to achieve a more coherent and biologically meaningful integration.
  • Scalability Matters: The development of the "Fast JADE" approach using hyperspots highlights the practical importance of computational efficiency for real-world applications of SRT data integration.

Key Results

  • Superior Clustering Accuracy: JADE consistently outperformed baseline methods (GraphST, STAligner, STAGATE) in recovering biologically relevant spatial domains, achieving higher Adjusted Rand Index (ARI) scores on both the DLPFC and axolotl brain datasets. For example, on the DLPFC dataset, JADE achieved median ARI scores of approximately 0.61 and 0.65 for adjacent slices, significantly improving upon other methods.
  • Improved Alignment Accuracy: JADE demonstrated superior alignment accuracy, especially in challenging scenarios like integrating slices with larger separations (e.g., B-C pairs in the DLPFC dataset). It significantly outperformed methods like PASTE, STAligner, and Seurat, with margins exceeding 50% in some cases.
  • Effective Batch Effect Mitigation: JADE effectively integrated data from different slices and developmental stages, as evidenced by high iLISI scores (e.g., 1.98 on DLPFC, 1.85 on axolotl brain), indicating excellent inter-slice mixing and superior batch correction performance compared to baselines.
  • Biological Validation: Differential expression analysis using JADE-identified domains revealed biologically meaningful marker genes with well-established region-specific expression patterns, confirming the biological relevance of the learned representations.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Slice-Specific Limitations: JADE's performance might be affected by extreme sparsity, highly unbalanced slice resolutions, or limited overlap in tissue regions across slices.
  • Pairwise Alignment: The current framework models pairwise slice alignment, which may limit its ability to fully reconstruct larger tissue volumes or resolve global correspondences across long serial sections.
  • Static Snapshot: JADE assumes a static snapshot of spatial organization and does not account for temporal variation, which is important for developmental or regenerative contexts.

Future work will focus on:

  • Extending JADE to handle fully joint alignment across multiple slices.
  • Incorporating spatiotemporal transcriptomics data and integrating additional data modalities (e.g., histological imaging, spatial epigenomics).
  • Improving scalability for ultra-high-resolution datasets and increasing robustness to technical noise or sample heterogeneity.

Summary updated

Abstract

As spatially resolved transcriptomics (SRT) datasets increasingly span multiple adjacent or replicated slices, effective joint analysis across slices is needed to reconstruct tissue structures and identify consistent spatial gene expression patterns. This requires resolving spatial correspondences between slices while capturing shared transcriptomic features, two tasks that are typically addressed in isolation. Multi-slice analysis remains challenging due to physical distortions, technical variability, and batch effects. To address these challenges, we introduce Joint Alignment and Deep Embedding for multi-slice SRT (JADE), a unified computational framework that simultaneously learns spatial location-wise alignments and shared low-dimensional embeddings across tissue slices. Unlike existing methods, JADE adopts a roundtrip framework in which each iteration alternates between alignment and embedding refinement. To infer alignment, we employ attention mechanisms that dynamically assess and weight the importance of different embedding dimensions, allowing the model to focus on the most alignment-relevant features while suppressing noise. To the best of our knowledge, JADE is the first method that jointly optimizes alignment and representation learning in a shared latent space, enabling robust multi-slice integration. We demonstrate that JADE outperforms existing alignment and embedding methods across multiple evaluation metrics in the 10x Visium human dorsolateral prefrontal cortex (DLPFC) and Stereo-seq axolotl brain datasets. By bridging spatial alignment and feature integration, JADE provides a scalable and accurate solution for cross-slice analysis of SRT data.
BibTeX
@misc{guo2025jade,
  author = {Y. P. Guo and Jun S. Liu and Huimin Cheng and YingLiang Ma},
  title = {{JADE: Joint Alignment and Deep Embedding for Multi-Slice Spatial Transcriptomics}},
  year = {2025},
  doi = {10.1101/2025.11.21.689823},
  howpublished = {bioRxiv},
}
Figure