Multi-modal domain adaptation for revealing spatial functional landscape from spatially resolved transcriptomics

Lequn Wang, Yaofeng Hu, Kai Xiao, Chuanchao Zhang, Qianqian Shi, Luonan ChenBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces stMDA, a novel multi-modal domain adaptation method that integrates gene expression, histology images, and spatial coordinates to map the functional landscape of tissues. By aligning spatial distributions across these modalities, stMDA significantly improves the accuracy of identifying distinct cellular regions and denoising gene expression data, offering a more robust tool for analyzing complex biological systems.

What Problem Does This Paper Address?

Spatially resolved transcriptomics (SRT) technologies have revolutionized our ability to study gene expression within its native tissue context, offering unprecedented insights into development and disease. However, SRT data is often sparse and noisy, making it challenging to accurately identify distinct cellular regions (spatial domains) and understand their functions. Existing methods often struggle to effectively integrate multiple data types (like histology images and spatial locations) with gene expression, or they rely on single models that may not generalize well across different platforms and species. This paper addresses the need for a robust, multi-modal approach that can overcome data sparsity and noise to reveal a more accurate and comprehensive spatial functional landscape.

Methods

The core innovation is stMDA (multi-modal unsupervised domain adaptation). It employs a multi-modal unsupervised domain adaptation strategy with deep spatial distribution alignment.

Here's a breakdown of the key components:

  • Modality-Specific Representation Learning: stMDA uses different neural network architectures tailored to each data type:
  • Gene Expression: A Variational Autoencoder (VAE) encoder learns representations from gene expression data.
  • Spatial Coordinates: A Graph Convolutional Network (GCN) leverages the spatial relationships between cells.
  • Histological Images: A pre-trained Convolutional Neural Network (CNN) extracts features from histology images.
  • Deep Spatial Distribution Alignment: This is a novel strategy to align the global and local spatial distributions of cells across the different modalities. It uses Maximum Mean Discrepancy (MMD) to minimize differences in how cells are distributed, ensuring that the learned representations are consistent across modalities and reflect true spatial organization. This alignment is crucial for robust clustering.
  • Joint Representation and Reconstruction: An attention mechanism integrates the modality-specific representations into a unified "joint representation." This joint representation is then used by a decoder to reconstruct the gene expression, effectively denoising the original data.
  • Spatial Domain Identification: The learned joint representation is used to build a k-nearest neighbor graph, which is then processed by the Leiden algorithm to identify spatial domains.

Datasets

The paper evaluates stMDA on a diverse set of 17 datasets from various SRT platforms:

  • 14 datasets from the 10x Visium platform.
  • 2 datasets from the Slide-seqV2 platform.
  • 1 dataset from the Stereo-seq platform.

Specific datasets highlighted for evaluation include:

  • Human Dorsolateral Prefrontal Cortex (DLPFC): Used for benchmarking against other methods, with known ground truth annotations for cortex layers.
  • Pancreatic Ductal Adenocarcinoma (PDAC) tumor section: Used to demonstrate stMDA's ability to handle heterogeneous tumor tissues and identify prognostic genes.
  • Mouse embryo slice (Stereo-seq): Used to showcase versatility across species and platforms.
  • Human embryo slice (10x Visium): Further validation of cross-platform and cross-species capabilities.
  • Mouse olfactory bulb and hippocampus (Slide-seqV2): Large-scale, high-resolution datasets used to assess scalability and precise anatomical structure identification.

Key Insights and Ideas

  • Multi-modality is Key: Integrating gene expression with spatial coordinates and histology images is crucial for overcoming the inherent sparsity and noise in SRT data.
  • Spatial Alignment Matters: The "deep spatial distribution alignment" strategy is vital for ensuring that representations learned from different data types are harmonized, leading to more accurate and consistent spatial domain identification.
  • Denoising Enhances Discovery: Reconstructing gene expression not only cleans up noisy data but also sharpens spatial expression patterns, making it easier to identify biologically relevant genes, including those with prognostic potential.
  • Generalizability: stMDA demonstrates robust performance across different SRT platforms, species, and tissue types (from well-structured brain tissue to heterogeneous tumors and developing embryos), highlighting its broad applicability.
  • Fine-Grained Anatomy: The method can accurately delineate intricate anatomical structures, even in high-resolution datasets, which is critical for understanding complex biological systems.

Key Results

  • Superior Spatial Domain Identification: stMDA achieved the highest mean Adjusted Rand Index (ARI) (0.572) on 12 DLPFC datasets compared to existing methods like BayesSpace, DeepST, GraphST, SEDR, SpaGCN, SpatialPCA, and STAGATE.
  • Improved Tumor Heterogeneity Analysis: On a PDAC tumor slice, stMDA identified spatial domains with higher purity (0.814) and better alignment with marker genes than other methods.
  • Prognostic Gene Identification: Denoising with stMDA significantly improved the identification of differentially expressed genes (DEGs) with prognostic value in PDAC tumors. For example, FAM83A, GJB3, and HAVCR1 showed stronger associations with patient survival after denoising.
  • Accurate Anatomical Structure Delineation: stMDA successfully identified distinct anatomical structures in mouse and human embryos and in large-scale olfactory bulb and hippocampus datasets, often outperforming other methods in capturing finer details and boundaries.
  • Scalability: The method was successfully applied to large-scale, high-resolution datasets (Slide-seqV2), demonstrating its scalability.

Limitations and Future Work

  • Parameter Tuning: The paper notes that the parameter 'a' (controlling the degree of spatial distribution alignment) is not universally applicable and requires tuning based on context. The authors suggest a grid search approach for optimization.
  • Future Work: While not explicitly detailed as future work, the paper's success across diverse datasets and its flexible architecture suggest potential applications in integrating even more modalities or handling emerging SRT technologies like SHARE-seq and CITE-seq. The denoising capability also opens avenues for more precise downstream functional analyses.

Summary updated

BibTeX
@article{wang2024multi,
  author = {Lequn Wang and Yaofeng Hu and Kai Xiao and Chuanchao Zhang and Qianqian Shi and Luonan Chen},
  title = {{Multi-modal domain adaptation for revealing spatial functional landscape from spatially resolved transcriptomics}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae257},
}
Figure