TL;DR
This paper introduces iSORT, a novel transfer learning framework that integrates single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) data to reconstruct the spatial organization of cells at single-cell resolution. iSORT not only accurately maps gene expression to spatial locations but also identifies key "spatial-organizing genes" (SOGs) and infers pseudo-growth trajectories (SpaRNA velocity), offering a powerful new tool for understanding tissue complexity and disease.
What Problem Does This Paper Address?
Understanding the intricate spatial organization of cells within tissues is crucial for deciphering biological processes and disease mechanisms. Single-cell RNA sequencing (scRNA-seq) provides high-resolution gene expression data but loses spatial information due to cell dissociation. Conversely, spatial transcriptomics (ST) captures gene expression within its spatial context but often has limitations in resolution or gene coverage. The challenge lies in effectively integrating these complementary datasets to reconstruct a comprehensive, single-cell resolution view of tissue organization. Existing methods struggle with accurately mapping scRNA-seq data to spatial locations, identifying the genes that truly drive this organization, or visualizing dynamic cellular processes in space.
Methods
iSORT employs a transfer learning approach using a neural network. The core idea is to leverage ST data as a reference to "teach" a model how to map gene expression from scRNA-seq data to spatial coordinates.
- Preprocessing: Both scRNA-seq and ST data undergo normalization, log transformation, and gene selection.
- iSORT Training:
- Co-embedding: A reference-based co-embedding approach (similar to Seurat) is used to map both scRNA-seq and ST data into a common latent space. This step aims to align the data distributions.
- Density Ratio Estimation: A key innovation is the estimation of the "density ratio" between scRNA-seq and ST data. This ratio helps account for differences in cell-type distributions and expression scales between the two data types.
- Neural Network Mapping: A neural network is trained to map the latent space representations to spatial coordinates. This network learns a function
f = g o h, wherehmaps gene expression to the latent space (z = h(x)) andgmaps the latent space to spatial coordinates (y = g(z)). The training minimizes a loss function that incorporates the density ratio. - Downstream Analysis: Once trained, the mapping
fis applied to scRNA-seq data to reconstruct spatial organization at single-cell resolution. This enables: - Spatial Pattern Revelation: Visualizing gene expression patterns in space.
- SOG Identification: Identifying "spatial-organizing genes" (SOGs) by quantifying their influence on spatial organization using a proposed index
Ig. - SpaRNA Velocity: Inferring pseudo-growth trajectories by projecting RNA velocity (from gene expression space) into physical space.
Datasets
iSORT was trained and evaluated on several publicly available datasets, including:
- Human Dorsolateral Prefrontal Cortex (DLPFC): Used for benchmarking and demonstrating robustness with different ST references.
- Mouse Embryo: Used for testing reconstruction accuracy.
- Mouse Brain: Used for reconstructing stratified structures.
- Drosophila Embryo: Used to reveal spatial strip patterns of gene expression.
- Human Arteries: Newly generated datasets (scRNA-seq and ST) were used to study atherosclerosis-related SOGs.
- Human Developmental Heart: Used to visualize SpaRNA velocity and pseudo-growth trajectories.
The paper also mentions using simulated coarse-grained ST data derived from high-resolution FISH-seq data for testing iSORT's performance with low-resolution references.
Key Insights and Ideas
- Density Ratio Transfer Learning: The core innovation is using density ratio estimation to bridge the distributional gap between scRNA-seq and ST data, enabling more accurate transfer learning.
- Spatial-Organizing Genes (SOGs): iSORT moves beyond identifying spatially variable genes to pinpointing genes that drive spatial organization, offering a deeper understanding of tissue development and function.
- SpaRNA Velocity: Integrating RNA velocity into physical space provides a novel way to visualize and understand dynamic cellular processes like growth and migration within tissues.
- Robustness to Data Variability: iSORT demonstrates remarkable robustness to variations in ST references, including different sample sources, cell-type distributions, and even distortions like rotation.
Key Results
- High Accuracy in Reconstruction: iSORT achieved high accuracy in reconstructing spatial organization across multiple benchmark datasets (e.g., average indicators of 0.89 for DLPFC).
- Superior Performance: iSORT outperformed several existing methods (scSpace, Tangram, novoSpaRc, CeLEry, CellTrek) in reconstructing spatial organization, particularly in capturing detailed structures and boundaries.
- Identification of Key Genes: iSORT successfully identified SOGs that, when "knocked out" in silico, significantly altered spatial organization, whereas genes identified by other methods (SVGs) had less impact.
- Disease Biomarker Potential: Analysis of human artery data revealed SOGs that could serve as biomarkers for atherosclerosis.
- Accurate Trajectory Inference: SpaRNA velocity successfully reconstructed known developmental trajectories in the DLPFC and human heart datasets, outperforming previous methods that struggled with discontinuities.
- Robustness Demonstrated: iSORT maintained high accuracy even with single or multiple heterologous ST references, and with rotated references, showcasing its resilience to common batch effects.
Limitations and Future Work
- Tissue Similarity Requirement: iSORT currently requires ST references to be from similar tissues as the scRNA-seq data. The authors suggest that training on more diverse samples (different organs, species) and leveraging large pre-trained models (like scGPT) could significantly improve its power and generalizability.
- Cell-Cell Communication (CCC): While iSORT infers spatial organization, it doesn't directly perform CCC inference. The authors suggest that iSORT's spatial information could be used in conjunction with matched scRNA-seq and ST data for CCC analysis.
- 3D Reconstruction: The current framework is primarily for 2D slices. The authors propose that iSORT could be generalized to incorporate 3D spatial reconstruction.
- SpaRNA Velocity Refinement: While SpaRNA velocity is a novel concept, further exploration and comparison with other RNA velocity methods are suggested.
Summary updated