TL;DR
This paper introduces SOCS, a novel method for inferring cell developmental trajectories from time-series spatial transcriptomics data. Unlike previous methods that often break apart biologically coherent structures, SOCS explicitly preserves these structures, leading to more accurate and biologically meaningful trajectory inferences.
What Problem Does This Paper Address?
Understanding how cells develop and differentiate over time is crucial for deciphering complex biological processes. While single-cell technologies have revolutionized this field, inferring these developmental trajectories, especially from spatial transcriptomics (ST) data, remains challenging. A key issue is that many tissues are composed of spatially contiguous, functionally integrated units (like kidney tubules or ovarian follicles) that develop as a whole. Existing computational methods, particularly those based on optimal transport (OT), often fail to respect these structural units, leading to biologically implausible trajectories that split these coherent structures over time. This makes it difficult to gain accurate insights into organogenesis and other dynamic biological processes.
Methods
The core of this paper is the Spatiotemporal Optimal transport with Contiguous Structures (SOCS) method. SOCS builds upon existing optimal transport frameworks but introduces a crucial new constraint: structural consistency.
Here's how it works:
- Inputs: SOCS takes spatial coordinates and gene expression data from two time points (t₁ and t₂) of a biological sample.
- Dimensionality Reduction: Principal Component Analysis (PCA) is used to reduce the dimensionality of the gene expression data.
- Optimization Problem: SOCS solves an optimization problem to find a mapping (a "transport map") between cells at t₁ and cells at t₂, representing ancestor-descendant relationships. This optimization balances three key requirements:
- Gene Expression Similarity: Descendants should have similar gene expression profiles to their ancestors.
- Geometric Consistency: The spatial relationships between cells should be preserved over time.
- Structural Consistency (The Novelty): This is the key innovation. SOCS explicitly penalizes mappings where cells belonging to the same predefined contiguous biological structure at t₁ end up in different structures at t₂. These structures are identified beforehand, either manually or automatically.
- Algorithm: The optimization is solved using a modified unbalanced Gromov-Wasserstein algorithm.
Datasets
The authors evaluated SOCS using two primary datasets:
- Mouse Organogenesis Spatiotemporal Transcriptomic Atlas (MOSTA): This dataset provides time-series spatial transcriptomics data from mouse embryos, allowing the study of organ development.
- Mouse Ovary MERFISH data: This dataset, generated by the authors, uses imaging-based MERFISH technology to capture high-resolution spatial transcriptomics of mouse ovaries during ovulation.
Key Insights and Ideas
- Structure Matters: The most significant insight is that preserving the spatial integrity of biological structures is paramount for accurate trajectory inference in many developmental processes. Ignoring this leads to biologically nonsensical results.
- Beyond Gene Expression and Geometry: While gene expression similarity and global geometric consistency are important, they are insufficient on their own. SOCS demonstrates that incorporating structural constraints significantly improves biological plausibility.
- Unlocking New Biological Questions: By providing more accurate trajectories, SOCS enables researchers to ask and answer questions about morphological changes, cell type transitions within structures, and the influence of the local environment on development, which were previously difficult or impossible with other methods.
Key Results
- Improved Trajectory Accuracy: SOCS consistently produced more biologically plausible trajectory inferences compared to existing OT-based methods (W-OT and Moscot) in both mouse lung and ovary datasets.
- Preservation of Biological Units: In the MOSTA dataset, SOCS successfully maintained the coherence of structures like lung airways, kidney glomeruli, and heart atria/ventricles, whereas W-OT and Moscot often split them.
- Accurate Cell Type Mapping: In the lung, SOCS mapped epithelial airway cells to descendant epithelial airway cells with 99.3% accuracy, significantly outperforming W-OT (79.2%) and Moscot (78.8%).
- Follicle Maturation Insights: In the mouse ovulation data, SOCS accurately mapped preantral follicles to small antral follicles, enabling the identification of differentially expressed genes and morphological changes associated with follicle maturation. It also revealed novel hypotheses about the role of stromal cells in neighboring maturing follicles.
- Robustness Across Technologies: SOCS performed well on both sequencing-based (Stereo-Seq) and imaging-based (MERFISH) spatial transcriptomics data.
Limitations and Future Work
- Cell Type Imbalances: SOCS, like other OT methods, can be sensitive to strong imbalances in cell type composition between time points. Future work may involve exploring partial optimal transport to mitigate this.
- Sampling Rate: The accuracy of SOCS is dependent on the sampling rate of the time-series data. If time points are too far apart, dramatic changes in tissue structure and gene expression can hinder accurate inference.
- Integration with Other Methods: The authors suggest future work could integrate SOCS with clonal barcoding lineage estimation or RNA velocity vectors to further enhance trajectory inference.
- Broader Applications: SOCS is proposed for studying not only organogenesis and homeostasis but also disease progression (e.g., tumor growth) and identifying diagnostic markers.
Summary updated