TL;DR
This paper introduces CellContrast, a novel deep learning method that leverages spatial transcriptomics (ST) data to reconstruct the spatial relationships of single-cell RNA sequencing (scRNA-seq) data. By outperforming existing methods in various benchmarks, CellContrast enables more accurate downstream analyses like cell-type co-localization and cell-cell communication, ultimately leading to new biological discoveries and reducing false positives.
What Problem Does This Paper Address?
Single-cell RNA sequencing (scRNA-seq) has generated a massive amount of gene expression data, offering unprecedented insights into cellular heterogeneity. However, a critical limitation of scRNA-seq is the absence of spatial information – we know what cells are present and their gene expression profiles, but not where they are located within a tissue. This lack of spatial context hinders our ability to understand complex biological processes like tissue organization, intercellular communication, and developmental patterning. While spatial transcriptomics (ST) technologies can provide this spatial information, they are often costly, technically challenging, and may have limitations in gene profiling depth or resolution. This paper aims to bridge this gap by developing a computational method that can infer spatial relationships in scRNA-seq data using ST data as a reference, thereby maximizing the utility of existing scRNA-seq datasets.
Methods
CellContrast employs a deep contrastive learning framework. The core idea is to train an encoder that maps gene expression profiles into a latent space where spatially proximate cells have similar representations. This is achieved by constructing training batches with "contrastive pairs":
- Positive Pairs: Spatially neighboring cells from the ST reference data.
- Negative Pairs: Spatially distant cells from the ST reference data.
An "encoder-projector" neural network architecture is used. The encoder extracts spatial semantic representations, and a projection head is used for calculating the loss. The model is optimized using an infoNCE loss, which encourages positive pairs to be closer in the latent space and negative pairs to be further apart. During inference, the projection head is removed, and the trained encoder generates representations for scRNA-seq data. These representations can then be used in two ways:
- sc-sc Similarity: To identify cell neighbors and estimate cell-cell distances within the scRNA-seq data itself.
- sc-ST Alignment: To project scRNA-seq cells onto the spatial coordinates of the ST reference, effectively predicting their locations.
The paper also introduces a confidence score for sc-ST assignments and a method for identifying spatially relevant genes based on their contribution to the latent representation.
Datasets
The study utilized a variety of datasets for training and benchmarking:
- Mouse Embryo Data:
- SeqFISH (single-cell resolution ST) for training and testing.
- Stereo-seq (spot-based ST) for training.
- scRNA-seq data from mouse embryos at E9.5 (GEO: GSE87038) and E8.5 for downstream analyses.
- Human Breast Data:
- 10X Visium (spot-based ST) for training and testing.
- Mouse Brain Data:
- MERSCOPE (single-cell resolution ST) for benchmarking.
The paper emphasizes benchmarking across diverse platforms (SeqFISH, Stereo-seq, 10X Visium, MERSCOPE) and scenarios, including matched and independent samples, and different developmental stages or even different donors (smoker vs. non-smoker for human lung data).
Key Insights and Ideas
- Transferability of Spatial Information: The core insight is that spatial relationships learned from ST data can be effectively transferred to scRNA-seq data, even when there are mismatches in platforms, developmental stages, or sample origins.
- Contrastive Learning for Spatial Context: Applying contrastive learning, which is typically used for image or text data, to gene expression data proves highly effective for capturing spatial proximity.
- Local vs. Global Reconstruction: The method excels at reconstructing local cellular neighborhoods, which is crucial for many downstream biological questions, while also providing a means for global spatial inference.
- Mitigating False Positives: By incorporating spatial information, CellContrast significantly reduces false positives in downstream analyses like cell-cell communication, a common issue with expression-only methods.
- Leveraging Existing scRNA-seq Data: CellContrast offers a powerful way to unlock the spatial potential of the vast amounts of scRNA-seq data already available, reducing the need for costly and complex ST experiments in some cases.
Key Results
- Superior Performance: CellContrast consistently outperformed existing methods (SageNet, CeLEry, CytoSpace, NovoSpaRc, Tangram) in reconstructing spatial relationships across multiple datasets and platforms.
- In mouse gastrulation cell neighbor reconstruction, CellContrast achieved the highest average hit numbers, identifying twice the number of true neighbors compared to some alternatives.
- Jessen-Shannon Distance (JSD) for neighborhood cell-type reconstruction was consistently lower with CellContrast, indicating better preservation of local cell-type heterogeneity.
- Robustness: The method demonstrated robustness across different testing scenarios, including independent samples and platforms, and even with a certain degree of mismatch between training and testing data.
- Downstream Applications:
- Cell-type Co-localization: CellContrast accurately detected cell-type co-localization patterns in mouse embryos, aligning well with reference patterns derived from multiple ST datasets.
- Cell-Cell Communication (CCC): When applied to CCC analysis, CellContrast, by incorporating spatial information, significantly reduced unsupported CCC signals and potential false positives compared to methods relying solely on gene expression. For example, in analyzing IGF2-IGF2R interactions, CellContrast provided spatially constrained evidence.
Limitations and Future Work
- "Cell Clumping" Issue: The SC-ST mapping approach can sometimes lead to multiple SC cells being assigned to the same ST location, potentially causing overlap and loss of predicted SC-SC distances. The authors suggest this can be mitigated by combining local and global distance estimations or by adjusting the number of SC neighbors used for alignment.
- Tissues with Repeating Elements: Tissues with highly repetitive structures might pose a challenge for spatial reconstruction, and the authors suggest pre-processing to reduce repeating elements as a potential solution.
- Future Directions:
- Refining the method's performance with improved ST data.
- Exploring new applications of spatial reconstruction in single-cell analysis.
- Investigating the contribution of specific genes to the latent representation for deeper biological insights into spatial gene patterns.
Summary updated