TL;DR
This paper introduces scMoMaT, a novel method for integrating diverse single-cell multi-omics datasets, even when they have different combinations of modalities (mosaic integration). scMoMaT not only achieves superior integration performance but also identifies cell-type-specific biomarkers across these modalities, leading to more accurate and higher-resolution cell type annotations.
What Problem Does This Paper Address?
Single-cell multi-omics technologies allow us to measure different biological features (like gene expression, chromatin accessibility, and protein levels) from the same cell. Integrating these diverse datasets is crucial for a comprehensive understanding of cellular identity and function. However, existing integration methods often struggle with "mosaic integration," where datasets have varying combinations of modalities and batches. This is a challenging scenario because it's difficult to find common ground for comparison. Furthermore, accurately annotating cell types based on these integrated datasets, especially when dealing with complex biological systems, remains a significant hurdle.
Methods
scMoMaT employs a matrix tri-factorization framework. For each input data matrix (representing a specific modality and batch), it decomposes the data into three latent factors: cell factors, feature factors, and an association matrix. This approach is extended to handle multiple data matrices simultaneously, allowing for the integration of diverse modalities and batches. A key innovation is its ability to handle mosaic integration by flexibly incorporating different combinations of modalities. Beyond cell embedding, scMoMaT also learns cell-type-specific bio-markers across modalities (genes, proteins, chromatin regions, and even motifs) by retraining the model with cluster labels. This allows for a more robust and interpretable annotation of cell types. The method is also designed to handle disproportionate cell type compositions between batches, a common issue in real-world datasets.
Datasets
The authors evaluated scMoMaT on:
- Simulated datasets: Generated to mimic various integration scenarios, including paired scRNA-seq and scATAC-seq data from multiple batches, with and without simultaneously profiled cells.
- Real datasets:
- Human PBMC dataset (4 batches, CITE-seq and ASAP-seq modalities)
- Mouse brain cortex dataset (5 batches, SNARE-seq, snRNA-seq, snATAC-seq, and Allen Brain Atlas data)
- Human bone marrow mononuclear cells (BMMC) dataset (2 batches, scATAC-seq and scRNA-seq)
- Mouse spleen dataset (2 batches, scRNA-seq and scATAC-seq, with a sub-sampling strategy to create disproportionate cell type compositions).
Key Insights and Ideas
- Mosaic integration is hard, but scMoMaT tackles it: The paper highlights mosaic integration as the most general and challenging scenario, and scMoMaT's matrix tri-factorization approach provides a flexible solution.
- Biomarkers are key for annotation: By learning modality-specific biomarkers after integration, scMoMaT offers a powerful way to interpret and annotate cell types with higher confidence and resolution than relying on single modalities alone.
- Handling batch imbalance is crucial: The ability to integrate datasets with unequal cell type proportions is a significant practical advantage, as this is a common issue in real-world single-cell experiments.
- Cross-modality insights: The method's ability to learn feature representations (genes, regions, etc.) alongside cell representations opens doors for understanding cross-modality relationships.
Key Results
- Superior performance: scMoMaT consistently outperformed existing methods (MultiMap, UINMF) on simulated and real datasets across various metrics like Graph Connectivity (GC), Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI). For instance, on the human PBMC dataset, scMoMaT showed better performance in preserving cell identity and mixing batches.
- Improved cell type annotation: scMoMaT's integrated cell embeddings combined with learned multi-modal biomarkers led to cell type annotations with higher quality and resolution, sometimes surpassing the original annotations provided with the datasets.
- Effective biomarker discovery: The method successfully identified known and novel cell-type-specific biomarkers across genes, proteins, and chromatin regions, which were crucial for accurate annotation.
- Robustness to batch composition: scMoMaT demonstrated robust performance in integrating datasets with disproportionate cell type compositions, a key advantage for real-world applications.
Limitations and Future Work
- Computational complexity: While not explicitly detailed as a limitation, matrix factorization methods can be computationally intensive, especially with very large datasets.
- Hyperparameter tuning: The method has several hyperparameters (latent dimension, regularization weight, neighbors, radius) that require careful tuning, though the paper provides suggestions.
- Future work: The authors suggest that incorporating feature representations into the integration framework can further enhance the understanding of cross-modality relationships. This hints at future directions for more sophisticated multi-modal integration models.
Summary updated