TL;DR
This paper introduces scMoMaT, a novel method for integrating single-cell multi-omics data across different batches and modalities, even when cell type compositions are uneven. Beyond just integrating data, scMoMaT excels at identifying multi-modal bio-markers that improve cell type annotation and resolution, outperforming existing methods on various simulated and real-world datasets.
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 individual cells. However, integrating data from different experiments (batches) and different types of measurements (modalities) is a significant challenge. This integration is crucial for a comprehensive understanding of cellular identity and the complex relationships between different biological layers. The authors categorize data integration into four scenarios: horizontal (batch effect removal), vertical (multiple modalities in one batch), diagonal (no shared cells or modalities), and mosaic (any combination). Mosaic integration is the most general and challenging, with few existing methods capable of handling it effectively, especially when dealing with unequal cell type distributions across batches. This paper aims to fill this gap by providing a robust method for mosaic integration that also leverages the multi-modal nature of the data to discover biologically meaningful markers.
Methods
The core of scMoMaT is a matrix tri-factorization framework. It treats each single-cell data matrix (representing cells from a specific batch and modality) as a relationship matrix between "cells" and "features." This matrix is decomposed into three components: a cell factor, a feature factor, and an association matrix. Crucially, scMoMaT shares these factors across different data matrices within the same integration task, allowing it to learn a unified representation.
Key novelties include:
- Mosaic Integration: scMoMaT is designed to handle arbitrary combinations of horizontal, vertical, and diagonal integration scenarios.
- Multi-Modal Bio-Marker Discovery: Beyond cell embedding, scMoMaT explicitly learns cluster-specific bio-markers across modalities (genes, chromatin regions, proteins, and even motifs). These markers are derived from the learned factors and are used to interpret and annotate cell clusters.
- Handling Unequal Cell Type Compositions: The method is designed to integrate cell batches that have disproportionate cell type distributions, a common issue in real-world datasets.
- Pseudo-count Matrix Generation: scMoMaT includes a method to generate pseudo-count matrices to fill in missing modalities, enabling integration even when not all modalities are present in every batch.
- Post-processing and Retraining: After initial factor learning, a post-processing step involves constructing a neighborhood graph for visualization and clustering. The learned cluster labels are then used to retrain the model, allowing for the extraction of feature scoring matrices that highlight bio-marker importance for each cluster.
Datasets
The authors evaluated scMoMaT on a combination of simulated and real-world datasets:
- Simulated Datasets: Used to quantitatively evaluate performance under various challenging integration scenarios, including unequal cell type compositions, imbalanced batch sizes, and rare cell types. These datasets were generated using a generalized version of the SymSim simulator.
- Real Datasets:
- Human Peripheral Blood Mononuclear Cells (PBMC) dataset (multiple batches, gene expression, and protein abundance).
- Mouse Brain Cortex dataset (multiple batches, chromatin accessibility, and gene expression).
- Human Bone Marrow Mononuclear Cells (BMMC) dataset (two batches, scATAC-seq and scRNA-seq).
- Mouse Spleen dataset (two batches, scATAC-seq and scRNA-seq, used to test handling of unequal cell type compositions).
The paper also introduces a benchmark by comparing scMoMaT's bio-marker detection accuracy against a baseline pipeline (UINMF + Wilcoxon rank-sum test) using simulated data with known ground truth markers.
Key Insights and Ideas
- Mosaic Integration is the Ultimate Challenge: The paper highlights that "mosaic integration" is the most general and difficult scenario, requiring methods that can flexibly combine data from various sources and modalities.
- Bio-markers are Key to Annotation: Simply integrating data to get a better cell embedding is not enough. The ability to derive biologically interpretable markers from the integrated data is crucial for robust cell type annotation and understanding.
- Unequal Cell Type Distributions are Common and Addressable: Real-world datasets often have batches with vastly different proportions of cell types. scMoMaT's ability to handle this is a significant practical advantage.
- Multi-modality Enhances Resolution: By integrating information from multiple biological layers, scMoMaT can achieve higher resolution cell type annotations, potentially identifying sub-types that are not apparent from single modalities alone.
- Matrix Tri-Factorization as a Unified Framework: The choice of matrix tri-factorization provides a flexible and powerful framework for jointly learning cell representations, feature representations, and their relationships across modalities and batches.
Key Results
- Superior Performance: scMoMaT consistently outperformed existing methods (MultiMap, UINMF, StabMap, LIGER, Seurat) in terms of graph connectivity (GC), Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI) scores across various simulated and real datasets.
- Improved Cell Type Annotation: scMoMaT's integrated cell embeddings, combined with learned multi-modal bio-markers, led to cell type annotations of higher quality and resolution compared to original annotations. For example, on the human PBMC dataset, scMoMaT's annotations showed better agreement with cell-cell variations within individual data matrices.
- Effective Bio-marker Detection: The bio-markers identified by scMoMaT were highly consistent with known markers and provided strong evidence for cell type annotations, even identifying sub-types within broader categories (e.g., naive vs. activated T cells).
- Robustness to Data Challenges: scMoMaT demonstrated strong performance in integrating datasets with unequal cell type compositions and imbalanced batch sizes.
- Quantitative Improvements: For instance, on simulated datasets, scMoMaT achieved higher ARI and NMI scores, indicating better separation of cell types. On the human PBMC dataset, scMoMaT's learned labels showed improved kNN agreement compared to the original labels.
Limitations and Future Work
- Hyperparameter Sensitivity: While generally robust, the paper notes that the latent dimension
dcan be sensitive to dataset complexity, and a smallerdmay lead to poorer results. The radius parameterrin post-processing can also impact performance, especially with significant cell type mismatches. - Computational Cost: While not explicitly detailed as a limitation, matrix factorization methods can be computationally intensive, especially for very large datasets.
- Future Work: The authors suggest that incorporating feature representations (e.g., gene representations) into the integration framework can further aid in learning cross-modality relationships. They also envision scMoMaT being widely applied to the growing number of single-cell multi-omics datasets. The inclusion of motif deviation matrices in the retraining step is also highlighted as a way to enhance bio-marker discovery.
Summary updated