Abstract
Droplet-based single-cell sequencing techniques rely on the fundamental assumption that each droplet encapsulates a single cell, enabling individual cell omics profiling. However, the inevitable issue of multiplets, where two or more cells are encapsulated within a single droplet, can lead to spurious cell type annotations and obscure true biological findings. The issue of multiplets is exacerbated in single-cell multiomics settings, where integrating cross-modality information for clustering can inadvertently promote the aggregation of multiplet clusters and increase the risk of erroneous cell type annotations. Here, we propose a compound Poisson model-based framework for multiplet detection in single-cell multiomics data. Leveraging experimental cell hashing results as the ground truth for multiplet status, we conducted trimodal DOGMA-seq experiments and generated 17 benchmarking datasets from two tissues, involving a total of 280,123 droplets. We demonstrated that the proposed method is an essential tool for integrating cross-modality multiplet signals, effectively eliminating multiplet clusters in single-cell multiomics data—a task at which the benchmarked single-omics methods proved inadequate.
Introduction
The "one cell, one droplet" paradigm is the cornerstone of droplet-based single-cell sequencing. However, the stochastic nature of cell encapsulation frequently results in multiplets—droplets containing two or more cells. These multiplets generate "ghost" data that can be misinterpreted as novel cell types or transitional states, confounding biological interpretation.
In the burgeoning field of single-cell multiomics, where gene expression (RNA), surface proteins (ADT), and chromatin accessibility (ATAC) are measured simultaneously, the multiplet problem is amplified. Integrating these disparate data layers can cause multiplets to aggregate into distinct, seemingly high-confidence clusters that are entirely artifactual. Current detection tools, primarily designed for single-modality RNA data, often rely on highly variable genes, which are susceptible to biological noise and can lead to false positives.
The COMPOSITE Framework
To address these challenges, we developed COMPOSITE (COMpound POiSson multIplet deTEction), a statistical framework that shifts the focus from variable to stable features.
The framework operates on three core principles:
- Stable Feature Selection: Unlike traditional methods that use highly variable genes, COMPOSITE identifies "stable features," such as housekeeping genes, which exhibit consistent expression across cell types. A significant spike in these features serves as a robust indicator of multiple cells within a droplet.
- Compound Poisson Modeling: The model assumes the number of cells per droplet follows a Poisson distribution. The contribution of each cell to a stable feature is modeled using a Gamma distribution (for RNA and ATAC) or a Gaussian distribution (for proteins).
- Cross-Modality Integration: COMPOSITE dynamically weights signals across modalities. By evaluating the goodness-of-fit (using KS statistics) and droplet-specific consistency, the model can prioritize cleaner signals (e.g., protein) over noisier ones (e.g., RNA) for any given droplet.
Benchmarking and Datasets
The performance of COMPOSITE was rigorously evaluated using a massive benchmarking effort. We generated 17 new trimodal DOGMA-seq datasets from peripheral blood and ileum tissue, encompassing 280,123 droplets.
To establish a "gold standard" for multiplet status, we utilized cell hashing (HTO labeling). This experimental technique allows for the identification of multiplets by labeling cells from different samples with unique oligonucleotide tags before pooling. This ground truth enabled a direct comparison between COMPOSITE and existing single-omics tools like scDblFinder and DoubletFinder.
Key Results and Discussion
COMPOSITE consistently outperformed state-of-the-art single-omics methods, achieving higher F1 scores across all benchmarking datasets. A critical advantage of COMPOSITE is its ability to eliminate spurious clusters in UMAP visualizations. For instance, clusters annotated as rare hybrid cell types (e.g., CD3+/CD19+ cells) were correctly identified as doublets and removed, leading to a cleaner representation of the underlying biology.
Interestingly, COMPOSITE occasionally outperformed the experimental "gold standard." It successfully identified "homotypic" multiplets—droplets containing two cells of the same type—which cell hashing cannot detect if the cells share the same tag. By modeling the statistical distribution of stable features, COMPOSITE can discern these events. Furthermore, the model accounts for a "decline effect," where resources per cell decrease as more cells enter a droplet, improving the accuracy of triplet detection.
Conclusion
COMPOSITE provides a unified, model-based solution for multiplet detection in multiomics data. By leveraging stable features and cross-modality weighting, it offers a significant improvement over single-modality tools. While the model relies on specific parametric assumptions, its efficiency—enhanced by GPU acceleration—makes it a practical and essential component of modern single-cell analysis pipelines. Future work may extend this framework to spatial transcriptomics, where cell count estimation per spot remains a significant challenge.
Summary updated