TL;DR
Single-nucleus RNA sequencing (snRNA-seq) is powerful for studying solid tissues but is plagued by "ambient RNA" contamination. This paper introduces DIEM, a novel machine learning approach that effectively filters out these contaminated droplets, leading to cleaner data and more accurate cell type identification in snRNA-seq experiments.
What Problem Does This Paper Address?
Single-cell and single-nucleus RNA sequencing (sc/snRNA-seq) have revolutionized our ability to understand cellular heterogeneity by measuring gene expression at the individual cell or nucleus level. snRNA-seq is particularly valuable for solid tissues where isolating intact cells is challenging. However, a significant hurdle in snRNA-seq is contamination from "ambient RNA" – RNA released from lysed cells or extracellular components. This ambient RNA can be captured in the same droplets as nuclei, leading to misleading gene expression profiles. If not properly filtered, this contamination can result in the identification of spurious cell types and biased downstream analyses. Existing methods, like hard count thresholds or the EmptyDrops algorithm, often struggle to effectively remove this contamination in snRNA-seq, especially from challenging samples like frozen tissues.
Methods
The core of this paper is the development of DIEM (Debris Identification using Expectation Maximization). DIEM is a semi-supervised machine learning approach designed to quantify and filter out debris-contaminated droplets in snRNA-seq data.
Here's a breakdown of the DIEM approach:
- Clustering and Initialization: DIEM first clusters all droplets using a multinomial mixture model. To initialize this model, it employs k-means clustering on droplets with at least 200 detected genes (to avoid fitting empty droplets).
- Semi-Supervised Expectation-Maximization (EM): The EM algorithm is then used to estimate the parameters of the mixture model. Crucially, droplets with very low UMI counts (below 100) are fixed as debris, making the process semi-supervised. This allows DIEM to model both debris and cell type RNA distributions.
- Debris Scoring: Instead of simply removing debris clusters, DIEM calculates a "debris score" for each individual droplet. This score is based on the expression of genes identified as being enriched in the debris set (determined through differential expression analysis between debris and cell type clusters).
- Filtering: Droplets are then filtered based on this debris score, with a user-defined threshold. This provides a more nuanced approach than hard cutoffs.
DIEM was also tested on single-cell RNA-seq (scRNA-seq) data, though its performance was less distinct compared to existing methods in that context.
Datasets
The authors evaluated DIEM using three types of snRNA-seq datasets:
- Human differentiating preadipocytes in vitro (DiffPA): A controlled experimental system.
- Fresh mouse brain tissue: A common and well-studied tissue type.
- Human frozen adipose tissue (AT): Six individuals' samples, representing a challenging and clinically relevant sample type where nuclei isolation is often preferred over cell isolation.
They also mention testing DIEM on a single-cell RNA-seq dataset of 68,000 PBMCs, though the primary focus and strongest results are for snRNA-seq.
Key Insights and Ideas
- Ambient RNA is a pervasive problem in snRNA-seq: The study provides clear evidence that extranuclear RNA contamination is common across various tissue types and sample preparations (fresh vs. frozen).
- Existing methods fall short for snRNA-seq: Hard count thresholds and even the EmptyDrops algorithm, while useful for scRNA-seq, are shown to be suboptimal for snRNA-seq, often removing valuable nuclear droplets or failing to adequately remove debris.
- Gene expression profiles differentiate debris from nuclei: The key insight driving DIEM is that debris and nuclear droplets have distinct, albeit sometimes subtle, gene expression profiles. This allows for a more sophisticated filtering approach than just relying on read counts.
- Debris scores offer finer control: Moving beyond binary classification (debris vs. nucleus), the debris score provides a quantitative measure of contamination, allowing for flexible filtering and potential use as a covariate in downstream analyses.
- Frozen tissues are particularly challenging: The study highlights that frozen tissues, while enabling snRNA-seq when cell isolation is impossible, present a greater challenge for debris removal due to potentially higher levels of ambient RNA. DIEM shows particular benefit here.
Key Results
- DIEM effectively removes debris-contaminated droplets: Across all tested datasets, DIEM consistently removed a higher proportion of background (debris) droplets compared to quantile-based filtering and EmptyDrops.
- DIEM preserves nuclear droplets: Importantly, DIEM removed fewer nuclear droplets than the other methods, leading to a higher proportion of nuclear droplets in the final filtered datasets. For example, in adipose tissue, DIEM resulted in 95.7% nuclear droplets, compared to 87.9% for EmptyDrops and 84.5% for quantile filtering.
- DIEM leads to cleaner clusters: UMAP visualizations and cluster analysis showed that DIEM-filtered data resulted in clusters with lower percentages of spliced reads (a proxy for extranuclear RNA), indicating less contamination. In some cases, DIEM removed entire "debris clusters" that persisted with other methods.
- DIEM's debris score correlates well with contamination metrics: The debris score calculated by DIEM showed a strong correlation (mean R=0.89) with the percentage of spliced reads, validating its effectiveness as a contamination metric.
- DIEM is computationally efficient: While not explicitly quantified in terms of speedup, the paper emphasizes DIEM's ability to perform fast and effective filtering.
Limitations and Future Work
- Debris score normalization: The authors note that debris scores are normalized relative to the best and worst cluster means within a sample. This means scores may not be directly comparable across independent experiments, especially if extranuclear RNA levels differ significantly. Integrating multiple samples might require further normalization strategies.
- Reliance on debris-enriched genes: The accuracy of the debris score depends on the accurate identification of debris-enriched genes. If there are few such genes, the score might be less reliable.
- PBMC scRNA-seq performance: While DIEM was tested on scRNA-seq data, its advantage was less pronounced compared to EmptyDrops. The authors suggest this might be due to smaller differences between background and cell type profiles in scRNA-seq.
- Future work: The authors suggest that DIEM's ability to quantify contamination could be leveraged as a covariate in downstream analyses like clustering and differential expression. They also highlight the potential for DIEM to enable more robust analysis of frozen tissue snRNA-seq data, which is crucial for leveraging biobanked samples.
Summary updated