Ambient RNA

Ambient RNA originates from mRNA molecules released into the cell suspension by cells that are stressed, damaged, or lysed during tissue dissociation and sample preparation. Unlike the intracellular transcriptome captured within an intact cell's droplet, these transcripts float freely in the solution. The primary sources include:

• Mechanical shear during tissue trituration or pipetting
• Enzymatic digestion stress from protocols like collagenase treatment
• Necrotic or apoptotic cells present in the original tissue sample This extracellular pool is then randomly partitioned into droplets alongside intact cells, creating a background signal that is independent of the cell's true biological state.

A defining computational signature of ambient RNA is its low and uniform UMI count profile across contaminated barcodes. Unlike real cells, which exhibit a wide dynamic range of counts with a clear separation between high-expressing and low-expressing genes, ambient-contaminated droplets show:

• Low total UMI counts that blend into the upper tail of the empty droplet distribution
• Absence of a distinct 'knee' in the cumulative count plot for these intermediate barcodes
• Flatter rank-expression curves without the steep drop-off characteristic of genuine cellular barcodes This pattern allows algorithms like EmptyDrops and CellBender to statistically distinguish ambient RNA from true cells by testing for deviations from the ambient expression profile.

Ambient RNA is not a uniform dilution of the bulk transcriptome. Certain transcripts are systematically overrepresented in the ambient pool due to differential RNA stability and cellular susceptibility to lysis. Key biases include:

• High expression of stress-response genes from fragile cell populations that lyse preferentially
• Enrichment for transcripts from large or structurally weak cells, such as hepatocytes or adipocytes
• Depletion of nuclear-enriched transcripts like Malat1 that remain sequestered in intact nuclei
• Overrepresentation of mitochondrial genes from cells undergoing apoptosis during preparation These biases mean that ambient RNA can mimic the expression signature of specific cell types, leading to false-positive differential expression calls if not corrected computationally.

The ambient RNA profile is mathematically modeled as a background 'soup' vector representing the average expression across all droplets classified as empty. This profile is estimated by:

• Aggregating counts from low-UMI barcodes that fall below the cell-containing threshold
• Calculating a weighted mean expression for each gene across these empty droplets
• Accounting for droplet-to-droplet variation in ambient capture rates using methods like the Good-Turing estimator Tools such as SoupX use this estimated soup profile to subtract the expected ambient contribution from each cell's count vector, while CellBender employs a Bayesian framework to jointly estimate the soup and true cellular expression. Accurate soup estimation is critical because errors propagate directly into corrected count matrices.

Uncorrected ambient RNA contamination introduces systematic artifacts that propagate through the entire single-cell analysis pipeline:

• False-positive marker genes: Ambient transcripts from one cell type can appear expressed in another, leading to incorrect cell-type annotation
• Spurious differential expression: Genes abundant in the ambient pool show artificially elevated expression in low-RNA-content cells
• Inflated cell-cell communication signals: Ligand-receptor analysis may detect interactions where ambient transcripts are mistaken for genuine expression
• Distorted trajectory inference: Pseudotime algorithms can be misled by ambient-driven expression gradients that mimic developmental progressions Rigorous ambient RNA removal is therefore a prerequisite for any publication-grade single-cell analysis, particularly in tissues with high cell fragility like brain or tumor samples.

Multiple algorithmic approaches exist to remove ambient RNA contamination, each with distinct assumptions and trade-offs:

• SoupX: Estimates the ambient profile from empty droplets and subtracts a per-cell fraction of this background using a contamination fraction estimated from cell-type-specific marker genes
• CellBender: Uses a deep generative model with a probabilistic latent variable framework to jointly infer true counts and remove background noise, operating directly on raw count matrices without requiring empty droplet data
• DecontX: A Bayesian method from the celda package that models each cell's expression as a mixture of true signal and ambient contamination, estimating both simultaneously
• EmptyDrops: While primarily a cell-calling algorithm, its statistical test for deviation from the ambient profile implicitly identifies and removes barcodes that are purely ambient The choice of method depends on the experimental design, tissue type, and whether empty droplet data was retained during sequencing.

Droplets in microfluidic systems that fail to encapsulate a cell but still contain cell-free mRNA from the suspension medium. These droplets generate barcodes with low total counts that are often filtered out during quality control (QC). However, the transcriptomic profile of empty droplets is the primary source of information for estimating the composition of ambient RNA, as they represent a pure sample of the background contamination without any endogenous cellular signal.

Two dominant approaches to ambient RNA removal with distinct trade-offs:

• SoupX: Fast, interpretable linear model; estimates contamination fraction from bimodal genes; subtracts background counts directly
• CellBender: Slower, deep learning-based; learns a full generative model of noise; produces probabilistic denoised counts
• Key difference: SoupX modifies existing counts, while CellBender learns to reconstruct the true signal, often removing more background but with higher computational cost

The estimated proportion of total UMIs in a cell's barcode that originate from ambient RNA rather than endogenous transcripts. This fraction typically ranges from 2% to 20% depending on cell concentration, sample quality, and tissue type. High contamination fractions are common in samples with fragile cells that lyse easily, releasing abundant cell-free mRNA into the suspension. Accurate estimation of this fraction is critical for downstream differential expression testing and marker gene identification.