Single Cell Ambient RNA Correction
Ambient RNA is RNA present in the cell suspension buffer that is not encapsulated within a cell's droplet. It can be captured and sequenced alongside true cellular transcripts, distorting expression profiles and leading to false cell types or spurious biological conclusions. This guide explains the origins, detection, and computational correction of ambient RNA in single cell RNA sequencing (scRNA seq) data. You should use this guide if you are a bioinformatics analyst, a bench scientist transitioning to analysis, or a student who needs a practical decision framework for handling ambient contamination in droplet based scRNA seq experiments. The concepts and workflows draw on published protocols and community recommendations from sources including the Galaxy Training Network [3] and EMBL EBI Training [2].
At a Glance
| Aspect | Summary |
|---|---|
| What is ambient RNA? | Free floating RNA from lysed or damaged cells present in the droplet solution. |
| Why correct it? | Prevents false positives in differential expression and cell type identification. |
| Common correction methods | SoupX, CellBender, DecontX, and empty droplet filtering. |
| Key decision factors | Level of contamination, expected cell types, and availability of empty droplets. |
| Main challenge | Distinguishing ambient from lowly expressed endogenous RNA without ground truth. |
Core Concepts
Ambient RNA arises from two primary sources. First, cells that lyse during tissue dissociation release their mRNA into the solution. Second, residual RNA in reagents or labware can contribute. In droplet based platforms such as 10x Genomics, each droplet ideally contains one cell and its captured mRNA. Empty droplets contain only barcoded beads and ambient RNA. A fraction of droplets labeled as cells may also contain considerable ambient RNA if the cell count is high or if lysis is extensive. This leads to expression signatures shared across many cells, most notably from highly expressed housekeeping genes or genes from abundant cell types. A study on ambient mRNA contamination in the context of Arabidopsis single cell data demonstrated that even low levels of ambient RNA can confound cluster interpretation [6].
Correction methods generally fall into two categories. The first uses empty droplets to estimate the ambient RNA profile. SoupX and CellBender are examples. The second uses a generative model to decompose observed expression into endogenous and ambient components without requiring empty droplets, as done by DecontX. The key assumption is that ambient RNA is uniformly distributed across droplets, whereas true cellular RNA follows a per cell distribution. The Bioconductor ecosystem provides several packages for these approaches [4]. Understanding these assumptions is critical for choosing the right tool.
Decision Points
1. Do empty droplets exist in your data?
If your experiment used a standard droplet based platform, you have barcodes that correspond to empty droplets. Check the barcode rank plot. A sharp drop in UMI counts often separates cells from empty droplets. If you can identify a clear empty droplet population, you can use SoupX or CellBender. If the separation is ambiguous, you may need DecontX or similar.
2. What is the expected contamination level?
A high contamination level (above 10% of total UMI) will distort differential expression more severely. In such cases a rigorous method like CellBender, which models background using a neural network, may outperform simpler subtraction approaches. For low contamination (under 5%), even basic filtering of highly expressed ambient genes might suffice.
3. Do you have prior knowledge of expected cell types?
If you know that certain genes should only appear in specific cell types, you can evaluate correction quality by checking whether those genes become restricted to the correct cells after correction. This is a useful sanity check as noted in the Galaxy Training Network materials [3].
Practical Workflow
Step 1: Generate raw count matrices
Align reads to the reference genome and produce gene expression count matrices. Use standard pipelines like Cell Ranger or STARsolo. Keep both the cell containing barcodes and empty droplet barcodes for downstream correction. The NCBI Sequence Read Archive [5] can serve as a repository for raw data if you need to access public datasets for testing.
Step 2: Estimate ambient RNA profile from empty droplets
If using SoupX, call autoEstCont(). The function examines expression in empty droplets and estimates a contamination fraction per cell. It also identifies markers for each cluster in the uncorrected data and assumes that those markers should be specific to their cluster. The tool then subtracts ambient contributions proportionally.
Step 3: Apply correction
Run adjustCounts() in SoupX to obtain corrected counts. Alternatively, use CellBender which requires a Python environment. CellBender processes the raw H5 file and outputs a corrected count matrix. For Bioconductor users, DecontX can be applied to a SingleCellExperiment object.
Step 4: Validate normalization and clustering
After correction, re normalize (e.g., scran or Seurat log normalization) and re cluster. Compare cluster markers before and after correction. Genes that were broadly expressed across clusters but are now confined to specific clusters likely were ambient. The transcription fidelity study [7] highlights that ambient contributions can mimic alternative splicing artifacts, so careful validation is warranted.
Step 5: Assess ambient RNA removal for CRISPR screens
If you are analyzing single cell CRISPR screens, ambient guide RNAs (gRNAs) can be a problem. A dedicated tool called CLEANSER was developed to filter ambient gRNAs [9]. It uses a similar principle of identifying guides present in empty droplets and then removing them from cell barcodes. The bioRxiv preprint provides a step by step protocol [11].
Quality Checks
After correction you must verify that the ambient signal has been reduced without removing genuine biological information.
- Check expression of known ambient genes. Housekeeping genes like GAPDH, ACTB, or ribosomal proteins are often highly expressed in lysed cells. Their average expression per cell should drop after correction, but not to zero if cells truly express them.
- Monitor library size changes. A large drop in total UMI per cell may indicate over correction. Ideally the reduction should be modest.
- Run differential expression between clusters. In the corrected data, cluster specific markers should be more distinct. If the same markers appear in multiple unrelated clusters, ambient contamination may still be present.
- Visualize contamination fraction. In SoupX, plot the estimated contamination fraction per cell. High fractions in many cells suggest poor data quality.
The BioConductor documentation [4] advises that no single metric can guarantee perfect correction, you must interpret results in the context of your biological system.
Common Mistakes
1. Forgetting to include empty droplets in the pipeline. Many pipelines discard empty barcodes early. You must retain them for ambient RNA estimation. If you lose that information you cannot use SoupX or CellBender. You will then have to rely on less accurate methods.
2. Using correction on data with very low contamination. If the ambient fraction is below 1%, correction can actually add noise. Always check the estimated contamination fraction before applying a method.
3. Assuming correction works for all gene types. Ambient RNA correction is less reliable for genes with very low expression or for genes that are highly expressed in many cell types. The correction can create false negatives.
4. Over interpreting corrected values as absolute. Corrected counts are estimates. They should not be used for quantitative comparisons like fold change without acknowledging uncertainty. The PLoS One paper on ambient RNA impact strongly advises against treating corrected values as ground truth [8].
5. Ignoring ambient RNA in CRISPR screens. Ambient gRNAs can produce false positive hits. Tools like CLEANSER are essential for those experiments but are often overlooked.
Limits and Uncertainty
Ambient RNA correction has inherent limits. No method can remove all contamination because the true endogenous expression of many genes overlaps with ambient profiles. For instance, a gene that is truly expressed at low levels in many cells cannot be distinguished from a low level of ambient RNA. The corrected data may still contain residual contamination.
Moreover, the assumption that ambient RNA is uniform across droplets is not always true. Droplet size variation, cell lysis during encapsulation, and differences in cell viability across samples can create non uniform ambient RNA distributions. Methods like CellBender partially address this by learning sample specific parameters, but they require high quality empty droplet data.
Interpretation of results must also account for biological RNA that is legitimately shared among cells. Some genes are truly expressed at low levels in many cells. Removing that signal as ambient would be a mistake. Therefore, you cannot rely solely on computational correction, you should combine it with experimental controls like cell hashing or spike in RNA.
The Arabidopsis single cell study [6] and the heat stress study [7] both emphasize that ambient RNA correction is especially challenging when studying stress responses because stress can induce widespread transcriptome changes that mimic ambient contamination. In such cases, validation with orthogonal methods like smFISH or flow cytometry becomes important.
Frequently Asked Questions
Q: Can I correct ambient RNA after clustering? Correcting after clustering is not recommended because ambient RNA can influence the clustering itself. You should correct counts before normalization and clustering for best results.
Q: Do I need empty droplets if I use DecontX? No. DecontX uses a Bayesian model and does not require explicit empty droplets. However, it still performs better when empty droplet information is available as a prior.
Q: How do I know if my correction was successful? Compare the expression of known marker genes for your cell types. If they become more specific to their respective clusters and lose signal in others, correction likely worked. Also check for a decrease in the average expression of suspected ambient genes.
Q: Does ambient RNA correction work for single nucleus RNA seq? Partially. Single nucleus RNA seq has less ambient mRNA from cytoplasm, but it can still have ambient nuclear RNA. The same methods apply but you may need to adjust parameters for lower contamination levels.
References and Further Reading
- Galaxy Training Network: scRNA seq workflow with ambient RNA correction 3
- EMBL EBI Training: Single cell RNA sequencing data analysis 2
- Bioconductor: Packages for single cell analysis including DecontX 4
- NCBI Bookshelf: Overview of RNA sequencing technologies 1
- Understanding and mitigating ambient mRNA contamination in scRNA seq 8
- CLEANSER: filtering ambient gRNAs in single cell CRISPR screens 9
- CLEANSER bioRxiv protocol 11
- Single cell transcriptomics in Arabidopsis with salicylic acid treatment 6
- Transcription fidelity and alternative splicing under heat stress 7
- NCBI Sequence Read Archive for raw sequencing data 5