Understanding UMI in Single-Cell Sequencing: What It Is and Why It Matters

Single-cell RNA sequencing (scRNA-seq) has transformed how researchers study cellular heterogeneity, but the technology introduces a fundamental problem: the amplification step that copies RNA molecules before sequencing does not amplify every molecule equally. Unique molecular identifiers (UMIs) are short random nucleotide sequences attached to individual RNA molecules before amplification, allowing researchers to count the original number of molecules instead of the number of amplified copies. This article explains what UMIs are, how they reduce amplification bias, and how they affect data interpretation for students, researchers, analysts, and life-science professionals working with single-cell datasets.

The Amplification Problem in Single-Cell Sequencing

The starting material in scRNA-seq is tiny. A single mammalian cell contains only picograms of RNA, far below the amount required for standard sequencing library preparation. To obtain enough material for sequencing, researchers must amplify the RNA through polymerase chain reaction (PCR) or in vitro transcription. This amplification step introduces bias because not all molecules amplify with equal efficiency. Some transcripts are amplified more than others due to sequence composition, length, or random variation in reaction conditions. The result is that the number of sequencing reads for a given gene does not directly reflect the number of RNA molecules that were originally present in the cell.

Without a correction mechanism, this amplification bias would make it impossible to distinguish true biological differences in gene expression from technical artifacts introduced during library preparation. The problem is particularly acute in single-cell experiments because the starting material is so limited that even small biases can have a large impact on the final count.

UMIs solve this problem by tagging each original RNA molecule with a unique sequence before amplification. After sequencing, reads that share the same UMI and map to the same gene are collapsed into a single count, representing one original molecule. This process, called UMI deduplication, removes the effect of amplification bias because the number of unique UMIs observed for a gene reflects the number of original molecules, regardless of how many times each molecule was amplified.

What Is a Unique Molecular Identifier

A UMI is a short random oligonucleotide sequence, typically 8 to 12 bases long, that is incorporated into the sequencing library during reverse transcription. Each RNA molecule in a cell receives a different UMI, so the UMI acts as a molecular barcode that identifies the original molecule. When the library is amplified and sequenced, reads that carry the same UMI and align to the same genomic position are considered copies of the same original molecule.

The design of UMIs requires careful consideration of the random sequence space. A 10-base UMI has 4 to the power of 10, or approximately one million, possible sequences. This is generally sufficient for the number of RNA molecules present in a single cell, but the probability of two different molecules receiving the same UMI, known as a collision, increases with the number of molecules and the length of the UMI. Most commercial protocols use UMIs that are long enough to make collisions rare for typical cellular RNA content.

UMIs are distinct from cell barcodes, which identify which cell a read came from in droplet-based or plate-based multiplexed experiments. Cell barcodes are constant within a cell but vary between cells, while UMIs vary within a cell but are unique to each molecule. Both types of sequences are read during sequencing and used during data processing to assign reads to cells and to count molecules.

How UMIs Reduce Amplification Bias

The key advantage of UMIs is that they allow the counting of molecules instead of reads. In a traditional RNA-seq experiment without UMIs, the number of reads mapped to a gene is used as a proxy for expression level. This assumes that amplification is uniform, which is not true. With UMIs, the number of unique UMI sequences observed for a gene is the count of original molecules, and the number of reads per UMI is irrelevant for quantification.

This distinction matters for data interpretation. A gene with 100 original molecules that are amplified unevenly might produce 1,000 reads, while another gene with 50 original molecules that amplify efficiently might also produce 1,000 reads. Without UMIs, both genes would appear equally expressed. With UMIs, the first gene would be counted as 100 molecules and the second as 50, revealing the true difference in expression.

The reduction of amplification bias also affects the statistical properties of the data. UMI-based counts are digital, meaning they are integer counts of molecules, and they follow a Poisson distribution more closely than read counts from non-UMI protocols. This property is important for downstream analysis because many statistical methods assume a particular distribution for the data. Research has shown that the Poisson distribution model fits UMI-based scRNA-seq data well, which supports the use of statistical methods that rely on this assumption [12].

UMI-Based Versus Full-Length Protocols

Single-cell sequencing protocols fall into two broad categories: UMI-based protocols and full-length protocols. UMI-based protocols, such as those using droplet-based methods commercialized by 10x Genomics, capture only the terminal portion of transcripts, typically the 3-prime end. These protocols can process thousands of cells in a single run and are well suited for identifying cell types and states through gene expression profiling. The tradeoff is that they do not provide information about the full transcript sequence, so they cannot easily detect isoforms or allele-specific expression [6].

Full-length protocols, such as Smart-seq and its derivatives, sequence the entire transcript and provide higher sensitivity for detecting lowly expressed genes and isoforms. However, these protocols are generally lower throughput and more expensive per cell. The choice between UMI-based and full-length protocols depends on the biological question. If the goal is to identify cell types in a complex tissue, a UMI-based protocol with many cells may be appropriate. If the goal is to study isoform usage or detect subtle differences in gene expression, a full-length protocol may be necessary despite the lower throughput [6].

The difference in read depth requirements between protocols is also important. UMI-based protocols require fewer reads per cell because the UMI collapses reads into molecule counts. Full-length protocols require higher read depth to achieve adequate coverage of the full transcript. Research comparing protocols for spatial reconstruction analysis found that gene expression profiles were highly reproducible between datasets generated by different protocols, but that the optimal balance between sequencing depth and number of cells depends on the specific biological question [11].

The Role of UMIs in Data Preprocessing

Data preprocessing for UMI-based scRNA-seq involves several steps that are not required for bulk RNA-seq. These steps include cell barcode identification, UMI deduplication, and gene count estimation. The choice of preprocessing workflow can affect the final count matrix and downstream analysis results.

A benchmarking study compared 10 end-to-end preprocessing workflows, including Cell Ranger, Optimus, salmon alevin, alevin-fry, kallisto bustools, dropSeqPipe, scPipe, zUMIs, celseq2, and scruff, using datasets generated by CEL-Seq2 and 10x Chromium platforms. The study found that the workflows varied in their detection and quantification of genes across datasets, but that after downstream analysis with performant normalization and clustering methods, almost all combinations produced clustering results that agreed well with known cell type labels. The choice of preprocessing method was found to be less important than other steps in the scRNA-seq analysis process [10].

UMI deduplication methods differ in how they handle reads that map to multiple genes or that have sequencing errors in the UMI sequence. Some tools discard gene-ambiguous reads, while others account for transcript-level constraints on the molecules from which UMIs may have arisen. Alevin, for example, considers both gene-unique reads and reads that multimap between genes, which improves the accuracy of gene abundance estimates compared to tools that discard ambiguous reads [17].

The scruff package for R and Bioconductor provides preprocessing for CEL-Seq and CEL-Seq2 protocols, including demultiplexing, alignment, counting, and UMI deduplication. It also provides visualization of quality metrics and read alignments for individual cells across multiple samples or runs [16].

At a Glance: UMI Workflow Decisions

Normalization of UMI Count Data

After preprocessing produces a count matrix, normalization is required to account for technical variation between cells. The number of molecules detected in each cell varies due to differences in cell size, RNA capture efficiency, and sequencing depth. This technical variation can confound biological heterogeneity if not properly addressed.

A modeling framework based on regularized negative binomial regression has been developed for normalization and variance stabilization of UMI-based scRNA-seq data. This approach uses cellular sequencing depth as a covariate in a generalized linear model and computes Pearson residuals that remove the influence of technical characteristics while preserving biological heterogeneity. The method pools information across genes with similar abundances to obtain stable parameter estimates, avoiding the overfitting that can occur with an unconstrained negative binomial model. This approach does not require pseudocount addition or log-transformation and improves downstream tasks such as variable gene selection, dimensional reduction, and differential expression [5].

The choice of normalization method matters for downstream analysis. Non-UMI data requires different normalization approaches than UMI data because the statistical properties of the counts differ. Methods developed specifically for UMI data, such as analytic Pearson residuals, are designed to account for the Poisson-like distribution of UMI counts [20]. For non-UMI data, quantile normalization and non-linear normalization methods have been developed to address the different noise characteristics [18][19].

Sequencing Depth and Number of Cells

One of the key decisions in designing a single-cell experiment is how to allocate sequencing resources between depth and cell number. Sequencing more cells increases the statistical power to detect rare cell types, while sequencing each cell more deeply increases the sensitivity to detect lowly expressed genes.

Research comparing protocols with different depth and cell number characteristics found that optimizing the balance between sequencing depth and number of cells within a protocol is necessary for efficient use of resources. UMI-based protocols such as 10x and MARS-seq allow for capturing more cells, while Smart-seq provides higher sensitivity and read depth for analysis of lower expressed genes and isoforms [11].

For UMI-based protocols, the relationship between sequencing depth and molecule detection is not linear. As sequencing depth increases, the number of new UMIs detected per additional read decreases because the same molecules are being re-sequenced. This saturation effect means that there is an optimal depth beyond which additional sequencing provides diminishing returns. The optimal depth depends on the complexity of the library, which is determined by the number of cells and the amount of RNA per cell.

Common Failure Patterns in UMI Analysis

Several common errors can compromise the quality of UMI-based single-cell analysis. Understanding these failure patterns helps researchers avoid them and interpret results correctly.

The first failure pattern is inadequate UMI length or complexity. If the UMI is too short or the number of molecules in a cell is too high, two different molecules can receive the same UMI sequence. This collision leads to undercounting of molecules and can distort gene expression measurements. Most commercial protocols use UMIs that are long enough to make collisions rare, but researchers should be aware of this limitation when working with cells that have very high RNA content.

The second failure pattern is sequencing errors in the UMI sequence. If a sequencing error occurs in the UMI, the read may be incorrectly assigned to a different UMI, leading to overcounting of molecules. Many preprocessing workflows include error correction steps that collapse UMIs that differ by a single base, but these methods can also introduce errors if they collapse UMIs that are genuinely different.

The third failure pattern is the use of inappropriate normalization methods. Applying normalization methods designed for non-UMI data to UMI data can introduce artifacts because the statistical properties of the counts differ. Similarly, applying UMI-specific methods to non-UMI data can produce incorrect results. Researchers should match the normalization method to the data type.

The fourth failure pattern is ignoring the impact of preprocessing choices on downstream analysis. While the benchmarking study found that preprocessing method choice was less important than other steps in the analysis process, the choice still affects the count matrix and can influence results for specific genes or cell types [10]. Researchers should document their preprocessing choices and consider how they might affect their conclusions.

Quality Control and Assessment

Quality control is essential for UMI-based single-cell analysis. The first step is to assess the quality of the raw sequencing data, including the read quality scores, the proportion of reads that contain valid cell barcodes and UMIs, and the alignment rate to the reference genome.

After preprocessing, quality metrics for individual cells should be examined. These metrics include the number of UMIs detected per cell, the number of genes detected per cell, the proportion of reads mapping to mitochondrial genes, and the proportion of reads mapping to ribosomal genes. Cells with very low UMI counts may be empty droplets or damaged cells, while cells with very high mitochondrial read proportions may be stressed or dying.

The scruff package provides visualization of quality metrics for cells from multiple experiments, including pre- and post-alignment metrics. It also supports visualization of quality metrics for sequence alignment files generated by Cell Ranger from 10x Genomics [16].

Simulation tools can be used to benchmark analysis methods and to generate ground truth data for method development. scReadSim is a simulator that generates synthetic sequencing reads by mimicking real scRNA-seq and scATAC-seq data. It provides ground truths, including UMI counts for scRNA-seq and open chromatin regions for scATAC-seq. Benchmark applications of scReadSim showed that UMI-tools achieved the top accuracy in scRNA-seq UMI deduplication [8].

Data Interpretation Limits

UMI-based counts provide a more accurate measure of gene expression than read counts, but they have limitations that affect data interpretation. The most important limitation is that UMI-based protocols capture only the terminal portion of transcripts, so they cannot distinguish between different isoforms of the same gene. If a gene has multiple isoforms that are expressed at different levels, the UMI count represents the total expression of all isoforms, not the expression of each isoform individually.

Another limitation is that UMI-based protocols have limited sensitivity for detecting lowly expressed genes. The capture efficiency of RNA molecules varies between protocols and between cells, and genes expressed at very low levels may not be detected at all. This dropout phenomenon creates a large number of zeros in the count matrix, which must be handled appropriately in downstream analysis.

The Poisson distribution model fits UMI-based scRNA-seq data well, but this does not mean that all variation in the data is technical. Biological variation between cells, including differences in cell cycle stage, metabolic state, and microenvironment, contributes to the observed heterogeneity. Distinguishing biological from technical variation is a central challenge in single-cell data analysis [12].

Reproducibility and Reporting

Reproducibility in single-cell analysis requires careful documentation of all steps in the workflow, from sample preparation through data analysis. This includes the protocol used for library preparation, the sequencing platform and depth, the preprocessing workflow and version, the normalization method, and the parameters used for downstream analysis.

The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable [4]. Applying these principles to single-cell data involves depositing raw sequencing data in public repositories, providing processed count matrices, and documenting the analysis workflow in sufficient detail for others to reproduce the results.

The National Center for Biotechnology Information provides data resources for depositing and accessing sequencing data [2]. The NIH Genomic Data Sharing Policy outlines expectations for data sharing and management for NIH-funded research [3]. Researchers should be aware of these requirements when planning their experiments and data management strategies.

The EMBL-EBI Training program offers resources for learning about bioinformatics methods, including single-cell analysis [1]. These resources can help researchers develop the skills needed to analyze UMI-based single-cell data correctly.

Practical Assessment Steps

When working with UMI-based single-cell data, researchers should follow a systematic assessment process to ensure data quality and correct interpretation.

First, assess the raw sequencing data quality. Check the read quality scores, the proportion of reads with valid cell barcodes and UMIs, and the alignment rate. Low alignment rates may indicate contamination or problems with the reference genome.

Second, examine the distribution of UMI counts across cells. A typical experiment should show a clear separation between cells with high UMI counts and empty droplets with low counts. The presence of many cells with intermediate UMI counts may indicate problems with cell capture or library preparation.

Third, evaluate the relationship between sequencing depth and UMI detection. Plot the number of UMIs detected per cell against the number of reads per cell. This plot should show a saturation curve, where the number of new UMIs detected per additional read decreases as depth increases. If the curve does not saturate, the sequencing depth may be insufficient.

Fourth, assess the performance of the preprocessing workflow. Compare the number of genes detected per cell and the total number of UMIs across cells. Check for batch effects between samples or runs that may indicate technical variation.

Fifth, validate the results of downstream analysis. If clustering reveals cell types, verify that the marker genes for each cluster are consistent with known biology. If differential expression analysis identifies genes, check that the effect sizes are biologically plausible.

Professional Escalation Criteria

Certain situations warrant escalation to a bioinformatics specialist or the protocol developer. If the alignment rate is unexpectedly low, if the UMI collision rate is high, or if the saturation curve does not reach a plateau at the planned sequencing depth, the experiment may need to be redesigned.

If preprocessing workflows produce substantially different count matrices for the same raw data, the choice of workflow should be investigated. While the benchmarking study found that preprocessing choice was less important than other steps, large discrepancies between workflows may indicate problems with the data or the analysis [10].

If normalization methods produce conflicting results, the statistical properties of the data should be examined. UMI data that does not follow the expected distribution may indicate problems with library preparation or sequencing.

If downstream analysis results are not reproducible across replicates or across different analysis methods, the biological interpretation should be reconsidered. Technical artifacts can produce spurious biological signals, and the interpretive rigor required for single-cell data analysis is similar to that required for other high-dimensional genomic technologies [13].

Frequently Asked Questions

What is the difference between a UMI and a cell barcode?

A cell barcode is a constant sequence that identifies which cell a read came from in a multiplexed experiment. All reads from the same cell carry the same cell barcode. A UMI is a random sequence that identifies which original RNA molecule a read came from. Each molecule in a cell receives a different UMI, so reads with the same UMI and the same gene alignment are considered copies of the same original molecule. Cell barcodes assign reads to cells, while UMIs count molecules within cells.

Why do UMI-based protocols require less sequencing depth than full-length protocols?

UMI-based protocols collapse reads into molecule counts, so multiple reads from the same molecule are counted once. This means that additional sequencing depth beyond the point of saturation provides little new information. Full-length protocols do not have this collapse step, so each read provides independent information about the transcript sequence. Higher read depth is needed to achieve adequate coverage of the full transcript and to detect lowly expressed genes and isoforms [11].

How does UMI length affect the accuracy of molecule counting?

The UMI length determines the number of possible UMI sequences. A longer UMI provides more unique sequences and reduces the probability that two different molecules receive the same UMI. If the UMI is too short relative to the number of molecules in a cell, collisions occur and molecules are undercounted. Most commercial protocols use UMIs that are long enough to make collisions rare for typical cellular RNA content, but cells with very high RNA content may require longer UMIs.

What is UMI deduplication and why is it necessary?

UMI deduplication is the process of collapsing reads that share the same UMI and map to the same gene into a single count. This process is necessary because amplification creates multiple copies of each original molecule, and without deduplication, the number of reads would reflect amplification efficiency instead of the original number of molecules. Deduplication produces a digital count of molecules that is not affected by amplification bias.

How do preprocessing workflows differ in their UMI handling?

Preprocessing workflows differ in how they handle reads that map to multiple genes, how they correct sequencing errors in UMIs, and how they assign reads to genes. Some tools discard gene-ambiguous reads, while others account for transcript-level constraints on the molecules from which UMIs may have arisen. A benchmarking study found that the choice of preprocessing method was less important than other steps in the analysis process, but the workflows varied in their detection and quantification of genes across datasets [10].

What normalization method should I use for UMI-based data?

Normalization methods developed specifically for UMI data, such as regularized negative binomial regression and analytic Pearson residuals, are designed to account for the Poisson-like distribution of UMI counts [5][20]. These methods use cellular sequencing depth as a covariate and remove technical variation while preserving biological heterogeneity. Methods designed for non-UMI data may not be appropriate because the statistical properties of the counts differ.

Can I use UMI-based data to study isoform expression?

UMI-based protocols capture only the terminal portion of transcripts, so they cannot distinguish between different isoforms of the same gene. The UMI count represents the total expression of all isoforms, not the expression of each isoform individually. Full-length protocols such as Smart-seq are required for isoform analysis because they sequence the entire transcript [6].

How do I know if my sequencing depth is sufficient?

The relationship between sequencing depth and UMI detection should show a saturation curve, where the number of new UMIs detected per additional read decreases as depth increases. If the curve does not saturate, the sequencing depth may be insufficient. The optimal depth depends on the complexity of the library, which is determined by the number of cells and the amount of RNA per cell. Subsampling analyses can be used to evaluate the tradeoffs between sequencing depth and number of cells within a protocol [11].

Related Bioinformatics Guides

• Single-Cell RNA Sequencing: From Bulk to Resolution
• Master Guide: Single-Cell RNA Sequencing Bioinformatics Workflows
• Foundation Models for Single-Cell Biology
• Single-Cell ATAC-Seq Bioinformatics
• The Role of ENCODE in Understanding the Non-Coding Genome

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression.. Genome biology, 2019.
• Full-Length Single-Cell RNA-Sequencing with FLASH-seq.. Methods in molecular biology (Clifton, N.J.), 2023.
• Single-cell RNA sequencing of human liver reveals hepatic stellate cell heterogeneity.. JHEP reports : innovation in hepatology, 2021.
• scReadSim: a single-cell RNA-seq and ATAC-seq read simulator.. Nature communications, 2023.
• Computational Analysis of Single-Cell RNA-Seq Data.. Methods in molecular biology (Clifton, N.J.), 2023.
• Benchmarking UMI-based single-cell RNA-seq preprocessing workflows.. Genome biology, 2021.
• Reproducibility across single-cell RNA-seq protocols for spatial ordering analysis.. PloS one, 2020.
• The Poisson distribution model fits UMI-based single-cell RNA-sequencing data.. BMC bioinformatics, 2023.
• The ctDNA Paradigm: Dynamic Observation, Quantitative Analysis, and Interpretive Limits in Precision Oncology.. 2026.
• Mining Two Decades of Soybean Genomics Literature Using Rule-Based Text Mining: Chromosome-Resolved Patterns of &lt,i&gt,Glyma&lt,/i&gt, Gene Mentions.. 2026.
• IRescue: single cell uncertainty-aware quantification of transposable elements expression. 2022.
• scruff: an R/Bioconductor package for preprocessing single-cell RNA-sequencing data. BMC Bioinformatics, 2019.
• Alevin efficiently estimates accurate gene abundances from dscRNA-seq data. Genome Biology, 2018.
• Quantile normalization of single-cell RNA-seq read counts without unique molecular identifiers. Genome Biology, 2020.
• Non-linear Normalization for Non-UMI Single Cell RNA-Seq. Frontiers in Genetics, 2021.
• Analytic Pearson residuals for normalization of single-cell RNA-seq UMI data. Genome Biology, 2021.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.