Unique Molecular Identifier (UMI)

UMIs enable digital counting of individual mRNA molecules rather than relative read abundance. Each transcript receives a unique random barcode before amplification, so after sequencing, reads sharing the same UMI and gene mapping are collapsed into a single count. This eliminates PCR amplification bias, where highly amplifiable transcripts would otherwise dominate the library. The result is a count matrix where each entry represents the number of distinct molecules detected per gene per cell, enabling valid cross-gene and cross-cell comparisons.

During library preparation, a single cDNA molecule can generate multiple identical sequencing reads through exponential amplification. Without UMIs, these PCR duplicates are indistinguishable from independent sampling events, inflating expression estimates. UMI-based deduplication works by:

• Grouping reads by cell barcode + UMI + gene
• Retaining only one representative read per unique combination
• Discarding reads likely arising from PCR errors within the UMI sequence itself This process dramatically reduces technical noise and improves the accuracy of downstream differential expression testing.

UMI sequences are short (typically 8-12 nucleotides) and susceptible to PCR and sequencing errors. A single nucleotide substitution can create a spurious UMI, artificially inflating molecular counts. Computational correction methods include:

• Hamming distance clustering: merging UMIs that differ by one base if one has substantially lower read support
• Directional adjacency: collapsing UMIs only when the more abundant UMI could have generated the less abundant one via a known error profile
• Network-based correction: constructing UMI similarity graphs and applying community detection Proper correction recovers 10-30% of lost molecules while minimizing false mergers.

UMI counts exhibit saturation behavior as sequencing depth increases. Initially, each additional read yields a new UMI, but eventually most reads represent re-sampling of already-observed molecules. The library complexity—the total number of distinct UMIs—plateaus, defining the point of diminishing returns. Key metrics:

• Saturation curve: plots unique UMIs detected vs. reads sequenced
• Sequencing saturation: the fraction of reads that are duplicates of previously observed UMIs
• Typical recommendations target 50-80% saturation, balancing cost against the marginal gain of deeper sequencing for rare transcript detection.

The information capacity of a UMI depends on its length and nucleotide composition. A fully random 10-nucleotide UMI provides 4^10 (~1 million) possible sequences, sufficient for most single-cell experiments where a few thousand transcripts are captured per cell. Design considerations include:

• Collision probability: the chance two distinct molecules receive the same UMI, calculated via the birthday paradox
• GC balance: avoiding extreme GC content that causes amplification bias
• Edit distance: ensuring the UMI set has sufficient pairwise differences to tolerate sequencing errors
• Partially random designs: combining random nucleotides with fixed spacers to improve base calling accuracy during sequencing.

UMI-based quantification fundamentally alters the statistical properties of single-cell data. Count matrices derived from UMIs follow a negative binomial distribution rather than a simple Poisson, due to biological and technical overdispersion. This affects:

• Normalization: methods like SCTransform or scran model UMI counts directly
• Differential expression: tests such as DESeq2 or MAST are designed for UMI count distributions
• Quality control: cells with very low UMI counts indicate poor capture efficiency or damaged membranes
• Data integration: UMI counts provide a common scale for merging datasets across platforms and laboratories.