What is a unique molecular identifier, and why do single-cell methods count UMIs instead of reads?

Why

A single cell holds very little RNA, so single-cell libraries need many PCR cycles. Some molecules amplify better than others, and read counts then reflect amplification luck as well as abundance. A UMI is added during reverse transcription, before any amplification.

Afterwards, all reads carrying the same UMI on the same gene in the same cell are collapsed into one count. The Bioconductor single-cell book describes UMI counts as interpreted like read counts but less affected by PCR artifacts. Droplet protocols such as 10x Genomics, and plate protocols such as CEL-seq2 and MARS-seq, use UMIs, while full-length Smart-seq2 data are counted as reads.

UMIs are not error-free. A sequencing error inside a UMI makes one molecule look like two. The UMI-tools authors showed such errors are common and introduced network methods that merge UMIs one edit apart, which improved quantification and clustering.

Counting molecules also does not fix capture efficiency. Most molecules in a cell are never captured, so a zero still often means not detected rather than not expressed.