Supplementary Figure 2: UMI efficiency as an alternative metric of sensitivity.

(A) Assuming that UMI counts correspond to a count of the fraction of molecules successfully captured by the RNA-sequencing process, in log-log space the efficiency corresponds to the offset from perfect correspondence between input molecules and counted UMIs. (B) With the exception of data from the MARS-Seq protocol, spike-in detection limits correspond well with UMI efficiency measures. The spike-in detection limit can however also be used for coverage based data quantified by TPM. (C) The assumption with UMI counting as a quantitative measurement is that efficiency is the only factor determining differences between real counts and observed counts. However, fitting a model with a non-one exponent on the number of input molecules shows this is almost in all cases < 1. This means UMI counts underestimate expression of highly expressed genes. (D) The saturation of UMI counts can be partially explained by short UMIs. If an experiment uses too short UMIs, eventually the number of possible observable UMIs plateau. However, even for very long UMIs, such as 10 base pairs, the mean molecule exponent is 0.8, indicating some additional unexplained factor is causing a saturation of UMI counts. (E) Averaged efficiency comparison of endogenous genes and ERCC spike-ins. The data by Grun et al had smFISH measurements for 9 genes in the same experimental conditions as the single-cell RNA-seq data. Assuming 100% capture rate for smFISH, we can compare average smFISH counts with average UMI counts. Round markers correspond to median value across cells, and bars correspond to 95% confidence interval across cells. The smFISH counts suggest UMI counts for endogenous transcripts are on the order of 5-10% on average, while ERCC spike-in UMI counts correspond to 0.5-1% efficiency on average.