Power Calculations for Detection of Rare Transcripts
We conducted the following statistical analysis to estimate the effects of various factors on our ability to make confident claims regarding the presence/absence of transcripts of interest (e.g., ACE2), both within individual cells and clusters (Figure S6). Specifically, we investigated the roles of capture/reverse transcription efficiency, ACE2 expression level, sequencing depth, and cell numbers. Taken together, the results of this power analysis are in agreement with other efforts to model biological and technical sources of zero-inflation within scRNA-seq data (e.g., https://satijalab.org/howmanycells and Kharchenko et al., 2014, Svensson, 2020).
We began by quantifying how likely we are to capture and transcribe at least one ACE2 mRNA molecule, as a function of the number ACE2 mRNA molecules per cell and a protocol’s efficiency (Figure S6A). Drop-Seq has a capture/transcription efficiency of ∼10% (as estimated using ERCC spike ins; see (Macosko et al., 2015), and the experimental platforms used in this study are either equivalent (e.g., Seq-Well v1, (Gierahn et al., 2017) or superior (e.g., 10-fold better unique molecule detection, 5-fold better gene detection using Seq-Well S3;(Hughes et al., 2019)). Most relevant to this context, inferior turbinate scrapings were processed using both Seq-Well v1 and Seq-Well S3 (Figure S3B). Importantly, Seq-Well S3 provided > two-fold increase in the detection frequency of rare ACE2 transcripts (i.e., ACE2+: 4.7% for v1 versus 9.8% for S3), making it reasonable to expect that such improvements in single-cell experimental technologies have yielded corresponding improvements in capture and transcription efficiency. Based on Drop-Seq’s 10% efficiency, even if ACE2 is expressed at the low level of 5 mRNA molecules per cell (a reasonable order-of-magnitude estimate, given that non-human primate ileum cells had a maximum of 10 ACE2 unique molecules per cell observed via sequencing and an average of 1.93 molecules per cell in expressing cells, see Figures 3B and 3C), our experimental platforms have a minimum likelihood of 41% to capture and reverse transcribe at least one ACE2 mRNA molecule in any given individual cell. This likelihood rapidly increases if we estimate higher efficiencies for improved scRNA-seq technologies (e.g., 67% likelihood within any individual cell at 20% capture/transcription efficiency, 76% likelihood at 25% efficiency, Figure S6A). Thus, while transcript drop-out may reduce the fraction of positive cells, with the capture and transcription efficiencies of improved single-cell technologies, the impact is likely to be minor (reads are likely underestimated by up to a factor of ∼2.5x), given a sufficient depth of sequencing (see below). We note that this impacts both clusters deemed to contain and not contain ACE2+ cells, and suggests our percentages are likely lower bounds for true expression (within a factor of ∼2.5x).
Next, we examined the probability of sequencing an ACE2 transcript as a function of read depth and ACE2’s fractional abundance in each single cell within our sequencing libraries. First, across two different tissues (non-human primate ileum and lung, representing a high expresser of ACE2 and low expresser, respectively), we calculated the proportion of unique ACE2 molecules in our ACE2+ cells (defined as any cell with at least 1 UMI aligning to ACE2) as a fraction of total reads within individual cells to provide an order-of-magnitude estimate for average ACE2 abundance in our single-cell sequencing libraries (i.e., the probability that a read within a cell corresponds to a unique molecule of ACE2, Figure S6B). We highlight that by calculating probabilities based on ACE2 unique molecules divided by an individual cell’s total reads, we are providing a conservative estimate for the probability of observing ACE2 as a function of sequencing depth (e.g., as compared to basing these probabilities on ACE2 non-UMI-collapsed reads divided by total reads). Next, we obtained information on the number of reads in these cell populations to provide estimates of average sequencing depths (Figure S6C). Using the mean fractional abundances of ACE2 from each tissue (Figure S6B) and the mean read depths for all genes (Figure S6C), we calculated the probability of detecting at least 1 ACE2 molecule (i.e., P(detecting > 0 ACE2 molecules) = 1 - (1 - ACE2 fractional abundance)Read depth). This results in a 93.7% probability in ileum-derived cell libraries that contain ACE2, and a 76.0% probability for lung-derived cell libraries, indicating that our sequencing depths are sufficient to detect ACE2+ cells (Figure S6D).
To further evaluate whether our ability to detect ACE2+ cells was an artifact of sequencing depth, we compared the number of ACE2+ cells in a cluster to the mean number of reads across all cells in that same cluster (Figure S6E). We did not observe any significant correlation: the ileum cell cluster with the highest number of ACE2+ cells had the lowest sequencing depth of all ileum clusters, and the lung cell cluster with the highest number of ACE2+ cells was approximately average in its read depth (on a log-log scale, Pearson’s r = −0.31, non-significant). Further, when comparing ACE2+ cells to ACE2- cells within a given tissue, we did not observe a positive correlation between read depth and ACE2 status (i.e., mean ± standard error of the mean, SEM, reads among all lung cells = 28,512 ± 344; mean ± SEM reads among ACE2+ lung cells = 28,553 ± 2,988; mean ± SEM reads among all ileum cells = 14,864 ± 288; mean ± SEM reads among ACE2+ ileum cells = 10,591 ± 441, full statistics on cell depth among ACE2+ cells compared to ACE2- cells of the same cell type can be found in Table S9). Thus, we can be confident that the observed differences in ACE2+ proportions across clusters are not driven by differences in sequencing depth.
Finally, we investigated how observed differences in ACE2+ proportions across clusters might be affected by cell sampling. Using the proportion of ACE2+ cells in a “typical” cluster annotated as being ACE2 positive (i.e., 6.8% in non-human primate type II pneumocytes, Figure 1), we calculated the cluster sizes needed to be confident that the probability of observing zero to a few positive cells is unlikely to have arisen by random chance (probabilities calculated under a negative binomial distribution with parameter p = 0.068, Figure S6E). We found that as cluster sizes approach and exceed 100 cells, the probability of observing zero to a few positive cells rapidly approaches zero, if we assume 6.8% of cells are positive. Further, to examine our confidence in estimating an approximate upper bound (ignoring the impact of protocol inefficiencies discussed above) for the fraction of cells positive in a cluster as a function of the number of cells in that cluster, we also calculated the probability of observing zero (and its complement, probability of observing at least 1) ACE2+ cells as a function of cluster size across true positive proportions ranging from 0.1% to 10% (probabilities calculated under a negative binomial distribution with parameter p = 0.001 to 0.1, representing hypothetical proportions of ACE2+ cells Figure S6F). Given our typical cluster sizes (on the order of hundreds of cells, exact values provided in Table S9), we find that for us to observe 0 ACE2+ cells in a cluster due to sampling artifacts, the fraction of true positives must be ∼1% or less. Thus, these complementary approaches demonstrate that our observed variations in ACE2+ cell proportions across clusters likely reflect underlying biological differences, rather than random chance.