When analysing a single-cell RNAseq dataset, one question I like to ask is: At a single-cell level, are these two genes correlated (expressed together) or anticorrelated (only one or the other is expressed)? However, I am concerned that - given the low read count per cell - it may falsely appear that two genes are anticorrelated when in fact it is just unlikely that both RNAs are detected at the same time.
This problem can partially be dealt with using an imputation technique such as SAVER, to estimate the true relative expression values of undetected genes. Additionally, a statistical test (e.g. permutation) can be performed to ask whether the incidence of cells with both transcripts detected is higher or lower than would be expected, given the total number of trancripts.
My concern, however, is that there may be additional technical aspects of the sequencing process that will exaggerate the apparent anticorrelation between genes. For example, in a cell where the total number of RNA transcripts at any one time is low, a single transcript may be read multiple times (due to multiple primers binding, resulting in multiple cDNAs).
In addition to this, are there any other technical or analytical aspects of ssRNAdeq likely to lead to false anti-correlations?
