Two different microphones are recording the same concert. One microphone stands by the stage and mostly picks up the loud sounds from the front row; the other hangs from the ceiling and hears the whole hall, but with the bass muffled. If you compare the recordings, you can figure out how much of the noise in them comes from the microphone and how much from the orchestra itself. Biologists compare two platforms for single-cell RNA sequencing in much the same way: 10x Chromium and Parse Evercode WT. Both read RNA molecules from thousands of individual cells, but each does it its own way, and until now it was unclear whether conclusions about "transcriptional noise" drawn from one of them could be trusted.
To test how much the noise depends on the recording method, the researchers took the GM12878 cell line (human lymphoblastoid cells) and isolated nuclei from it. Each platform was run four times, as if the same concert had been recorded four times with each microphone. Then the recordings were compared on three parameters: how many RNA molecules were captured in total, which genes were seen better, and how strongly the activity of the same gene varied from cell to cell. The first thing that stood out: Parse predicted sequencing depth and cell number almost perfectly, whereas 10x sometimes overshot by a factor of two and sometimes undershot by a factor of two. At the same time, both platforms detected around 3,500 genes per nucleus at 50,000 reads, but 10x saw polyadenylated transcripts (that is, RNA with an adenine tail) better, while Parse saw non-polyadenylated ones better, such as the long noncoding RNA TALAM1.
The biggest surprise came in the noise analysis. It turned out that if, after removing the dependence on mean expression level, you measure the residual variance (that is, how much a gene's activity bounces around its mean), this quantity is reproducible across platforms. The rank correlation between experiments ranged from 0.46 to 0.70. In other words, genes that are noisy in the recording from one microphone are also noisy in the recording from the other. What's more, the noise turned out to be linked to gene length: the longer the gene, the greater the variation. The total length of exons (the regions that remain in mature RNA), however, did not affect the noise. Another detail: genes with a CpG island in their promoter (a DNA region rich in cytosine-guanine pairs) were less noisy than genes without such an island.
This result means that transcriptional noise is not an artifact of a particular technology but a property of the gene itself that can be carried over between platforms. Previously, researchers working with different datasets could not be sure they were comparing like with like. Now there is a foothold: if noise is reproducible across different capture chemistries, it can be studied on combined data from public archives. In practice, this means that the choice of platform and sequencing depth can be made more deliberately, rather than at random. Admittedly, the mechanism linking gene length and CpG islands to noise has not yet been uncovered. It may have to do with topological stress as the transcript lengthens, or with regulatory elements inside introns.
To return to the microphone metaphor: we have found that the hiss in the recording is not a defect of a particular microphone but a feature of the music itself. Now we can listen to different recordings of the same concert without fearing that we are hearing only the distortions of the equipment. The cell keeps playing its part, and we can finally make out its own style.
The Parse Evercode WT platform uses combinatorial barcoding: cells are sequentially mixed and split, receiving unique combinations of barcodes, as if each cell gets its own postal code.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.