Single-cell RNA sequencing (scRNA-seq) has become the standard for studying cellular heterogeneity, but different platforms use different capture and barcoding chemistries. This creates technical biases that are hard to separate from biological variation. Nearby in the archive are studies on DNA sequencing, neural network-based DNA reading, gene network analysis, and cell clustering. This paper stands apart: it focuses on comparing platforms and quantifying transcriptional noise, which is not covered by the neighboring works.
The researchers isolated nuclei from the lymphoblastoid cell line GM12878 and performed four technical replicates for each platform. For 10x Chromium they used microfluidic barcoding with oligo(dT) primers, while for Parse Evercode WT they used combinatorial split-pool barcoding with random hexamers and oligo(dT). Additionally, they ran an experiment on whole cells with 10x. Libraries were sequenced on Illumina, and data were processed with Cell Ranger (10x) and Split-Pipe (Parse). To assess noise, they used two independent methods: a stratified z-score and the overdispersion parameter from the BASiCS package.
Parse Evercode WT achieved nearly 100% of the target sequencing depth and cell number, whereas 10x Chromium missed in every experiment (overshooting by 2-fold or undershooting by 2-fold). Both platforms detected around 3500 genes per nucleus at 50,000 reads, but 10x had a pronounced 3' bias, while Parse had a 5' bias and a preference for long genes and exons. Parse captured non-polyadenylated transcripts better (for example, the lncRNA TALAM1), while 10x captured polyadenylated B-cell markers better (IGHD, CD79A, CD81). Cells clustered by platform, indicating a strong technical bias. The key result: transcriptional noise is reproducible across platforms (rank correlation 0.46–0.70 between experiments), decreases in the G2M phase of the cell cycle, and positively correlates with gene length (genomic span) but not with total exon length. Genes with a CpG island in their promoter are less noisy than those without one.
The work shows that transcriptional noise, measured as residual variance after removing the dependence on the mean, is a gene property that is portable across platforms. This differs from neighboring studies, which either focus on a single method or do not test the reproducibility of noise between chemistries. It is the first proof that the noise metric is not an artifact of a particular platform, which opens the way to combining heterogeneous public datasets for noise analysis.
The topic could develop toward studying the mechanisms linking gene length and promoter architecture to noise. Possible roles are played by topological stress during elongation, regulatory elements in introns, or the kinetics of transcriptional bursts. It would also be promising to study noise in synchronized cells to separate the contribution of the cell cycle from other factors.
The results will influence the design of scRNA-seq experiments and the interpretation of data from different platforms. They are also important for studies of transcriptional regulation and cellular heterogeneity.
The next step could be to test the link between noise and gene length in other cell lines and tissues, as well as to study the role of CpG islands in reducing noise through targeted experiments.
In the field of single-cell transcriptomics, the question of which biological factors determine the level of gene expression noise remains unresolved. This work links noise to gene length and promoter architecture but does not reveal the molecular mechanism. There is also the problem of integrating data from different platforms due to technical biases.
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.