The research team, led by Valeryia Shydlouskaya, S. M. Mansour Haeryfar, and Tallulah S. Andrews at Western University in Canada, analyzed matched 3-prime and 5-prime data from 36 subjects across six publicly available datasets spanning multiple tissues, including thymus, liver, bone marrow, kidney, skin, and spleen. Because each subject had been profiled with both chemistries from the same tissue, the researchers could isolate the technical effect of the protocol from genuine biological variation, a rare and powerful experimental design in the single-cell field.

The origin of the two chemistries explains why both exist. Early single-cell protocols attached barcodes to poly(dT) oligonucleotides, capturing the 3-prime end of transcripts during reverse transcription. This worked well for general expression profiling but was poorly suited to immune research, because sequencing the constant region of recombined T-cell and B-cell receptors provides little information about their antigen specificity. To read the variable regions of these receptors, protocols were redesigned to add unique molecular identifiers and cell barcodes at the 5-prime end instead. As a result, 5-prime chemistry dominates immune-focused studies, while 3-prime chemistry remains the workhorse of large-scale cell atlases thanks to its lower cost.

Using a Wilcoxon rank-sum test on each subject separately, the team identified genes consistently differentially expressed between the two chemistries and compiled a consensus list of 867 protocol-biased genes, significant in at least five subjects at a stringent threshold. The number of biased genes ranged from 609 to 2,598 depending on the significance cutoff, but removing even the smallest set markedly improved the alignment of 3-prime and 5-prime data in low-dimensional space. Cosine similarity between matched cell-type expression profiles dropped to 0.25 for the biased genes, compared with 0.5 to 0.75 for random gene sets matched in number and expression level, showing that these genes captured a disproportionate share of the protocol-associated differences.

Crucially, the biased gene set generalized. In seven additional matched 3-prime and 5-prime comparisons held out from the discovery analysis, removal of the predefined 867 genes improved mixing between the two chemistries in every case, and the number of significantly biased genes matched what was seen in the discovery datasets. Most of the biased genes were highly expressed housekeeping genes involved in cellular translation and oxidative phosphorylation, and most showed higher expression in 5-prime than in 3-prime libraries. Only nine were specific to a non-immune cell type, indicating the analysis had power across a broad range of tissues.

The researchers then probed what might cause the bias. A natural suspect is GC content, since PCR amplification is known to favor GC-rich sequences, and the 5-prime ends of transcripts are generally more GC-rich than the 3-prime ends. But the biased genes showed no unusual GC disparity between their two ends, undermining that explanation. Gene length was also largely ruled out, with the slight shortness of biased genes likely explained by their housekeeping function. The team also tested whether counting intronic reads could account for the differences, remapping one matched spleen sample with and without introns. Including intronic reads reduced the number of significant biased genes from 525 to 390, but most remained significant and the 5-prime-skewed directionality persisted, showing that exon-only quantification explains only part of the effect.

With the biased genes characterized, the team benchmarked eleven widely used normalization and batch correction methods, including ComBat, limma, mnnCorrect, fastMNN, Scanorama, M3Drop, SCTransform, Z-transformation, scVI, and scArches. Performance was assessed with statistical measures such as Pearson correlation, cosine similarity, mean squared error, Euclidean distance, and Jensen-Shannon divergence, alongside biological measures like UMAP mixing and marker-gene consistency. ComBat and mnnCorrect consistently produced the largest improvements in statistical alignment, while the housekeeping-gene-based Z-transformation performed worst. Notably, SCTransform applied separately to each protocol outperformed the version that explicitly regressed out the 3-prime versus 5-prime difference, likely because the protocol bias is gene-specific rather than consistent across genes of similar expression levels.

The most striking finding emerged when the methods were tested in a biologically realistic scenario. The researchers constructed synthetic use cases from thymus, liver, and bone marrow datasets in which only about 25 percent of cells were shared between the 3-prime and 5-prime samples, with some cell types unique to each chemistry. They then asked which method best recovered the true differentially expressed genes between two cell types, using a ground truth defined by concordant results in both assays. Log-normalized data with the 867 biased genes removed achieved the highest Matthews correlation coefficient, followed by uncorrected log-normalized data. Among correction methods, fastMNN performed best but still fell short of simple log-normalization, while scVI inflated differential expression counts most dramatically. Methods that smoothed and imputed expression values, such as fastMNN and scVI, sometimes over-smoothed genes with overlapping distributions, while limma and M3Drop altered zero values in ways that contradicted the expected expression patterns.

The authors attribute this pattern to a fundamental distinction between statistical alignment and biological accuracy. Methods that excel at making datasets look similar can inflate both true and false positives in differential expression analysis, particularly when biological differences are subtle or partially confounded with batch structure. Deep generative models like scVI and scArches are explicitly designed to denoise data, a process previously shown to inflate false-positive differential expression. The study also highlights a gap in the field: most integration methods operate only in a reduced-dimensional space and do not return corrected gene expression values, and those that do rarely optimize the accuracy of their reconstructed expression.

The practical message for the single-cell community is that protocol bias between 3-prime and 5-prime chemistries is limited in scope, and targeted removal of the affected genes offers a simpler and often more reliable alternative to aggressive batch correction. The authors caution that their biased gene list, while broadly representative of typical tissues for 5-prime sequencing, is not exhaustive, and that lowly expressed genes and tissues such as brain and reproductive organs remain underrepresented. They also note that the findings may extend to other closely related technologies that differ only in enzyme formulation or capture strategy, but that more divergent methods, such as probe-based quantification, may carry more pervasive biases. As publicly available datasets increasingly mix 3-prime and 5-prime chemistries, knowing that most genes can be compared directly, with a short list of known offenders set aside, may prove one of the most useful tools yet for building reliable cross-protocol single-cell analyses.

Subject of Research: Protocol-dependent gene expression bias between 3-prime and 5-prime single-cell RNA sequencing chemistries and its correction

Article Title: Characterizing and mitigating protocol-dependent gene expression bias in 3′ and 5′ single-cell RNA sequencing

Article References: Shydlouskaya, V., Haeryfar, S. M. M., & Andrews, T. S. (2026). Characterizing and mitigating protocol-dependent gene expression bias in 3′ and 5′ single-cell RNA sequencing. Frontiers in Bioinformatics, 6, Article 1931243. https://doi.org/10.3389/fbinf.2026.1931243

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1931243

Keywords: single-cell RNA sequencing, 3-prime chemistry, 5-prime chemistry, protocol bias, batch correction, normalization, differential gene expression, fastMNN, scVI, data integration, 10X Genomics, bioinformatics