Hidden Biases in Brain Single-Nucleus Sequencing Could Distort Brain Cell Atlases, Study Warns

Single-nucleus RNA sequencing has become one of the most powerful tools in modern neuroscience, allowing researchers to catalog the dizzying diversity of cells that make up the mammalian brain without having to dissociate fragile neurons into single living cells. But a new systematic benchmarking study suggests that the technology may be quietly distorting the very biology it aims to measure. By rigorously testing how tissue harvest methods, chemical fixation, and sequencing platforms shape the resulting data, researchers have uncovered critical technical biases that could ripple through brain cell atlases worldwide, potentially masking some of the most interesting transcripts involved in advanced cognition.

The study, published in BMC Neuroscience by a team at Xi’an Jiaotong University led by corresponding author Shiquan Sun, generated single-nucleus RNA sequencing data from adult mouse brains under deliberately varied experimental designs. Rather than testing one variable at a time in isolation, the researchers dissected the individual and interactive effects of three major workflow choices: whether tissue was harvested post-mortem or in vivo, whether nuclei were fixed in methanol or left fresh, and whether the library was sequenced on the widely used 10x Genomics platform or the SeekOne digital droplet system. Each of these decisions is made routinely in labs around the world, often with little consideration of how they might reshape the final dataset.

The most striking finding concerns the standard post-mortem harvest, the default approach in most single-nucleus studies because it is technically simpler than extracting nuclei from living tissue. The researchers found that post-mortem collection triggers a rapid and selective loss of metabolically active glutamatergic neuron subpopulations, the excitatory neurons that form the backbone of cortical computation. This loss is not a uniform shrinkage of all cell types; it preferentially eliminates specific vulnerable subpopulations, meaning the resulting cell composition no longer faithfully reflects the living brain. Intriguingly, the global transcriptome appeared largely preserved, which means the damage can hide beneath standard quality-control metrics that look only at overall gene detection.

That selective vulnerability carries a troubling implication for brain atlases. The study reports that while the global transcriptomic profile survives the post-mortem interval, transcripts related to advanced cognitive functions become obscured. In other words, a researcher comparing healthy and diseased brains might attribute differences in cognition-related gene expression to disease, when in fact the differences stem from how long the tissue sat before nuclei were extracted. Because glutamatergic neurons are central to nearly every model of cortical function, from memory circuits to the excitatory-inhibitory balance implicated in autism and schizophrenia, their disproportionate sensitivity to harvest stress is a bias that no atlas builder can afford to ignore.

The second variable the team examined was fixation. Methanol fixation is popular because it stabilizes samples for shipping and storage, allowing labs to process tissue collected at distant sites and to pause experiments mid-workflow. The benchmarking showed that methanol fixation minimally altered cell-type compositions, which is reassuring for anyone worried that the chemical treatment might preferentially destroy certain nuclei. However, the treatment was not transcriptionally neutral. It introduced cell-type-specific gene expression shifts, preferentially stabilizing structural and core functional transcripts in both neurons and glia. The practical consequence is that the measured abundance of particular genes depends on whether the sample was fixed, even when the underlying biology is identical.

Perhaps the most consequential discovery, however, involves the sequencing platform itself, a variable most researchers assume is merely a matter of cost and throughput. The team found that platform choice shaped biological interpretation in a profound way: the 10x Genomics and SeekOne digital droplet platforms captured largely non-overlapping differential expression signatures in response to harvest stress. When the same biological perturbation was profiled on the two systems, each platform reported a different set of genes as changing. The 10x platform showed a bias toward detecting synaptic transcript degradation, while SeekOne leaned toward preserved metabolic and glial programs. Two labs studying the same question with the same tissue could therefore reach substantively different molecular conclusions based on platform alone.

This platform dependence strikes at the heart of the reproducibility crisis in single-cell genomics. Large-scale initiatives, including efforts to build comprehensive cell atlases of the human and mouse brains, aggregate data generated on different instruments and with different preservation protocols. If the differential expression signatures triggered by a stressor are largely platform-specific, then integrating datasets without accounting for these effects risks conflating technical artifacts with genuine biological signal. The study’s authors argue that platform choice must be treated as a critical, non-neutral variable rather than an interchangeable commodity, and that meta-analyses should model platform effects explicitly rather than assuming they wash out in large datasets.

To separate artifact from biology, the researchers deployed a battery of quantitative benchmarks. Their evaluation toolkit included clustering concordance measures such as the Adjusted Rand Index and Normalized Mutual Information, cell-type separation metrics like the cell-type average silhouette width, and integration diagnostics including batch silhouette widths and Local Inverse Simpson’s Index variants that quantify how well cell types mix or remain distinct after computational integration. By applying these metrics across the factorial experimental designs, the team could attribute specific distortions to specific workflow steps. This framework, they emphasize, is not just a catalog of problems but a practical recipe: it gives other labs a way to audit their own pipelines and to design studies in which technical variables are measured and controlled rather than left unexamined.

The study also maps the cellular geography of vulnerability in fine detail. Drawing on established brain cell taxonomy, the team characterized glutamatergic neuron subclasses including intratelencephalic and extratelencephalic projecting cortical neurons, layer 6b near-projecting corticothalamic neurons, dentate gyrus immature glutamatergic neurons, and glutamatergic populations of the hypothalamus, thalamus, and midbrain, alongside GABAergic neurons, oligodendrocyte lineage cells, and astrocyte-ependymal cells. This granularity matters because the selective loss of metabolically active glutamatergic subpopulations is not visible at the level of broad cell classes. An analysis that lumps all excitatory neurons together would miss the fact that particular subclasses are disappearing, and with them the transcripts that define their specialized functions.

For the field at large, the message is a call for radical transparency in methods reporting. The authors conclude that precise reporting and control of workflow variables are prerequisites for accurate biological discovery in single-cell transcriptomics, particularly as the community moves toward integrative brain cell atlases meant to serve as durable community resources. Journals and consortia increasingly demand detailed metadata, and this study supplies a concrete reason why: harvest method, fixation chemistry, and sequencing platform each leave measurable fingerprints on the data, and those fingerprints can masquerade as biology. Until such variables are standardized or explicitly modeled, the researchers warn, differences between datasets should be interpreted with caution, and claims of disease-related or experience-related transcriptional changes in the brain should be validated across harvest conditions and platforms before being accepted as fact.

Subject of Research: Technical biases in single-nucleus RNA sequencing of the mouse brain

Article Title: A systematic multi-factorial benchmarking reveals critical technical biases in single-nucleus RNA sequencing of the brain

Article References: A systematic multi-factorial benchmarking reveals critical technical biases in single-nucleus RNA sequencing of the brain. (n.d.). https://doi.org/10.1186/s12868-026-01058-7

Image Credits: AI Generated

DOI: 10.1186/s12868-026-01058-7

Keywords: single-nucleus RNA-seq, benchmarking, brain cell atlas, glutamatergic neurons, post-mortem harvest, methanol fixation, sequencing platform bias, 10x Genomics, SeekOne, transcriptomics, neuroscience, reproducibility