Your Single-Cell Cell Types Are Probably Wrong: Why scRNA-seq Pipelines Fail Silently

The Figure That Changed When You Rerun the Pipeline
A reviewer asks for one thing: the upstream analysis. Not the model, not the mechanism, not the polished network diagram — just the count matrix and the script that turned it into cell types.
You send it. Three weeks later, they write back with a table of their own. Their clustering has nine cell types. Yours has eleven. One of them is called "proliferating T cells." In their rerun, it does not exist. In your rerun — run on the same machine, from the same archive, with the same container — it does.
This is the quiet failure mode of single-cell analysis, and it is far more common than the loud ones. A crashed pipeline is free: it costs a day and tells you the truth. A pipeline that runs cleanly and produces a subtly wrong biological claim costs a paper, a grant cycle, and sometimes a program.
We have now run enough of these analyses for pharma and academic groups that the pattern is remarkably consistent. The problems are almost never in the biology. They are in four decisions that nobody writes down: what the unit of replication is, how much signal to remove during integration, how many clusters to claim, and how the annotation was validated.
Where These Analyses Actually Break
Pseudoreplication: The Most Expensive Error in Single-Cell
If you treat 30,000 cells from 4 donors as 30,000 independent replicates, your p-values are fiction. The effective sample size is four. Everything downstream — differential expression, machine-learning classifiers, marker claims, "this cell type is depleted in condition X" — inherits the inflation.
This is not a subtle effect. Benchmarking work on scRNA-seq differential analysis showed that statistical tests using cells as replicates produce false-positive rates high enough to reject the null hypothesis for the majority of truly null comparisons, at rates that would be considered indefensible in any other field [1][2]. The recommended fix is well established: aggregate cells into pseudobulk profiles per sample — per donor, per patient, per replicate — and run the count-based differential expression methods designed for that structure [1].
The part that gets skipped is the design decision before sequencing: donors are the replicates, cells are the measurement. If you need 20 donors per arm for a specific test, that is a sample-size calculation, not a wish. Discovering this at analysis time, after the data is on a freezer, is how studies die.
Batch Correction That Removes Your Biology
Batch integration is not one operation. It is a family of operations with different assumptions, and one benchmark comparing them found that no single method performs well across all datasets and all biological signals you might want to preserve [3]. There is a real tension, and it is well documented: aggressive correction of technical effects can also erase genuine biological differences that happen to correlate with batch structure [4].
The failure looks like this. You run correction. Your condition separation weakens. You either accept the loss, which quietly reduces your statistical power, or you reduce the correction strength, which brings the batches back — and now you cannot tell whether the residual batch structure or the residual biology is what you are seeing. Either way you have learned less than you think, and the decision was made by a parameter you never tuned on purpose.
Over-Clustering: Every Cluster Is a Discovery Until Proven Otherwise
Clustering resolution is a hyperparameter, and the number of cell types it produces is a direct, monotonic function of it. That makes it trivially easy to produce whatever result the story requires. The methodological literature has been blunt about this: a cluster is not a cell type until it has been validated as one, and cluster identity at different resolutions is not comparable across studies [5].
A cleaner framing: if you report one cell type and it holds up, you have one cell type. If you report eleven, you have eleven claims, each of which needs independent support — marker coherence, absence of technical artifacts, reproducibility across integration runs and donors, ideally orthogonal validation. The teams that get this right treat resolution as a sweep: they report the panel of results and make an argument for where the biologically meaningful boundary sits, rather than presenting a single number that happened to make the story work.
Circularity: Using the Same Data to Define and Validate
Cell type annotation is usually done by eyeballing canonical marker genes in a dot plot. That is a legitimate starting point and a legitimate final step. What it cannot be is both the definition and the validation.
If you pick the markers that make your clusters look canonical and then report "our cell types were validated by canonical marker expression," you have validated nothing. You looked at the data, chose criteria consistent with what you already saw, and then checked that the data satisfied your criteria. Circularity of that kind does not fail loudly; it produces a figure with a nice legend and a claim that a careful reader can dismantle.
The defensible version looks different: annotation derived independently of the hypothesis being tested, marker sets fixed from reference atlases or literature before looking at the new data, and validation against orthogonal evidence — spatial context, protein-level measurement, sorted-cell or flow confirmation, or replication in an independent cohort.
Provenance: The Pipeline That Changed Under You
The example that opens this post is not exotic. Dependencies move, default parameters shift, seeding behavior differs across backends, and a package major release can change clustering output without anyone touching your code. If your pipeline does not pin its environment, log its parameters, and record the exact commit and container for every run, then "we reran it" is not a reproducible operation — it is a new experiment.
This is the least glamorous item on the list and the one that most reliably determines whether a reanalysis six months later is a week of work or an archaeology project.
What a Defensible Single-Cell Workflow Actually Looks Like
Across the engagements we run, the analyses that survive external scrutiny share a shape:
Design before data. Sample size is computed at the donor level against the specific statistical test, with room for donor-to-donor variability. Cell yield per donor is planned for, not hoped for.
Pseudobulk as the default for differential work. Cell-level analysis is used for what it is good at — defining states, discovering markers, visualizing the landscape. Inference about differences between conditions runs on per-sample pseudobulk with appropriate count-based methods [1].
Integration benchmarked on your data. Run more than one correction method, evaluate against known technical structure and known biology, then fix one method and record its parameters. Do not shop for the method that maximizes your preferred result.
Resolution presented as a sweep. Report results across a range, argue for the boundary you claim, and describe cell states with the confidence your evidence supports.
Annotation independent of the hypothesis. Marker sets from literature or reference atlases, fixed in advance. Validation from an orthogonal modality where one is available.
Provenance as a build artifact. Pinned environment, containerized, parameterized config committed alongside the code, parameters and seeds logged per run, and outputs versioned.
The Quiet Strategic Point
Notice what every item above has in common: none of them are about better algorithms. The mature analysis tools in this space are good, and open-source single-cell stacks are excellent. What separates a publishable result from a fragile one is a handful of decisions made deliberately, documented, and defended — and those decisions live in workflow design and review process, not in software selection.
That is where we do most of our work. We design and run single-cell and multi-omic analysis pipelines for pharma and research teams, build the QC and validation gates that catch silent failures before they reach a figure, and set up the provenance and quarterly operational review that keeps results reproducible after the first analysis is done. If your single-cell results have never been independently rerun, we would argue that is the highest-leverage next step — and a good place to start.
Pii Data Science Solutions builds production-grade bioinformatics workflows for single-cell, spatial, and multi-omic programs. We work with computational biologists, translational teams, and platform leadership on pipeline design and rerunnable environments; validation gates for cell type annotation and differential expression; integration benchmarking and resolution strategy; and the operational cadence that keeps NGS and single-cell analysis stable quarter over quarter. If your team is at the point where reproducibility is no longer theoretical, let's talk about what a defensible pipeline would take to build in your environment.
Sources
[1] Squair, J.W. et al. "Orchestrating single-cell sequencing projects: a guideline for reproducible scRNA-seq analysis" — https://doi.org/10.1186/s13059-021-02404-0
[2] Zimmerman, S., et al. "False discoveries in scRNA-seq data" — https://doi.org/10.1038/s41467-021-22509-y
[3] Luecken, M.D. et al. "Benchmarking atlas-level data integration in single-cell genomics" — https://doi.org/10.1038/s41592-021-01336-8
[4] Nygaard, J. et al. "Methods that remove batch effects while retaining group differences may lead to exaggerated confidence in downstream analyses" — https://doi.org/10.1186/s13059-016-0971-7
[5] Dou, Y. et al. "A guideline for the controlled analysis of scRNA-seq data" — https://doi.org/10.12688/f1000research.1497.1
