Evidence checked October 31, 2026. Historical methodological studies retain their original dates. Numerical scenarios are hypothetical; this is an analysis of research evidence, not clinical advice.
1. An atlas is built from samples
The Human Cell Atlas's publications page, checked for this article, describes its November 2024 collection of more than forty papers as progress toward a first draft atlas. The consortium's record also lists subsequent research. That is an extraordinary mapping effort, but the word atlas should not make the reader imagine that every cell in every kind of person has been observed. A reference map is assembled from particular samples using particular measurement and analytical choices.
Single-cell RNA sequencing makes a powerful change in perspective possible. Instead of reducing a tissue to one average expression profile, researchers can examine variation among individual captured cells. A rare population or a difference between cell states may become visible where a bulk measurement obscured it. The resulting plots can be so vivid that their points seem to be a direct census of the tissue.
They are better understood as the visible end of a selection process. People contribute samples; tissue regions are chosen; material is processed; some cells or nuclei are captured; molecules are detected; records are filtered; and an analysis organizes what remains. At each step, the relationship between the dataset and the original tissue can change.
This does not make the atlas untrustworthy. It identifies what trust requires. The strongest interpretation connects the final pattern back through the sampling and measurement chain. It asks which biological distinctions the method can resolve and which distinctions could have been introduced, erased, or distorted along the way. More cells can improve a map, but the cell count alone cannot answer those questions.
2. The person is not interchangeable with the cell
A dataset containing one hundred thousand cells can sound much larger than a study involving twenty people. Both descriptions may refer to the same experiment. They count different units of information. Individual cells are valuable observations, but cells from the same person share biological background and often share handling conditions. They are not independent people.
A 2021 methodological study by Zimmerman and colleagues examined pseudoreplication in single-cell analysis: the problem of treating correlated observations as independent evidence. It found that accounting for the experimental unit improved control of false-positive findings in its evaluations. The primary paper supports a basic rule of interpretation: the number of captured cells does not replace the number of independently sampled biological units.
Consider a hypothetical comparison involving three people with a condition and three without it. Each contributes ten thousand cells. There are sixty thousand cellular profiles, but only six donor backgrounds. A difference seen across many cells could still be driven by an unusual feature of one donor. Treating all sixty thousand cells as independent representatives of their groups can make uncertainty look much smaller than it is.
The design question follows directly. If the aim is to characterize a rare cell population within one sample, more cells may be exactly what is needed. If the aim is to generalize a disease-associated difference across people, more independent donors may be more informative. There is no universal exchange rate between cells and people. The unit that supports the claim must be identified before the headline sample size is celebrated.
3. A piece of tissue is a location, not the whole organ
An organ contains regions with different structures and functions. Sampling one region does not guarantee representation of another. Even a carefully collected specimen may reflect a particular anatomical location, depth, lesion boundary, or surgical context. When a dataset is described simply as lung, liver, or brain, that geographical information can disappear from the reader's mental model.
Imagine a hypothetical study comparing tissue near a lesion in one group with tissue farther from a lesion in another. A cell-state difference might reflect the condition under study, the local environment, or both. The correct interpretation requires knowing where the material came from. A computational adjustment cannot recover location information that was never recorded.
Sampling geography also affects cell proportions. A region rich in one structure may naturally contain more of the cells associated with that structure. If the study samples regions unevenly, an apparent group difference in cell abundance could arise without a corresponding change throughout the organ. This is a question about the target population of tissue, not merely the target population of people.
A strong report therefore describes the anatomical unit represented by each sample. It distinguishes a local finding from an organ-wide claim and explains how samples were selected. When multiple regions from one person are available, their relationship should be preserved in the analysis rather than flattening them into unrelated observations. The result becomes more useful, not less, when its geography is explicit. Another investigator can then ask whether the same pattern appears elsewhere, whether it follows a gradient, or whether it is specific to a particular tissue neighborhood.
4. Preparing the sample can change the state being measured
To observe individual cells, many workflows separate them from their original tissue environment. That processing is not guaranteed to leave every molecular state unchanged. A measurement can contain a response to the preparation as well as a record of the tissue's prior biology. Distinguishing those contributions is essential when a newly identified cluster is interpreted as a naturally occurring cell state.
Van den Brink and colleagues demonstrated this issue in a 2017 study of mouse muscle stem cells. They found expression changes associated with dissociation and used observations in intact tissue to investigate whether an apparent subpopulation existed before processing. The author-hosted primary paper is a specific demonstration, not proof that every stress-related cluster in every dataset is an artifact.
The broader analytical lesson is to ask what else could create the observed signature. A hypothetical inflammatory-looking profile might reflect the disease, the tissue's response to handling, or a combination. If cases and controls were processed differently, the alternative explanation becomes especially difficult to dismiss. A gene-expression pattern can be biologically real in the measured cells while still being induced outside the body.
Independent observations in the original tissue can help distinguish these possibilities. So can study designs that avoid confounding the biological comparison with preparation conditions. The goal is not to remove every signal associated with stress automatically. That could erase genuine biology. It is to determine whether the signal belongs to the phenomenon the study claims to measure.
5. The cells that survive are a selected population
A single-cell dataset contains the cells that made it through the workflow and passed the analysis criteria. If some kinds of cells are more difficult to recover or more likely to be excluded, the final dataset can underrepresent them. Increasing the number of captured cells does not necessarily correct a systematic recovery difference.
Consider a hypothetical tissue containing equal numbers of two cell types. Suppose the workflow retains half of one type and nearly all of the other. The observed proportions will differ from the original proportions even if the measurement of every retained cell is technically excellent. The experiment has accurately measured a selected sample, not automatically a census of the tissue.
This distinction becomes important when comparing conditions. If diseased tissue is harder to process or contains cells with different fragility, recovery can differ between groups. A change in the observed fraction of a cell type might reflect biological abundance, differential recovery, or both. An interpretation focused only on the final count matrix may miss the competing explanation.
The answer is not to assume that every abundance result is unreliable. It is to seek evidence about the sampling process and, when the claim warrants it, compare with measurements that preserve more of the original tissue context. Reports should explain the starting material, exclusions, and recovery-related limitations. The relevant denominator is the tissue population the researchers want to describe, not simply the set of records remaining after filtering. A large dataset can provide a precise estimate of the wrong denominator if the selection process is ignored.
6. A zero is not always an absence
Single-cell RNA measurements are sparse: many genes have zero recorded counts in many cells. A zero can arise because the relevant RNA was not present at the time of measurement, because the assay did not capture or detect it, or because of other features of the measurement process. The count matrix does not automatically distinguish those explanations for every entry.
A 2020 methods paper, Embracing the dropouts in single-cell RNA-seq analysis, investigated how patterns of detected and undetected expression can themselves contain information for identifying populations. The primary study illustrates why sparsity is not merely empty space to fill. It also explains that low molecular input and detection limitations contribute to the observed pattern.
A hypothetical researcher might see no detected transcript for a receptor in a small group of cells and conclude that the receptor is biologically absent. That conclusion may exceed the measurement. A more defensible statement distinguishes non-detection under the assay from demonstrated absence. The strength of the inference depends on the assay's sensitivity, the number and quality of observations, and corroborating evidence.
The opposite error can occur when missing values are computationally filled. An imputed value is an estimate based on assumptions and neighboring information; it is not a newly observed molecule. Such estimates can support some analyses, but their provenance should remain clear. Replacing zeros may make a figure smoother without making every cell-level statement more certain. The useful question is whether the analytical procedure improves the specific inference, not whether it makes the matrix look less sparse.
7. A detected transcript may have come from elsewhere
A recorded RNA signal is not always evidence that the assigned cell produced the molecule. In droplet-based workflows, free RNA in the surrounding material can contribute background signal. This creates a different problem from non-detection: the dataset may contain an apparent expression event that does not belong to the cell to which it is assigned.
Young and Behjati's 2020 SoupX paper described ambient RNA contamination and evaluated a method for estimating and reducing it. The authors' abstract and publication record explain why background contamination can distort biological interpretation. That is evidence for the problem and the evaluated method, not a guarantee that one correction procedure resolves every dataset.
Imagine a hypothetical tissue containing a highly abundant cell type that expresses a distinctive transcript strongly. If RNA from that population contributes to the background, other cell profiles may show weak traces of the same transcript. Without considering contamination, a researcher might propose a surprising new function for those cells. The discovery could be real, but the background explanation needs examination.
The relevant evidence may include how the signal is distributed, whether it follows the expected contamination pattern, and whether another method detects it in the proposed cells within tissue. Correction should also be evaluated carefully: an overly aggressive procedure could remove genuine low-level expression. The purpose is not to force the data to match familiar biology. It is to distinguish a novel biological signal from a measurement mixture before giving the novelty a name.
8. One barcode can conceal more than one cell
A record labeled as a single cell may sometimes contain material from more than one cell. Such mixtures can create profiles that appear to combine features of different populations. That can be particularly tempting to interpret as a transitional state, an unusual hybrid identity, or a newly discovered cell type.
The National Cancer Institute's single-cell training material discusses doublets, ambient RNA, and quality-control trade-offs. It cautions that aggressive filtering can remove valuable populations and that biological differences can resemble quality problems. The NCI guide supports an important principle: quality control requires biological judgment, not merely a universal threshold copied from another experiment.
In a hypothetical dataset, a small group of records expresses markers associated with two distinct cell types. One explanation is a real intermediate state. Another is mixed material. A third is imperfect annotation of a legitimate population. The correct response is to investigate these alternatives rather than deciding from the novelty of the cluster alone.
The analysis should also preserve uncertainty when the alternatives cannot be fully resolved. A label such as candidate mixed-profile population is less dramatic than announcing a new cell identity, but it can be scientifically more honest. Additional observations may later establish a stable biological population. Until then, the computational cluster is a pattern in the data, not a guarantee that every point corresponds to one coherent cell with the proposed properties. The single-cell label describes the intended measurement resolution; quality assessment determines how closely the actual records meet that intention.
9. Filtering changes the question if it changes the population
Quality-control decisions are necessary, but they are not neutral. Removing records with unusual molecular counts or other properties can improve the dataset while also changing which biological populations remain. The challenge is to distinguish evidence of poor measurement from legitimate biological variation that happens to produce an unusual measurement profile.
A hypothetical tissue may contain both small, relatively quiet cells and large cells with more RNA. A single count threshold applied without context could remove one group disproportionately. The remaining dataset might be technically cleaner yet less representative of the tissue. If the excluded group is related to the condition being studied, the filtering decision can alter the apparent biological result.
This is why sensitivity analysis matters. Does the principal conclusion persist under reasonable alternative filtering choices? Which cell populations are most affected? Are the exclusions similar across donors and groups? These questions can reveal whether the finding depends on one narrow analytical decision. They do not require retaining every questionable record; they require understanding the consequences of removal.
Reports should also distinguish prospective acceptance rules from choices made after inspecting the result. Adjusting filters to rescue a desired cluster can create an unnoticed selection problem. Adjusting them because an initial rule clearly mishandled a known biological population may be justified, but the reason should be documented. Quality control is part of the scientific analysis, not a preliminary cleaning step that can be omitted from the story. The final map is partly defined by the gate through which its points passed.
10. Batch correction can remove biology as well as noise
Combining datasets from different laboratories, technologies, or sampling conditions can reveal patterns at a scale no single study could achieve. It also creates a difficult separation problem: which differences are unwanted technical variation, and which are biology worth preserving? An integration method must make assumptions about that distinction.
Luecken and colleagues' atlas-level integration benchmark, published in the January 2022 issue of Nature Methods, evaluated methods using both batch-removal and biological-conservation criteria. The primary benchmark showed why successful mixing alone is not an adequate definition of success. An integrated map can look smooth while losing meaningful distinctions.
Consider a hypothetical study in which all cases were processed in one laboratory and all controls in another. The disease effect and laboratory effect are confounded. An algorithm cannot reliably discover their separate contributions from the labels alone. Removing the difference might erase disease biology; preserving it might retain a technical artifact. The design has created uncertainty that attractive integration cannot simply dissolve.
A more informative evaluation asks whether expected shared populations align while credible biological differences remain. It examines multiple criteria and the assumptions behind them. A plot in which every batch color is evenly mixed may be reassuring for one purpose and concerning for another. The goal is not maximum visual blending. It is a representation that supports the intended comparison without manufacturing similarity or difference where the evidence cannot justify it.
11. A cluster is an analytical result before it is a cell type
Clustering organizes similar profiles into groups. The resulting boundaries can help researchers recognize populations and generate hypotheses. But the existence of a cluster does not automatically establish a new biological type. The grouping depends on the measurements, selected features, similarity definition, and resolution of the analysis.
A hypothetical population might split into several clusters when the analysis is tuned to finer resolution. Those clusters could correspond to stable identities, temporary states, technical variation, or a mixture of these. Conversely, a broad cluster might combine biologically distinct populations because the relevant distinction is poorly measured. The algorithm has found a partition; interpreting that partition is a separate scientific task.
The distinction between type and state is particularly important. A cell may change its expression as it responds to its environment without becoming a different lineage. A label that treats every response state as a new type can inflate the apparent number of populations and complicate comparison across studies. Yet collapsing all variation into familiar types can obscure important biology. There is no substitute for defining what the label is intended to mean.
A strong annotation uses multiple lines of evidence and makes uncertain assignments visible. It explains whether names are based on established markers, reference mapping, functional evidence, spatial location, or an exploratory pattern. Readers should be able to distinguish a recognized population from a provisional cluster. The latter can still be valuable. Scientific discovery often begins with a pattern that does not yet have a secure name.
12. A reference atlas can help and still mislabel a new sample
Mapping new cells to a reference can make annotation more consistent and efficient. It can also encourage the analysis to assign a familiar label to a population that is poorly represented in the reference. The result may be the closest available category rather than the correct biological identity.
Imagine a hypothetical reference built largely from healthy adult tissue being used to label cells from a different developmental stage or a substantially altered disease environment. Some populations may map well. Others may share enough features with a reference group to receive its label while differing in ways central to the study. A confident-looking assignment should not conceal the mismatch between the reference's coverage and the new sample.
The appropriate response is to examine uncertainty and out-of-reference behavior. Does the method allow an unknown assignment? Are the distinguishing features preserved? Do the labels remain plausible when considered alongside the sample's anatomy and other measurements? A reference should constrain interpretation without preventing discovery of what the reference does not contain.
Versioning matters too. An atlas can change as new data and annotations are added. A label produced by one version may not match a later version's categories exactly. Research records should identify the reference and preserve the mapping procedure. This allows readers to understand whether a disagreement reflects new biological evidence, a revised taxonomy, or a different computational assignment. An atlas is a scientific resource that evolves; its authority should come from traceable evidence rather than the permanence implied by a familiar name.
13. Cell proportions are a compositional measurement
When results are expressed as percentages of captured cells, the categories share a fixed total. An increase in one group's fraction necessarily changes the fractions of others, even if their absolute abundance in the tissue has not changed. This mathematical constraint can complicate biological interpretation.
Suppose a hypothetical sample contains one hundred cells of type A and one hundred of type B. Type A accounts for half the total. In a second sample, type A remains at one hundred while type B rises to three hundred. Type A now accounts for one quarter. Its fraction has fallen, but its absolute count has not. Describing the result simply as loss of type A would be incorrect in this constructed example.
Real single-cell datasets add further uncertainty because capture and filtering may differ among populations. A reported fraction is conditional on the measurement process. It may be a useful comparison, but the intended claim should match that conditional quantity unless additional evidence supports an inference about absolute tissue abundance.
The analysis can be clearer by stating the denominator explicitly and distinguishing relative composition from absolute quantity. If a biological story depends on actual depletion or expansion, it may require corroborating measurements designed for that question. The point is not that percentages are bad. They are often informative. The point is that a percentage is a relationship among categories, not an independent count of each category in the original organ. Understanding that relationship prevents a mathematically inevitable change from being mistaken for a separate biological event.
14. Differential expression needs a comparison that respects the design
After populations are identified, researchers often ask which genes differ between conditions. The calculation can produce long lists with very small significance values. Those values are meaningful only in relation to the statistical model and the independent units supporting the comparison. Large numbers of correlated cells can otherwise produce an exaggerated sense of certainty.
Squair and colleagues' 2021 study compared approaches to single-cell differential expression and found strong performance from methods that accounted for biological replicates through aggregation in their evaluated datasets. The primary analysis illustrates why respecting donor structure is central to inference. It does not imply that one named method is universally best for every design or that all cell-level analyses are inherently invalid.
A hypothetical disease comparison should therefore ask whether the expression difference appears across donors or is dominated by a few samples with many cells. It should consider the magnitude of the difference, its consistency, and uncertainty, not only whether a threshold is crossed. A tiny but precisely estimated difference may have a different biological meaning from a large, variable one.
The comparison also needs a clear target. Is the study asking about average expression within a defined population, the fraction of cells expressing a gene, or a shift in the distribution of states? Those are related but distinct questions. A method suited to one may not answer another. The scientific claim should be written before the list of significant genes becomes the story, so that the analysis serves the question rather than allowing the available output to define it retrospectively.
15. A pathway label is another layer of inference
A list of differentially expressed genes is often summarized using pathway or functional categories. This can help interpret a complicated result. It can also compress uncertainty into a familiar biological phrase that sounds more direct than the evidence. An enriched category is not automatically proof that the corresponding process has increased, decreased, or caused the condition under study.
Consider a hypothetical set of genes associated with a stress-response category. Their enrichment might reflect a genuine response in the tissue, preparation effects, a change in the mixture of states, or overlapping annotation among related genes. The pathway label organizes the evidence; it does not independently resolve those alternatives.
The direction of a process can be especially difficult to infer from expression alone. A gene's RNA abundance is one part of a larger system. A transcriptional response may indicate compensation rather than the primary change. A category containing both activating and inhibitory components may not translate into a simple up-or-down statement. The interpretation needs to match what was actually measured.
A stronger account identifies the specific genes driving the category, the consistency across samples, and independent observations that support the proposed process. It also distinguishes exploratory interpretation from a tested mechanism. The purpose of pathway analysis is to generate and organize hypotheses, not to turn a statistical list into a causal explanation by renaming it. A useful summary makes the next biological question clearer while preserving the uncertainty that remains in the underlying measurements.
16. A trajectory is not automatically a time-lapse
Single-cell analyses can arrange profiles along a proposed progression, often described as a trajectory or pseudotime. Such arrangements can illuminate gradual changes and suggest developmental or response pathways. But a set of cells measured at one moment is not the same as watching the same cells change over time.
A hypothetical dataset might contain profiles ranging from state A through intermediate patterns to state B. An algorithm could organize them into a smooth path. That path may reflect a real transition, but it could also combine distinct populations or reflect a gradient unrelated to the proposed direction. The arrangement alone does not establish where the process begins, how fast it proceeds, or whether individual cells traverse the entire path.
Direction usually requires additional information or assumptions. Time-resolved sampling, lineage-related evidence, or other measurements may help, depending on the question. Those contributions should be identified rather than allowing the visual arrow on the plot to imply that direction was directly observed. An inferred ordering can be useful without being a literal history of each cell.
The same caution applies to branching diagrams. A branch can be a hypothesis about alternative fates, not proof that a particular ancestor produced both outcomes. The reader should ask which parts of the trajectory are data-supported relationships and which parts rely on the model. This is not a reason to avoid trajectory analysis. It is a reason to describe it accurately: an analytical reconstruction that can guide experiments capable of testing the proposed biological sequence.
17. Removing cells from tissue removes relationships
A dissociated cell profile contains information about the cell's measured molecular state, but it may no longer retain the exact neighborhood in which the cell existed. Spatial relationships can be central to a biological explanation. Two populations present in the same sample are not necessarily adjacent, and apparent compatibility of their expression programs does not prove that they interacted.
A hypothetical analysis might identify one population expressing a signaling molecule and another expressing a corresponding receptor. That is a basis for a communication hypothesis. It does not establish that the cells were close enough, that the relevant proteins were present and active, or that the proposed signal produced the observed response. Co-occurrence in a dataset is a weaker relationship than demonstrated communication in tissue.
Spatially resolved measurements can add important constraints, but they have their own resolution and measurement limits. The correct question is what relationship the additional method actually resolves. A spatial region containing several cells does not necessarily identify the exact source of every signal. Combining methods is powerful when their complementary information is understood rather than when one is treated as a universal validation stamp.
For interpretation, distinguish four claims: the populations are present, they occupy related locations, they have compatible signaling machinery, and they causally influence one another. Each claim requires a different level of evidence. A clear paper can build through those levels. A compressed headline may skip directly from expression to communication. Restoring the missing steps makes the biological story more credible and the next experiment more focused.
18. More modalities can expose disagreement rather than erase it
Measuring RNA alongside proteins, chromatin features, or spatial information can help distinguish cell identities and states. The additional data do not guarantee that every modality will tell the same story. Differences may be informative because the measurements capture different aspects of biology and may operate on different timescales or with different sensitivity.
Imagine a hypothetical population with elevated RNA for a particular molecule but no corresponding increase in the measured protein. The discrepancy could reflect biological regulation, timing, technical limitations, or a combination. It should not automatically be resolved by declaring one modality correct and the other defective. The question is which explanation fits the measurement properties and the biological context.
Integrated analyses can also conceal missingness. If one modality is measured in one set of cells and another in a different set, the combined representation may rely on statistical matching rather than direct joint observation. That can be useful, but a cell-level statement should distinguish measured combinations from inferred combinations. The integrated map is not always a table in which every feature was observed in every cell.
A robust interpretation treats disagreement as something to investigate. It records which data support the annotation, which support the proposed mechanism, and where the evidence remains incomplete. The benefit of multiple modalities is not simply a richer-looking figure. It is the opportunity to test explanations against observations with different limitations. That opportunity is lost if integration smooths away every inconsistency before anyone asks whether the inconsistency contains the most interesting biology.
19. The rare population has a discovery and a confirmation problem
Rare cells can be scientifically important, and single-cell methods can make them visible. Their rarity also makes them vulnerable to sampling variation and artifacts. A handful of unusual profiles may be a genuine population, a technical mixture, or an idiosyncratic feature of one sample. The burden is to distinguish these possibilities without discarding novelty simply because it is unfamiliar.
A hypothetical study finds twelve unusual cells among twenty thousand. If all twelve come from one donor or one processing batch, the pattern has a different evidentiary status from a similar population recovered independently across many donors. The total cell count can obscure that concentration. Reporting the distribution across samples is essential.
Confirmation should match the claim. Reproducing the same cluster with a slightly different computational setting tests analytical stability. Finding related profiles in an independent dataset tests broader reproducibility. Locating the proposed cells in tissue using another measurement adds a different kind of evidence. Demonstrating a distinctive function goes further still. These steps should not be treated as interchangeable confirmations.
The naming decision can wait for the evidence to mature. A provisional label preserves the finding and invites investigation without forcing it into an established category or elevating it prematurely to a new cell type. Rare-population research benefits from this restraint because false discoveries can consume substantial follow-up effort, while overzealous filtering can erase real discoveries. The goal is a transparent path from unusual observations to a biological entity whose existence and significance others can test.
20. Sequencing depth and sampling breadth buy different information
A research budget can be spent on more cells, more molecular information per cell, more tissue regions, or more independent donors. These choices are not substitutes in a simple arithmetic sense. Each reduces a different uncertainty. A study should identify the uncertainty that limits its central claim before treating a larger total as an automatic improvement.
In a hypothetical search for a rare population, observing more cells from an appropriate sampling frame may improve the chance of encountering it. If the problem is weak detection of a particular transcript within already captured cells, a different allocation may be needed. If the central question concerns variability among people, additional donors can be more important than an increasingly detailed account of one person's cells. The optimal design depends on the scientific target.
This distinction also affects reuse of an existing atlas. A dataset built to classify common populations may not support a precise estimate of rare-cell abundance. A deeply measured set of cells from one region may not describe anatomical diversity. The data can remain excellent for their original purpose while being insufficient for a new claim. Reuse requires evaluating that fit rather than assuming that a large public resource is universally informative.
An honest limitations statement should therefore identify what additional observation would most improve the conclusion. More data is too vague. More independent donors, better coverage of a tissue boundary, or a measurement that resolves an ambiguous signal are actionable descriptions. They connect uncertainty to study design and help readers distinguish a genuine evidence gap from a generic call for further research.
21. A useful atlas makes its gaps legible
The most valuable cell atlas is not necessarily the one with the largest headline count. It is the one whose coverage, measurements, and uncertainties are clear enough to support reliable reuse. A million profiles from a narrow set of donors and tissue locations answer different questions from a smaller but more deliberately varied collection. Both can be useful if their scope is visible.
For a new single-cell claim, begin with the sampling frame: who contributed, what tissue was taken, and how the samples relate. Then follow the measurement chain: what was recovered, what was detected, what was filtered, and how the remaining profiles were organized. Finally, examine the inference: which comparisons respect independent biological units, which labels are provisional, and which mechanisms have evidence beyond expression patterns.
The methodological studies discussed here span 2017 through 2022, while the Human Cell Atlas resource was checked in October 2026. Their age does not make the underlying problems obsolete, nor does their existence imply that current studies ignore them. They provide a set of questions for evaluating new work rather than a blanket verdict on the field.
Single-cell sequencing has made biological variation visible at a scale that can transform a research question. Its limitations are most productive when treated as part of that question. The missing cell, the uncertain zero, the donor effect, and the lost neighborhood are not footnotes to a finished map. They define how the map relates to the tissue. Making those relationships explicit is what turns a beautiful collection of points into evidence another scientist can use.