Contents
48 — Single-cell and spatial genomics Part 10 — Functional genomics

48 — Single-cell and spatial genomics

Before this: Ch 47 · Ch 41 · Ch 01 §5 · Time: ~50 min

Statistics needed: S2 Distributions · S4 Hypothesis testing · S7 High-dimensional data

What you'll be able to do

The core idea

A bulk measurement of a tissue is a weighted mean over a mixture whose weights you did not observe:

b_g = Σ_k π_k · μ_gk          π = cell-type proportions (unknown)
                              μ = per-cell-type expression (unknown)

You observe the left-hand side. Both factors on the right are free. So when b_g changes between two conditions, π changed, μ changed, or both — and bulk data cannot tell you which. This is not a subtle statistical caveat; it is a non-identifiability, and it silently invalidates a large fraction of the mechanistic conclusions ever drawn from tissue expression data.

Single-cell methods fix it by measuring the mixture components instead of the mixture. The price is exact and unavoidable: a fixed sequencing budget divided across many more units means each unit is measured shallowly. Every difficulty in this chapter — the sparsity, the zeros, the fragile clustering, the abused embeddings — descends from that one trade.


1. The confound, with numbers

A tissue containing proliferating cells (type A) and quiescent stroma (type B). Gene MKI67 sits at 100 units per cell in A and 20 in B.

CompositionBulk value
Control60% A / 40% B0.6(100) + 0.4(20) = 68
Treated, story 1: composition shifts, no cell changes30% A / 70% B0.3(100) + 0.7(20) = 44
Treated, story 2: composition fixed, A downregulates to 6060% A / 40% B0.6(60) + 0.4(20) = 44

Story 1 says the drug killed proliferating cells. Story 2 says it stopped cells proliferating without killing them. These are different mechanisms with different clinical implications, and the bulk measurement is 44 in both cases. No amount of replication, no better statistical model, no deeper sequencing separates them. The information is not in the data.

The same non-identifiability produces gene–gene correlations that exist only in the mixture. Two genes each marking cell type A correlate perfectly across bulk samples that vary in A content, with no regulatory relationship whatsoever — a Simpson's paradox at the level of the tissue. Bulk co-expression networks are, to a first approximation, cell-composition networks.

2. Measuring one cell at a time

Three isolation strategies, each a different point on the same trade-off.

Plate-based (full-length). Sort one cell per well, build a full-length library per well. Hundreds to low thousands of cells; hundreds of thousands of reads each; across the whole transcript body, so isoforms, -specific expression and receptor sequences are readable. Expensive per cell.

Droplet-based (3′ or 5′ counting). A microfluidic junction co-encapsulates a cell and a barcoded bead in an oil droplet. Cells are loaded at low concentration so most droplets are empty. Tens of thousands to millions of cells, at ~10²–10⁴ reads each. Only the transcript end near the capture primer is sequenced, so you get a count per gene and almost no isoform information.

Combinatorial indexing (split-pool). No physical isolation at all. Fixed cells or nuclei are distributed across a plate, tagged with a well-specific barcode, pooled, redistributed, tagged again — repeat r times. A cell's identity is the concatenation of the barcodes it collected. With W wells and r rounds there are Wʳ combinations; the collision rate (two cells sharing the full barcode) is a birthday problem you tune by keeping the number of cells far below Wʳ. Cheap per cell and scales to millions, at the cost of shallow depth.

The barcoding scheme

Droplet and split-pool methods both encode identity in the read itself:

Read 1  ┌────────────────┬────────────┐
        │  cell barcode  │    UMI     │        e.g. 16 bp CB + 12 bp UMI
        │    16 bp       │   12 bp    │
        └────────────────┴────────────┘

Read 2  ┌──────────────────────────────────────────────┐
        │  cDNA — aligned to the transcriptome          │
        └──────────────────────────────────────────────┘

count[cell, gene] = number of DISTINCT UMIs observed for that (cell, gene)

The two barcodes do different jobs and confusing them is a common error.

UMIs collide. With a 12 bp UMI there are 4¹² ≈ 1.68 × 10⁷ tags; if m molecules of one gene in one cell are tagged, the expected number of distinct UMIs seen is N(1 − e^{−m/N}), which for m ≪ N is essentially m. For typical counts (tens per gene per cell) collisions are negligible; for a very highly expressed gene with a short UMI they are not, and the occupancy formula above is the correction.

Sequencing saturation makes the depth trade-off concrete: as you add reads, the fraction that are of an already-observed UMI rises. Past saturation, extra reads buy almost no new molecules, and the budget is better spent on more cells. Below it, extra cells buy you noisier cells.

Plate, full-lengthDroplet, 3′ countingCombinatorial indexing
Cells per experiment10²–10³10³–10⁶10⁴–10⁷
Reads per cell10⁵–10⁶10³–10⁴10²–10³
Transcript coveragefull body3′ (or 5′) end onlyend only
Isoforms / allele-specificyesessentially nono
Best formechanism in a known populationdiscovery, compositionatlas-scale surveys

Choose by question. Finding a cell type that is 0.1% of the tissue is a cell-number problem (you need ≥ ~10⁴ cells to see 10 of them); quantifying a splice-isoform switch is a depth problem. The two are in direct competition for the same budget.

3. The three characteristic artefacts

graph LR S["cell suspension<br/><i>dissociation stress,<br/>some cells lyse</i>"] --> D["droplet<br/>encapsulation"] D --> E1["<b>empty droplet</b><br/>ambient RNA only"] D --> E2["<b>singlet</b><br/>1 cell + ambient"] D --> E3["<b>doublet</b><br/>2 cells + ambient"] E1 --> M["count matrix"] E2 --> M E3 --> M

Doublets. Two cells in one partition share a barcode and appear as one cell with a summed profile. Cells load into droplets approximately Poisson with mean λ, so among occupied droplets the multiplet fraction is P(≥2)/P(≥1) ≈ λ/2 — linear in loading concentration, and therefore linear in the number of cells you recover. The vendor rule of thumb on the current droplet chemistry sits around 0.4% multiplets per 1,000 cells recovered — roughly half the ~0.8–1% of the preceding generation, and a fast-moving specification: check current figures rather than inheriting a number from an old protocol. At 0.4% a 10,000-cell run carries roughly 4% doublets; on the older chemistry the same run carried 8–10%, so a rate copied forward from a 2023 methods section is off by a factor of two. Detection exploits the fact that a doublet is a sum: simulate synthetic doublets by adding randomly chosen pairs of observed profiles, embed real and simulated cells together, and score each real cell by how many simulated doublets are in its neighbourhood. When several donors are pooled in one run, cross-donor doublets are found directly from — a barcode showing heterozygous reads at sites where both donors are homozygous for different is two cells (Ch 46). Same-type doublets remain nearly undetectable, which is why a "cluster" sitting exactly between two real clusters and co-expressing both marker sets should be treated as an artefact until proven otherwise.

Ambient RNA. Cells lysed during dissociation release mRNA into the suspension. Every droplet — occupied or not — receives a draw from this shared "soup" whose composition is the tissue average, dominated by whatever cell type is most abundant and most fragile. The effect on the matrix is a low-level, systematic false signal: haemoglobin transcripts in every cell of a blood-rich tissue, insulin in every cell of a pancreas islet prep. Correction estimates the soup profile from the empty barcodes and subtracts a per-cell contamination fraction, either by a simple ratio estimator using genes known to be off in a given cell type, or by an explicit generative model that decomposes each cell's counts into endogenous plus ambient. The failure mode this creates — apparent co-expression of markers from different lineages — is indistinguishable at the level of a single cell from real biology, which is why it must be addressed by design and correction rather than by interpretation.

Empty droplets. Barcodes with only ambient RNA vastly outnumber real cells. The naive filter is a hard threshold on total counts, found from the knee of the barcode-rank curve:

 total
 counts   ┤ ▓▓▓▓▓▓▓▓
 (log)    ┤         ▓▓▓▓                     real cells
          ┤             ▓▓▓
          ┤                ▓▓  ← knee
          ┤                  ▓▓▓▓▓
          ┤                       ▓▓▓▓▓▓▓▓▓▓ ambient-only barcodes
          └────────────────────────────────────────────
                    barcode rank (log)

A hard threshold discards every small, RNA-poor real cell — and RNA content varies by an order of magnitude between cell types. The better test asks a different question: is this barcode's expression profile distinguishable from the ambient profile? Model the ambient composition as a Dirichlet-multinomial estimated from the deep tail, and test each low-count barcode's profile against it. A barcode with 300 counts that look nothing like the soup is a cell; one with 800 counts that look exactly like the soup is not.

4. The count matrix and the zeros

The output is an integer matrix, cells × genes, typically 90–95% zero, stored sparse and usually backed by a chunked on-disk format because it will not fit in memory at atlas scale.

The zeros have three distinct causes, and conflating them has generated a lot of bad method:

CauseWhat it isWhat to do
Biological absencethe gene is not transcribed in this cellnothing — this is signal
Samplingtranscribed, but no molecule was captured and sequencedmodel it; it is ordinary count noise
Technical failurethe cell/droplet failed (poor lysis, degraded RNA)QC, per-cell

The sampling contribution is much larger than intuition suggests. Capture efficiency in droplet protocols is on the order of 10–30% of a cell's mRNA molecules, and the surviving molecules are then subsampled by sequencing. Take a cell yielding n = 5,000 total and a gene at relative abundance p = 2 × 10⁻⁴ in that cell. Expected count = np = 1, and

P(count = 0) = (1 − p)^n ≈ e^(−np) = e^(−1) = 0.37

37% of cells that genuinely express this gene, at the same level, record zero. For p = 2 × 10⁻⁵, e^(−0.1) = 0.90. Nine in ten zeros, from a gene expressed in every single cell, with no special process at all.

Statistics: the Poisson limit of the binomial that turns (1 − p)^n into e^(−np) is S2 §2; the that appears below — a Poisson whose rate itself varies — is S2 §5, and why a plain Poisson is not enough for count data is S7 §7.

That calculation is the reason the field's earlier framing has been substantially revised. Around 2015–2017 scRNA-seq was routinely modelled with zero-inflated distributions — a negative binomial plus an extra point mass at zero representing "dropout" — and a small industry of methods grew up to fill those zeros back in. Careful re-examination of UMI-based data showed the extra component is largely unnecessary: once you account for the distribution of per-cell depths and per-gene means, the observed zero fractions are close to what a plain multinomial/Poisson–gamma (i.e. negative binomial) model predicts. The consensus position now:

The practical upshot, once you have S2 §5 in hand: use a count model with an offset. Do not transform-and-pretend-Gaussian if you can avoid it, and do not add machinery for a phenomenon that the sampling model already explains.

5. Quality control, and the cell type it deletes

Three per-cell statistics carry most of the signal: total , number of detected genes, and the fraction of counts from mitochondrial genes. The last is a leakage indicator: when the plasma membrane is compromised, cytoplasmic mRNA escapes while mitochondria — which have their own membranes (Ch 01) — are retained, so the mitochondrial proportion rises even though nothing about the mitochondria changed.

The danger is that all three statistics are genuine biological properties of cell types.

ThresholdWhat it is meant to removeWhat it actually removes
min_genes > 500failed dropletsplatelets, erythrocytes, neutrophils, small resting lymphocytes — all genuinely RNA-poor
mito_pct < 5%dying cellscardiomyocytes, hepatocytes, kidney proximal tubule, skeletal muscle — mitochondria-rich by function
max_counts capdoubletslarge secretory and polyploid cells

A fixed threshold applied globally is a filter on biology wearing the costume of a filter on quality, and it fails silently: the missing cell type simply never appears in any figure. Two disciplines fix most of it. Set thresholds adaptively per sample (e.g. median ± k·MAD on the log statistic) rather than as universal constants. And inspect what you removed: cluster the discarded cells and look for markers. Debris clusters incoherently; a deleted cell type clusters tightly and expresses a coherent programme.

6. Normalisation, features, and the neighbour graph

Total counts per cell vary by an order of magnitude, driven by both capture efficiency (technical) and cell size and RNA content (biological) — and those two are not separable from the data alone. Every choice therefore takes a position on how much of the size factor is nuisance. This is the compositional problem from Ch 47, sharpened: a single-cell library is a small multinomial draw, so only relative abundances are measured, and a genuine global change in transcriptional output is invisible without a spike-in.

Two workable approaches:

  1. Scale and log. Divide by total counts, multiply by a constant, log(x + 1). Cheap and ubiquitous. The pseudocount interacts with depth, and low-count genes end up with depth-dependent variance, so downstream is partly ordering cells by library size.
  2. Model, then take residuals. Fit a per-gene GLM with the cell's total count as an offset and use Pearson or deviance residuals as the normalised values. Variance stabilisation comes from the model rather than from a fixed transform, and the residual scale is comparable across genes of very different means.

Then feature selection: rank genes by variance in excess of the fitted mean–variance trend and keep ~1,000–5,000 highly variable genes. This is not merely a speed optimisation — carrying 20,000 mostly-noise genes into PCA is a signal-to-noise decision, not a completeness decision.

Then PCA to 10–50 components. This is the real dimensionality reduction; everything downstream — the neighbour graph, the clustering, the embedding, the trajectory — operates in PC space, not in the 2-D picture you eventually publish.

Statistics: PCA as eigenvectors of a covariance matrix, and the measured extent to which t-SNE and distort the PC-space distances they are built from, are covered in S7 §5 — the quantitative backing for §7 below.

Then a k-nearest-neighbour graph in PC space (k ≈ 10–50), built with approximate NN search at scale, with edges usually reweighted by shared-neighbour overlap so that spurious long edges are down-weighted. This graph is the substrate for the next two sections.

7. UMAP: what it preserves and what it does not

and t-SNE lay that neighbour graph out in two dimensions by minimising a loss that rewards keeping neighbours close and pushing non-neighbours apart. They are neighbourhood layout algorithms. They are not projections, they are not distance-preserving, and they have no metric interpretation.

Read from a UMAPDo not read from a UMAP
Which cells are near which cells (locally)The distance between two clusters — set by the repulsion term and the initialisation, not by expression divergence
Whether a population is connected or discrete-ishThe size or area of a cluster — a function of local density and repulsion, not of cell number or heterogeneity
That a labelling is coherent in neighbour spaceThe relative arrangement of clusters — left/right, above/below, angles between arms
Continuity worth investigating properlyDensity — UMAP does not preserve it, and t-SNE actively distorts it

Three facts that should end any quantitative argument made from an embedding:

The correct role: an embedding is a legend for a clustering, a way of showing that a labelling is spatially coherent. Every quantitative claim — how distinct, how many, how related — must be made in PC space, on the graph, or on expression itself, and only illustrated on the embedding.

8. Clustering and annotation

Community detection on the kNN graph — Louvain, or Leiden, which adds a refinement step guaranteeing the communities it returns are internally connected (Louvain can return disconnected "communities", which is a real bug, not a technicality).

The thing to internalise: the number of clusters is a parameter, not a discovery. Both algorithms optimise modularity at a chosen resolution γ, and the number of communities is monotone in γ. There is no data-driven optimum, and modularity has a resolution limit — below a size scale set by the total edge count, genuinely separate communities merge regardless of how distinct they are. So no single resolution can be correct for both an abundant cell type and a rare one in the same dataset.

What to do instead of hunting for the right γ: sweep it, build the tree of how clusters split and merge across the sweep, and justify each split independently — does it have markers that are not ambient or doublet artefacts, does it survive bootstrapping cells or genes, does it reproduce in a second sample? Granularity is then a per-branch decision.

maps clusters to known biology, two ways. Manually, from marker genes and prior knowledge (PTPRC for leukocytes, EPCAM for epithelium, CD3E for T cells). Or by reference-based label transfer: given an annotated reference, project query cells into a shared space, find each query cell's nearest reference neighbours, and transfer labels with a confidence weight. Fast, reproducible, and dangerous in one specific way — a classifier can only emit labels the reference contains, so a genuinely novel cell type is assigned to its nearest reference relative, often with high confidence. Always keep and inspect the uncertainty, and treat a cluster with uniformly low transfer confidence as interesting rather than as a failure.

Note also that a cluster is not an ontological category. It is a density feature at one resolution. The distinction between a cell type (stable, heritable identity) and a cell state (transient — cycling, stressed, activated, interferon-responding) is biological, and clustering does not respect it: a strong state signature will split every type in the dataset along the same axis.

9. Testing on the data you clustered

Two statistical errors are near-universal in this field, and both matter more than the choice of test.

Statistics: what a p-value's null distribution is, and how choosing an analysis after seeing the data destroys it, are S4 §3 and §8; the donor-level random effect behind the pseudobulk and mixed-model fixes is S7 §8.

Double-dipping (post-selection inference). The standard marker workflow clusters cells to maximise separation in expression space, then tests each gene for a difference between clusters using that same expression matrix. The null hypothesis "gene g has equal mean in group 1 and group 2" was used to construct the groups. The resulting p-values are wildly anticonservative: take a single homogeneous population with no structure whatsoever, force k = 2 clusters, and you will obtain markers at p < 10⁻¹⁰. This is not a small correction factor — it is a test whose null distribution is wrong by construction.

Two classes of fix actually work:

Note why the obvious fix fails: splitting cells into train and test does not work, because assigning cluster labels to held-out cells uses those cells' own expression, which reintroduces exactly the dependence you were trying to break. You have to split the counts, not the cells.

. When comparing conditions — disease vs control, treated vs untreated — the independent unit is the donor, not the cell. Ten thousand cells from three patients are three replicates, not ten thousand. A cell-level test treats them as independent, so the standard error shrinks by roughly √(cells per donor): with 2,000 cells per donor, the reported standard error is ~45× too small, and the false-positive rate approaches one. The fix is pseudobulk: sum counts within each (cell type, donor) pair, then run the ordinary bulk differential-expression pipeline from Ch 47 on a matrix with one column per donor per cell type. Cell-level mixed models with a donor random effect are the other defensible option. Naive Wilcoxon across cells is not.

10. Integration, and the biology it can erase

Cells from different donors, chemistries, sites or days separate by batch before they separate by cell type. Integration methods align them, in three broad families: removing batch as a linear factor in PC space with iterative centroid correction; matching cells between batches by mutual nearest neighbours and using the matched pairs to define correction vectors; or fitting a latent-variable model (typically a variational autoencoder) with batch as a covariate and using the latent space as the corrected representation.

Every one of them is instructed that batches ought to look alike. So if a batch genuinely contains a cell type the others lack, or a real condition-driven shift in state, they will remove it — and the usual diagnostic rewards exactly that failure, because batch-mixing entropy is maximised by destroying the difference. Evaluate with paired metrics that score biological against batch mixing, and test explicitly that two populations known to be distinct are not merged.

The real fix is experimental. Multiplex donors within a run — antibody hashing, or genetic demultiplexing from the donors' own SNPs — so that donor is not confounded with batch, and the correction has something to hold constant.

11. Trajectories, pseudotime, and RNA velocity

Some processes are asynchronous and continuous: differentiation, activation, cell cycle. A single snapshot then contains cells at every stage, and the ordering can be recovered from geometry. Pseudotime fits a graph or principal curve through the cells in PC space, designates a root, and orders cells by distance along it.

Three honest caveats. Pseudotime is defined only up to a monotone reparametrisation: it has no units and no calibration to wall-clock time. Cell density along the trajectory reflects both dwell time in a state and sampling and survival biases, so "cells accumulate at stage X" is confounded. And the root is supplied by you, not by the data — the geometry is symmetric.

attempts to break that symmetry. Reads mapping to report unspliced pre-mRNA; reads spanning exon junctions report mature mRNA. A minimal kinetic model:

du/dt = α(t) − β·u        u = unspliced,  α = transcription
ds/dt = β·u   − γ·s        s = spliced,    β = splicing, γ = degradation

Under a steady-state assumption, s* = (β/γ)u, so fitting a line through the (u, s) cloud across cells estimates β/γ; a cell above the line has ds/dt < 0 (gene switching off), below it ds/dt > 0 (switching on). Pooling the sign across genes gives each cell a displacement vector, drawn as an arrow on an embedding.

Treat the arrows as a hypothesis, for reasons that are structural rather than incidental:

The confirmatory experiments are lineage tracing with heritable barcodes, metabolic labelling that marks newly transcribed RNA directly, and actual time courses.

12. Multi-modal single cell

Once a cell is barcoded, anything you can convert into a barcoded sequencing library can be measured alongside its RNA.

AssaySecond modalityStatistical character
CITE-seqsurface protein, via antibodies carrying oligo tagscounts are dense, not sparse; high background from unbound antibody; needs its own , often using empty droplets to estimate the background level
single-cell ATACopen chromatin (Ch 49)near-binary: a cell has at most 2 copies of a locus, so per-peak counts are 0/1/2. ~10⁵ features, extremely sparse. TF-IDF weighting plus truncated SVD replaces
MultiomeATAC and RNA in the same nucleuslets you link a peak to a gene by covariation across cells, rather than by proximity — the data-driven route to –gene assignment used in Ch 52

A practical note that affects every tissue study: for frozen material, fibrous tissue, or large cells that do not survive dissociation, the assay operates on nuclei rather than cells. Single-nucleus data has a higher intronic fraction, loses cytoplasmic and mitochondrial transcripts, and recovers a different cell-type composition than single-cell data from the same tissue. Neither composition is the truth; both are protocol-dependent.

13. Spatial genomics

Dissociation destroys position. It also biases composition — cell types differ enormously in how well they survive being pulled out of a tissue — so the loss is not only of coordinates but of the denominator. Spatial methods keep the tissue intact and split into two families that make opposite trade-offs.

Imaging / in situ. Fluorescent probes are hybridised to chosen transcripts over multiple rounds, each transcript species assigned an error-correcting binary code across rounds (Hamming distance ≥ 4 between codewords, so single-round errors are correctable). Imaging after each round reads the code out. Alternatively, sequencing chemistry is run directly on the tissue section.

Capture / sequencing-based. A slide carries an array of spatially barcoded capture features. Tissue is placed on it, permeabilised, and released mRNA binds the feature beneath it; the barcode encodes (x, y).

capture spot (≈50 µm)              what the counts actually are
 ┌───────────────┐
 │  ◍  ◍   ◍     │   3 tumour     y_g = depth · Σ_k π_k μ_gk
 │    ◍   ●  ●   │   2 T cells                └── the §1 problem, again
 │  ◍   ●        │   1 fibroblast
 └───────────────┘

is therefore §1 run deliberately in reverse: given a single-cell reference of per-cell-type profiles μ, estimate the proportions π in each spot — non-negative regression, or a topic model, with a count likelihood and a sum-to-one . The inverse-problem warnings all apply: the reference must contain every type actually present (anything missing is absorbed into the nearest available profile), and cell types with collinear expression profiles are poorly identified no matter how deep the data.

Imaging / in situCapture / sequencing
Resolutionsubcellularfeature-limited: sub-µm to tens of µm
Genesfixed panelwhole transcriptome
Key artefactsegmentation errormulti-cell mixing → needs deconvolution
Detection sensitivityhigh per targeted genelower, capture-limited
Best fortesting a hypothesis about known genesdiscovery, tissue-domain mapping

What spatial data buys, once you have it: neighbourhood analysis (which cell types abut which, tested against a null of random placement), spatially variable genes (spatial autocorrelation — Moran's I, or a Gaussian process with a spatial kernel against a no-structure null), tissue domain segmentation (clustering that uses coordinates as well as expression), and candidate cell–cell signalling from ligand–receptor — where colocalisation is a necessary and very far from sufficient condition for signalling.

14. Atlases

A cell atlas is a large, annotated, harmonised reference: many donors, many tissues, common ontology terms, standardised processing. Its value is not that it contains an answer, but that it is the reference object several other analyses require — a shared vocabulary and coordinate system, a target for label transfer (§8), a reference for spot (§13), and the denominator that lets you ask which cell type expresses the gene at a locus (Ch 51, Ch 52).

Consortium efforts have assembled data on the order of tens of millions of cells across dozens of human tissues, with integrated draft releases appearing through the mid-2020s; cell counts and tissue move quickly enough that a current figure should be taken from the consortium's own release page rather than from a textbook.

The caveat is structural and easy to forget: an atlas inherits every bias of its donors, dissociation protocols, and QC thresholds. A cell type that does not survive dissociation, or that falls below a gene-count threshold, is absent from the atlas — and therefore absent from every label transfer and every deconvolution that uses it, in a way that produces no error message and no warning.


Common misconceptions

What people thinkWhat's actually true
A bulk expression change means cells changed their expressionIt is equally consistent with a change in cell-type proportions. The two are not separable from bulk data — see §1
Zeros in scRNA-seq are "dropout" needing For data most zeros are ordinary sampling and are predicted by a plain count model. The zero-inflation framing has been substantially retired, and imputation manufactures correlation structure
Clusters far apart on a are more different than clusters close togetherBetween-cluster distance in a UMAP carries no information. Neither does cluster size, density, or relative arrangement
The clustering found the right number of cell typesThe number of clusters is set by the resolution parameter. There is no data-driven optimum, and modularity's resolution limit means one setting cannot suit both abundant and rare populations
A marker gene with p = 10⁻³⁰⁰ is definitively a markerThe clusters were defined using that gene's expression. Uncorrected post-clustering p-values are anticonservative by construction, even on data with no structure at all
More cells means more statistical power for a condition comparisonOnly if the cells are independent. Across conditions the unit is the donor; cell-level tests pseudoreplicate and inflate significance by roughly the square root of cells per donor
Batch correction removes technical variationIt removes between-batch variation, technical or biological. When batch is confounded with condition or donor, it removes the finding
arrows show the direction of differentiationThey show the sign of a steady-state residual under strong kinetic assumptions, projected onto a non-metric embedding. Documented cases point backwards
gives per-cell expression with coordinatesCapture-based methods give per-spot mixtures needing ; imaging-based methods give per-molecule positions needing segmentation. Neither hands you clean single-cell profiles

Worked example: one gene, three errors avoided

A trial biopsy, before and after treatment. Bulk shows MKI67 falling from 68 to 44 normalised units, p = 0.001 across replicates.

Step 1 — what bulk can support. Using the model of §1 with two cell types:

Both fit the observation exactly. The bulk result supports "MKI67 signal fell" and nothing more. Any mechanistic sentence beyond that is unsupported.

Step 2 — single-cell, and expecting the zeros. Run droplet scRNA-seq, recover 12,000 cells across 3 donors per arm. Type A cells yield a median 5,000 ; MKI67 is 2 × 10⁻⁴ of transcripts in a proliferating cell. Expected count = 5,000 × 2 × 10⁻⁴ = 1.0, so

P(zero | genuinely expressing, type A) = e^(−1.0)  = 0.37
P(zero | genuinely expressing, type B) = e^(−0.2)  = 0.82     (p = 4×10⁻⁵)

37% of A cells and 82% of B cells record zero. If you had reached for a zero-inflated model or an step you would have been correcting a phenomenon that is fully explained by np being small. Note also what this does to a per-cell "expressed/not expressed" call: it is mostly a readout of sequencing depth.

Step 3 — doublet budget. The multiplet rate in §3 is linear in the cells recovered per run, so budget per arm, not on the pooled total: 6,000 cells recovered at ~0.4% multiplets per 1,000 recovered implies roughly 2.4% doublets, ~144 cells per arm — ~290 across the study. An A/B doublet has ≈ (100 + 20)/2 = 60 units of MKI67 and expresses both lineages' markers — it will form its own small cluster sitting between A and B, and if annotated naively it becomes a spurious "intermediate transitional state". Remove it by simulation-based scoring, and by demultiplexing since donors were pooled.

Step 4 — the answer, computed correctly. Cluster (on split counts, §9), annotate, then:

ControlTreated
Fraction type A0.610.29
Pseudobulk MKI67 in A, per donor98, 103, 99101, 97, 104

Two separate tests, on two separate quantities. A composition test on the proportions (3 vs 3 donors, on a logit or a Dirichlet-multinomial to respect the ) is significant. A pseudobulk state test within type A — 3 values vs 3 values, mean 100.0 vs 100.7 — is not. Story 1 is right: the drug depleted proliferating cells and did not slow the survivors.

Had you instead run a Wilcoxon test across all type-A cells (≈3,700 vs ≈1,700 cells), the standard error would be understated by roughly √(1,200 cells per donor) ≈ 35×, and that 0.7% difference would have returned a small p-value and the opposite mechanistic conclusion.

Connections