When a mutation's signature is noise, not a shift

What the work claims

This is a primary single-cell result from an Edinburgh group.¹ The 16p11.2 microdeletion, a loss of roughly 574 kilobases across a gene-dense interval of chromosome 16, is one of the strongest common genetic risk factors for neurodevelopmental and metabolic conditions; the paper cites odds ratios of about 43-fold for obesity and 40-fold for autism in carriers. Its penetrance and expressivity are famously variable, and the standard explanation attributes that variability to the rest of the genome or to environment. The paper makes a more radical claim: that the deletion itself increases variability, that the 16p11.2 locus normally acts to stabilize gene expression between cells, and that losing one copy lets transcription drift apart cell to cell across the whole genome.

That is a claim about variance layered on top of the more familiar change in means, and the variance part is the kind most disease-modeling pipelines are not built to detect. If it holds, unusual transcriptional variability is not measurement noise to be averaged away; it is a candidate mechanism.

How it works

The design is the part to trust. The authors used human induced pluripotent stem cell (iPSC) lines that are isogenic, meaning identical in genetic background, except for whether they carry a heterozygous 16p11.2 deletion. Three deletion lines, each made by a separate CRISPR targeting event, and two engineered controls, all derived from one ancestral line, were confirmed by SNP array to differ only at the intended locus. That isolates the deletion from the usual confounder in any variability study, which is that different people's cells differ for a thousand reasons. The lines were differentiated in two-dimensional culture toward ventral telencephalic interneuron progenitors, the inhibitory-neuron precursors implicated in the excitation-inhibition imbalance long hypothesized in autism, and profiled by single-cell RNA sequencing, yielding a progenitor population of 12,713 cells.

Two analyses carry the argument. A conventional differential-expression test found hundreds of genes changed on average, enriched for cell signaling, chromatin regulation, and autism and obesity gene sets, the expected result. The novel step was Expression Variation Analysis (EVA), a method borrowed from cancer heterogeneity work that compares the variance of transcriptional profiles rather than their averages. By EVA, deletion progenitors were significantly more variable than controls, and crucially this held not only for functionally coherent gene regulatory networks (regulons defined by a transcription factor and its targets, 94 of them) but also for randomly chosen gene sets. Variability that shows up even in random gene collections is, on the authors' reading, the signature of a genome-wide effect rather than a few dosage-sensitive genes, though as the next section argues it is also the signature of a technical one. The increase in variability was only weakly related to how much a gene set's mean expression had changed (correlation about 0.25), which argues the scatter is not simply a byproduct of the shift in averages. The regulons that gained the most variability were enriched for cell-cycle genes, echoing the same group's earlier finding of increased cell-cycle variability between whole deletion organoids.

Where a skeptic should push

The most load-bearing worry is that increased cell-to-cell variability in a dish can be manufactured by uneven differentiation rather than by biology. If the deletion lines simply differentiate a bit less synchronously, their cells will occupy a wider spread of states, and any variance metric will rise, most obviously for cell-cycle genes if the cultures are less synchronized in the cycle. The isogenic design removes genetic-background noise but not this differentiation-asynchrony confound, and it is exactly the axis the paper's strongest enrichment (cell cycle) would predict. The authors point to the effect on random gene sets, the weak coupling to the mean shift, and a convergent organoid result, but none decisively excludes a technical origin. A genome-wide rise in scatter that appears even in random genes is equally the fingerprint of a mundane artifact: losing one copy of a block of genes can lower average expression and raise dropout across the transcriptome, and uneven capture depth, ambient RNA, or a partial confound between genotype and sequencing run (the samples were split across two runs) all inflate apparent variance without any regulatory destabilization. Deeper still, variation analysis on pooled cells cannot separate two very different things: a genuine rise in single-cell stochastic noise, which is what a stability-restoring drug would target, and a wider mixture of cell substates, in which each substate is just as tightly controlled but the population spans more of them. These carry opposite therapeutic readings, and only within-substate or pseudotime-matched analysis, not done here, can tell them apart. Until then the honest description is increased heterogeneity, not demonstrated single-cell instability, and differentiation asynchrony is simply the time-axis version of the same ambiguity.

The evidence is also bounded: two-dimensional culture rather than three-dimensional tissue, progenitors rather than mature interneurons, one sex (male), one ancestral background, and only three deletion and two control lines, which is thin for a claim about variance specifically, since between-line differences can leak into an apparent between-cell spread unless the variance is measured within each line. And "the locus stabilizes transcription genome-wide" is a strong causal statement resting on association; the deletion carries plausible mediators (chromatin and proliferation regulators such as those in the interval), but the specific gene responsible for the destabilization is not identified here. The right reading is that a genome-wide increase in transcriptional variability is a real and interesting candidate, demonstrated in a specific in-vitro window, not a settled property of the human disorder.

Screening for a phenotype made of variance

For organoid disease models and the drug discovery built on them, this is a warning about the shape of the phenotype, and it lands on a familiar sore spot: reproducibility. If increased variability is part of a genotype's molecular signature, then two habits are worth rethinking. First, sampling. A disease model built from one or a few iPSC lines or organoids will under-represent a phenotype whose essence is spread, and the resulting run-to-run inconsistency will look like sloppy protocol when it is actually the biology. If the deletion encodes the variability, it is exportable across labs as irreproducibility. That inverts the usual generalization trap: instead of one line being wrongly promoted to a property of the tissue, here the property of the tissue is precisely that no single line represents it.

Second, and more consequential for screening, the statistics, with an important caveat. This genotype also moves means, in hundreds of genes, so a conventional screen is not helpless here. But a mean-based readout is blind to the part of the effect that lives in the spread, and if any of that spread is real, a treatment could tighten or widen the cell-to-cell distribution without moving its center and a mean-based assay would score nothing on that axis (the variance difference survives on random gene sets and tracks the mean shift only weakly, correlation about 0.25). Separately, an elevated baseline spread, whether biological or technical, raises the noise floor, so a screen on this genotype needs more independent lines and replicates to reach the same power, a direct cost and part of why iPSC CNS screens so often fail to replicate.

The opportunity is the flip side of the same coin, genuinely non-obvious but stated as a hypothesis with a trap attached. If the excess variability is real, a compound that makes cells more uniform, whether by damping true single-cell noise or by narrowing the mixture of states, could in principle help even if it never shifts the average, and single-cell dispersion is measurable with the sequencing the field already runs. The trap is that a screen which simply rewards a narrower distribution will preferentially pick cytostatic or mildly toxic compounds that collapse cell-state diversity nonspecifically, a classic false-hit generator, and there is as yet no evidence that narrowing dispersion does any good in 16p11.2. So dispersion is worth carrying as a secondary readout anchored to a functional or clinical outcome, not promoted to a standalone endpoint. The accompanying hype-correction stands on its own: for high-variance genotypes a tidy single-line phenotype may not generalize, and a hit called against a mean difference in an underpowered, high-variance landscape may not survive more lines.

The bottom line

Established, within its window: in an isogenic human iPSC interneuron-progenitor model, the heterozygous 16p11.2 deletion is associated with hundreds of average expression changes and, more distinctively, a genome-wide increase in cell-to-cell transcriptional variability that is only weakly explained by the mean shift and is strongest in cell-cycle regulons. Still hypothesis: that variability itself is a disease mechanism, and that it persists beyond two-dimensional progenitors in three male lines. The confounds to exclude first are technical: a wider mixture of cell states rather than higher single-cell noise, differentiation asynchrony, and mean or dropout differences across genotype and sequencing run masquerading as biological variance. What would confirm the stronger reading is within-substate or pseudotime-matched variance analysis across many independent lines and at mature neuronal stages, plus a perturbation that restores a deleted gene and lowers variability; what would break it is showing the excess scatter is a batch, dropout, or differentiation-efficiency artifact. The defensible drug-discovery upshot is narrower than a slogan: for copy-number-variant models, sample enough independent lines to see the spread, treat cell-to-cell dispersion as a complementary axis a mean-only screen can miss, and, if you ever screen on it, anchor it to a functional readout so you are not merely rewarding whatever compound flattens diversity by being toxic.