When a stretch of chromosome is duplicated or deleted, the genes inside it usually change their expression. Current single-cell methods treat that effect as one fixed number per gene: three copies give roughly 1.5 times the transcripts of two copies, whatever the cell is doing. Copy number is a gene-dosage change, meaning a change in how many physical copies of a gene the cell carries. Cells, however, pass through differentiation and activation programs, and there is good reason to expect some dosage changes to be buffered at one stage and exposed at another. We propose a statistical test for exactly that question. Along a developmental ordering of single cells, does the expression effect of a copy-number alteration change, and where? The test has to stay honest about two things that break naive versions of it. The first is statistical power: testing every gene at every stage drowns in multiple testing. The second is confounding: genetically distinct clones often sit at different stages, so a stage difference can look like a dosage effect.
The central move is to change the unit of testing. Genes inside one copy-number segment share the same dosage change in the same cells. We therefore pool them and ask a single question per segment: is the dosage response a flat line or a curve over developmental time? The shape of that curve is summarized with a handful of smooth basis functions. This turns roughly $10^4$ genes times dozens of time bins into a few hundred segments times two or three curve coefficients. Two borrowed ideas make it work. The first is adjacent-bin merging from CNV association studies, which lengthens segments to accumulate weak signal. The second is functional summaries of expression along pseudotime from dynamic eQTL work. Pseudotime is an ordering of cells by transcriptional similarity that stands in for developmental progress.
Core Claim
We claim that time-varying dosage sensitivity is detectable at the level of the copy-number segment, and that it is not detectable at the level of the gene, in datasets of realistic size. We also claim that it can only be estimated without bias if pseudotime is computed without the genes being tested and if clone identity is decoupled from developmental position. For gene $g$ in cell $c$ with pseudotime $t_c$ and copy number $n_{gc}$, we model expected expression $\mu_{gc}$ as
$$ \log \mu_{gc} = \log L_c + a_g(t_c) + \beta_{s(g)}(t_c)\,\log\frac{n_{gc}}{2} + \gamma_g \log\frac{n_{gc}}{2}, $$
In this model, $L_c$ is library size, $a_g(t)$ is the gene's baseline trajectory shared by all clones, $s(g)$ is the segment containing $g$, $\gamma_g$ is a gene-specific constant deviation, and $\beta_s(t)$ is the segment's dosage-response curve. We write $\beta_s(t) = b_{s0} + \sum_{k=1}^{K} b_{sk}\,\phi_k(t)$ with $K \le 3$ smooth basis functions $\phi_k$. The null hypothesis of a constant dosage effect is $b_{s1}=\dots=b_{sK}=0$, which gives one $K$-degree-of-freedom test per segment. When $\beta_s(t)$ is flat, the model collapses to the constant dosage response that clonealign uses to assign cells to clones. That nesting is deliberate. The proposal succeeds only if the curve is measurably not flat.
Positioning Against Prior Work
Three lines of work each hold one piece of this, and each supplies a specific method we reuse.
The first line joins copy number to expression. clonealign assigns RNA-profiled cells to DNA-defined clones using a fixed dosage response. SCATrEx maps expression onto a fixed copy-number evolutionary tree and finds within-clone expression states. Both treat residual expression variation beyond dosage as noise or as discrete sub-states. Neither lets the dosage effect itself depend on continuous developmental position, and neither addresses genome-wide testing. These papers motivate our question and supply our clone assignments. SCATrEx's reported XIST underexpression in X-deleted clones also serves as a positive control.
The second line is trajectory-aware genetic association. ti-scMR shows that summarizing expression along pseudotime with functional principal components rescues power under uneven cell density. Its genetic variants, however, are SNPs only, and it assumes a diploid genome. We reuse its functional-summary idea but apply it to the dosage coefficient rather than to expression.
The third line concerns power in CNV association. A robust statistical method for Genome-wide association analysis of human copy number variation merges adjacent bins with similar state distributions under an empirical Bayes prior, and reports near-zero false discovery where circular binary segmentation spikes. We reuse the merging logic, applied here to adjacent segments whose dosage curves are statistically indistinguishable.
The closest existing combination is the HCC single-cell study of SF3A3 and metabolic reprogramming. It found that inferred CNV burden rises toward the end of an epithelial pseudotime. That design is also the clearest warning. It infers CNV from the same expression used to build pseudotime, so the association is partly circular, and it has no formal test. We therefore treat expression-inferred CNV as a secondary setting only. Our primary analyses use copy number measured from DNA.
Research Program
The work has three parts, each written as software and evaluated separately.
The first part addresses leakage and confounding. Pseudotime is computed leave-one-segment-out: when we test segment $s$, the trajectory is inferred with Slingshot, as in ti-scMR, from genes outside $s$. This keeps a dosage change from bending the time axis it is tested against. Clone-by-time confounding is handled by estimating $a_g(t)$ jointly across all clones. We also restrict inference to the pseudotime range where clones overlap, and we report overlap explicitly, because outside that range dosage and stage cannot be separated.
The second part is segment-level curve testing. The model above is fit per segment as a negative binomial mixed model with the $\gamma_g$ shrunk toward zero. The $K$-degree-of-freedom test uses a permutation null that shuffles copy-number labels within pseudotime bins. This preserves the density imbalance that ti-scMR identified as the source of false positives. Adjacent segments whose fitted curves do not differ are merged before the final test, following the dynamic-merging logic. Merging reduces the number of tests and pools cells across neighbouring alterations.
The third part is a calibrated simulator. It generates clonal copy-number profiles, a branching trajectory with uneven cell density, gene-level baseline dynamics, and dosage curves. The dosage curves are either flat or buffered early and exposed late. The simulator includes a tunable correlation between clone identity and pseudotime. It is built on top of the SCATrEx synthetic design of 1,000 cells so that results are comparable to that setting.
Baselines
The naive reference point is clone-level differential expression. We run the same clone assignments, pool all cells per clone, and test gene expression differences between clones with a standard negative binomial test and Benjamini–Hochberg correction. This ignores time entirely and can only detect constant effects. Its role is to show how much of the signal is time-varying at all.
The strongest existing method for the dosage component is clonealign's constant dosage model. It is fit on the same cells with the same clone labels, and we test its per-gene dosage coefficients. The number to beat on its own task comes from SA501, where clonealign reports predicted clone proportions of 80.7%, 16.7% and 2.6% against DNA ground truth of 82.3%, 10.8% and 6.9%. In simulation it reports a median assignment AUC above 0.8, as stated in clonealign. Our method must not degrade clone assignment when used as a drop-in replacement. It beats clonealign on the new question if, on held-out data, the time-varying curve improves predicted expression of segment genes. This is measured as held-out log-likelihood per cell, and the improvement must be significant under a clone-stratified cross-validation split.
For the temporal component, the strongest competitor is a gene-level dosage-by-pseudotime interaction model. It is ti-scMR's dynamic interaction eQTL with the SNP genotype replaced by copy number. It gets identical pseudotime, covariates and permutation budget, and Benjamini–Hochberg correction across genes. No reported result was found on Lacuna for any method testing time-varying copy-number dosage effects, so the comparison is made on our simulator and on real data at matched false discovery rate.
Two ablations isolate the key ingredients. The first keeps segment pooling but computes pseudotime on all genes, which isolates leakage. The second keeps leave-segment-out pseudotime but tests genes individually, which isolates pooling. We beat the gene-level baseline if we recover more true time-varying segments at 10% false discovery rate in simulation while keeping type I error nominal under the flat null.
Why not just…?
Why not just test each gene's dosage-by-pseudotime interaction and correct with FDR? This is the real competitor and it may be adequate in large amplifications, where many cells carry the change. Its weakness is that each gene contributes one noisy, dropout-prone interaction estimate. Correcting across about $10^4$ genes imposes the same attrition seen when the MoBa methylation study found no single site significant across more than 760,000 CpGs, even though grouped analyses recovered signal. The per-gene ablation measures this directly. If per-gene testing matches segment testing at equal false discovery rate, pooling is unnecessary and we will report that.
Why not just run clonealign or SCATrEx and then stratify each clone into pseudotime bins? Binning after the fact reintroduces the problems ti-scMR documents. Sparse bins at the ends of trajectories lose power, and bin boundaries are arbitrary. Binning also leaves the confounding untreated: if one clone dominates late pseudotime, bin-wise contrasts compare clones at unmatched densities. We include a binned variant as a secondary baseline on the confounded simulations to show this failure, or to show that it does not occur.
Why not just infer copy number from expression, as in the HCC and glioblastoma pericyte studies? Expression-smoothed copy number, as in the SF3A3 HCC study and the tumor-derived pericyte study, is estimated from the very expression whose dosage response we want to measure. Any stage-dependent expression shift therefore leaks into the copy-number call. On SA501 we will compare curves estimated from DNA-derived clones with curves from expression-inferred copy number, and report how often inferred copy number manufactures time-varying effects.
Why not just assume dosage effects are constant, as clonealign does? This is the null our method is built to reject. clonealign's success with only a fraction of genes being dosage-responsive shows that constant dosage is a strong first-order signal. If our test finds essentially no non-flat segments in real data at calibrated error rates, the conclusion is that time-varying dosage is rare. That is a negative result worth publishing, and in that case we would not recommend the extra model.
Proposed Experiments
Simulation comes first. On the SCATrEx-style 1,000-cell design, scaled to 5,000 and 20,000 cells, we vary three things: the fraction of segments with non-flat dosage curves, cell-density imbalance along pseudotime, and clone–pseudotime correlation from 0 to 0.8. Metrics are type I error under the flat null, power at 10% false discovery rate, and the false-positive rate specific to the confounded regime. The decisive check here is that the full method holds nominal error at high clone–pseudotime correlation where the leakage ablation does not.
Primary real data is the SA501 triple-negative breast cancer xenograft. It has independent single-cell DNA and RNA and published clone assignments, so copy number is measured rather than inferred. The X-chromosome deletion with XIST underexpression reported by SCATrEx is a positive control for a nonzero $b_{s0}$. Segments containing the MHC class I genes that both clonealign and SCATrEx found downregulated are candidates for time-varying effects. We will test whether that downregulation deepens along the epithelial trajectory rather than being uniform. Clone overlap in pseudotime will be reported first, because the analysis is only interpretable where overlap exists.
Secondary data is the public HCC single-cell compendium. The goal there is to quantify the circularity of expression-inferred copy number, not to make discovery claims.
The decisive empirical check on real data has two conditions. At least a few merged segments must show non-flat dosage curves under the within-bin permutation null. Those curves must also improve held-out expression prediction over clonealign's constant model in clone-stratified cross-validation. If both hold, dosage sensitivity is shown to be stage-dependent and measurable. If neither holds, the constant-dosage assumption underlying current clone-mapping methods gains its first direct test.
References
- clonealign: statistical integration of independent single-cell RNA and DNA sequencing data from human cancers
- Mapping single-cell transcriptomes to copy number evolutionary trees
- ti-scMR: trajectory-inference-based dynamic single-cell Mendelian randomization identifies causal genes underlying phenotypic differences
- A robust statistical method for Genome-wide association analysis of human copy number variation
- Exploring the level of metabolic reprogramming and the role of prognostic factor SF3A3 in hepatocellular carcinoma through integrated single-cell landscape analysis
- Effects of prenatal exposure to (es)citalopram and maternal depression during pregnancy on DNA methylation and child neurodevelopment
- Immunoregulatory Roles of Tumor‐Originated Pericytes Identified by Single‐Cell Analysis in Glioblastoma