scPower accelerates and optimizes the design of multi-sample single cell transcriptomic studies

The study addresses the lack of fast, accurate power analysis tools for multi-sample single-cell RNA-seq experiments that compare individuals (e.g., differential expression and eQTL analyses at cell-type resolution). Existing approaches rely largely on simulations or bulk RNA-seq power models that do not capture scRNA-seq-specific sparsity, cell-type heterogeneity, or design trade-offs across sample size, cells per sample, and sequencing depth. The authors propose scPower, an analytic framework that models overall detection power as a function of (i) the probability that a gene is expressed within a cell type and across individuals and (ii) the statistical power to detect DE or eQTL effects, under user-specified significance control and realistic priors. The purpose is to enable rapid evaluation and optimization of experimental designs under budget constraints, providing actionable recommendations on how to allocate resources (samples, cells, reads) to maximize discoveries. This is important because cell-type–specific inter-individual comparisons require careful balancing of depth versus breadth to achieve sufficient detection and statistical power, especially for large-scale eQTL studies.

Prior work on power and design for scRNA-seq has focused on simulations (e.g., powsimR, muscat) or targeted questions like cell-type identification or effective sample size for eQTLs. Pseudobulk methods have emerged as top performers for multi-sample scRNA-seq DE analysis and negate zero-inflation at the aggregate level. Earlier analytic work on single-cell eQTL power considered effective sample size but did not model gene-level effect sizes or expression levels, nor did it provide a generalizable tool or cover DE. Bulk RNA-seq power methods exist for negative binomial and linear models and FDR/FWER control, but they do not explicitly handle scRNA-seq-specific expression sparsity or the dependence of expression detection on the number of cells measured per individual. scPower integrates these strands by combining cell-type–specific expression priors with analytic power models and multiple testing control, extending to budget optimization and generalization across platforms (10x Genomics, Drop-seq, Smart-seq2).

Overview: scPower analytically estimates overall detection power for multi-sample scRNA-seq designs by modeling (1) gene expression detectability within a cell type across individuals and (2) DE/eQTL test power, conditional on experimental parameters and prior distributions of expression and effect sizes. It supports two-group comparisons (extendable via GLMs) and integrates a cost model for budget-constrained optimization of sample size, cells per sample, and read depth. Data model and pseudobulk: For each cell type c, single-cell UMI counts per gene i are modeled as negative binomial (NB) with mean μ_ci and dispersion φ_ci. Counts are aggregated to pseudobulk by summing UMI counts over all cells of that type within each individual, yielding pseudobulk counts per gene per individual per cell type that follow an NB with mean and dispersion scaled by the number of cells per individual of that cell type. Expression priors and expression probability: Cell-type–specific expression priors are learned from pilot scRNA-seq data (e.g., PBMCs). The distribution of gene means across genes is modeled as a mixture of a zero component plus two left-censored gamma components to capture high-expression outliers. Mixture parameters vary linearly with average UMI counts and are linked to read depth (reads per cell) via empirically fitted saturation curves; dispersion is modeled as a mean–dispersion relationship (DESeq-like) per cell type. For a planned experiment, a gene’s prior expression rank determines its expected mean at the planned read depth via the fitted mixture; the pseudobulk NB parameters scale with the number of measured cells per individual. Expression is defined via a flexible threshold (default: minimal pseudobulk count > threshold in at least a specified fraction of individuals). The probability a gene exceeds the count threshold in an individual is obtained from the NB CDF; the probability it is expressed in at least k% of n_I individuals is from a binomial model. Summing over genes yields the expected number of expressed genes. DE power: For expressed genes, DE power is calculated analytically for NB regression (two-group comparison), with inputs: sample size per group, mean expression (from the prior rank mapping), dispersion, fold change (from prior DE studies), significance level α adjusted for multiple testing (FDR via Jung’s method or FWER Bonferroni). Pseudobulk was verified to be well described by NB without zero inflation. eQTL power: For expressed genes, eQTL power is calculated using linear regression on transformed counts. For genes with sufficiently large means, an F-test–based analytic calculation is used with effect sizes parameterized by R² (from prior eQTL studies) and FWER control (e.g., GTEx approach). For low-mean genes, scPower switches to a simulation-based calibration that samples genotypes (HWE), generates NB counts with effect β matching target R² after log transformation, and estimates power via linear regression p-values. The analytic and simulation methods are reconciled at a mean-count threshold determined by benchmarking. Overall detection power: For each prior DE/eQTL gene, overall detection power is the product of its expression probability and its DE/eQTL test power. The reported overall power is the mean across all prior genes. Multiple testing is handled as: DE via FDR (Jung) or FWER; eQTL via FWER with an assumed number of independent SNPs per gene (default 10, user-configurable). Cost model and optimization: Total cost C for 10x Genomics is modeled as library preparation cost (function of number of samples and cells per sample) plus sequencing cost (reads required as a function of read depth per cell). Given a fixed budget, scPower searches grids of cells per individual and read depth; sample size is determined from the budget. Doublets are modeled as increasing linearly with cells per lane (factor from 10x user guide), reducing usable singlets and shifting reads because doublets carry more reads (assumed factor 1.8). Mapping efficiency is applied (default 80%). The tool returns the parameter combination maximizing overall detection power. Analogous platform-specific adaptations are implemented for Drop-seq (constant doublet rate; per-cell costs) and Smart-seq2 (full-length reads; gene-length adjustments; expression thresholds per kb; dispersion modeled versus read depth; constant doublet rate). Validation and implementation: Expression probability predictions were validated by subsampling read depth and comparing predicted vs observed numbers of expressed genes across PBMC batches and an independent dataset (r² > 0.9). DE power estimates were benchmarked against powsimR and muscat (with adaptations for multi-sample pseudobulk), showing close agreement; eQTL analytic power was benchmarked against a custom simulation. Runtime and memory usage were compared, showing orders-of-magnitude efficiency gains. The framework is implemented as an R package with a Shiny web app; priors for 25 cell types across three tissues are provided, and users can generate custom priors.

- Expression detectability modeling: Predicted vs observed numbers of expressed genes per cell type matched closely across read depths and cell numbers (training and validation PBMC datasets), with r² > 0.9. - DE overall detection power: Using priors from a CLL subtype study (84 DEGs; median |log2FC| ≈ 2.8; sample size 6), maximal overall detection power reached ~74% at 3,000 cells per cell type and individual, total sample size 20 (10 per group), with FDR control. In this setting, DE test power approached ~98% but overall was limited by expression probability (~74%). - eQTL overall detection power: Using priors from Blueprint T cells (N=192; 5,132 eGenes; median absolute β ≈ 0.89), maximal detection power ~64% at sample size 200 and 3,000 cells per individual; similar results for monocytes (~65%). Power increased strongly with sample size; increasing cells per individual boosted both expression probability and eQTL power by raising mean expression. - Design trade-offs and optimization: For a DE study budgeted at €10,000, optimal parameters were ~1,200 cells per individual, 4 samples, 30,000 reads per cell. For an eQTL study budgeted at €30,000, optimal parameters were ~1,500 cells per individual, 242 samples, 10,000 reads per cell. Generally, shallow sequencing with many cells provides higher overall power than deep sequencing of fewer cells; optimal read depths were often ~10,000 reads per cell. - Parameter evolution with budget: Across prototypic and data-driven scenarios, the number of cells per individual was the primary determinant of power, typically increased first as budget rises. For DE scenarios with strong effects, read depth tended to be increased before sample size; for eQTL, increasing sample size was more beneficial than deeper sequencing (read depth often ~10,000). - Simulations versus analytics: scPower’s analytic estimates agreed well with powsimR and muscat across expressed genes, DE power, and overall power; differences reflected method-specific assumptions and DE tools. Batch effects reduced power if unmodeled but were recoverable when included as covariates. Unbalanced cell-type proportions reduced power; conservative estimates can be obtained by using the lower group’s frequency. - Efficiency: For analyses analogous to Fig. 3, powsimR required ~8 days and ~48 GB RAM; muscat ~3 days and ~35 GB; scPower <1 minute and a few MB of RAM. - Generalization: The expression probability model achieved r² ≈ 0.995 (Drop-seq lung) and r² ≈ 0.991 (Smart-seq2 pancreas). Budget-optimized designs reflected platform cost structures; Smart-seq2 required higher read depths and yielded lower power under equal budgets due to higher per-cell cost. - Practical guidance: Overloading 10x lanes (with robust doublet detection) increases usable cell numbers and overall power despite increased doublets and slight depth reduction per singlet. Recommended designs generally favor many cells at modest depth (~10k reads/cell).

scPower addresses the central design question for multi-sample scRNA-seq: how to allocate a fixed budget among samples, cells per sample, and read depth to maximize detectable DEGs/eQTLs at cell-type resolution. By explicitly modeling cell-type–specific expression detectability and combining it with analytic test power under realistic effect-size and expression priors, the framework converts prior knowledge and user constraints into quantitative design recommendations. Results show that increasing the number of measured cells per individual substantially improves overall detection power by boosting the probability that genes are expressed above thresholds in enough individuals; sample size has a strong additional effect on eQTL power. Consequently, shallow sequencing of many cells is generally more efficient than deep sequencing of fewer cells, with optimal read depths often around 10,000 reads per cell. The analytic approach enables rapid exploration of broad parameter spaces with minimal computational resources, making it feasible to tailor designs to specific biological contexts and platforms. Validation against independent datasets and simulation-based tools supports the fidelity of scPower’s predictions, and cost optimization analyses provide actionable, budget-aware settings for typical DE and eQTL studies. The framework’s generalization to 10x Genomics, Drop-seq, and Smart-seq2 and its integration with standard multiple-testing controls make it broadly applicable across tissues and technologies.

The paper introduces scPower, an analytic framework and R package/web tool for designing and powering multi-sample scRNA-seq DE and eQTL studies at cell-type resolution. By modeling expression detectability with cell-type–specific priors and combining it with DE/eQTL power under multiple-testing control, scPower enables rapid optimization of sample size, cells per sample, and read depth under budget constraints. Empirical validations and benchmarking demonstrate accurate predictions and orders-of-magnitude runtime and memory advantages over simulation-based approaches. Broad recommendations emerge: prioritize measuring many cells per individual at modest read depth (~10k reads/cell), and for eQTLs, allocate budget to increase sample size. Future work can expand priors using growing single-cell atlases, extend to other applications (e.g., co-expression, variance QTLs), and incorporate continuous cell-state annotations or additional experimental complexities. The tool provides a foundation for rational, data-driven experimental design across organ systems and platforms.

- Dependence on priors: Optimal designs and power estimates depend on effect-size and expression priors; mismatched or small pilot datasets bias power, particularly at larger sample sizes where smaller effects become detectable but are absent from priors. - Threshold choices: Expression detection thresholds (counts and fraction of individuals) influence power and optimal parameters; more lenient thresholds can inflate false positives. FDR optimization of thresholds can increase discoveries but may raise false-positive rates. - Doublet modeling: Default 10x-based doublet rates are lower-bound estimates; real experiments may have higher doublet rates, reducing usable cells and read depth per singlet. Benefits of overloading assume high doublet-detection accuracy. - Model scope: The framework assumes discrete cell-type annotations and pseudobulk aggregation; continuous trajectories (e.g., pseudotime) require discretization. Co-expression or other endpoints may need different optimal designs not addressed here. - eQTL analytics at low expression: Analytic power is adjusted by simulations for low-mean genes; accuracy depends on simulation calibration and assumptions (e.g., transformations, residual normality). - Multiple testing control: eQTL FWER assumes a default number of independent SNPs per gene (e.g., 10); deviations affect thresholds. DE FDR control uses Jung’s method since p-values are not explicitly computed. - Platform-specific costs and efficiencies: Smart-seq2 incurs higher per-cell costs, limiting power under fixed budgets; estimates rely on cost parameters that may vary across labs. - Unmodeled factors: While batch effects can be accommodated analytically by covariates in downstream models, uncorrected batch or group-wise differences in cell-type proportions reduce power; guidance is provided for conservative estimates but not fully modeled within the analytic core. Variance QTL power is not addressed due to lack of priors.