Bio Experimental Design Power Analysis

Version Compatibility

Reference examples tested with: RNASeqPower 1.42+, PROPER 1.34+, powsimR 1.2+ (GitHub), DESeq2 1.42+, edgeR 4.0+, pwr 1.3+.

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws an error, introspect the installed package and adapt to the actual API. Notes: RNASeqPower::rnapower() solves for whichever of n or power is omitted; PROPER is a multi-step pipeline (RNAseq.SimOptions.2grp -> simRNAseq -> runSims -> comparePower); powsimR is GitHub-only and its estimateParam/Setup/simulateDE signatures drift — pin a commit SHA for reproducible work. Verify each against the installed help before relying on argument names.

Power Analysis for Genomics Experiments

"How many replicates does my sequencing experiment need?" -> Compute the probability of detecting a biologically meaningful effect given replicate number, sequencing depth, and biological variability — modeling counts as negative-binomial and recognizing that power is a per-gene quantity, not one number for the whole transcriptome.

• R: RNASeqPower::rnapower() — closed-form NB power/sample size; PROPER, powsimR — simulation from the mean-dispersion trend

The Single Most Important Modern Insight -- Genomics Power Is Per-Gene; Simulate, and Never Report Observed Power

Power in a sequencing experiment is not a single number. It is a per-gene quantity that depends on that gene's mean expression and dispersion, so the honest summary is the marginal (average) power across the expression distribution at a target FDR — the expected discovery rate. A single coefficient of variation plugged into a closed-form formula mis-states power for low- and high-expressed genes alike, because dispersion varies systematically with the mean; the defensible default for count data is simulation from the empirical mean-dispersion trend (PROPER, Wu 2015 Bioinformatics 31:233; powsimR, Vieth 2017 Bioinformatics 33:3486). The second rule is negative: observed (post-hoc) power is information-free. Computed from the effect a study actually estimated, it is a one-to-one function of the p-value and cannot explain a null result (Hoenig & Heisey 2001 Am Stat 55:19). Power is a design-stage quantity, computed for hypothesized effects before data exist. Underpowering does not merely miss true effects — it makes the significant ones overstate magnitude (Type-M) and sometimes reverse sign (Type-S), lowering the chance a significant call is real (Button 2013 Nat Rev Neurosci 14:365; Gelman & Carlin 2014 Perspect Psychol Sci 9:641).

Algorithmic Taxonomy

Decision Tree by Scenario

Closed-Form NB Power -- RNASeqPower

Goal: Get a fast, transparent power or replicate number for bulk RNA-seq from depth, biological CV, and fold change.

Approach: Supply per-gene depth, biological coefficient of variation, the fold change to detect, and alpha; supply n to get power, or power to get the required n. Treat the result as a single-gene approximation and sanity-check against simulation.

library(RNASeqPower)
# depth = reads/gene; cv = biological coefficient of variation; effect = fold change
rnapower(depth = 20, n = 5, cv = 0.4, effect = 2, alpha = 0.05)          # solves for POWER
rnapower(depth = 20, cv = 0.4, effect = 2, alpha = 0.05, power = 0.80)   # solves for n per group

Simulation-Based Power -- the Honest Default for Counts

Goal: Estimate marginal power and the true realized FDR across the whole expression distribution, accounting for the mean-dispersion trend.

Approach: Build (or fit from pilot) a simulation model of counts with a realistic dispersion-mean relationship and DE-effect distribution, simulate many datasets at each candidate sample size, run the intended DE test, and read the average power at the target FDR.

library(PROPER)
sim_opts <- RNAseq.SimOptions.2grp(ngenes = 20000, p.DE = 0.05,
                                   lOD = 'cheung', lBaselineExpr = 'cheung')  # empirical dispersion/expr priors
sims <- runSims(Nreps = c(3, 5, 8, 12), sim.opts = sim_opts, nsims = 50,
                DEmethod = 'edgeR')
powr <- comparePower(sims, alpha.type = 'fdr', alpha.nominal = 0.05,
                     stratify.by = 'expr', delta = log(1.5))          # delta is NATURAL-log lfc in PROPER; marginal power by expression stratum
summaryPower(powr)

Depth vs Replicates -- the Budget Question

For bulk RNA-seq differential expression, sequencing depth shows diminishing returns once it is adequate — Liu, Zhou & White 2014 (Bioinformatics 30:301) found the inflection near ~10 million mapped reads in MCF7 (commonly generalized to a 10-20M band) — whereas adding biological replicates improves power across the whole range. Under a fixed budget, allocate to more biological units before more depth. ATAC/ChIP have their own depth floors (library complexity, peak detection), but the principle holds: biological variance, not read count, limits discovery once depth is adequate.

CV / Dispersion Guidelines (estimate from pilot when possible)

These are starting points, not substitutes for a pilot estimate; real dispersion is study-specific and a literature CV can be off by a factor of two (estimate via DESeq2/edgeR estimateDispersions — see experimental-design/sample-size).

Per-Method Failure Modes

Single CV for the whole transcriptome

• Trigger: one cv plugged into rnapower() for all genes.
• Mechanism: dispersion varies with mean expression; a single CV mis-states low/high-expressed genes.
• Symptom: simulation gives materially different power than the closed form.
• Fix: simulation-based power (PROPER/powsimR) from the mean-dispersion trend.

Observed (post-hoc) power

• Trigger: "non-significant, but observed power was 0.3, so add samples."
• Mechanism: observed power is a monotone function of the p-value (Hoenig-Heisey 2001).
• Symptom: circular reasoning that adds nothing to the CI.
• Fix: report effect size + CI; do prospective power for the next study.