Before you start, establish that the studies are independent and address the same directionally defined comparison in one organism and compatible gene identifiers. Re-deposited samples are not an independent replication.
1. Add study membership#
On Samples, use Add/review study column and assign each sample its dataset study of origin. Choose the comparison factor and groups in Analysis settings. The live eligibility text previews current editor values, including counts and exclusions; authoritative run validation checks saved inputs. Study labels do not prove independent recruitment or compatible biology.
2. Set the meta-analysis controls#
| Setting | What it changes | When to change it / example |
|---|---|---|
| Multi-study meta-analysis: off | Adds separate study fits and combination alongside the joint analysis. | Enable for independent studies of the same contrast; it is not a way to turn technical repeats into replication. |
| dataset | Separates the per-study fits; inspect the recorded joint formula to determine whether study adjustment was fitted. | Illustration: study_A and study_B each include treated and control samples. |
| Per-study enrichment: off | Adds enrichment for each constituent study. | Enable if those separate functional comparisons are required; it adds computation and service use. |
| Meta GO ontology: BP | Selects biological process, molecular function (MF) or cellular component (CC). | Choose the annotation aspect matching your question, not the one yielding the most terms. |
| Meta figure labels 10 / heatmap genes 50 / enrichment terms 6 | Limits displayed content in comparative figures. | Changing display counts does not change which genes pass the meta-analysis criteria. |
3. Understand what is combined#
In the current workflow, each eligible study is fitted separately with DESeq2 using the selected contrast factor only. The requested meta design may also name additive dataset, which is constant within each study and is omitted from those separate fits. The meta input gate refuses extra covariates, interactions, transformations and no-intercept formulas before a per-study fit; a requested covariate cannot be silently dropped. The joint fit uses the recorded engine and actual design and must be interpreted separately. P-values are combined with a replicate-weighted inverse-normal method; unshrunken effects are also combined with an effect-size model. The effect model uses a common-effect fixed-effect fit for two studies and DerSimonian–Laird random effects for three or more. The workflow deliberately omits heterogeneity fields for two-study results; this does not mean heterogeneity is mathematically inestimable.
A combined-FDR hit requires matching nonzero signs across retained studies, but does not require individual per-study significance or guarantee independent replication. The combined-p BH family contains matching-sign genes; the pooled-effect BH family contains every estimable pooled test, including opposite-sign and neutral rows. The latter adjusted value is rem_padj and does not decide the combined-p call. The ledger results/meta/meta_eligibility.json records both family sizes, post-filter losses and direction categories. Missing-statistic reasons can overlap. An old result without the corrected method marker needs recomputation before pooled-effect adjustment is interpreted. New runs record the executed factor, numerator, denominator, per-study formula and combined-p result-call alpha in the eligibility ledger. A figure-setting alpha can change a displayed overlap classification without changing stored result calls; the report distinguishes them. An older result without execution fields reports its comparison orientation or threshold as not recorded even when its pooled-effect marker is valid.
Cross-study enrichment applies the configured per-study FDR and absolute log2-fold-change thresholds before restricting each up/down selection to the shared tested universe. The FDR boundary is strict, the effect boundary is inclusive, and an exact zero effect remains neutral when the effect threshold is zero. It then uses the main enrichment route's identifier policy for official identifiers, the configured keytype and aliases. Exactly one accepted Entrez mapping is used for both the shared universe and every study or convergent foreground. Unresolved one-to-many mappings are excluded rather than selected by database row order. Inspect results/meta/meta_enrichment_mapping.tsv for accepted, unmapped and ambiguous identifiers and check 18 for universe and foreground coverage. An ambiguity finding remains review-required when enrichment later skips for a small set or no terms.
4. Review the separate report#
Inspect meta_analysis_report.html, the complete linked tables under results/meta/, constituent-study results and meta-specific checks. Review excluded studies, opposite or neutral signs, duplicate-study warnings and how many studies reach per-study FDR for each reported gene.
Advanced: design limits
If the question needs within-study covariate adjustment beyond the supported meta formula, the meta switch does not implement that model. Keep the needed covariate and use an appropriate analysis rather than removing it to enable a run. Review historical meta-analysis notices before reusing older outputs.