Public data deserves primary-evidence discipline
Secondary analysis is not second-class science, but it must respect cohort design, provenance, and consent boundaries.
Our position
Public genomic, single-cell, and phenotype resources make questions testable that no single laboratory could address alone. Their scale creates scientific leverage. It does not remove the original study design, participant context, measurement error, access conditions, or consent boundary.
Secondary analysis should therefore be conducted with primary-evidence discipline. The source cohort, assay, tissue acquisition, processing, reference versions, exclusion rules, and permitted uses remain part of every derived result. Public availability is a distribution condition, not a quality certificate or blanket ethical permission.
The source is part of the result
A cell atlas is not simply a matrix. It is a collection of donors, tissues, protocols, capture efficiencies, quality thresholds, and author-defined labels. A variant database is not simply an allele count. It has ascertainment, ancestry composition, coverage, calling, filtering, and annotation assumptions. A phenotype file is not simply a graph. It reflects curation sources, evidence codes, frequency, and release history.
Pooling sources without retaining these structures creates synthetic certainty. The combined sample size increases while the ability to explain what was measured decreases. We prefer federated summaries or study-aware models when direct pooling would erase the unit of evidence.
Provenance before harmonization
Every source should enter with an accession, release, download date, license or access terms, file identity, cohort description, sample exclusions, and source citation. Every transformation should record identifier mapping, normalization, filtering, label changes, reference builds, and software environment. Derived objects should be traceable to inputs and code state through immutable identifiers or checksums.
Harmonization must preserve disagreement. If two sources assign different cell labels or variant consequences, the mapping should carry confidence and alternatives. A unified table that discards uncertainty is easier to analyze and harder to trust.
Cohort and donor structure cannot be repaired by cell count
Single-cell resources can contain millions of cells from a modest number of donors. The cells increase resolution; they do not create millions of independent biological replicates. Analyses that ignore donor structure underestimate variance and can produce false confidence. Similar issues arise when multiple tissues, timepoints, or family members are treated as independent without acknowledging hierarchy.
Metadata limitations should change the claim. If critical covariates are absent, an analysis may support transportability or hypothesis generation but not a causal contrast. Calling the same model “adjusted” because a batch variable was available does not solve unmeasured confounding.
Open access does not erase participant context
Data can be legally accessible and still require ethical restraint. Researchers should use the minimum resolution necessary, avoid re-identification attempts, honor controlled-access and downstream-use limitations, and consider whether group-level interpretations could stigmatize communities. Rare disease increases this responsibility because combinations of phenotypic and genomic features can be identifying.
Communication matters. Aggregate public evidence should not be narrated as though the secondary analyst recruited the cohort, generated the assay, or established a clinical finding. Source investigators and participants are part of the evidence history, and attribution should reflect that.
Quality control must follow the question
A generic quality-control pipeline is not enough. The relevant checks depend on the claim: coverage and transcript consequence for rare variants; donor composition and pseudoreplication for single-cell contrasts; ontology release and term influence for phenotype similarity; tissue handling and postmortem variables for state programs.
Negative controls should also be source aware. A control signature derived from fresh blood may not diagnose dissociation effects in brain tissue. A frequency threshold appropriate for one ancestry or inheritance model may be misleading in another. Primary-evidence discipline means matching the control to the generating process.
Release what another group needs to disagree
A reusable release includes source manifests, environment locks, transformation records, analysis code, sensitivity summaries, and interpretive limits. When data access prevents redistribution, the release should still provide enough identifiers and checks for an authorized group to reconstruct the work.
The purpose is not ceremonial openness. It is adversarial usability: another group should be able to rerun the analysis, substitute a source, challenge a mapping, and understand why the conclusion changes.
Positions we apply in review
- Public availability does not remove study design or consent context.
- Donor and cohort structure must remain visible after integration.
- Harmonization should preserve disagreement and mapping confidence.
- Minimum-necessary use and non-reidentification are baseline obligations.
- A public-data release should enable reconstruction and informed disagreement.
Our conclusion
Our editorial conclusion is that public-data science should be held to the same inferential and ethical standards as newly generated evidence. REELD will treat provenance as part of the result, use the biological replicate as the unit of inference, preserve source boundaries, and narrow claims when metadata cannot support more.