Publish the negative space around a result
Failed replications, null tissues, and unstable rankings define where a mechanism does not travel.
Our position
Positive enrichment is easier to narrate than the conditions under which it disappears. Yet those conditions often determine whether another group can use the result. A mechanism without its null tissues, failed transfers, unstable parameters, and incompatible phenotypes is a map without boundaries.
REELD treats negative space as part of the primary result. We mean more than publishing non-significant p-values. We mean documenting where the signal was measured adequately and not observed, where the analysis lacked power, where a control explained the effect, and where an independent reference changed the interpretation.
Not all nulls mean the same thing
A null result can represent true absence, inadequate power, low feature detection, incompatible measurement, wrong timing, incorrect cell context, poor annotation, or a model that absorbed the signal. These explanations have different scientific implications. “Not significant” is not a sufficient description.
An interpretable null requires an estimand, expected direction, measurement sensitivity, uncertainty interval, quality controls, and comparison with a positive or technically informative reference. If the assay could not have detected a biologically relevant effect, the result defines a measurement limit rather than a mechanism boundary.
Failed transfer is a localization tool
When a signal appears in one atlas and not another, the first question should not be which atlas is correct. Differences in tissue, developmental stage, donor composition, platform, disease state, and cell taxonomy may localize the effect. The failed transfer can identify the condition required for the mechanism.
This requires source-aware comparison. If two references do not support the same biological contrast, non-replication should not be framed as contradiction. The transfer test must distinguish biological restriction from incompatible design.
Parameter instability is evidence about dependence
A ranking that changes under a reasonable ontology depth, background universe, normalization, or label resolution is telling us what the result depends on. The correct response is to identify the responsible candidate, term, gene set, or source—not to average across settings until the instability disappears.
We favor stability maps that show which conclusions persist, which reverse, and which are unresolved. This preserves local structure. A global similarity statistic can look reassuring while the top experimental priority changes completely.
Controls can convert a positive into a narrower positive
A negative control that explains part of a signal does not necessarily invalidate the study. It can refine the claim. A disease module that overlaps generic interferon response may retain a smaller disease-specific residual. A cell-state enrichment explained by expression level may still contain a context-specific neighborhood after matching.
Publishing only the raw positive or only the post-control residual hides the scientific transition. Both belong in the record, together with the reason the claim changed.
Negative space changes resource allocation
Experimental follow-up is expensive. If a computational mechanism fails in adult tissue, requires a developmental window, or disappears under donor-aware inference, those facts should influence whether and how a model is built. Hidden negative space causes other groups to repeat avoidable failures or select the wrong context.
The value is cumulative. Well-characterized failures can support meta-analysis, benchmark development, power calculations, and better control libraries. Unreported failure preserves a cleaner story for one paper and a dirtier evidence base for the field.
A release format for boundaries
We recommend a boundary table for every mechanism claim: tested source or condition, expected signal, observed direction and uncertainty, measurement adequacy, interpretation, and consequence for scope. Sensitivity outputs should be machine readable where possible so another group can compare boundaries across studies.
Licensing and governance can limit what is released. They do not prevent a clear description of which tests were performed and what they imply.
Positions we apply in review
- A null result requires an interpretable measurement context.
- Failed transfer can localize a mechanism rather than simply refute it.
- Parameter instability should be traced, not averaged away.
- Controls often narrow a claim and that change should be visible.
- Boundary tables belong beside primary findings.
Our conclusion
Our editorial conclusion is that the negative space around a result is part of its meaning. REELD will publish where a signal fails, why the test was informative or limited, and how the failure changes the claim. A bounded mechanism is more transferable than an apparently universal mechanism whose boundaries were never tested.