From Correlation to Mechanism: What Strong Experimental Design Actually Requires

One of the most important lessons from my own research is that a result becomes compelling not when it is merely statistically significant, but when the experimental logic makes competing explanations progressively harder to sustain.

This distinction matters across molecular biology, cell biology, biotechnology and plant science. A change in gene expression may accompany a phenotype without causing it. A protein may relocate after treatment without mediating the response. A candidate identified by an association study may be biologically relevant, or it may be travelling with the real causal variant. The data can be reproducible and still support a more limited conclusion than the manuscript claims.

Strong research strategy and experimental design begin by deciding what the study must distinguish. Methods, controls and sample size matter, but they become meaningful only after the inferential problem has been defined.

A correlation can be valuable without being a mechanism

My research on plant adaptation to altitude began with a simple environmental pattern: the partial pressure of oxygen declines as altitude increases. That observation suggested that oxygen could provide plants with information about absolute elevation.

But altitude is not a one-variable experiment. Temperature, radiation, humidity, growing season and other ecological factors also change across mountain systems. Finding that a molecular or physiological trait varies with elevation would therefore be interesting, but it would not establish that oxygen sensing caused the pattern.

In our first-author Nature study, An oxygen-sensing mechanism for angiosperm adaptation to altitude, we had to connect several levels of evidence: controlled changes in atmospheric oxygen, natural variation across altitudinal gradients, activity of the PCO N-degron oxygen-sensing pathway, genetic perturbation of ERFVII transcription factors, chlorophyll-precursor regulation and patterns across diverse angiosperm species.

No single result carried the full conclusion. The mechanism emerged from the way the experiments constrained one another.

1. Define the competing explanations before selecting the methods

A study is easier to design when the principal alternatives are stated explicitly. If treatment X changes phenotype Y, at least several explanations may remain:

• X acts through the proposed molecular pathway.
• X acts through a different pathway that produces the same readout.
• The effect reflects toxicity, altered growth or another non-specific consequence.
• The apparent difference comes from batch, developmental stage or sampling structure.
• Y changes with X but is not functionally involved in the biological response.

The purpose of experimental design is not simply to produce a positive result. It is to create comparisons that discriminate among explanations.

This changes how controls are chosen. A vehicle control may establish that the delivery method did not cause the effect, but it cannot show that the proposed pathway did. A wild-type comparison may show that a mutant differs, but it may not separate the target gene from background variation or developmental effects. The correct control is determined by the claim, not by convention.

2. Build a chain of evidence rather than searching for one decisive experiment

Mechanistic conclusions usually require several forms of evidence with different weaknesses. A useful sequence is:

1. Observation: establish that the phenomenon exists and is reproducible.
2. Association: show that the candidate pathway changes with the phenomenon.
3. Perturbation: alter the candidate pathway and test whether the response changes.
4. Specificity: distinguish the proposed mechanism from plausible alternatives and non-specific effects.
5. Restoration or convergence: use rescue, complementation or an independent perturbation where feasible.
6. Context: determine whether the relationship persists across relevant genotypes, cell types, environments or models.

Not every project needs every element, and the order will vary. The principle is that independent approaches should converge on the same explanation. Two assays that share the same hidden assumption are not as independent as they appear.

3. Use perturbation carefully: necessity is not sufficiency

Loss-of-function evidence can show that a component is required under the conditions tested. It does not automatically show that the component is sufficient to produce the phenotype, that it acts directly, or that no parallel pathway contributes.

Likewise, overexpression can demonstrate that a system is capable of producing an effect, but expression outside the normal range, tissue or developmental window may create behaviour that is not physiologically relevant.

A stronger design asks what each perturbation establishes:

• Does loss of the candidate reduce or abolish the response?
• Does restoration recover it?
• Does an independent allele, reagent or perturbation give a compatible result?
• Can indirect effects on viability, development or global physiology explain the phenotype?
• Is the timing of the perturbation compatible with the proposed mechanism?

The language of the conclusion should follow these answers. Required, sufficient, direct, mediates and contributes to are not interchangeable descriptions.

4. Keep measurements and claims at the same biological level

Modern projects often combine transcriptomics, imaging, biochemical assays, genetic screens and organismal phenotypes. This breadth can be powerful, but it also creates opportunities to jump between levels of evidence.

For example, a transcript increase does not prove greater protein activity. Protein localisation does not establish target engagement. An in-vitro interaction does not show that the interaction determines a phenotype in vivo. A cellular response does not automatically predict an organismal, clinical or agricultural outcome.

The gap is not a writing problem. It is an experimental one. Before expanding the conclusion, identify the bridge that is missing between the measurement and the intended biological claim.

5. Treat context as part of the mechanism

Biological effects depend on genotype, cell state, developmental stage, environment, dose and time. A mechanism demonstrated in one context can be real without being universal.

Natural variation was important in our altitude work because adaptation is a population-level proposition. Controlled experiments could establish oxygen-dependent molecular responses, but contrasting accessions and altitudinal clines were needed to connect those responses to adaptation.

In other systems, the relevant test of context may involve:

• more than one cell line or genetic background;
• primary material alongside an established model;
• a developmental or disease-relevant state;
• a realistic dose and exposure duration;
• independent cohorts or experimental batches; or
• an environmental condition under which the pathway is expected to matter.

Generality should be earned by the design. It should not be added later through broader wording.

6. Decide in advance what a negative result would mean

An experiment is informative when its possible outcomes lead to different scientific decisions. If every result can be explained as consistent with the preferred hypothesis, the experiment has limited discriminatory value.

Before committing resources, ask:

• What result would weaken the proposed mechanism?
• Would a negative outcome be interpretable, or could assay sensitivity and statistical power explain it?
• Which alternative would become more plausible?
• What is the most informative next experiment after each possible outcome?

This is particularly important in large omics studies and complex screening programmes. Generating more data is not always the same as reducing uncertainty.

7. Let the design determine the claim—not the other way around

By the time a manuscript is written, researchers are deeply familiar with the project and understandably attached to its main explanation. This can make the evidence chain look more complete than it does to an independent reader.

A useful final exercise is to map every major claim to the experiment that supports it, then ask which alternative explanation that experiment excludes. If a claim has evidence but no discriminating comparison, it may need qualification or another experiment.

The same discipline should shape the manuscript. As discussed in What Scientific Editing Should—and Should Not—Change, stronger language cannot convert association into mechanism. The conclusion has to remain proportionate to the design.

A practical experimental-logic check

Before beginning a critical experiment—or before deciding that a study is complete—ask:

• What exact scientific decision must this experiment support?
• What are the strongest competing explanations?
• Which comparison distinguishes among them?
• What is the correct experimental unit, and what constitutes biological replication?
• Does the perturbation test necessity, sufficiency, directness or only association?
• Is there an independent method with different assumptions?
• Are the measurement and the conclusion at the same biological level?
• Which context limits the generality of the result?
• What negative result would change the interpretation?
• Do the verbs in the proposed conclusion match what the design can establish?

Good experimental design does not remove uncertainty from research. It makes uncertainty visible, places it in the right part of the argument and directs resources towards the evidence most likely to change the scientific decision.