Divergent Results Are Often Signal — Not Failure
The reproducibility problem in biology is treated as a quality issue — something to be solved with stricter protocols, tighter controls, and better standardization. But reproducibility failures are not noise. They are data. The difference between what one lab observes and what another observes, running the same biology under nominally identical conditions, is itself a biological signal — one that tells us more about the system than either result alone.
A genetic circuit characterized in one chassis strain and failing in another reveals context-dependent expression that the original characterization was too narrow to capture. A microbiome composition study diverging across sequencing facilities exposes how extraction protocols interact with community structure in ways no single protocol controls for. A strain performing inconsistently across sites is not unstable — it is sensitive to variables that no one is measuring, and each site’s divergence is pointing at what those variables might be.
This is the information the field discards. When results do not replicate, we investigate protocol deviations, tighten controls, or deprioritize the candidate. The strain gets shelved. The process gets redesigned around narrower conditions. Research directions close — not because the biology was wrong, but because the data was too thin to explain what happened. Work that could have revealed something fundamental about how the system behaves under different conditions becomes a failed experiment instead of a dataset.
The reason this information is lost is structural. Every lab captures a different slice of the system — different instruments, different sampling intervals, different filtering pipelines — and stores data in formats that serve local analysis but are incomparable to data from anywhere else. The observation is fragmented. Not because labs are careless, but because no infrastructure exists to make cross-lab observation possible at the level where it matters: before filtering, before aggregation, before the biology is compressed into a reported metric.
What the field needs is not standardized results. It is connected observation. Distributed sensing that captures raw biological signal across sites. A shared data layer where different instruments, different processes, and different experimental contexts become structurally comparable. Computational methods that treat cross-lab variability as information to be resolved — not error to be eliminated.
When that infrastructure exists, the question changes. A strain that behaves differently across sites is no longer a reproducibility failure — it is a map of what conditions actually matter. A process that diverges between facilities is no longer inconsistent — it is exposing the variables the monitoring framework was never designed to capture. Irreproducibility stops being a dead end and becomes the most information-dense dataset in biology.
The problem was never that biology is irreproducible. It is that we built no infrastructure to learn from the differences.
— Pegah Farr