Biological variability is not noise — it is unresolved structure
We often assume that signal and noise are properties of the data. They are not. They are properties of how we measure.
In biological systems, variability is expected. Signals fluctuate, patterns are inconsistent, measurements drift. We label much of this as noise — something to be filtered, averaged, or removed. But what we call noise is often not random. It is unresolved structure.
A sensor does not passively capture reality. It defines what can be seen. A dissolved oxygen probe in a fermenter reads bulk concentration — it does not resolve the localized oxygen gradients that drive metabolic heterogeneity within a dense cell population. An OD measurement captures average turbidity — it does not distinguish between uniform growth and two subpopulations diverging under different metabolic strategies. The sensor is not failing. It is resolving what it was designed to resolve.
Filtering compounds this. To make data usable, we smooth fluctuations, remove outliers, average across time or populations. In a chemical process with stable kinetics, these assumptions hold. In a co-culture where competitive dynamics produce transient electrochemical shifts that precede any measurable change in growth rate, those same assumptions remove the most informative signal the system produces.
Fermentation monitoring makes this visible. Process control relies on stable, aggregated signals — pH, dissolved oxygen, optical density, metabolite concentration. These are useful. They track trajectory and catch large-scale deviations. But they represent bulk behavior, not system dynamics. In a mixed-culture fermentation for organic acid production, the point where one population begins to outcompete another does not appear in the pH trend. It appears as subtle, irregular variation in the electrochemical baseline — variation that standard signal processing removes. By the time the shift shows up in monitored outputs, the competitive dynamics that caused it are already resolved.
The same pattern holds in upstream development. In cell line screening for biologics, clones are selected on titer, growth rate, and product quality — all endpoint measurements taken after the biology has settled. But clone instability and sensitivity to process perturbation are dynamic properties. They manifest during transitions — passage to passage, scale-up, media adaptation — as variability. That variability is averaged across replicates or dismissed as assay noise. The clone that looked stable in characterization becomes unstable in manufacturing. The information was there. It was classified as noise.
This is not a processing failure. It is a sensing framework built for stable systems applied to systems where variability is the signal. Electrochemical impedance spectroscopy captures what the system does when forced — not what it does on its own. Cyclic voltammetry sweeps a voltage range and reads current response, but the measurement itself drives reactions that would not have occurred without the applied potential. In a living system continuously producing its own electrochemical signal, an applied excitation does not add information. It overwrites it.
If the sensor defines the signal, then every downstream step — processing, modeling, interpretation — inherits that definition. No amount of processing recovers what the sensor never captured. No model reconstructs dynamics filtered at the source. The limitation is upstream.
Signal and noise are not fixed categories. They are outcomes of how we observe. Until the measurement captures what the system is doing rather than what it does when forced, we will continue to clean the data and remove the system in the process.
— Pegah Farr