We optimize results — but the biology that produced them is not in the data
We design biological systems around outputs. Yield, growth rate, final concentration — these are what we optimize, and what we understand least.
Outputs are resolved states. They reflect what the system has already completed. A measured titer at the end of a fermentation run does not describe how the culture behaved — whether it shifted metabolic strategy midway, whether a subpopulation dominated early and collapsed late, whether a stress response temporarily redirected carbon flux. The titer describes where the system ended. The path that produced it is not in the measurement.
This is the optimization trap. When outputs become targets, systems are shaped to reach them. Parameters are tuned, conditions are stabilized around the desired endpoint. In industrial enzyme production, if two batches produce the same activity per liter — one through steady expression, another through early overexpression followed by partial metabolic collapse and recovery — the output does not distinguish them. Two systems can produce the same output through entirely different internal behavior. Optimization hides this difference.
That difference surfaces when the process moves. A strain optimized for yield in a controlled 5-liter fermenter encounters different oxygen gradients, mixing dynamics, and nutrient distributions at 5,000 liters. If the internal behavior that produced the yield was never observed, there is no basis for predicting what will hold and what will break. The process worked. Why it worked is not in the data.
The same blind spot appears in cell line development. Clones are ranked by titer and growth rate — endpoint metrics taken after the biology has settled. But two clones with identical titer can differ in metabolic stability, sensitivity to media variation, and behavior under scale-up stress. One holds. One does not. The ranking did not capture the difference because outputs compress behavior into a single number.
In synthetic biology, a genetic circuit is characterized by its transfer function — input mapped to output. The characterization confirms the circuit works. It does not confirm how it works inside a host under production conditions, where metabolic burden and growth-rate competition reshape behavior in ways the transfer function never captured. The output matched the design. The system that produced it did not.
Each of these is the same structural problem. Outputs are discrete. Behavior is continuous. What we measure is not where the system lives.
If behavior becomes the focus, design shifts — from capturing stable signals to resolving dynamic ones, from aggregating outputs to observing interactions, from enforcing consistency to preserving the structure that consistency obscures. Understanding does not come from better outputs. It comes from observing how those outputs are formed — not at the end of the process, but as the system evolves.
Behavior is not a secondary layer. It is the system itself.
— Pegah Farr