I just finished reading “Science in the Age of Computer Simulation” by Eric Winsberg and want to write about a specific claim of the book. In the conclusions, Winsberg writes: “[…] simulations sometimes have great epistemic power – there are indeed some questions we can ask for which simulations provide much more reliable answers than any experiment can provide.”

Are there biological questions where simulations are more reliable than experiments? Most biologists’ intuition would probably be no. Simulations are seen as tools that can complement experiments, but they enjoy little trust on their own. Given a disagreement between an experiment and a simulation, we would certainly err on the side of the experiment. But where exactly does that intuition come from and are there cases where we could justifiably trust the simulation over the experiment?

I think the intuition comes from the fact that all models are wrong. If the model behind a simulator is by default wrong, any of its conclusions could be just as wrong. Another argument is that biological systems are too complex to model. This makes superficial sense. If the systems were simpler, it would be easier to build a less-wrong simulator. But in the contest between experiments and simulations I will turn the argument around and claim that biological systems are too complex to experiment on. The best example of that are compensatory mechanisms.

Imagine we wanted to know if variable A had a causal effect on C. In a perfect experiment, we randomly assign values to A and measure C. If we find no change of C at all or any variability in C is completely random, we conclude there is no causal influence from A to C. But this conclusion could be incorrect. Another variable, B, could be monitoring A and holding C constant by compensating for the change in A. This could be resolved with a simulator, where B is held constant. In that case, the simulated result would give the correct answer. The experimental result would provide the wrong answer.

A biological counterpoint might be that causal conclusions should respect the compensatory mechanism. However, discovering causal mechanisms that are obscured by compensation is important. Maybe there are ways to disable the compensatory monitoring and use the causal relationship therapeutically.

In summary, there are cases where simulations are more reliable than experiments. That alone is interesting and I would have probably disagreed with this not too long ago. However, trust in experiments and simulations is built on a case-by-case basis. And in most cases we will be correct to trust experiments.