This is why I think the quote "All models are wrong, but some are useful" is so meaningful. Lots of bright eyed comp sci majors come into statstical genomics and bioinfomatics PhD programs every year, and all learn the hard lesson that wet lab still exists for a reason.

Not understanding your underlying conceptual model (be it biology here, or maybe healthcare, or natural resource exploration) will lead a lot of poor assumptions, limiting the impact of all the hard work.

I would argue we can't even model those things. Non-linear, dynamic systems are not predictable beyond a certain point even if you have a perfect model. Uncertainty with models of dynamic systems like biological ones, increases exponentially as a function of time. And they are highly sensitive to initial conditions, so if your input is off from reality by even the tiniest amount, reality will veer off dramatically from your expected results, even if you have the perfect model. Chaos theory is a thing. We can use models to learn about systems and how they might work, or how we might think they work, but we can't use them to model reality effectively enough to determine treatments.

While individual trajectories of non-linear systems are not predictable, they all end up on an attractor and usually all we are interested in, is the attractor.

That is when performing a test on a mouse we are not measuring what a particular mouse would do, but what the mice do on average. So as long as model behaves like a mouse it can be useful, even though it would not behave like any particular mouse.

Also it's important to remember that the question is not about all or nothing, when we get a model most likely instead of 100 experiments on real mice we will be doing millions on virtual mice, and then 10 on real mice to verify the results.

> Programs and algorithms can only encode models (which are themselves imperfect) of what we already know.

The Game Of Life seems to contradict this claim. Also worth considering, computational physics and weather/climate simulations.

Scientists of all sorts already do learn new insights by creating simulated models of systems. The simulated systems will continue to get more accurate.

> As bad as animal (and human) studies and trials can be, the beauty of them is we don't really have to know how or why things work, we can just do the thing and if the results are good, we have a new treatment.

If you didn't have at least a broad idea of why something might work, you would most likely never have "done the thing". You need to narrow the search space. Simulations can help here.

ML models sometimes cannot be explained by humans and contains information we do not know. But they can be contained in a program or algo.