The problem with this machine-learned “predictive biology” framework is that it doesn’t have any prescription for what to do when your predictions fail. Just collect more data! What kind of data? As the author notes, the configuration space of biology is effectively infinite so it matters a great deal what you measure and how you measure it. If you don’t think about this (or your model can’t help you think about it) you’re unlikely to observe the conditions where your predictions are incorrect. That’s why other modeling approaches care about tedious things like physics and causality. They let you constrain the model to conditions you’ve observed and hypothesize what missing, unobserved factors might be influencing your system.

It’s also a bit arrogant in presuming that no other approaches to modeling cells cared about “prediction”. Of course, systems and mathematical biologists care about making accurate predictions, they just also care about other things like understanding molecular interactions *because that lets you make better predictions*

Not to be cynical but this seems like an attempt to export benchmark culture from ML into bio. I think that blindly maximizing test set accuracy is likely to lead down a lot dead end paths. I say this as someone actively doing ML for bio research.