In this case I’m not sure if the reasoning for what molecules work would make sense even if we had an “oracle” to explain it. Why a molecule works is a complicated interplay of fundamental physics and emergent properties. The explanation is likely not human interpretable without a ton of equations in the first place.

That said in broader medical cases of ML where particular symptoms and measurements drive a diagnosis for example, I tend to agree.

I’m not sure about the exact mechanisms of drug efficiency. But there are well-known patterns in molecular biology, auch as specific sequences in proteins reliably forming specific 3D structures.

If these models produce novel but (somewhat) effective structures, it must be because they pick up on less obvious patterns in the data. To be able to describe these would seem to super effective.