Under which biological conditions should intervening on a target help a patient?
In her recent essay, Drug Discovery Has No Magic Wands, Daphne Koller argues that better molecular design cannot substitute for disease understanding.
From a systems biology perspective, a therapeutic mechanism needs its conditions spelled out: cell state, tissue environment, disease stage, and patient population. Calling a target “correct” leaves those conditions unresolved.
This suggests a concrete test for AI models: can they predict intervention responses in biological contexts excluded from training, and identify where those predictions become unreliable?
Passing that test would support a defined scope of predictive validity. Clinical benefit and causal mechanism would still require separate evidence.
AI could also help distinguish competing explanations and select experiments that could falsify them. That contribution deserves explicit evaluation alongside predictive accuracy, especially when deciding which therapeutic hypothesis to advance.
Original essay by Daphne Koller:
https://www.a16z.news/p/drug-discovery-has-no-magic-wands
What evidence would convince you that a model has learned a transferable intervention response?
A therapeutic mechanism is meaningless without its biological boundary conditions—cell state, tissue niche, and disease stage. Computational models must be stress-tested on their ability to predict intervention outcomes in unseen biological contexts and identify where predictions become unreliable.