There is something missing in the conversation about AI drugs
You can always count on Derek Lowe for an incisive (meaning, with tooth) analysis of a half-baked paper with a promising title (link). While Derek focuses on the paper's poorly defined categories (can a target really be "discovered by AI" if it was known before?), there is one blaring omission from the paper: the possible "doubling" of pharma R&D productivity based on success rates alone does not take little things into account such as cost and cycle time. Even with a similar success rate in Phase I and Phase II, if AI can help cut discovery and development time and cost (cost to success as well cost to failure), that can make a big difference. I am not sure whether the data is available, but that would be useful to know.
Of course, we still need to see how "AI drugs" advancing into Phase III fare.
For those interested, the reference article on pharma R&D productivity (and actually the first reference of the paper) is the illuminating 2010 piece by Steven Paul, Stacy Lindborg, my partner in mischievousness Aaron Schacht and their colleagues (link): How to improve R&D productivity: the pharmaceutical industry's grand challenge.