FDA: AI drugs show 80–90% Phase 1

Drug discovery is the business of finding molecules that do two things at once — look promising on a computer and survive contact with an actual human body. That second part is where a lot of candidates die. The new wrinkle is that AI-designed drugs seem to be clearing the first human safety hurdle much more often than older industry averages. And the FDA has moved from watching that trend to building a framework around it. (sciencedirect.com) ### What’s the actual claim here? The number people keep citing comes from a 2024 Drug Discovery Today review of AI-discovered molecules already tested in humans. That paper said the small set of AI-derived candidates it tracked showed an 80–90% success rate in Phase 1, which is the first stage mainly focused on safety, dosing, and whether the drug behaves in the body the way resea(sciencedirect.com)in the broad neighborhood of 40–65%. (sciencedirect.com) ### Why is Phase 1 such a big deal? Because Phase 1 is where a candidate stops being a theory. A molecule can look elegant in silico, bind beautifully in a dish, and still fail once it hits metabolism, off-target effects, or plain old human biology. So if AI systems are really improving Phase 1 outcomes, that suggests they may be getting better at choosing compounds with cleaner sa(sciencedirect.com)AI may be helping companies make fewer obviously bad bets. (sciencedirect.com) ### Does that mean AI is solving drug development? Not quite. The catch is that Phase 1 is the easiest clinical bar to clear. It mostly asks, “Is this safe enough to keep going?” not “Does this meaningfully help patients?” Even the optimistic review noted that the limited Phase 2 data looked much more ordinary — around 40%, which is much closer to normal biotech attrition. So the headline is real, but it is also early and narrow. (sciencedirect.com) ### How good is the historical comparison? Useful, but messy. Older benchmark studies put overall Phase 1 transition rates around the low-50% range across large industry datasets, and success rates vary a lot by disease, modality, and era. A 2016 BIO/Biomedtracker analysis became one of the standard references for that baseline, while newer academic work argues clinical success rat(sciencedirect.com)ough shorthand, not a settled law of nature. (go.bio.org) ### Why is the FDA involved now? Because AI is no longer just helping write code or sort papers — it is showing up inside actual drug submissions. FDA says CDER saw more than 500 submissions with AI components from 2016 through 2023, spanning nonclinical work, clinical development, postmarketing, and manufactur(go.bio.org)tors start formalizing expectations, the technology has crossed from hype cycle into operating reality. (fda.gov) ### What is the FDA worried about? Mostly the boring but crucial stuff — data quality, model credibility, bias, drift, and whether a sponsor can show that an AI-generated output is reliable for the specific decision it supports. In other words, the FDA does not care that a model sounds smart. It cares whether the evidence chain is auditable and stable enough to trust when safety, efficacy, or manufacturing quality is on the line. (federalregister.gov) ### So what changed? The story is no longer “AI might help discover drugs someday.” The story is that early clinical signals are strong enough to get repeated, investors have latched onto them, and the FDA has started writing the rulebook around real submissions. But the finish line has not moved yet — no flashy Phase 1 statistic can substitute for late-stage proof that patients actually get better. (sciencedirect.com) ### Bottom line? AI looks increasingly good at finding molecules that deserve a first shot in humans. The harder part — proving those molecules become real medicines — is still where the industry will be judged. (sciencedirect.com)