Two new research papers this week asked the same question: can AI actually design drug molecules that work? Their answers are opposite.
The first paper, from a team at Nanyang Technological University and partners ↗, tested general-purpose chatbot-style AIs (ChatGPT, Meta's Llama, DeepSeek) on molecule design. You give the AI a plain-language description of the molecule you want, it produces a candidate. The finding: these AIs are decent on broad requests (make a molecule that behaves like a drug) but they collapse when the request is specific (make a molecule that fights this particular protein). Across three specific disease targets, their zero-shot success rates were essentially zero. One of those targets, called JNK3, was hit by the best AI in just 0.6% of attempts.
Show the same AI ten examples of what the right answer looks like, and the success rate on that target jumps to 32%. But the examples change the behavior in another way: the molecules the AI produces become far less varied, often repeating shapes close to the examples.
The second paper takes the opposite approach. Instead of a general-purpose AI, the PepMorph authors ↗ built a specialized tool narrowly for one job: designing peptides that assemble into specific shapes — fibers or spheres, which matter for drug-delivery materials. The payoff is big. When asked to design sphere-forming peptides, their tool found 76 working candidates out of 4,800 attempts. A general unconditioned approach tried the same task and found zero.
The bigger picture. For the last couple of years headlines have promised AI-designed drugs as one general solution. What both papers suggest, from opposite directions, is that AI works when matched carefully to its task. A broad chatbot-style system handles broad questions. A narrow specialized model handles narrow questions. Neither does the other job well.
The caveats. Neither result has been validated in a wet lab. The chatbot paper measures success with software that predicts whether a molecule is chemically valid and hits its target, not with experiments on real cells or patients. The peptide paper validates using physics simulations and visual inspection of the simulated shapes. Both are steps toward real drugs, not proof of them.
A note from our side. We host peptide candidates at different maturity levels (designed, computed, reproduced, synthesized, bioassayed). The practical implication of this week's papers is that the AI tool someone reaches for depends on where they are in that ladder. Early-stage exploration can use broader AI. Late-stage specific work needs specialized models.
One sentence to take away. AI for drug design is no longer one technology. It is at least two, and knowing which one to reach for at which stage is the new skill.