AI in drug discovery matures

Drug discovery usually fails for a simple reason: a medicine has to fit a protein in the body like a key fits a lock, but the lock is moving the whole time. University of Virginia researchers said on April 1 that most computer methods still treat proteins like frozen snapshots, which is one reason so many drug ideas die later in testing. (news.med.virginia.edu) The moving part has a name: proteins change shape when a drug grabs them. University of Virginia’s team compared older software to designing a key for a lock sitting still, while their system tries to design the key while the lock is still jiggling. (news.med.virginia.edu) That is what the new University of Virginia package does. YuelDesign proposes new molecules, YuelPocket finds the exact pocket on the protein where a drug can attach, and YuelBond checks whether the chemical bonds in the proposed molecule actually make sense. (med.virginia.edu) The underlying method is a diffusion model, which is the same family of artificial intelligence systems used to generate images by refining noise into a final picture. University of Virginia said YuelDesign uses that approach to generate both the shape of the protein pocket and the small molecule together, so each can adapt to the other during the design step. (news.med.virginia.edu) This is a very different kind of artificial intelligence story from the 2023 wave of “type a prompt, get a miracle drug.” The University of Virginia group is using one model to suggest candidates, a second to locate the binding site, and a third to reject chemically impossible answers before a lab team wastes time on them. (med.virginia.edu) Insilico Medicine’s April 9 update shows the same shift from demo to grind. Its Chemistry42 system was used to design inhibitors for a cancer target called protein kinase membrane associated tyrosine and threonine 1, or PKMYT1, where the hard part is not finding any binder but finding one that hits the right target without hitting many close cousins. (news-medical.net) Selectivity is the medicinal chemistry version of threading a needle. Insilico said existing clinical leads against PKMYT1 had less than 10-fold selectivity over off-target kinases such as BRAF, RAF1, and SRC, and those off-target hits were linked to dose-limiting toxicities including severe skin rashes. (news-medical.net) Its answer was not a chatbot-style leap but a geometry fix. The team replaced an older core scaffold with a thiazolyl-pyrazole ring system that locks the molecule into the shape PKMYT1 prefers, and the reported lead A4-ent1 reached an inhibitory concentration of 2.2 nanomolar with more than 100-fold selectivity over the related kinase WEE1. (news-medical.net) Insilico had already shown the same pattern in a November 28, 2025 Nature Communications paper. In that study, the company used its generative platform to build a bifunctional molecule that both inhibits PKMYT1 and tags it for destruction, and the lead compound showed oral bioavailability and tumor responses in xenograft models. (nature.com) The third piece of this week’s story comes from Korea Advanced Institute of Science and Technology, which announced on April 9 that Gyu Rie Lee and David Baker designed proteins from scratch that recognize specific small compounds. Instead of searching nature for a protein that already does the job, they built new binders de novo, which means from nothing but the design rules. (kaist.ac.kr) Their test case was cortisol, the stress hormone. Korea Advanced Institute of Science and Technology said the team built a protein that selectively binds cortisol and turned it into a biosensor, and the group reported experimentally verified binding proteins for six compounds including metabolites and small-molecule drugs. (kaist.ac.kr) Put those three projects together and the shape of the field looks different. Artificial intelligence is increasingly doing the unglamorous middle work of drug discovery — proposing molecules, checking pockets, fixing geometry, and handing experimental teams candidates that are a little less wrong before the expensive part begins. (news.med.virginia.edu) (news-medical.net) (kaist.ac.kr)