AI-designed molecules block a human enzyme viruses use, in the lab

Viruses mutate fast, and a drug that attacks one of their proteins can stop working once the virus learns to resist it. A team at Korea University tried another route: instead of hitting the virus, it went after a human enzyme that the virus needs. The researchers, led by Kwang Yeon Hwang of the College of Life Sciences and Biotechnology and Hong-Rae Kim of the College of Medicine, call this host-directed therapy, an emerging approach that treats an infection by targeting the body's own biological pathways.

The enzyme is DHPS, deoxyhypusine synthase. It takes part in a chemical modification called hypusination, which activates a protein known as eIF5A. That process, the researchers explain, is important for the replication of a broad range of viruses, including HIV-1 and SARS-CoV-2. To find molecules that block DHPS, the team benchmarked 13 machine-learning models for predicting binding scores and 10 methods for generating new molecules, all against the same target. Among the prediction models, OOF-Stacking performed best overall. Among the generators, RL-Design reached a predicted hit rate of 92.02%, and 98.0% of its top-ranked molecules passed a standard drug-likeness threshold. These are computed predictions, not laboratory results.

The laboratory test is where the predictions met reality. The researchers picked six fragments, small compounds of 140 to 190 daltons, a unit of molecular mass: three predicted to be active and three predicted to be inactive. All three predicted-active fragments inhibited DHPS, and none of the predicted-inactive ones did. Two of the active fragments gave measurable IC50 values of 31.2 and 49.1 μM, a measure of inhibition strength that the team places within the typical range for initial fragment hits. An X-ray co-crystal structure, which shows how a molecule sits inside a protein, also placed one compound at the entrance of the enzyme's substrate and NAD-binding pocket.

Kim is careful about what these results are. The best-known DHPS inhibitor closely mimics a natural molecule in the cell, which carries a risk of acting on the wrong targets. The new hits are built on unrelated scaffolds, and small fragments of this kind leave room to be grown into more potent compounds. "However, these molecules are early-stage hits rather than drug candidates," he said. The team says further studies are needed on the inhibition mechanisms, selectivity, cellular activity, toxicity and performance in animal models.

For now, the gain is a method. The pipeline is freely available, and its modest computing requirements could let laboratories with limited computational resources run experimentally grounded AI-assisted drug searches. The team says it could be applied to other disease targets. For patients, the study offers no timeline and no drug: it is a proof of concept that a predicted molecule can be confirmed in a test tube, and the work that separates that from a treatment has not been reported.