A team at Korea University has developed an AI-guided drug-discovery pipeline that designs potential antiviral compounds targeting deoxyhypusine synthase (DHPS), a human host enzyme. DHPS is involved in the hypusination of eIF5A, a process that supports the replication of a broad range of viruses, including HIV-1 and SARS-CoV-2, making it a potential target for host-directed antiviral drug discovery. The paper was published in the Journal of Cheminformatics.
The researchers systematically benchmarked 13 machine-learning binding-score prediction models and 10 molecule-generation methods: OOF-Stacking performed best at prediction, while RL-Design achieved the highest generation hit rate at 92.02%, with 98.0% of its top-ranked molecules passing a standard drug-likeness threshold.
Unlike most AI drug-discovery efforts that stop at computational evaluation, the team experimentally validated the results: six AI-generated fragments (140–190 daltons) were tested — all three predicted-active fragments inhibited DHPS, while all three predicted-inactive ones did not, showing complete agreement between computational predictions and experiments. Two of the active fragments gave measurable IC50 values of 31.2 and 49.1 μM, and an X-ray co-crystal structure confirmed one compound binding at the entrance of the DHPS substrate and NAD-binding pocket.
Notably, the new hit fragments are built on scaffolds unrelated to the known DHPS inhibitor — the best-known inhibitor closely mimics a natural cellular molecule, which brings a risk of off-target activity. The researchers stress these are early-stage hits rather than drug candidates; further studies will be needed on mechanism of action, selectivity, cellular activity, toxicity, and animal-model performance. The pipeline is freely available and has modest computing requirements, which could let more laboratories pursue experimentally grounded AI drug discovery.
