Key Takeaways

  • Korea University researchers experimentally confirmed that three AI-selected fragments inhibited deoxyhypusine synthase, while three predicted-inactive fragments did not.
  • The study connected model benchmarking, molecular generation, laboratory assays and X-ray structural analysis in one reproducible workflow.
  • The compounds remain early-stage research hits, but the pipeline could give smaller laboratories a more accessible route into AI-assisted drug discovery.

Korea University researchers have taken an AI drug-discovery project beyond computational predictions, experimentally validating small molecules designed to inhibit a human enzyme associated with viral replication. The results offer an early demonstration of how machine learning, molecular generation and laboratory testing can be assembled into a practical discovery workflow.

The study, published online in the Journal of Cheminformatics on August 05, 2026, focused on deoxyhypusine synthase, or DHPS. Researchers from Korea University's College of Life Sciences and Biotechnology and its College of Medicine led the research.

DHPS helps activate eukaryotic translation initiation factor 5A, known as eIF5A, through a modification called hypusination. That biological process supports the replication of multiple viruses, including HIV-1 and SARS-CoV-2. Targeting DHPS could therefore provide a host-directed therapy strategy rather than attacking a viral protein.

That distinction matters. Drugs aimed at viral proteins can lose effectiveness as viruses mutate, although the degree and pace vary by pathogen. Host-directed therapies instead interfere with cellular functions that a pathogen uses to reproduce. The approach may present a different barrier to resistance, but it also raises difficult questions about safety and selectivity because the target performs functions in human cells.

To construct their workflow, the Korea University team compared 13 machine-learning models that predict molecular binding scores and 10 methods for generating new molecules. Out-of-Fold Stacking, or OOF-Stacking, produced the strongest overall results among the prediction models tested. RL-Design recorded the highest predicted hit rate among the generation methods at 92.02%, while 98.0% of its top-ranked molecules cleared a standard drug-likeness threshold.

Predicted performance alone does not establish biological activity. The researchers addressed that familiar gap by selecting six molecular fragments weighing between 140 and 190 daltons. Three were classified as predicted-active and three as predicted-inactive.

All three predicted-active fragments inhibited DHPS in the experimental assessment. None of the predicted-inactive fragments did. Two active fragments produced measurable IC50 values of 31.2 and 49.1 μM, concentrations consistent with the range where fragment hits often begin.

The sample was tiny, so the perfect classification across six fragments should not be treated as broad evidence of model accuracy. Still, including negative controls strengthened the experiment. The researchers were not simply testing a handful of highly ranked molecules and reporting whichever ones worked.

Structural evidence added another layer. An X-ray co-crystal structure indicated that one AI-generated compound bound at the entrance to the DHPS substrate and NAD-binding pocket. This can help medicinal chemists understand how a fragment interacts with the enzyme and where chemical modifications might improve potency or selectivity.

The new fragments also use scaffolds unrelated to the established DHPS inhibitor, which closely resembles a natural cellular molecule and may create off-target concerns. Because fragments are intentionally small, they provide room for optimization into potential antiviral medicines.

Further work would need to examine inhibition mechanisms, target selectivity, activity in cells, toxicity and performance in animal models. The researchers explicitly characterized the molecules as early-stage hits rather than drug candidates.

For B2B technology and life-sciences organizations, the larger point is the pipeline. It is freely available, reportedly has modest computing requirements and could be adapted to other disease targets. That may broaden participation beyond heavily funded pharmaceutical AI programs, particularly for academic groups and smaller biotechnology companies that can combine computational screening with laboratory access.

The work also arrives as South Korea expands its wider AI infrastructure base. IDC Korea forecasts that the domestic server market will grow at a 7.8% compound annual rate through 2029, reaching ₩7.037 trillion in annual revenue. Although drug discovery has specialized data and validation requirements, greater domestic compute capacity can support molecular modeling, research collaboration and larger experimental pipelines.

For Korea University College of Medicine, the immediate result is narrower but concrete: AI-generated chemistry produced experimentally confirmed DHPS-binding fragments. The commercial and clinical distance remains substantial. Even so, the study shows a credible route from algorithmic design to physical evidence, which is where AI drug discovery starts becoming more useful to research partners, biotechnology investors and pharmaceutical development teams.