AI-designed fragments inhibit a human enzyme targeted in antiviral research

Korea University researchers have moved AI-generated molecules from computer predictions into laboratory tests against a human enzyme linked to viral replication. In an October 19 account, the university described six selected fragments whose experimental results matched their predicted activity. The finding offers an early starting point for antiviral research—not evidence of a working antiviral drug.

The account, supplied by Korea University College of Medicine and published by Phys.org, describes research published online in the Journal of Cheminformatics on August 5, 2026. The team was led by Kwang Yeon Hwang of the university’s College of Life Sciences and Biotechnology and Hong-Rae Kim of its College of Medicine.

A human enzyme becomes the target

The starting problem is resistance: viruses evolve rapidly and can become resistant to medicines aimed directly at viral proteins. The researchers instead pursued host-directed therapy, an approach that targets biological pathways or proteins in the infected person rather than the virus itself. Their chosen target was deoxyhypusine synthase, a human enzyme usually shortened to DHPS.

DHPS helps activate a protein called eIF5A through a chemical modification known as hypusination. Hwang said this process supports replication in a broad range of viruses, including HIV-1 and SARS-CoV-2. That biological role makes the enzyme a potential target for drug discovery; it does not establish that these particular fragments can treat either infection.

From competing models to six fragments

Before testing compounds, the team compared 13 machine-learning models for predicting molecular binding scores and 10 methods for generating molecules. All were evaluated against DHPS. A prediction approach called out-of-fold stacking, or OOF-Stacking, performed best overall among the prediction models tested. RL-Design led the generation methods with a predicted hit rate of 92.02%.

The team also reported that 98.0% of RL-Design’s top-ranked molecules passed a standard drug-likeness threshold. Both figures describe computational screening results, not a laboratory success rate or proof that the molecules are medicines. The physical test was much smaller: six selected fragments, each weighing between 140 and 190 daltons.

Three fragments were predicted to be active and three inactive. All three predicted-active fragments inhibited DHPS; none of the predicted-inactive fragments did. Two active fragments produced measurable potency results, reported as IC50 values of 31.2 and 49.1 micromolar. The researchers described those values as within the typical range for initial fragment hits.

The team also obtained an X-ray co-crystal structure showing one AI-generated compound bound at the entrance to the enzyme’s substrate and NAD-binding pocket. That supplied structural evidence of binding alongside the inhibition measurements, rather than relying solely on the computer’s prediction of how a molecule would interact with DHPS.

Our hits are fragments built on scaffolds unrelated to the known inhibitor. In fragment-based discovery, this is exactly where you want to start: they are small and bind efficiently for their size, which leaves considerable room to grow them into more potent compounds. However, these molecules are early-stage hits rather than drug candidates.

Hong-Rae Kim, Korea University, in the university account published by Phys.org

The next work is biological

Kim contrasted the fragments with the best-known DHPS inhibitor, which closely resembles a natural cellular molecule and carries a risk of activity against unintended targets. Different molecular structures offer a starting point, but the team identified several questions that further experiments must resolve:

• How the fragments inhibit DHPS, and how selectively they act.
• Whether they show useful activity in cells, and what toxicity they have.
• How they perform in animal models.

The university says the pipeline is freely available and has modest computing requirements, potentially making it accessible to laboratories with limited computing resources. Its account, also published by News-Medical, proposes applying the workflow to other disease targets. That broader use remains a possibility, not a result demonstrated here.