AI Achieves Perfect Hit Rate in Hunt for Next-Generation Antiviral Compounds

Artificial intelligence has delivered a rare perfect score in drug discovery. Researchers at Korea University report that all three molecular fragments predicted by their AI pipeline to inhibit a key human enzyme involved in viral replication showed activity in laboratory tests, while all three predicted-inactive compounds failed to inhibit—a 100% concordance between computational prediction and experimental reality.

The achievement, published in the Journal of Cheminformatics in August 2026, represents more than a technical milestone. It demonstrates that AI-guided drug discovery can work with modest computational resources while targeting host-directed antiviral therapy—an approach that attacks the human cellular machinery viruses hijack, rather than the rapidly mutating pathogens themselves.

Targeting the Host, Not the Virus

The research team, led by Professor Kwang Yeon Hwang from the College of Life Sciences and Biotechnology and Associate Professor Hong-Rae Kim from the College of Medicine, focused on deoxyhypusine synthase (DHPS), a human enzyme involved in a process called eIF5A hypusination. This cellular modification supports the replication of multiple viruses, including HIV-1 and SARS-CoV-2.

By targeting a host enzyme rather than viral proteins, the approach could potentially circumvent the resistance mechanisms that plague traditional antiviral drugs. Viruses mutate rapidly; human enzymes do not.

The researchers systematically benchmarked 13 machine-learning models for predicting binding scores and 10 methods for generating new molecular structures. The best-performing combination—Out-of-Fold Stacking for prediction and RL-Design for molecular generation—achieved a predicted hit rate of 92.02%, with 98.0% of the top-ranked molecules passing standard drug-likeness criteria.

From Silicon to Laboratory

The true test came in experimental validation. The team selected six small molecular fragments (140-190 daltons)—three predicted to be active and three predicted to be inactive—and tested them against DHPS in the laboratory.

All three predicted-active fragments inhibited the enzyme, with two showing IC50 values of 31.2 μM and 49.1 μM, typical for early-stage fragment hits. None of the predicted-inactive fragments showed inhibitory activity. X-ray crystallography confirmed that the active compounds bound at the entrance to DHPS's substrate and NAD-binding pocket, validating not just activity but the predicted binding mode.

Democratizing AI Drug Discovery

Perhaps as significant as the scientific results is the accessibility of the approach. The researchers emphasize that their pipeline requires only modest computing requirements and is freely available, potentially enabling smaller laboratories and biotech startups to leverage AI-guided drug discovery without massive infrastructure investments.

The work represents a proof of concept for host-directed antiviral therapy using AI-driven methods. However, substantial development remains before these early-stage hits could become therapies. The fragments require mechanism studies, selectivity testing, evaluation of cellular activity and toxicity, and validation in animal models—the standard gauntlet of drug development that typically takes years and significant resources.

The perfect concordance between prediction and experiment suggests that AI-guided approaches have matured beyond generating interesting hypotheses to producing experimentally validated starting points for drug development, particularly for targets where host-directed therapy offers advantages over traditional pathogen-focused strategies.

Sources

1. Researchers use AI to identify early-stage inhibitors for antiviral therapy — News-Medical.net