Researchers at Korea University used AI to discover early-stage inhibitors of DHPS, a human enzyme critical to viral replication in HIV-1 and SARS-CoV-2. Unlike most AI drug-discovery efforts, the team experimentally validated their findings — and all three AI-predicted active compounds successfully inhibited the enzyme. The work offers a promising, accessible blueprint for host-directed antiviral therapy.
Viruses are notoriously tricky to treat because they mutate fast, quickly developing resistance to drugs that target them directly. Researchers at Korea University are taking a different approach: instead of attacking the virus, they're targeting a human enzyme the virus needs to replicate. Using an AI-guided pipeline, they identified early-stage inhibitors of deoxyhypusine synthase (DHPS), a human enzyme that viruses like HIV-1 and SARS-CoV-2 hijack to fuel their own replication.
The team benchmarked 13 machine-learning models for predicting molecular binding and 10 methods for generating new drug-like molecules. Their top-performing generation method, RL-Design, achieved a predicted hit rate of 92%, and 98% of its top molecules passed standard drug-likeness filters. Crucially, the researchers didn't stop at the computer — they tested six AI-selected fragments in the lab, and all three predicted-active compounds successfully inhibited DHPS, with zero false positives among the predicted-inactive ones.
By the Numbers:
Why it matters: This study offers a reproducible, computationally modest pipeline that bridges AI-generated molecules with real lab validation — a gap that often goes untested. Because the pipeline is freely available and doesn't require heavy computing resources, it could open the door to AI-assisted drug discovery for labs that previously lacked the infrastructure to participate.
