An AI agent designed its own molecular dynamics pipeline, and the binder it picked turned out to be real.
The system, called MDForge, is built around a large language model agent that treats pipeline design as open-ended code generation rather than picking from a fixed toolbox of steps. Because running a molecular dynamics simulation is too expensive for trial and error, the agent's feedback signal is sparse, so the researchers added a multi-agent debate among simulated physics experts to turn that sparse feedback into richer in-context guidance. On three SAMPL host-guest binding free-energy benchmarks, the agent-designed pipelines matched pipelines built by human experts. The team then pointed MDForge at a library of untested candidate molecules for a host compound called CB[7], and its pipeline flagged a promising binder that wet-lab competition NMR testing confirmed as a genuine high-affinity, picomolar-strength binder.
Molecular dynamics pipelines are usually hand-tuned by specialists, which limits how many labs can run this kind of atomistic modeling in the first place. MDForge's result matters less for the benchmark scores and more because the agent's pick survived contact with an actual wet-lab assay, a step most AI-for-science papers skip.
One confirmed CB[7] binder from one candidate library is encouraging, not evidence this generalizes to messier drug-discovery targets like protein pockets.