Physics-based docking holds up in an antiviral blind challenge, but receptor flexibility remains hard to score
- Two open-source, physics-based entries joined the 2025 ASAP-Polaris-OpenADMET antiviral pose prediction challenge, using templated ligand growth with hybrid ML/MM potentials.
- Both ranked among the stronger physics-based approaches; a revised, template-informed workflow succeeded on most SARS-CoV-2 and MERS-CoV main protease cases.
- Side-chain ensembles improved pose generation, yet scoring could not reliably pick the correct receptor conformation.
💡 Points Based on the abstract
Mpro is a rigid-ish, well-templated pocket, so I'd read this as a good-case result for physics-based docking, not proof it generalizes. The sharper finding is that side-chain ensembles help generation but scoring can't pick the right conformer. On a real project, I'd dock into an ensemble, cluster poses, and benchmark against a co-folding model rather than trust one score.