Evaluation of Physics-Based Pose Prediction and the Role of Receptor Flexibility in a Community Antiviral Blind Challenge

Abstract Prediction of protein–ligand binding modes is a key step in computer-aided drug design, since incorrect pose prediction inevitably limits the achievable accuracy of downstream physics-based affinity predictions. The ASAP-Polaris-OpenADMET antiviral competition was launched in 2025 as a means to prospectively test computational methods across a range of measurements that guide early prioritization during a drug discovery campaign including pose prediction. Here, we describe our two open-source, physics-based entries to the pose prediction challenge, which are based on templated molecular growth with hybrid machine learning/molecular mechanics interatomic potentials. We show that our approach was among the top-performing physics-based methods, and with a retrospective change to the workflow to make greater use of the training data as templates, we obtain a success rate of 74% across SARS-CoV-2 and MERS-CoV main protease target protein structures. Finally, we investigate the importance of receptor side-chain flexibility on retrospective performance. We find that building poses against an ensemble of feasible side-chain conformers is useful in generating improved binding poses but that the tested scoring functions are unable to successfully pick out native structures from the generated ensemble.