Deep-Learning vs Physics-Based Docking Tools for Future Coronavirus Pandemics

The COVID-19 pandemic revealed the urgent need for faster
and
more accessible drug discovery for future coronavirus outbreaks. We
participated in the Polaris Antiviral Drug Discovery 2025 competition
to evaluate how AI can accelerate this process, comparing three molecular
docking approaches: AutoDock Vina (physics-based), GNINA (deep learning-assisted),
and Boltz-2 (data-driven), using SARS-CoV-2 and MERS-CoV data sets.
Our results demonstrate that AI has transformed drug discovery: Boltz-2
reached over 80% accuracy, greatly improving docking precision. However, accurate prediction of binding potency remains challenging and requires further methodological advances.
Traditional methods like Vina and GNINA are faster, making them suitable
for large-scale screening, and their physical basis is easier to interpret,
with room for future improvement. By combining modeling and analysis
tools into easy-to-use platforms like ChemOrchestra, drug discovery
can become more decentralized and accessible, allowing
people to use advanced models quickly.

新冠疫情（COVID-19 pandemic）暴露了针对未来冠状病毒暴发，开发更快速、更具可及性的药物研发手段的迫切需求。本研究团队参与了北极星抗病毒药物发现2025（Polaris Antiviral Drug Discovery 2025）竞赛，旨在评估人工智能（AI）如何加速该进程，并采用SARS-CoV-2与MERS-CoV数据集，对比了三种分子对接方法：基于物理原理的AutoDock Vina、深度学习辅助的GNINA以及数据驱动的Boltz-2。研究结果表明，人工智能已重塑药物研发范式：Boltz-2的准确率突破80%，大幅提升了分子对接精度。然而，精准预测结合效能仍存在挑战，有待进一步的方法学突破。诸如AutoDock Vina与GNINA这类传统方法运行速度更快，适用于大规模虚拟筛选，且其物理基础更易于阐释，仍具备后续优化空间。通过将建模与分析工具整合为ChemOrchestra这类易用平台，药物研发将更加去中心化、更具可及性，使使用者能够快速运用先进模型。