QuantaMind MD enables protein modeling with ab initio accuracy

Abstract Accurate biomolecular simulations are essential for understanding chemical processes and advancing applications such as protein engineering and drug design. While quantum mechanical (QM) methods can provide chemical accuracy, their computational cost limits their scalability. Machine learning force fields (MLFFs) offer a promising alternative by achieving similar accuracy with vastly improved efficiency, but their effectiveness is often constrained by limited conformational and chemical diversity in training. We introduce QuantaMind, a robust MLFF workflow designed to extend chemical and conformational coverage, particularly in transition-state regions. QuantaMind enables quantum-accurate molecular dynamics simulations, successfully capturing complex phenomena such as proton diffusion, water dissociation, and acid-base neutralization. We show that QuantaMind can be applied to large biomolecular systems, enabling accurate protein structure optimization and improving the prediction of residue contacts and hydrogen bonds. A pocket-centric simulation strategy further allows QuantaMind to efficiently model protein-ligand interactions with high structural accuracy. These results establish QuantaMind as a versatile and scalable tool for atomistic simulations at ab initio accuracy. Competing Interest Statement The authors declare the following competing interests: A patent application related to the QuantaMind training workflow has been submitted (Application No. 202510725979.7, pending); A patent application related to the QuantaMind structure optimization has been submitted (Application No. 202510725999.4, pending). A patent application related to the QuantaMind training using transition-state data has been submitted (Application No. 202511174067.1, pending). The applications are filed by MoleculeMind and includes contributions from Deqiang Zhang, Song Xia and Jinbo Xu. The authors declare no other competing interests. Footnotes Section "Protein dynamic structure prediction with ab initio accuracy" updated. Data Availability The source code used in this study is available at https://github.com/MoleculeMindOpenSource/QuantaMind. The molecular dynamics trajectories generated in this work are accessible via Zenodo at https://zenodo.org/records/16910424.