In this work, we constructed a high-quality torsional fragment dataset with DFT-level conformational accuracy, providing a strong foundation for model training. By pretraining a deep learning model on this dataset, we established a framework capable of rapidly evaluating large numbers of small-molecule 3D conformations while maintaining quantum mechanical–level accuracy.

The paper is now available online: https://www.nature.com/articles/s41467-026-69303-5