XDXD: End-to-end crystal structure determination with low resolution X-ray
1. A groundbreaking study introduces XDXD, the first end-to-end deep learning framework that can determine complete atomic models directly from low-resolution single-crystal X-ray diffraction data. This innovation bypasses the need for manual map interpretation and significantly simplifies the process of crystal structure determination.
2. XDXD achieves a remarkable 70.4% match rate for structures with data limited to 2.0 ̊A resolution, with a root-mean-square error (RMSE) below 0.05. This demonstrates its robustness and accuracy, even for complex systems with up to 200 non-hydrogen atoms.
3. The model is trained on a diverse set of 395,117 simulated diffraction patterns and validated on 24,000 experimental structures from the Crystallography Open Database (COD). It shows strong performance across various space groups and chemical compositions, highlighting its broad applicability.
4. A key innovation of XDXD is its diffusion-based generative model, which iteratively refines atomic coordinates. This approach, combined with cross-attention mechanisms, allows the model to effectively leverage geometric information and produce chemically plausible crystal structures.
5. Case studies on small peptides demonstrate XDXD's potential for extension to more complex biological systems, such as proteins and nucleic acids. This suggests that the model could revolutionize structural biology by providing automated structure solutions for previously intractable cases.
6. The study also includes systematic ablation studies that quantify the critical dependence of structure determination performance on diffraction signal quality. Higher-resolution diffraction data significantly enhance predictive accuracy, confirming the importance of rich reciprocal-space information.
7. The authors plan to make the source code publicly accessible via GitHub upon acceptance of their paper, promoting further research and development in this exciting field.
📜Paper:
arxiv.org/abs/2510.17936v1
#XrayDiffraction #DeepLearning #CrystalStructure #StructuralBiology #AI #Science