深度学习引导的高通量 3D 冷冻电镜断层扫描技术在天然细胞中解析病毒衣壳

A fully automated, high-throughput workflow couples cryo-electron tomography (cryo-ET) with deep-learning-based reconstruction and structural analysis to generate sub-nanometer 3D models of viral capsids within their native cellular context. This pipeline integrates a physics-aware generative data-augmentation module for tilt-series synthesis, a multi-scale U-Net for particle picking and segmentation, and a graph-neural-network (GNN) refinement engine for solvent-model interaction mapping. Furthermore, a Bayesian hyper-parameter optimization loop dynamically adjusts electron dose, tilt-step, and reconstruction algorithm settings in real time. Leveraging a curated dataset, this methodology efficiently and accurately resolves the intricate structures of viral capsids, elucidating their assembly mechanisms and interactions within cells. The approach provides unprecedented high-resolution structural information for virology research, promising to accelerate the development of antiviral drugs and vaccines by offering detailed insights into viral architecture and host interactions. This comprehensive system significantly advances the capabilities of structural biology, enabling a deeper understanding of viral pathogenesis and facilitating targeted therapeutic interventions.