On-lamella dual-axis cryo-electron tomography and modelling of lamella stability

Cryogenic electron tomography (cryo-ET) of cryogenic focused ion beam (cryo-FIB)–milled lamellae enables visualization of three-dimensional cellular architecture in near-native conditions at nanometer-scale resolution. However, cryo-FIB milling is typically performed using rectangular patterns that limit the accessible tilt range, precluding dual-axis tomography. In addition, cryo-lamellae must remain mechanically connected to the cell body and are prone to fracture during preparation and transfer to the cryo-transmission electron microscope (cryo-TEM), reducing experimental throughput. We developed a cryo-FIB trapezoid milling strategy that permits dual-axis cryo-ET by enabling grid loading with the tilt axis oriented perpendicular to the milling direction. Dual-axis and single-axis cryo-ET datasets were acquired and compared, including single-axis tomograms processed with deep learning–based missing wedge restoration. To address lamella stability, we performed finite element–based mechanical simulations to calculate the build-up of von Mises stresses within different lamella geometries and milling designs. Dual-axis cryo-ET provided additional three-dimensional information compared to conventional single-axis tomography and to single-axis tomograms after deep learning–based missing wedge restoration. This resulted in improved tomogram interpretability and segmentation quality. Furthermore, the mechanical simulations identified stress accumulation and fracture-prone regions within cryo-lamellae, allowing the rational design of milling patterns with enhanced mechanical stability. The trapezoid cryo-FIB milling strategy enables routine dual-axis cryo-ET of cellular lamellae, improving 3D information content and segmentation quality. In addition, the finite element–based mechanical framework provides a predictive tool for optimizing lamella design and stability. Together, these methods increase cryo-ET data quality and throughput and offer valuable training datasets for the further development of deep learning–based restoration approaches.