Abstract: CryoEM and cryoET enable imaging of biological specimens frozen in vitreous ice, revealing 3D molecular or cellular structures at high resolution and in their native state. However, cryoET is limited by the “missing-wedge” problem due to restricted tilt angles, and cryoEM often suffers from preferred orientation, resulting in uneven sampling of angular views and leaving parts of Fourier space poorly covered. We developed IsoNet and spIsoNet to address these inverse problems through data-driven, self-supervised deep learning: both methods learn from collected data alone, using well-sampled orientations to infer under-represented ones. IsoNet restores isotropy in tomograms by reconstructing missing information; spIsoNet adapts these principles to single-particle and subtomogram averaging workflows, improving angular coverage and alignment. Together, our deep learning methods reliably mitigate previously challenging physical constraints in cryoEM/cryoET.
a joint institute of 
and 
and