Abstract
We introduce IsoNet2, an end-to-end self-supervised deep-learning method that directly reconstructs high-quality 3D densities from cryogenic electron tomography. A unified network simultaneously performs denoising, contrast transfer function correction, and missing-wedge restoration, achieving ~20 Å resolution without averaging. A feature-rich GUI enables rapid, dataset-specific fine-tuning for end-users. IsoNet2 resolves domain organization in HIV capsid proteins, tRNA occupancy in individual ribosomes, and in situ architectures of mitochondrial respiration-related complexes, enabling atomic-level interpretation of cellular environments.