The article discusses the application of deep learning to improve the alignment and reconstruction of noisy images in electron cryo-microscopy (cryo-EM) for macromolecular structure determination. The authors introduce a method called Blush regularization, which uses a denoising convolutional neural network (DCNN) to enhance the quality of reconstructions. This approach leverages prior knowledge about biological macromolecular structures, which is difficult to express mathematically, to improve image alignment. The DCNN is trained on pairs of half-set reconstructions from the Electron Microscopy Data Bank (EMDB) and is used as an alternative to a smoothness prior. The results demonstrate that Blush regularization yields better reconstructions, particularly for data with low signal-to-noise ratios. The method is tested on various datasets, including a protein-nucleic acid complex with a molecular weight of 40 kDa, which was previously intractable. The study shows that denoising neural networks can expand the applicability of cryo-EM structure determination to a wider range of biological macromolecules. The authors also discuss the potential for overfitting and the use of spectral trailing to prevent it, as well as the broader applicability of Blush regularization to different types of structures and refinement methods.The article discusses the application of deep learning to improve the alignment and reconstruction of noisy images in electron cryo-microscopy (cryo-EM) for macromolecular structure determination. The authors introduce a method called Blush regularization, which uses a denoising convolutional neural network (DCNN) to enhance the quality of reconstructions. This approach leverages prior knowledge about biological macromolecular structures, which is difficult to express mathematically, to improve image alignment. The DCNN is trained on pairs of half-set reconstructions from the Electron Microscopy Data Bank (EMDB) and is used as an alternative to a smoothness prior. The results demonstrate that Blush regularization yields better reconstructions, particularly for data with low signal-to-noise ratios. The method is tested on various datasets, including a protein-nucleic acid complex with a molecular weight of 40 kDa, which was previously intractable. The study shows that denoising neural networks can expand the applicability of cryo-EM structure determination to a wider range of biological macromolecules. The authors also discuss the potential for overfitting and the use of spectral trailing to prevent it, as well as the broader applicability of Blush regularization to different types of structures and refinement methods.