PREreview of CryoNeRF: reconstruction of homogeneous and heterogeneous cryo-EM structures using neural radiance field

In this study, Qu H et al. present “CryoNeRF”, a novel approach for addressing heterogeneity in single-particle cryo-EM data, a challenge that has become an area of intense study and development in the field. CryoNeRF is an intriguing application of Neural Radiance Field (NeRF), which is a deep learning-based method commonly used in computer vision research to generate 3D objects or scenes from numerous 2D images. One of the proposed benefits of this methodology is that it enables 3D reconstruction to be performed entirely in Euclidean space, removing the need for Fourier transforms. This aspect of CryoNeRF distinguishes the methodology from other deep neural network-based methods that describe the heterogeneity within a given dataset, such as CryoDRGN, 3DFlex, and DynaMight. By avoiding Fourier transformations, CryoNeRF claims to mitigate potential artifacts such as phase errors, ringing effects, interpolation artifacts, and loss of high-frequency information associated with some implementations of Fourier-based reconstruction approaches.

The authors first demonstrate the capacity of CryoNeRF to produce high-resolution structures from homogeneous datasets, with purported improvements of ~1 Å resolution over current methods, although this improvement is debatable, based on the presented structures. CryoNeRF was also used to describe conformational and compositional heterogeneity present in simulated CryoBench datasets, as well as in experimental EMPIAR datasets. Based on the output presented in the manuscript, CryoNeRF’s overall performance in resolving heterogeneity is on par with existing deep-learning-based methods, leaving its broader advantages yet to be fully realized. Further, CryoNeRF’s algorithm incurs a significant increase in computational cost compared to conventional methods, which raises uncertainties about whether its benefits outweigh its drawbacks for routine use. Generally, the application of this deep neural network framework with a Euclidian reconstruction approach is worth further consideration and development. However, the data presented in this manuscript are not particularly compelling.

Major Issues- 

·       The main text section and figure legend regarding multi-resolution hash encoding (Fig. 1D) provide an abstract overview of the approach, but are unclear on how it is specifically implemented in the context of heterogeneity modeling.

·       Despite claims that CryoNeRF outperforms CryoSPARC for homogenous reconstruction (Fig. 2A), the density maps obtained from CryoNeRF of the RAG1-RAG2 complex, as well as secondary structural elements of the 80S ribosome maps, appear to be equal to or lower in resolution than the maps generated by CryoSPARC. The authors should include panels showing detailed representative density with equivalent contour levels from a selection of specific domains or residues to enable a more rigorous comparison of local resolution. In the absence of improved structural features, the alignment and reconstruction algorithms may be influencing the FSC without actually improving the quality of the density. Along these lines, the CryoNeRF FSC curve does not reach zero for the ribosomal dataset (Fig. 2C, right panel), which is indicative of overfitting.

·       The authors have uploaded their maps and input training files to Zenodo, though they appear to be missing the sharpened maps obtained from CryoSPARC and CryoDRGN which can be used for direct comparison to the CryoNeRF sharpened maps. Additionally, it is not clear whether sharpened or unsharpened maps are compared in Fig 2, making it difficult to interpret observed differences.

·       While the Cryo-EM density maps have been uploaded to Zenodo, they have not yet been deposited into the Electron Microscopy Data Bank (EMDB). We encourage the authors to deposit all maps to the EMDB to enable further validation and quality assessments, as well as to ensure accessibility within the cryo-EM community.

·       Supplementary videos are not available in the supplementary materials section that was posted to bioRxiv.

·       We suggest that the authors compare the CryoNeRF results from the simulated datasets to other deep-learning-based approaches, as this data is available on CryoBench and would inform potential users on the advantages of CryoNeRF over other methods.

Minor issues-

·       Citations after Ref. 17 are misaligned with their corresponding references, requiring a thorough revision to ensure accuracy.

·       The heatmap coloring scheme used in Fig 2C. is challenging to interpret, particularly for the high-resolution features. We recommend using an established colorblind-friendly palette.

·       The difference in FSC curves between the 10k training step and subsequent steps is vastly different for the RAG1-RAG2 complex and the ribosomal dataset. The authors could provide a brief explanation for this.

·       In Fig. 3C, the angle of rotation between views is not provided, making it difficult to appreciate the orientations. Additionally, a comment on the stationary vs rotational regions should be included in the figure legend to justify the selection of views. 

·       In Fig. 6A, the class IDs are mislabeled or absent, making this panel difficult to interpret.

·       There are no statistical analyses of CryoNeRF, making it difficult to assess significance and reproducibility.

·       The authors did not consider any datasets with low-resolution or highly flexible domains, so its generalizability remains uncertain.

Competing interests

The authors declare that they have no competing interests.