Cryo-electron microscopy (cryo-EM) has become one of the major technologies for high-resolution structure determination of biological macromolecules, after the so-called “resolution revolution”1, 2. Indeed, resolution — here referring to the size of the finest structural details that can pass a consistency test — is understood to be of central importance for quality assessment of a cryo-EM density map3, 4. Technically, resolution can be calculated by the Fourier shell correlation (FSC) procedure, one of the most widely used consistency tests, which measures the normalized cross-correlation between the two ‘half maps’ (assessed shell-by-shell) in the Fourier space. Of note, half maps here are obtained by randomly splitting the entire dataset of particle images into two halves and then independently refining and reconstructing two density maps.Table 1.

A cryo-EM density map can have varying extents of resolvability in different local regions, owing to factors including sample heterogeneity and radiation damage[5]. Accordingly, resolution is best assessed locally. One of the first methods in this area, blocres[3], estimates local resolution based on FSC calculations of two half maps within a moving window. However, although researchers are encouraged to submit half maps in the latest EMDB deposition requirments, to date roughly half of the deposited maps in EMDB still do not include half maps, making the requirement for two half maps untenable for local resolution assessment for many structures. It follows that methods that can estimate local resolution using a single final map are substantially more convenient in practice. Methods in this category include ResMap and MonoRes. ResMap[6] estimates the local resolution as the minimum wavelength of a 3D sinusoidal wave that is detectable above a noise threshold for each voxel of the map. MonoRes[7] follows a similar signal detection strategy, but is based on the calculation of the monogenic signal. These two methods suffer from the requirement for accurate estimation of noise variance, which is highly sensitive to parameters used in the pre-processing steps (e.g., B-factor correction and map sharpening)[8].

In addition, some tools (blocres and one running mode of MonoRes) require that users provide a mask enclosing the macromolecule in the input density map; and an inaccurate mask is known to affect the estimation accuracy of local resolution[3]. It bears emphasis that the molecular mask, which is a minimum volume surrounding the biological macromolecule in a density map[9], is also essential for many cryo-EM computational tasks beyond local resolution estimation. For example, in density map reconstruction, a mask is recommended to be applied at every iteration of reconstruction if the macromolecule is far from spherical[10]. Currently, a molecular mask can be estimated with a single map, but it requires users to tune parameters manually[11].

Owing to the excellent ability for patten recognition, deep learning has enabled major breakthroughs in computer vision[12], natural language processing[13], and biological data mining14, 15, 16. To date, thousands of experimental cryo-EM density maps have been deposited in EMDB. These maps can be used to train deep learning models, as they contain rich information corresponding to resolution-related density patterns. The recently reported DeepRes[8] method uses deep learning to learn density pattens from macromolecular structures of different resolutions. However, it is notable that the DeepRes model was trained on simulated cryo-EM density maps, so it suffers from low generalization ability for real experimental cryo-EM data.

We here present a deep learning-based algorithm, CryoRes, for estimating local resolution for cryo-EM density maps with only a single map as input. We trained CryoRes on 1,174 experimental cryo-EM density maps, and it achieved an average root-mean-square error (RMSE) of 2.26Å for local resolution estimation relative to the currently most reliable FSC-based method blocres. Furthermore, CryoRes can also output a molecular mask with accuracy 12.12% higher than a ResMap-generated mask. CryoRes does not require half maps, masks, or any manually tuned parameters, and is ultra-fast, fully automatic, and applicable to cryo-EM subtomogram data.