Abstract
Cryogenic electron microscopy (cryoEM) is widely used to study biological macromolecules that comprise regions with disorder, flexibility or partial occupancy. For example, membrane proteins are often kept in solution with detergent micelles and lipid nanodiscs that are locally disordered. Such spatial variability negatively impacts computational threedimensional (3D) reconstruction with existing iterative refinement algorithms that assume rigidity. We introduce nonuniform refinement, an algorithm based on crossvalidation optimization, which automatically regularizes 3D density maps during refinement to account for spatial variability. Unlike common shiftinvariant regularizers, nonuniform refinement systematically removes noise from disordered regions, while retaining signal useful for aligning particle images, yielding dramatically improved resolution and 3D map quality in many cases. We obtain highresolution reconstructions for multiple membrane proteins as small as 100 kDa, demonstrating increased effectiveness of cryoEM for this class of targets critical in structural biology and drug discovery. Nonuniform refinement is implemented in the cryoSPARC software package.
Similar content being viewed by others
Main
Singleparticle cryoEM has transformed rapidly into a mainstream technique in biological research^{1}. CryoEM images individual protein particles, rather than crystals and has therefore been particularly useful for structural studies of integral membrane proteins, which are difficult to crystallize^{2}. These molecules are critical for drug discovery, targeted by more than half of drugs today^{3}. Membrane proteins pose challenges in cryoEM sample preparation, imaging and computational 3D reconstruction, as they are often of small size, appear in multiple conformations, have flexible subunits and are embedded in a detergent micelle or lipid nanodisc^{2}. These characteristics cause strong spatial variation in structural properties, such as rigidity and disorder, across the target molecule’s 3D density. Traditional cryoEM reconstruction algorithms, however, are based on the simplifying assumption of a uniform, rigid particle.
We develop an algorithm that incorporates such domain knowledge in a principled way, improving 3D reconstruction quality and allowing singleparticle cryoEM to achieve higherresolution structures of membrane proteins. This expands the range of proteins that can be effectively studied and is especially important for structurebased drug design^{4,5}. We begin by formulating a crossvalidation (CV) regularization framework for singleparticle cryoEM refinement and use it to account for the spatial variability in resolution and disorder found in a typical molecular complex. The framework incorporates general domain knowledge about protein molecules, without specific knowledge of any particular molecule and critically, without need for manual user input. Through this framework we derive a new algorithm called nonuniform refinement, which automatically accounts for structural variability, while ensuring that key statistical properties for validation are maintained to mitigate the risk of overfitting during 3D reconstruction.
With a graphics processing unitaccelerated implementation of nonuniform refinement in the cryoSPARC software package^{6}, we demonstrate improvements in resolution and map quality for a range of membrane proteins. We show results on a 48kDa membrane protein in lipid nanodisc with a Fab bound, a 180kDa membrane protein complex with a large detergent micelle and a 245kDa sodium channel complex with flexible domains. Nonuniform refinement is reliable and automatic, requiring no change in parameters between datasets and is without reliance on handmade spatial masks or manual labels.
Iterative refinement and regularization
In standard cryoEM 3D structure determination^{6,7,8}, a generative model describes the formation of twodimensional (2D) electron microscope images from a target 3D protein density (Coulomb potential). According to the model, the target density is rotated, translated and projected along the direction of the electron beam. The 2D projection is modulated by a microscope contrast transfer function (CTF) and corrupted by additive noise. The goal of reconstruction is to infer the 3D density map from particle images, without knowledge of latent 3D pose variables, that is, the orientation and position of the particle in each image. Iterative refinement methods formulate inference as a form of maximum likelihood or maximum a posteriori optimization^{6,9,10,11}. Such algorithms can be viewed as a form of blockcoordinate descent or expectationmaximization^{12}, each iteration comprising an Estep, estimating the pose of each particle image, given the 3D structure, and an Mstep, regularized 3D density estimation given the latent poses.
Like many inverse problems with noisy, partial observations, the quality of cryoEM map reconstruction depends heavily on regularization. Regularization methods, widely used in computer science and statistics, leverage prior domain knowledge to penalize unnecessary model complexity and avoid overfitting. In cryoEM refinement, regularization is needed to mitigate the effects of imaging and sample noise so that protein signal alone is present in the inferred 3D density and so accumulated noise does not contaminate latent pose estimates.
Existing refinement algorithms use an explicit regularizer in the form of a shiftinvariant linear filter, typically obtained from Fourier shell correlation (FSC)^{6,10,13,14,15,16,17}. Such filters smooth the 3D structure using the same kernel, and hence the same degree of smoothing, at all locations. As FSC captures the average resolution of the map, such filters presumably under and overregularize different regions, allowing noise accumulation in some regions and a loss of resolvable detail in others. This effect should be pronounced with membrane proteins that have highly nonuniform rigidity and disorder across the molecule. As a motivating example, Fig. 1 shows a reconstruction of the TRPA1 membrane protein^{18} with a relatively low density threshold to help visualize regions with substantial noise levels, which indicate overfitting (e.g. the disordered micelle and the flexible tail at and bottom of the protein). We hypothesize that underfitting occurs in the core region where overregularization attenuates useful signal. As such, accumulated noise and attenuated signal will degrade pose estimates during refinement, limiting final structure quality. For inference problems of this type, the amount and form of regularization depends on regularization parameters. Correctly optimizing these parameters is often critical, but care must be taken to ensure that the optimization itself is not also prone to overfitting.
Results
We next outline the formulation of an adaptive form of regularization and with it, a new refinement algorithm called nonuniform refinement. We discuss its properties and demonstrate its application on several membrane protein datasets.
Adaptive crossvalidation regularization
With the aim of incorporating spatial nonuniformity into cryoEM reconstruction, we formulate a family of regularizers denoted r_{θ}, with parameters θ(x) that depend on spatial position x. Given a 3D density map m(x), the regularization operator, evaluated at x, is defined by
where h(x; ψ) is symmetric smoothing kernel, the spatial scale of which is determined by parameter ψ.
This family provides greater flexibility than shiftinvariant regularizers, but in exchange, requires making the correct choice of a new set of parameters, θ(x). We formulate the selection of the regularization parameters as an optimization subproblem during refinement, for which we adopt a twofold CV objective^{19,20}. The data are first randomly partitioned into two halves. On each of two trials, one part is considered as training data and the other is treated as heldout validation data. To find the regularizer parameters θ, one minimizes the sum of the pertrial validation errors (e), measuring consistency between the model and the validation data; that is,
where m_{1} and m_{2} are reconstructions from the two folds of the data. Similar objectives have been used for image denoising^{21}. We also introduce constraints on the parameters θ(x) to control degrees of freedom. The optimization problem is solved using a discretized search algorithm (Methods).
The resulting CV regularizer automatically identifies regions of a protein density with differing structural properties, optimally removing noise from each region. Fig. 2 illustrates the difference between uniform filtering (FSCbased) and the CVoptimal adaptive regularizer in nonuniform refinement. Shiftinvariant regularization smooths all regions equally, while the adaptive regularizer preferentially removes noise from disordered regions, retaining highresolution signal in wellstructured regions that is essential for 2D−3D alignment.
Nonuniform refinement algorithm
The nonuniform refinement algorithm takes as input a set of particle images and a lowresolution ab initio 3D map. The data are randomly partitioned into two halves, each of which is used to independently estimate a 3D halfmap. This ‘goldstandard’ refinement^{17} allows use of FSC for evaluating map quality and for comparison with existing algorithms. The key to nonuniform refinement is the adaptive CV regularization, applied at each iteration of 3D refinement. The regularizer parameters θ(x) are estimated independently for each halfmap (Methods), adhering to the ‘goldstandard’ assumptions. In contrast, in conventional uniform refinement the two halfmaps are not strictly independent as the regularization parameters, determined by FSC, are shared by both halfmaps. Finally, the optimization and application of the adaptive regularizer cause nonuniform refinement to be approximately two times slower than uniform refinement in cryoSPARC.
Validation with membrane protein datasets
We experimentally compared nonuniform refinement with conventional uniform refinement with three membrane protein datasets, all processed in cryoSPARC (see Supplementary Information for an additional membrane protein with no soluble region). Both algorithms were given the same ab initio structures. Except for the regularizer, all parameters and stages of data analysis, including the 2D−3D alignment and backprojection, were identical in uniform and nonuniform refinements. The default use of spatial solvent masking during 2D−3D alignment in cryoSPARC was used for both algorithms (Supplementary Information). No manual masks were used and no masking was used to identify or separate micelle or nanodisc regions. The same nonuniform refinement default parameters (other than symmetry) were used for all datasets.
When computing goldstandard FSC during the analysis of the reconstructed 3D density maps, for each dataset we used the same mask for uniform and nonuniform refinement. The masks were tested using phase randomization^{13} to avoid FSC bias. Also, the same Bfactor was used to sharpen both uniform and nonuniform refinement maps for each dataset. This consistency helped to ensure that visible differences in 3D structure were due solely to algorithmic differences. To color maps using local resolution, we used a straightforward implementation of Blocres^{22} for resolution estimation. No local filtering or local sharpening was used for visualization. Uniform and nonuniform refinement density maps were thresholded to contain the same total volume for visual comparison.
We also note that the FSCbased regularizer in conventional uniform refinement is equivalent to a shiftinvariant regularizer optimized with CV. That is, one can show that the optimal shiftinvariant filter under CV with squared error has a transfer function equivalent to the FSC curve. Thus, the experiments below also capture the differences between adaptive and shiftinvariant regularization.
STRA6CaM: CV regularization yields improved pose estimates and FSC
Zebrafish STRA6CaM^{23} is a 180kDa C2symmetric membrane protein complex comprising the 74kDa STRA6 protein bound to calmodulin (CaM). STRA6 mediates the uptake of retinol in various tissues. We processed a dataset of 28,848 particle images of STRA6CaM in a lipid nanodisc, courtesy of O. Clarke and F. Mancia (O. Clarke, personal communication). Nonuniform refinement provides a substantial improvement in nominal resolution from 4.0 Å to 3.6 Å (Fig. 3a), indicating an improvement in the average signaltonoise over the entire structure. However, different regions exhibited different resolution characteristics (Fig. 3c), as is often observed with protein reconstructions^{22}. There was substantial improvement in structural detail in most regions, while peripheral and flexible regions remained at low resolutions. Differences in structure quality, clearly visible in detailed views of αhelices (Fig. 3d), are especially important during atomic model building, where in many cases, protein backbone and side chains can only be traced with confidence in the nonuniform refinement map. Improvements in structure quality coincided with changes in particle alignments (Fig. 3b), approximately 3° on average. While disorder in the lipid nanodisc and nonrigidity of the CaM subunits are problematic for uniform refinement, adaptive regularization in nonuniform refinement reduces the influence of noise on alignments and produces improved map quality, especially in the periphery of the protein.
PfCRT: enabling atomic model building from previously unusable data
The Plasmodium falciparum chloroquine resistance transporter (PfCRT) is an asymmetric 48kDa membrane protein^{24}. Mutations in PfCRT are associated with the emergence of resistance to chloroquine and piperaquine as antimalarial treatments. We processed 16,905 particle images of PfCRT in lipid nanodisc with a Fab bound (EMPIAR10330 (ref. ^{24})). For PfCRT, the difference in resolution (Fig. 4a) and map quality (Fig. 4c) between uniform and nonuniform refinement is striking.
Using uniform refinement, reaching 6.9 Å, transmembrane αhelices are barely resolvable. Nonuniform refinement recovers signal up to 3.6 Å and provides a map from which an atomic model can be built with confidence. Transmembrane αhelices can be directly traced, including side chains. In contrast, the uniform refinement map does not show helical pitch and does not separate βstrands in the Fab domain. Indeed, an early version of the nonuniform refinement algorithm was essential for reconstructing the published highresolution density map and model^{24}.
On the spectrum of proteins studied with cryoEM, the PfCRTFab complex (100 kDa) is small. The lipid nanodisc (~50 kDa) also accounts for a large fraction of total particle molecular weight (see Fig. 2). We hypothesized that disorder in this relatively large nanodisc region leads to over and underregularization in uniform refinement. Most particle images exhibited orientation differences >6° between the two algorithms (Fig. 4b), suggesting that a large fraction of particle images were grossly misaligned by uniform refinement.
Na_{v}1.7 channel: improvement in highresolution features
Na_{v}1.7 is a voltagegated sodium channel found in the human nervous system^{25}. It plays a role in the generation and conduction of action potentials and is targeted by toxins and therapeutic drugs (e.g. for pain relief). We processed data of Na_{v}1.7 bound to two Fabs, forming a 245kDa C2symmetric complex solubilized in detergent^{25}. Following preprocessing (Methods), as described elsewhere^{25}, we detected both active and inactive conformational states of the channel. We obtained reconstructions with resolutions better than the published literature for both, but here we focus on 300,759 particle images of the active state.
Compared to the preceding datasets, the Na_{v}1.7 complex has a higher molecular weight and, in relative terms, a smaller detergent micelle. However, other regions are disordered or flexible, namely, a central fourhelix bundle, peripheral transmembrane domains and the Fabs. For Na_{v}1.7, uniform refinement reaches 3.4 Å resolution and nonuniform refinement reaches an improved 3.1 Å (Fig. 5a). This result is also an improvement over the published result of 3.6 Å (EMDB0341)^{25}, where the authors performed all processing in cisTEM^{10}.
With nonuniform refinement, map quality was clearly improved in central transmembrane regions, while some flexible parts of the structure (Fab domains and fourhelix bundle) remained at intermediate resolutions (Fig. 5c). In detailed views of αhelices (Fig. 5d) closer to the periphery of the protein, improvement in map quality and interpretability was readily apparent in the nonuniform refinement map, allowing modeling of sidechain rotamers with confidence. Central αhelices showed less improvement, but map quality remained equal or slightly improved, indicating that reconstructions of protein regions without disorder were not harmed by using nonuniform refinement.
Increased data efficiency
Examining map quality as a function of dataset size further helped to explore data efficiency. Figure 6 plots inverse resolution versus the number of particles. Such plots are useful as they relate to image noise and computational extraction of signal^{15}. Across all datasets, nonuniform refinement reached the same resolution as uniform refinement with fewer than half the number of particle images. It has also been argued that higher curves indicate more accurate pose estimates^{26}. Notably, the resolution gap between uniform and nonuniform refinement persists over a wide range of dataset sizes. While this resolution gap may decrease for much larger datasets than those considered here, the collection of more data alone may not allow uniform refinement to match the performance of nonuniform refinement for membrane protein targets until resolution is saturated as one nears fundamental limits.
Discussion
With adaptive regularization and effective optimization of regularization parameters using CV, nonuniform refinement achieves reconstructions of higher resolution and quality from singleparticle cryoEM data. It is particularly effective for membrane proteins, which exhibit varying levels of disorder, flexibility or occupancy. Here we focused on one specific family of adaptive regularizers (Methods), but it is possible within the same framework to explore other families that may be even more effective. For example, one could look to the extensive denoising literature for different regularizers.
Spatial variability of structure properties and the existence of over and underregularization in uniform refinement are well known. For instance, cryoEM methods for estimating local map resolution once 3D refinement is complete leverage statistical tests on the coefficients of a windowed Fourier transform^{22} or some form of wavelet transform^{27,28}. Although nonuniform refinement does not estimate resolution per se, the regularizer parameter θ(x) is related to a local frequency band limit. As such it might be viewed as a proxy for local resolution, but with important differences. Notably, θ(x) is optimized with a CV objective and it does not depend on a specific definition of ‘local resolution’ nor on an explicit resolution estimator.
Local resolution estimates^{22,29} or statistical tests^{30} have also been used to adaptively filter 3D maps, for visualization, to assess local map quality or to aid molecular model building. The family of filters and frequency cutoffs are typically selected to maximize visual quality. They are not optimized for the local resolution estimator nor is consideration given to the number of degrees of freedom in filter parameters or the strict independence of halfmaps, all critical for reliable regularization. Thus, while local resolution estimation followed by local filtering is useful for postprocessing, its use for regularization during iterative refinement (e.g. EMAN2.2 documentation^{9}), can yield overfitting (Methods). Map artifacts such as spikes or streaks are especially problematic for datasets with junk particles, structured outliers or small difficulttoalign particles. Nonuniform refinement couples an implicit resolution measure to the choice of regularizer, with optimization designed to control model capacity and avoid overfitting of regularization parameters (Methods).
Another related technique, used in cisTEM^{10} and Frealign^{7} and among the first to acknowledge and address under and overfitting, entails manual creation of a mask to label a local region one expects to be disordered (e.g. detergent micelle), followed by lowpass filtering in that region to a preset resolution at each refinement iteration. While this shares the motivation for nonuniform refinement, it relies on manual interaction during refinement, often necessitating a tedious trial and error process that can be difficult to replicate.
SIDESPLITTER^{31} is a recently proposed method to mitigate overfitting during refinement. Based on estimates of signaltonoise ratios from voxel statistics, it smooths local regions with low signaltonoise more aggressively than indicated by a global FSCbased resolution. The method does not directly address underfitting nor does it maintain independence of halfmaps or control for the number of degrees of freedom in the local filter model. More generally, the CV framework here is robust to different noise distributions and model misspecification, which can be problematic for methods with parametric noise models. Empirical results indicate that SIDESPLITTER mitigates overfitting and shows improvement in map quality relative to uniform refinement.
Bases other than the Fourier basis may also enable local control of reconstruction. Wavelet bases have been used for local resolution estimation, but less commonly for reconstruction^{32}. Kucukelbir et al.^{33} proposed the use of an adaptive wavelet basis and a sparsity prior. While similar in spirit to the goal of nonuniform refinement, this method has a single regularizer parameter for the entire 3D map, computed from noise in the corners of particle images. This may not capture variations in noise due to disorder, motion or partial occupancy.
Nonuniform refinement has been successful in helping to resolve several new structures. Examples include multiple conformations of an ABC exporter^{34}, mTORC1 docked on the lysosome^{35}, the respiratory syncytial virus polymerase complex^{36}, a GPCRG proteinβarrestin megacomplex^{37} and the SARSCoV2 spike protein^{38}.
Methods
Regularization in iterative refinement
In the standard cryoEM 3D reconstruction problem setup, the target 3D density map m is typically parameterized as a realspace 3D array with density at each voxel, in a Cartesian grid of box size N, and a corresponding discrete Fourier representation, \(\hat{m}=Fm\). The goal of reconstruction is to infer the 3D densities of the voxel array, called model parameters. Representing the 3D density, its 2D projections, the observed images and the noise model in the Fourier domain is common practice for computational efficiency, exploiting the wellknown convolution and Fourierslice theorems^{6,7,8,9}. The unobserved pose variables for each image are latent variables.
Algorithm 1
Iterative refinement (expectationmaximization)
Require: Particle image dataset \({\mathcal{D}}\) and ab initio 3D map  
1:  Use smoothed ab initio 3D map array as the initial model parameters m^{(0)} 
2:  while not converged do 
3:  Estep: Given current estimate of model parameters, m^{(t−1)}, from step t − 1, estimate (via marginalization or maximization) the latent variables: \({z}^{(t)}\leftarrow f({m}^{(t1)},{\mathcal{D}})\) 
4:  Mstep: Given the latent variables z^{(t)}, compute raw estimates of the model parameter \({\tilde{m}}^{(t)}\) (without regularization): \({\tilde{m}}^{(t)}\leftarrow h({z}^{(t)},{\mathcal{D}})\) 
5:  Regularize: Given noisy model parameters \({\tilde{m}}^{(t)}\), apply the regularization operator, r_{θ}, with regularization parameters θ: \({m}^{(t)}\leftarrow {r}_{\theta }({\tilde{m}}^{(t)})\) 
6:  end while 
Iterative refinement methods (Algorithm 1), which provide stateoftheart results^{6,9,10,11}, can be interpreted as variants of blockcoordinate descent or the expectationmaximization algorithm^{12}. In cryoEM, and more generally in inverse problems with noisy, partial observations, a critical component that modulates the quality of the results is regularization.
One can regularize problems explicitly, using a prior distribution over model parameters or implicitly, by applying a regularization operator to the model parameters during optimization. Iterative refinement methods tend to use implicit regularizers, attenuating noise in the reconstructed map at each iteration. In either case, the separation of signal from noise is the crux of many inference problems.
In the cryoEM refinement problem, like many latent variable inverse problems, there is an additional interplay between regularization, noise buildup and the estimation of latent variables. Retained noise due to underregularization will contaminate the estimation of latent variables. This contamination is propagated to subsequent iterations and causes overfitting.
This paper reconsiders the task of regularization based on the observation that common iterative refinement algorithms often systematically underfit and overfit different regions of a 3D structure simultaneously. This causes a loss of resolvable detail in some parts of a structure and the accumulation of noise in others. The reason stems from the use of frequencyspace filtering as a form of regularization. Some programs, such as cisTEM^{10}, use a strict resolution cutoff, beyond which Fourier amplitudes are set to zero before alignment of particle images to the current 3D structure. In RELION^{16}, regularization was initially formulated with an explicit Gaussian prior on Fourier amplitudes of the 3D structure, with a handtuned parameter that controls Fourier amplitude shrinkage. Later versions of RELION^{17} and cryoSPARC’s homogeneous (uniform) refinement^{6} use a transfer function (or Wiener filter) determined by FSC computed between two halfmaps^{13,14,15}.
Such methods presume a Fourier basis and shiftinvariance. Although well suited to stationary processes, they are less well suited to protein structures, which are spatially compact and exhibit nonuniform disorder (e.g. due to motion or variations in signal resolution). FSC, for instance, provides an aggregate measure of resolution. To the extent that FSC under or overestimates resolution in different regions, FSCbased shiftinvariant filtering will oversmooth some regions, attenuating useful signal and undersmooth others, leaving unwanted noise. To address these issues we introduce a family of adaptive regularizers that can, in many cases, find better 3D structures with improved estimates of the latent poses during refinement.
Crossvalidation regularization
We formulate a new regularizer for cryoEM reconstruction in terms of the minimization of a CV objective^{19,20}. CV is a general principle that is widely used in machine learning and statistics for model selection and parameter estimation with complex models. In CV, observed data are randomly partitioned into a training set and a heldout validation set. Model parameters are inferred using the training data, the quality of which is then assessed by measuring an error function applied to the validation data. In kfold CV, the observations are partitioned into k parts. In each of k trials, one part is selected as the heldout validation set and the remaining k − 1 parts comprise the training set. The pertrial validation errors are summed, providing the total CV error. This procedure measures agreement between the optimized model and the observations without bias due to overfitting. Rather, overfitting during training is detected directly as an increase in the validation error. Notably, formulating regularization in a CV setting provides a principled way to design regularization operators that are more complex than the conventional, isotropic frequencyspace filters. The CV framework is not restricted to a Fourier basis. One may consider more complex parameterizations, the use of metaparameters and incorporate cryoEM domain knowledge.
Given a family of regularizers r_{θ} with parameters θ, the minimization of CV error to find θ is often applied as an outer loop. This requires the optimization of model parameters m to be repeated many times with different values of θ, a prohibitively expensive cost for problems like cryoEM. Instead, one can also perform CV optimization as an inner loop, while optimization of model parameters occurs in the outer loop. Regularizer parameters θ are then effectively optimized onthefly, preventing under or overfitting without requiring multiple 3D refinements to be completed.
To that end, consider the use of twofold CV optimization to select the regularization operator, denoted r_{θ}(m) in the regularization step in Algorithm 1 (note that k > 2 is also possible). The dataset \({\mathcal{D}}\) is partitioned into two halves, \({{\mathcal{D}}}_{1}\) and \({{\mathcal{D}}}_{2}\) and two (unregularized) refinements are computed, namely m_{1} and m_{2}. For each, one half of the data is the ‘training set’ and the other is held out for validation. To find the regularizer parameters θ we wish to minimize the total CV error E, that is,
where e is the negative log likelihood of the validation halfset given the regularized halfmap. The second line simplifies this expression by using the raw reconstruction from the opposite halfset as a proxy for the actual observed images. Note that assumptions for ‘goldstandard’ refinement^{17} are not broken in this procedure. With the L2 norm, equation (3) reduces to a sum of pervoxel squared errors, corresponding to white Gaussian noise between the halfset reconstructions. When the choice of θ causes r_{θ} to remove too little noise from the raw reconstruction, the residual error E will be unnecessarily large. If θ causes r_{θ} to overregularize, removing too much structure from the raw reconstruction, then E increases as the structure retained by r_{θ} no longer cancels corresponding structure in the opposite halfmap. As such, minimizing E(θ) provides the regularizer that optimally separates signal from noise.
We note that similar objectives have been used for general image denoising^{21} and adapted for cryoEM images^{39,40,41}; however, in these methods the aim was to learn a general neural network denoiser, whereas the goal here is to optimize regularization parameters on a single data sample. It is also worth noting that this formulation can be extended in a straightforward way to compare each halfset reconstruction against images directly (dealing appropriately with the latent pose variables) or to use error functions corresponding to different noise models.
Regularization parameter optimization
The CV formulation in equation (3) provides great flexibility in choosing the family of regularizers r_{θ}, taking domain knowledge into account. For nonuniform refinement, we wish to accommodate protein structures with spatial variations in disorder, motion and resolution. Accordingly, we define the regularizer to be a spacevarying linear filter. The filter’s spatial extent is determined by the regularization parameter θ(x), which varies with spatial position. Here, we write the regularizer in operator form:
where h(x; ψ) is symmetric smoothing kernel, the spatial scale of which is specified by ψ. In practice we let h(x; ψ) be an eighthorder Butterworth kernel. The eighthorder kernel provides a middle ground between the relatively poor frequency resolution of the Gaussian kernel and the sharp cutoff of the sinc kernel, which suffers from spatial ringing. We have experimented with different orders of Butterworth filters and found that an eighthorder kernel performs well on a broad range of particles.
When equation (4) is combined with the CV objective for the estimation of θ(x), one obtains
With one regularization parameter at each voxel, that is, θ(x), this reduces to a large set of decoupled optimization problems, one for each voxel. That is, for voxel x one obtains
With this decoupling, θ(x) can transition quickly from voxel to voxel, yielding high spatial resolution. On the other hand, the individual subproblems in equation (6) are not well constrained as each parameter is estimated from data at a single location, so the parameter estimates are not useful. In essence, our regularizer design has two competing goals, namely, reliable signal detection and high spatial resolution (respecting boundaries between regions with different properties). Signal detection improves through aggregation of observations (e.g. neighboring voxels), while high spatial resolution prefers minimal aggregation (equation (6)).
To improve signal detection, we further constrain θ^{*} to be smooth. That is, although in some regions θ should change quickly (solvent–protein boundaries), in most regions we expect it to change slowly (in solvent and regions of rigid protein mass). Smoothness effectively limits the number of degrees of freedom in θ, which is important to ensure that θ itself does not overfit during iterative refinement. One can encourage smoothness in θ by explicitly penalizing spatial derivatives of θ in the objective (equation (5)), but this yields a Markov random field problem that is hard to optimize. Alternatively, one can express θ in a lowdimensional basis (e.g. radial basis functions), but this requires prior knowledge of the expected degree of smoothness. Instead, we adopt a simple but effective approach. Assuming that θ is smoothly varying, we treat measurements in the local neighborhood of x as additional constraints on θ(x). A window function can be used to give more weight to points close to x. We thereby obtain the following leastsquares objective:
where w_{ρ}(x) is positive and integrates to 1, with spatial extent ρ. This allows one to estimate θ at each voxel independently, while the overlapping neighborhoods ensure that θ(x) varies smoothly.
This approach also provides a natural way to allow for variable neighborhood sizes, where ρ(x) depends on location x, so both rigid regions and transition regions are well modeled. Notably, we want ρ(x) to be large enough to reliably estimate the local power of the CV residual error to estimate θ(x) correctly, but small enough to enable rapid local transitions. A reasonable balance can be specified in terms of the highest frequency with substantial signal power, which is captured by the regularization parameter θ(x) itself. In particular, for regularizers that are close to optimal, we expect the residual signal to have its power concentrated at wavelengths near θ(x). In this case a good measure of the local residual power is to aggregate the squared residual over a small number of wavelengths θ(x)^{42}. Thus we can reliably estimate both θ(x) and ρ(x) as long as ρ(x) is constrained to be a small multiple of θ(x).
We therefore adopt a simple heuristic, that ρ(x) > γ θ(x) where γ, the adaptive window factor (AWF), is a constant. With this constraint we obtain the final form of the computational problem solved in nonuniform refinement to regularize 3D electron density at each iteration; that is,
We find that as long as γ > 3 we obtain reliable estimates of the local power of the residual signal. For γ < 2, estimates of residual power are noisy and optimization of the regularization parameters therefore suffers. The algorithm is relatively insensitive to values of γ > 3, but there is some loss in the spatial resolution of the adaptive regularizer as γ increases.
Nonuniform refinement algorithm
Algorithm 2
Regularization step for nonuniform refinement
Require Particle image dataset \({\mathcal{D}}\) with pose estimates z  
1:  Randomly partition \({\mathcal{D}}\) into halves, \({{\mathcal{D}}}_{1}\) and \({{\mathcal{D}}}_{2}\) with corresponding poses z_{1} and z_{2} 
2:  Reconstruct \({\tilde{m}}_{1}\) and \({\tilde{m}}_{2}\), the raw (noisy) 3D maps from each halfset 
3:  Estimate regularization parameters θ^{*} by solving equation (8) 
4:  Reconstruct a single map from \({\mathcal{D}},\ z\) and apply the optimal regularizer \({r}_{{\theta }^{* }}\) 
Given a set of particle images and a lowresolution ab initio 3D map, nonuniform refinement comprises three main steps, similar to conventional uniform (homogeneous) refinement (Algorithm 1). The data are randomly partitioned into two halves, each of which is used to independently estimate a 3D halfmap. This ‘goldstandard’ refinement^{17} allows use of FSC for evaluating map quality, and for comparison with existing algorithms. The alignment of particle images against their respective halfmaps, and the reconstruction of the raw 3D density map (the E and M steps in Algorithm 1) are also identical to uniform refinement.
The difference between uniform and nonuniform refinement is in the regularization step. First, in nonuniform refinement, regularization is performed independently in the two halfmaps. As such, the estimation of the spatial regularization parameters in Algorithm 2 effectively partitions each halfdataset into quarterdatasets. We often refer to the raw reconstructions in Algorithm 2 as quartermaps. The nonuniform refinements on halfmaps are therefore entirely independent, satisfying the assumptions of a ‘goldstandard’ refinement^{17}. In contrast, conventional uniform refinement uses FSC between halfmaps to determine regularization parameters at each iteration, thereby sharing masks and regularization parameters, both of which contaminate final FSCbased assessment because the two halfmaps are no longer reconstructed independently.
Most importantly, nonuniform refinement uses equation (8) to define the optimal parameters with which to regularize each halfset reconstruction at each refinement iteration. Figure 2 shows an example of the difference between uniform filtering (FSCbased) and the new CVoptimal regularizer used in nonuniform refinement. Uniform regularization removes signal and noise from all parts of the 3D map equally. Adaptive regularization, on the other hand, removes more noise from disordered regions, while retaining the highresolution signal in wellstructured regions that is critical for aligning 2D particle images in the next iteration.
In practice, for regularization parameter estimation, equation (8) is solved on a discretized parameter space where a relatively simple discrete search method can be used (e.g. as opposed to continuous gradientbased optimization). The algorithm is implemented in Python within the cryoSPARC software platform^{6}, with most of the computation implemented on graphics processing unit accelerators. An efficient solution to equation (8) is important in practice because this subproblem is solved twice for each iteration of a nonuniform refinement.
Finally, the tuning parameters for adaptive regularization are interpretable and relatively few in number. They include the order of the Butterworth kernel, the discretization of the parameter space and the scalar relating ρ(x) and θ(x), called the AWF. In all experiments, we use an eighthorder Butterworth filter and a fixed AWF parameter γ = 3. We discretize the regularization parameters into 50 possible values, equispaced in the Fourier domain to provide greater sensitivity to smallscale changes at finer resolutions. We find that nonuniform refinement is approximately two times slower than uniform refinement in our current implementation.
Overfitting of regularizer parameters
As mentioned in the Discussion, local resolution estimates or local statistical tests have commonly been used to adaptively filter 3D maps. While these methods are generally satisfactory as onetime postprocessing steps for visualization, in our experience they can lead to severe overfitting when used iteratively within refinement as a substitute for regularization. (Supplementary Fig. 1 illustrates one example with the Na_{v}1.7 dataset.) During iterative refinement, small misestimations of local resolution at a few locations (due to high estimator variance^{22}) cause subtle over or underfitting, leaving slight density variations. Over multiple iterations of refinement, these errors can produce strong erroneous density that contaminate particle alignments and the local estimation of resolution itself, creating a vicious cycle. A related technique using iterative local resolution and filtering was described briefly in EMAN2.2 documentation^{9} and may suffer the same problem. The resulting artifacts (e.g. streaking and spikey density radiating from the structure) are particularly prevalent in datasets with junk particles, structured outliers or small particles that are already difficult to align. To mitigate these problems, the approach we advocate couples an implicit resolution measure to a particular choice of local regularizer, with optimization explicitly designed to control model capacity and avoid overfitting of regularizer parameters.
Data preprocessing
Experimental results for STRA6CaM and PfCRT datasets were computed directly from particle image stacks, with no further preprocessing. The original data for the Na_{v}1.7 protein comprise 25,084 raw microscope movies (EMPIAR10261) from a Gatan K2 Summit direct electron detector in counting mode, with a pixel size of 0.849 Å. We processed the dataset through motion correction, CTF estimation, particle picking, 2D classification, ab initio reconstruction and heterogeneous refinement in cryoSPARC v.2.11 (ref. ^{6}). A total of 738,436 particles were extracted and then curated using 2D classification yielding 431,741 particle images (pixel size 1.21 Å, box size 360 pixels). As described elsewhere^{25}, we detected two discrete conformations corresponding to the active and inactive states of the channel, with 300,759 and 130,982 particles. We obtained reconstructions with resolutions better than the published literature for both states, but for the results in this work we focus solely on the active state (refined with C2 symmetry yielding 601,518 asymmetric units).
Reporting Summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
Data availability
No new datasets were created in this study. The raw datasets analyzed in this study were either downloaded from the EMPIAR repository (EMPIAR10024, EMPIAR10330, EMPIAR10261) or were provided by authors of other studies, cited in the main text. Density maps and atomic coordinates from EMDB0341 and Protein Data Bank 6N4Q were used for visualization. All other results and outputs of data analysis in this study are available from the corresponding author on reasonable request.
Code availability
The cryoSPARC software package is freely available in executable form for nonprofit academic use at www.cryosparc.com.
References
Cheng, Y. Singleparticle CryoEM at crystallographic resolution. Cell 161, 450–457 (2015).
Cheng, Y. Membrane protein structural biology in the era of single particle cryoEM. Curr. Opin. Struct. Biol. 52, 58–63 (2018).
Overington, J. P., AlLazikani, B. & Hopkins, A. L. How many drug targets are there? Nat. Rev. Drug Discov. 5, 993–996 (2006).
Renaud, J.P. et al. CryoEM in drug discovery: achievements, limitations and prospects. Nat. Rev. Drug Discov. 17, 471–492 (2018).
Scapin, G., Potter, C. S. & Carragher, B. CryoEM for small molecules discovery, design, understanding, and application. Cell Chem. Biol. 25, 1318–1325 (2018).
Punjani, A., Rubinstein, J. L., Fleet, D. J. & Brubaker, M. A. CryoSPARC: Algorithms for rapid unsupervised cryoem structure determination. Nat. Methods 14, 290–296 (2017).
Grigorieff, N. FREEALIGN: High resolution refinement of single particle structures. J. Struct. Biol. 157, 117–125 (2007).
Scheres, S. H. W. & Bayesian, A. view on cryoem structure determination. J. Mol. Biol. 415, 406–418 (2012).
Bell, J. M., Chen, M., Baldwin, P. R. & Ludtke, S. J. High resolution single particle refinement in EMAN2.1. Methods 100, 25–34 (2016).
Grant, T., Rohou, A. & Grigorieff, N. cisTEM, userfriendly software for singleparticle image processing. eLife 7, e35383 (2018).
Zivanov, J. et al. New tools for automated highresolution cryoem structure determination in RELION3. eLife 7, e42166 (2018).
Dempster, A. P., Laird, N. M. & Rubin, D. B. Maximum likelihood from incomplete data via the EM algorithm. J. R. Stat. Soc. Ser. B 39, 1–38 (1977).
Chen, S. et al. Highresolution noise substitution to measure overfitting and validate resolution in 3D structure determination by single particle electron cryomicroscopy. Ultramicroscopy 135, 24–35 (2013).
Harauz, G. & van Heel, M. Exact filters for general geometry three dimensional reconstruction. Optik 73, 146–156 (1986).
Rosenthal, P. B. & Henderson, R. Optimal determination of particle orientation, absolute hand, and contrast loss in singleparticle electron cryomicroscopy. J. Mol. Biol. 333, 721–745 (2003).
Scheres, S. H. W. RELION: implementation of a Bayesian approach to cryoEM structure determination. J. Struct. Biol. 180, 519–530 (2012).
Scheres, S. H. W. & Chen, S. Prevention of overfitting in cryoEM structure determination. Nat. methods 9, 853–854 (2012).
Paulsen, C. E., Armache, J. P., Gao, Y., Cheng, Y. & Julius, D. Structure of the TRPA1 ion channel suggests regulatory mechanisms. Nature 520, 511–517 (2015).
Golub, G. M., Heath, M. & Wahba, G. Generalized crossvalidation as as a method for choosing a good ridge parameter. Technometrics 21, 215–223 (1979).
Wahba, G. A comparison of GCV and GML for choosing the smoothing parameter in the generalized spline smoothing problem. Ann. Stat. 13, 1378–1402 (1985).
Lehtinen, J. et al. Noise2Noise: learning image restoration without clean data. Proc. Mach. Learn. Res. 7, 4620–4631 (2018).
Cardone, G., Heymann, J. B. & Steven, A. C. One number does not fit all: mapping local variations in resolution in cryoem reconstructions. J. Struct. Biol. 184, 226–236 (2013).
Chen, Y. et al. Structure of the stra6 receptor for retinol uptake. Science 353, aad8266 (2016).
Kim, J. et al. Structure and drug resistance of the Plasmodium falciparum transporter PfCRT. Nature https://doi.org/10.1038/s415860191795x (2019).
Xu, H. et al. Structural basis of Nav1.7 inhibition by a gatingmodifier spider toxin. Cell https://doi.org/10.1016/j.cell.2018.12.018 (2019).
Stagg, S., Noble, A., Spilman, M. & Chapman, M. Reslog plots as an empirical metric of the quality of cryoEM reconstructions. J. Struct. Biol. 185, 418–426 (2014).
Kucukelbir, A., Sigworth, F. J. & Tagare, H. D. Quantifying the local resolution of cryoEM density maps. Nat. Methods 11, 63–65 (2014).
Vilas, J. L. et al. Monores: automatic and accurate estimation of local resolution for electron microscopy maps. Structure 26, 337–344 (2018).
Felsberg, M. & Sommer, G. The monogenic signal. IEEE Trans. Signal Process. 49, 3136–3144 (2001).
Ramlaul, K., Palmer, C. M. & Aylett, C. H. S. A local agreement filtering algorithm for transmission EM reconstructions. J. Struct. Biol. 205, 30–40 (2019).
Ramlaul, K., Palmer, C. M., Nakane, T. & Aylett, C. H. S. Mitigating local overfitting during single particle reconstruction with SIDESPLITTER. J. Struct. Biol. 211, 107545 (2020).
Vonesch, C., Wang, L., Shkolnisky, Y. and Singer, A. Fast waveletbased single particle reconstruction in cryoEM. Proc. IEEE Int. Symp. Biomed. Imaging https://doi.org/10.1109/ISBI.2011.5872791 (2011).
Kucukelbir, A., Sigworth, F. J. & Tagare, H. D. A Bayesian adaptive basis algorithm for single particle reconstruction. J. Struct. Biol. 179, 56–67 (2012).
Hofmann, S. et al. Conformation space of a heterodimeric ABC exporter under turnover conditions. Nature 571, 580–583 (2019).
Rogala, K. B. et al. Structural basis for the docking of mTORC1 on the lysosomal surface. Science 366, 468–475 (2019).
Gilman, M. S. A. et al. Structure of the respiratory syncytial virus polymerase complex. Cell 179, 193–204 (2019).
Nguyen, A. H. et al. Structure of an endosomal signaling GPCRG proteinβarrestin megacomplex. Nat. Struct. Mol. Biol. 26, 1123–1131 (2019).
Wrapp, D. et al. CryoEM structure of the 2019nCoV spike in the prefusion conformation. Science 367, 1260–1263 (2020).
Bepler, T., Noble, A. & Berger, B. TopazDenoise: general deep denoising models for cryoEM. Nat. Commun. 11, 5208 (2020).
Tegunov, D. & Cramer, P. Realtime cryoelectron microscopy data preprocessing with warp. Nat. Methods 16, 1146–1152 (2019).
Buchholz, T.O., Jordan, M., Pigino, G. & Jug, F. cryoCARE: contentaware image restoration for cryotransmission electron microscopy data. Preprint at arXiv https://arxiv.org/abs/1810.05420 (2018).
Potamianos, A. & Maragos, P. A comparison of the energy operator and the Hilbert transform approach to signal and speech demodulation. Signal Process. 37, 95–120 (1994).
Acknowledgements
We are extraordinarily grateful to O. Clarke, F. Mancia and Y. Zi Tan for providing valuable cryoEM data early in this project and for sharing their experience as early adopters of nonuniform refinement and cryoSPARC. We thank the entire team at Structura Biotechnology Inc., which designs, develops and maintains the cryoSPARC software system in which this project was implemented and tested. We thank J. Rubinstein for comments on this manuscript. Resources used in this research were provided, in part, by the Province of Ontario, the Government of Canada through NSERC (Discovery Grant RGPIN 201505630 to D.J.F.) and CIFAR and companies sponsoring the Vector Institute.
Author information
Authors and Affiliations
Contributions
A.P. and D.J.F. designed the algorithm, A.P. and H.Z. implemented the software and performed experimental work. A.P. and D.J.F. contributed expertise and supervision. D.J.F. and A.P. contributed to manuscript preparation.
Corresponding authors
Ethics declarations
Competing interests
A.P. is CEO of Stuctura Biotechnology, which builds the cryoSPARC software package. D.J.F. is an advisor to Stuctura Biotechnology. Novel aspects of the method presented are described in a patent application (WO2019068201A1), with more details available at https://cryosparc.com/patentfaqs.
Additional information
Peer review information Allison Doerr was the primary editor on this article and managed its editorial process and peer review in collaboration with the rest of the editorial team.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Supplementary Information
Supplementary Figs. 1–4 and Supplementary Notes.
Supplementary Video 1
Density maps comparing uniform and nonuniform refinement. The video shows 3D density maps and detail views from three different membrane protein datasets. For each dataset, the entire structure is shown using uniform (conventional) and nonuniform refinement on the same data. Density maps are sharpened with the same Bfactor and have thresholds set to enclose the same volume. Detail views show the improvement of reconstructed map quality for individual helical segments within each protein.
Rights and permissions
About this article
Cite this article
Punjani, A., Zhang, H. & Fleet, D.J. Nonuniform refinement: adaptive regularization improves singleparticle cryoEM reconstruction. Nat Methods 17, 1214–1221 (2020). https://doi.org/10.1038/s41592020009908
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1038/s41592020009908
This article is cited by

RNA targeting and cleavage by the type IIIDv CRISPR effector complex
Nature Communications (2024)

Stepwise activation of a metabotropic glutamate receptor
Nature (2024)

From structural polymorphism to structural metamorphosis of the coat protein of flexuous filamentous potato virus Y
Communications Chemistry (2024)

Architecture of the RNF1 complex that drives biological nitrogen fixation
Nature Chemical Biology (2024)

Structural insights into IL11mediated signalling and human IL6ST variantassociated immunodeficiency
Nature Communications (2024)