CryoEM-based Hybrid Modeling Approaches for Structure Determination

Abstract

Recent advances in cryo-electron microscopy (cryoEM) have dramatically improved the resolutions at which vitrified biological specimens can be studied, revealing new structural and mechanistic insights over a broad range of spatial scales. Bolstered by these advances, much effort has been directed towards the development of hybrid modeling methodologies for the construction and refinement of high-fidelity atomistic models from cryoEM data. In this brief review, we will survey the key elements of cryoEM-based hybrid modeling, providing an overview of available computational tools and strategies as well as several recent applications.

Introduction

The application of cryoEM to biomolecular structure determination has become a widespread and lucrative approach. Enhanced by new cameras that directly detect electrons [1], improved microscopes with more stable optics [2], and advances in image processing software [3], cryoEM is now used to determine high-resolution macromolecular structures of broad biological significance in their native and functional contexts [4••]. At the same time, the applicability of cryoEM to biological systems spanning a wide range of length scales has helped to give rise to a more cohesive picture of microbiological processes, one that now ranges from the single molecule all the way to the whole cell [5].

CryoEM may be used in several modalities depending on the nature of the specimen under investigation; including its stability, size, and conformational flexibility. At low resolution (20–100 Å), three-dimensional (3D) images called tomograms may be reconstructed using cryo-electron tomography (cryoET) to reveal structures of pleomorphic objects, including small bacterial cells, isolated viruses, and elements of cellular ultrastructure such as mitochondria and other organelles. In cases where multiple instances of a molecule are present, sub-tomograms may be computationally extracted from these reconstructions for further alignment and classification (cryoSTAC) [6]. Upon averaging, these aligned sub-populations lead to intermediate resolution (5–20 Å) structures of biological assemblies such as virus capsids and microtubules as well as molecular motors and other macromolecular machines. Particularly exciting are biomolecular structures solved to high resolution (2–5 Å) via single-particle-analysis (SPA) [7•] or electron crystallography [8], which now routinely rival the resolutions obtained through X-ray crystallography (XRC) and NMR spectroscopy [9•]. Moreover, SPA and cryoSTAC, in addition to their increased resolution capabilities, greatly relieve constraints on specimen size and heterogeneity imposed by more traditional high-resolution techniques that rely on ensemble-based measurements, permitting the study of many previously intractable systems, including membrane-bound proteins and large, dynamic biomolecular complexes [10].

Accordingly, hybrid modeling methodologies, which seek to extend the chemical interpretability of cryoEM data through the derivation of high-fidelity atomistic models, have become an area of fervent development [11,12]. How suitable a particular hybrid modeling protocol is for solving a given structure, depends largely on the quality and resolution of the cryoEM data used as input, assumed here to take the form of a cryoEM density map: a 3D image in which the Coulomb potential distribution of the specimen is represented by linearly proportional gray-scale intensities. The Electron Microscopy Data Bank (EMDB) [13•], created to publicly archive cryoEM maps, now boasts over 5000 entries, with new maps being deposited at an ever-increasing rate. In the following sections, we will survey computational strategies for constructing, refining, and validating atomic models based on cryoEM maps, highlighting specific tools and procedures for both the intermediate-resolution and high-resolution regimes (summarized in Figure 1 and Table 1). It should be noted at the outset of our discussion that the division of hybrid modeling tools into the specific resolution ranges presented here is, of course, not strictly inherent to the tools themselves, but rather should be thought of as a fuzzy demarcation centered on resolutions where the tool may be applied with greatest confidence.

Figure 1.

General workflow for cryoEM-based hybrid modeling as a function of map resolution. An initial atomic model may be obtained, for instance, from an existing, rigidly-docked structure or through the use of de novo modeling techniques. Depending on the resolution of the map, model refinement strategies range from broad conformational refinement using flexible-fitting methodologies to more detailed secondary structure and side chain refinement using, for example, Rosetta-based approaches or XRC software suites. The accuracy of a refined model may be subsequently assessed by way of stereochemical or goodness-of-fit measures as well as cross-validation procedures to identify over-fitting.

Table 1.

List of commonly used methods and tool suites for various hybrid modeling tasks.

	Task	Method/Tool Suite**
Input	Comparative Modeling	Modeller [15], RosettaCM [16]
	Ab initio Structure Prediction	MELD [22], I-TASSER [19]
	Model Selection/Scoring	MolProbity [26], DOPE [21]
	Map Segmentation	Chimera [37], Segger [38]
EM-based Modeling	Rigid Fitting	Chimera [37], Situs [31], HADDOCK [32]
	Flexible Fitting	MDFF [52], Flex-EM [49], DireX [40]
	De novo Modeling (<3.5 Å)	Coot [77], Phenix [78], REFMAC [81]
	De novo Modeling (3.5-5 Å)	Gorgon [88], Rosetta de novo [95, 96], RosettaES [97]
	Structural Refinement (<5 Å)	Rosetta de novo [95, 96], RE-MDFF [69]
Validation	Stereochemical Assessment	MolProbity [26], EMRinger [109]
	Goodness-of-fit Measure	CCC [108], iFSC [96], ResMap [112]
	Feature Stability	All-atom MD [123]
	Over-fitting Detection	Cross-validation [117,118]

^**The authors emphasize that the cited examples are based on their personal user preference and represent but a fraction of tools available for any given task.

Hybrid Modeling at Intermediate Resolution

At intermediate resolutions (5-20 Å), which are now routinely reached by both SPA and cryoSTAC, hybrid modeling strategies range from the rough placement of rigid models within a cryoEM map to their all-atom, density-driven conformational refinement. As the information present at these resolutions is, in general, not sufficient to unambiguously discern a structure’s full topology or, in lower resolution cases, even its tertiary organization, intermediate-resolution approaches typically require the user to provide a complete, atomistic model as input. Molecular coordinates may be taken directly from one or more of the 100,000-plus structures in the Protein Data Bank (PDB) [14] or derived from the solved structure of a suitably close homologue using comparative modeling suites [15,16] or web-based homology tools [17–20]. A particularly comprehensive suite of tools for comparative modeling, Modeller [15] takes as input a sequence alignment and automatically constructs, taking into account multiple structural templates as well as user-provided structural restraints, an ensemble of complete atomic models ranked according to DOPE score [21]. For proteins with less than 150 residues, a model may be constructed using ab initio structure prediction [22–24], although it is currently difficult to ensure model reliability [25]. When possible, the prudent modeler will compare multiple preliminary models constructed using several different tools. For the purpose of model selection, a number of scoring functions are available to quantify the stereochemical quality and energetics of a given model [26–30], the most popular of which, MolProbity [26], is discussed in more detail below.

Whether the goal is to assemble known components into an unknown complex (i.e., quaternary assembly) or to derive a model of a new functional state, the initial model(s) must be fit to the map to facilitate direct comparison. Fitting strategies fall into two classes: rigid fitting (commonly called “docking”) and flexible fitting. Rigid-fitting algorithms search for the optimal positioning of one or more high-resolution models within a map without allowing changes in model conformation. Many programs exist for single and multi-body rigid fitting, which utilize different searching and scoring strategies to increase efficiency or accuracy at different map resolutions [31–36]. To reduce the complexity of the fitting problem, cryoEM maps are often segmented into separate densities corresponding to individual biomolecules or molecular domains. UCSF-Chimera [37] a feature-rich suite for the visualization and general manipulation of cryoEM data, enables both rapid rigid fitting through its “Fit in Map” routine as well as high-precision segmentation using the Segger-based [38,39] “Segment Map” routine.

While useful for the analysis of low resolution maps or construction of preliminary models, rigid fitting is often not sufficient for obtaining the optimal fit to a map owing to the inherently dynamic nature of many biomolecules and their complexes. To improve model-map overlap, flexible-fitting methods allow a rigidly docked model to be plastically deformed while (ideally) maintaining proper stereochemistry. Several methods have been developed that utilize reduced representations of biomolecular mobility or structure to assess flexibility and increase fitting efficiency, including those based on elastic network models [40], normal modes analysis [41–43], constrained geometries [44], and Bayesian statistics [45]. More accurate, albeit computational more expensive, methods employ detailed force fields combined with Monte Carlo (MC) [46,47] or molecular dynamics (MD) [48,49•,50,51] simulation. Of these, Molecular Dynamics Flexible Fitting (MDFF) [48,52] has become one of the most widely used, assisting in the structural determination of the HIV-1 virus capsid [53,54], Ribosome [55–57], and many others [58–61,62••,63••].

MDFF approaches the flexible fitting problem through the addition of a density-derived term to the MD potential energy function, thereby generating forces that explicitly drive the conformations of the substrate biomolecules towards regions of high density within the map while taking into account all-atom electrostatic and solvent effects [52]. Furthermore, MDFF’s tight integration with the molecular visualization program VMD [64] and high-performance MD package NAMD [65] provides increased functionality, including tools for setting up MDFF simulations [52,66] and efficiently analyzing quality-of-fit [67] as well as combining structural refinement with enhanced sampling methods [68,69••] or additional restraints based on symmetry [70] or complimentary experimental data [71•,72]. Importantly, NAMD’s scalability to very large system sizes, having already enabled the application of MDFF for the determination of structures tens-of-millions of atoms in size [11,53], provides a promising platform for the future development of methods to tackle even larger sub-cellular complexes [73–75]. Figure 2 highlights a recent application in which Cassidy et al. used MDFF and other hybrid modeling tools to derive an atomic model of the bacterial chemosensory array 20 million atoms in size based on an 11.7 Å map determined via cryoSTAC [63••]. All-atom MD simulations with NAMD were subsequently performed to further refine a portion of the model and study its dynamical behavior, allowing for the identification of novel interfaces and tertiary motions in a signaling protein ubiquitous throughout bacterial motility.

Figure 2.

Quaternary structure of bacterial chemosensory array determined by hybrid modeling and intermediate-resolution cryoSTAC. Combing high resolution, XRC-derived component structures (A) with near-sub-nanometer resolution maps of chemosensory array (B) Cassidy et al. used an assortment of hybrid modeling tools (grey box) to construct an atomistic model of the extended signaling array 20-million-atoms in size (C). Distinct symmetry centers used for sub-tomogram classification are circled in black and grey in (B) along with their corresponding atomistic regions in (C). Further investigation of a portion of the model using all-atom MD simulations confirmed its stability and revealed tertiary conformational changes, along with stabilizing contacts (shown in licorice representation), that impact kinase control within the fundamental signaling unit (D). An atomic model of the MDFF-refined symmetric signaling unit was deposited in the PDB under accession code 3JA6. Figure panels adapted from [63••].

Hybrid Modeling at High Resolution

At high resolution (2-5 Å), the degree of structural detail present within a given map typically permits de novo modeling techniques to construct a primary sequence directly into the density without the use of a structural template [76]. If the resolution is sufficiently high, usually better than 3.5 Å, structure topology will be unambiguous, secondary structure features will be clearly resolved, and the majority of side chains will be chemically identifiable. In such cases, standard manual-modeling tools from software suites originally developed for XRC such as Coot [77], Phenix [78], and others [79–81] may be utilized to obtain a refined atomic model “from scratch” while also correcting for small discrepancies in model-map overlap [82,83]. Though well established, the use of XRC tools to interpret near-atomic resolution maps can be labor intensive and care must still be taken to avoid misinterpretations of the density.

As map resolution decreases below 3.5 Å, precise sequence registration and overall topology become more ambiguous, and de novo modeling requires the addition of structural information. If a homologous structure is not known, most de novo workflows build up protein structure hierarchically, beginning with the identification of pre-defined secondary structural elements within the density [84–87] and from these constructing an alpha carbon trace using pathwalking [88,89•] or other pattern-recognition strategies [90,91]. Given an alpha carbon trace, a fully atomistic model, including unresolved loop segments and side chains, can be constructed and optimized using routines from the Rosetta software suite [46,92,93,95••]. A broad source of tools for general de novo modeling, Rosetta uses an iterative, fragment-assembly approach, choosing small fragments from the PDB based on local sequence homology and employing MC sampling with an empirical force field [30] to conformationally optimize and assemble fragments.

In many cases, a mostly complete structural template will be available, which can be conformationally refined using a high-resolution cryoEM map in lieu of large-scale de novo modeling. The Rosetta approach has been particularly successful in the use of 3.5-5 Å and higher resolution cryoEM maps for structural refinement [94,95••,96]. Above the 5 Å threshold, however, the structural information within the map is usually not enough to correctly direct fragment positioning for large protein segments [96]. To help relieve this constraint, RosettaES, a sampling protocol based on the above fragment-assembly approach, but using fragment ensembles pruned in a “beam search” fashion, was recently shown to efficiently and automatically construct, without a template, accurate models of missing segments up to 100 residues using high-resolution maps while also allowing for the modeling of multiple interacting segments [97••]. In addition, improvements to the MDFF protocol, namely Cascade MDFF and Resolution Exchange MDFF (RE-MDFF), have been shown to produce models of comparable quality to Rosetta using sub-5 Å maps [69••], though, it should be noted that methods based on MDFF alone do not currently provide functionality for de novo modeling.

In general, when interpreting the structural information present in high-resolution cryoEM maps, a combination of several hybrid modeling tools, each with their unique strengths, will be required to arrive at an intact and well-fitting atomic model. A case in point, Wehmer et al. recently used an iterative combination of Rosetta and MDFF (similar to that described in [98]) as well as homology modeling and XRC tools to construct atomic models of the 26S proteasome using maps of four distinct conformers, including the inactive and active complex at 4.1 and 4.5 Å resolution respectively [62••]. In another recent example, shown in Figure 3, Chen et al. applied a wide range of hybrid modeling tools and MD simulation to derive atomic models of the AMPA receptor-TARP complex in its closed, active, and desensitized states, revealing neurotransmitter-induced conformational changes in the receptor’s ligand-binding and transmembrane regions [99,100••]. Though beyond the scope of this review, these applications also emphasize the importance of advanced computational tools for image processing, which enable the isolation of distinct conformational states within SPA data [101–104].

Figure 3.

Hybrid modeling provides insight into the mechanisms of activation and desensitization in AMPA receptor-TARP complex, a membrane-bound assembly responsible for excitatory synaptic transmission in mammals. Using a wide range of hybrid modeling tools (grey box) with a high-resolution XRC structure of AMPA receptor (A), atomic models of the AMPA receptor-TARP complex corresponding to the closed state (PDB: 5KK2), the activated (PDB: 5VOT) and desensitized (PDB: 5VOV) states were constructed based on respective cryoEM density maps (B) [99,100••], revealing the large-scale conformational changes undergone by the receptor in response to neurotransmitter-binding as well as residue-level information suggesting potential gating mechanisms (C). The central region of each model is shown using a pipe representation to highlight the coupling of conformational changes in the ligand binding (LBD) and transmembrane (TMD) domains (C). Reproduced with permission (Copyright 2017, Yuhang Wang).

Model Validation

The final and most important step in any hybrid-modeling pipeline is a critical assessment of the accuracy of the resulting atomic model. Numerous methods have been developed to quantify either the stereochemical quality of a model or its goodness-of-fit to a cryoEM map. MolProbity [26], a widely used tool for judging the integrity of biomolecular structures, provides a detailed breakdown of model quality by comparing certain geometric quantities (e.g., backbone bond lengths and dihedral angles, side chain rotamers, etc.) to corresponding statistics derived from the PDB, allowing the user to diagnose structural “outliers” within a model. Meanwhile, the global and local goodness-of-fit between a model and corresponding density can be measured in real-space by metrics based on Pearson’s cross-correlation coefficient (CCC) [105,106] (although, a number of other measures exist that may perform better in certain situations [107,108]) or in reciprocal space using measures such as the integrated Fourier shell correlation (iFSC) or the estimated phase error [96]. Complimentary to the above tools, EMRinger [109•] can be utilized to assess both the stereochemistry and quality-of-fit of high-resolution features in cryoEM maps, in particular, side chain conformation.

Regardless of map resolution or the hybrid modeling protocol used, special care must be taken not to over-fit the cryoEM data [110,111]. Over-fitting occurs when the density is allowed to determine features of the model for which an appropriate amount of information is not present, for instance, if the map is insufficiently resolved or if sample heterogeneity (e.g., in a flexible loop or dynamic domain region) has led to local regions of reduced resolution or missing density. Tools such as ResMap [112•] can be used for detecting spatially variable resolution within a map, helping to pinpoint local regions of reduced information that may potentially affect model accuracy [113,114]. Globally, over-fitting must be minimized by reducing the number of refined parameters (e.g., by the use of secondary-structure constraints or selective exclusion of certain atoms) or through the addition of supplemental structural information (e.g. by employing a physics-based force field) [115]. In addition, several cross-validation procedures, in which the original cryoEM data are split into multiple, ideally independent sets that are separately used for model construction/refinement and accuracy assessment, have been proposed to identify over-fitting to high-resolution maps [82,96,116,117•,118].

Though, the above validation tools cannot yet attest to the absolute accuracy of a model (i.e., how far it deviates from the “real” biomolecular structure), such considerations can help to evaluate how consistent a model is with other known structures as well as whether or not it represents the cryoEM data well. Towards this end, it is critical that the modeler clearly highlight the degree of confidence with which a feature has been determined, for example, by annotating a single model [119•] or reporting an ensemble of models that capture feature variability [120]. Moreover, the results of multiple fitting or model construction protocols, using different parameters, can be compared to help lend confidence to a given solution [121,122]. Finally, the subsequent application of unbiased MD simulation provides powerful means by which to gauge the stability of certain model features while simultaneously exploring the post-refinement free energy landscape [123].

Outlook

Rapid advances in the high-resolution capabilities of cryoEM have opened the door to the study of previously inaccessible biomolecular systems, ranging from membrane-bound proteins to flexible and heterogeneous macromolecular complexes. At the same time, steady progress at lower to intermediate resolutions continues to produce quality images of numerous biomolecular structures central to diverse microbiological processes. Here, we have surveyed the state-of-the-art in computational methods for constructing, refining, and validating atomic models using intermediate and high-resolution cryoEM maps. Nevertheless, while the development of new and improved methodologies has been key to the success of the hybrid modeling community highlighted in this review, the current state of the field is such that its most pressing issues involve an improvement in the general usability of already existing software, tools, and protocols. In particular, many hybrid modeling tools currently exist in the wild, scattered over different sources and computing platforms with user interfaces and documentation of widely varying quality, reducing their value to non-experts or those not specializing in scientific computing. We suggest that efforts be made to ahistorically centralize hybrid modeling tools and software, including source code, files for conducting clear, positive-control test cases (called unit tests), and documentation detailing basic user-procedures for each tool, thereby greatly reducing their “black box” nature. Apart from increased visibility, such a tool bank would facilitate direct method comparison and protocol customization as well as provide a space for cataloging invaluable community feedback regarding the application of tools to novel biomolecules and extensions to code functionality, helping to ensure that tools can be made compatible with the new trends and standards developed within the rapidly evolving cryoEM community. Moreover, general usability may be increased by minimizing the technical overhead or computational cost required to conduct a given protocol, for instance, through the use of affordable, cloud-based infrastructures [69••], which allow wider community access to specialized computational resources (e.g., clusters with high-memory or GPU-accelerated nodes). Finally, to ensure that tools are used correctly and in a way that can be reproduced, standard validation and annotation procedures should be adopted and, where possible, made a prerequisite for model deposition in public databases [111,124•]. Community-wide challenges issued by the EMDB (http://challenges.emdatabank.org) have contributed greatly to a raised awareness of these and other issues [13•] and will become increasingly important as both the cryoEM and hybrid-modeling communities continue to pursue the development of advanced methods for determining the structures of increasingly complex biological assemblies.

Highlights.

• CryoEM now provides near-native structural information for a broad spectrum of biomolecules and their complexes.

• Hybrid modeling approaches enable the construction, refinement, and validation of atomic models using cryoEM data over a wide range of resolutions.

• CryoEM-based hybrid modeling is a powerful and rapidly evolving approach for gaining unique, atomistic insight into biomolecular structure and function.