Practical Considerations for Atomistic Structure Modeling with Cryo-EM Maps

Abstract

We describe common approaches to atomistic structure modeling with single particle analysis derived cryo-EM maps. Several strategies for atomistic model building and atomistic model fitting methods are discussed, including selection criteria and implementation procedures. In covering basic concepts and caveats, this short perspective aims to help facilitate active discussion between scientists at different levels with diverse backgrounds.

Graphical Abstract

Introduction

Single particle analysis is a popular cryo-electron microscopy (cryo-EM) method that often elucidates higher resolution than tomography¹. The ultimate goal of single particle analysis is an atomistic structure model that enables insight into mechanism, molecular dynamics, structure comparison and structure-based drug design. In many cases, constructing the atomistic model from the 3-D cryo-EM reconstruction maps represents a significant bottle neck². Therefore, easier access to convenient modeling tools is of interest. Recently, Malhotra et al. provided an excellent review of atomistic structure modeling that introduces comprehensive lists of programs with respect to different cryo-EM map resolutions and validation³. Here, we focus on several programs with respect to homologous sequence and structure availability and initial correlation between map and model that can be applied to other programs as well. These programs often aim to automate most modeling processes to reduce modeling time and error. In fact, as cryo-EM biological sample preparation and grid optimization become more automated^1,4, computational modeling procedures are becoming more automated^4,5 relative to manual modeling methods^6,7 as well. For example, although the Segger segments cryo-EM maps efficiently, overcoming limitations of labor-intensive work⁸, phenix.segment_and_split_map intends to integrate the segmentation and symmetry analysis with automated modeling-building⁹. Throughout this perspective article, we highlight roles of molecular dynamics (MD) for cryo-EM based atomistic structure modeling. However, the role of MD simulation for other purposes such as simulation of protein-ligand association events and dissociation rates¹⁰ and deciphering off-target effects in CRISPR-Cas9¹¹ should be recognized as well.

Atomistic model building

Once a 3-D cryo-EM reconstruction map is obtained, one needs to either find or build the atomistic molecule that will fit into the map. There are several methods of building atomistic structural models. The choice of building method depends on availability of sequence and structure (Figure 1). Availability of sequence can be ascertained from BLAST (Basic Local Alignment Search Tool) search¹². When our target is a protein, PSI-BLAST (Position-Specific Iterative BLAST), rather than BLASTp (basic search for protein sequence against a protein database), finds distant relatives for the query protein¹³. PSI-BLAST has improved sensitivity over sequence–sequence comparison methods, as a sequence profile (built from a multiple alignment of homologous sequences) contains more information about the sequence family than a single sequence¹⁴. Although the PSI-BLAST takes longer to run since it searches iteratively, it retrieves more relevant and useful sequences for us to do homology modeling. Practically speaking, PSI-BLAST with the NCBI nr (non-redundant protein sequences) database is quite memory intensive, especially in light of the rapid growth of sequence databases in recent years¹⁵. The UniRef50 database, which is further clustered from UniRef100, is more manageable while maintaining high search speed and quality, where identical sequences and sub-fragments with 11 or more residues are placed into a single record¹⁶.

Figure 1.

Atomistic structure modeling method flow chart.

Availability of sequence whose 3-D structure is already determined can be identified by a BLAST search toward the PDB (select database to ‘protein data bank’ rather than ‘nucleotide collection’ or ‘non-redundant protein sequences’). Here, if the BLAST search returns a match nearly identical to our target sequence, we can use the previously determined 3-D structure. In this case, the two 3-D structures may originate from a different organisms or conditions (often ligand existence) and are quite similar. Therefore, it is relatively straightforward to fit into a similar map conformation either by rigid-body fitting or flexible fitting¹ apart from a few exceptional cases (e.g., domain swapping)¹⁷.

Homology modeling.

In the case that the atomistic molecule for target sequence does not exist, a novel structure building procedure is needed. Among structure prediction methods, homology modeling (comparative modeling, template-based modeling) is most accurate¹⁸. This procedure is based on the assumption that two different but homologous proteins still share similar 3-D shape, provided that they have not diverged substantially during evolution. A structure for a target protein can be therefore calculated based on an experimentally established structure of another protein, called a template. The selection of a template (or templates) for modelling and their alignment with a target sequence are the most important factors influencing final results. If the sequence identity is > 30%¹⁸ or the protein:protein alignment has an E-value (the probability of obtaining the alignment by chance) < 0.001 or the DNA:DNA alignment has an E-value < 10⁻¹⁰, homology modeling can be considered. Certainly, the threshold that characterizes whether the templates have adequate sequence similarity is variable among researchers¹⁹. For example, some scientists use retrieved sequences as long as the E-value < 0.01. Others choose an E-value < 0.00001¹⁹. To make matters more confusing (but more promising when the available sequences are lacking), homologous sequences do not always share significant sequence similarity²⁰. Even when the overall similarity is low, some use multiple templates partially, as long as their local similarities are sufficiently high. All these prerequisites are summarized at the new Robetta webserver²¹, where it states that the accuracy of homology modeling mainly depends on whether similar sequences (homologs) exist in available sequence databases and in the PDB. This availability of similar sequences and structures correlates with the actual global distance test (GDT) to native structure and is conveniently provided as Robetta result files²¹. In order to extend the applicability of homology modelling, a broad variety of methods have been proposed to detect more distant homologs and to improve quality of alignments²². Classical alignment of two protein sequences can be a good solution when sequence identity is relatively high (say, above 50%). In a more difficult cases, one can align sequence profiles²³, Hidden Markov Models^14,24 or deep neural-network based contact maps¹⁸. Finally, one can align a target sequence profile with a template 3-D structure using a 3-D threading approach²⁵. Several of these methods have been also published as web services, such as a very convenient-to-use Bioinformatics Toolkit²⁶.

There are many methods to build a homology model based on a selected template. Here, we focus on Rosetta Comparative Modeling (RosettaCM) that more accurately modeled structures relative to other methods when the sequence identity is greater than 15%²⁷ and has been continued to be the most consistent top performing method²⁸ since it joined CAMEO (Continuous Automated Model EvaluatOn)²⁹. Like other Rosetta applications (such as Ab initio folding and some loop modeling methods), RosettaCM requires fragments (see below). Once a user has sequence information, all procedures of RosettaCM can be run automatically from the new Robetta webserver²¹. If one wants to run locally either for transparent understanding of procedures or high performance modeling of thousands of molecules, all procedures of RosettaCM (including clustering of decoys³⁰ and energy based selection after FastRelax³¹) can run automatically as long as a user provides sequence information³² with a help from BioPython³³.

Ab initio modeling.

If we cannot find homologous sequences to target sequence, Ab initio modeling can be considered. Among many Ab initio modeling methods, here we describe a Rosetta Ab initio procedure which itself ranked high in recent CASP (Critical Assessment of Protein Structure Prediction) competitions and whose part of its score functions and sampling procedures have been used by many scientists including recent protein structure prediction using multiple deep neural networks³⁴ and prediction of differences in free-energies³⁵. All steps of Rosetta Ab initio procedures can be run automatically from the new Robetta webserver²¹. Local running of Rosetta Ab initio in either high performance computing or Linux workstation is also fairly straightforward as long as a user manually selects the final model³⁶. The caveat of the Rosetta Ab initio protein structure prediction is that generally it is accurate up to 100 amino acids long and more accurate with low contact order proteins (e.g. α-helices) than high contact order ones (e.g. β-sheets). To overcome this limitation, co-evolution based distance constraints are needed³⁷. Ab initio approaches are by far more computationally demanding than homology modeling³⁸. By exploiting information from a template, one can calculate a few plausible structural models and accomplish modeling within one hour. In the case of Ab initio modeling however, one has to create at least 10⁴ – 10⁵ structural models, spending from 10 minutes to 1 hour of Rosetta computations for each of them. Therefore, while homology modelling can be easily done on a personal laptop, in-house Ab initio modelling can require access to considerable computational resources.

If one wants to incorporate cryo-EM map based structural information beginning in atomistic model building stage, RosettaES (enumerative sampling)³⁹, phenix.map_to_model⁴⁰, and pathwalker⁴¹ can be used. By using additional structural guidance from the cryo-EM map, these de novo modeling methods are able to model challenging regions which have been impossible to model with other methods. Therefore, relatively “high” (better) resolution of map (< 4 Å) is required.

Common features of homology modeling and Ab initio modeling.

Both RosettaCM and Rosetta Ab initio modeling use structural fragments to build a full 3-D structure of a target protein. Fragments, i.e., short backbone conformations, are extracted from known structures that were deposited to the PDB. Rosetta utilizes two sets of fragments: long and short ones, to introduce large and small changes to a modelled conformation, respectively. Traditionally 3-mers and 9-mers have been used^42,43,44, but one can apply also longer or shorter fragments⁴⁴, depending on the goal. Fragments can be computed by the fragment_picker⁴⁵ application, which is also available online as a part of the Robetta web-server⁴⁶. The program assesses how every segment from a large library of structures fits to a particular location in a target protein. The fitness (also known as score) function takes into account secondary structure match: predicted for a target vs observed for a fragment (SecondarySimilarity score) as well as similarity between the two sequence profiles (ProfileSimilarity score). A score based on a Ramachandran map (RamaScore) is used to prevent quite common mistakes with beta-branched amino acids. SecondarySimilarity is the most important part of the fragment scoring scheme while ProfileSimilarity has only minor effect on quality of fragments. The fragments are sequence-specific and must be computed independently for every protein sequence subjected to modelling. A set contains 400 fragments (200 short and 200 long ones) for each of the sequence position results.

For more accurate modeling, additional constraints from various sources (including chemical probing) can be used (Figure 1). For example, refinement using both NMR chemical shift and cryo-EM density outperforms cryo-EM density only refinement⁴⁷. In another example, we used the Situs package⁴⁸ to conveniently transform SAXS (small angle x-ray scattering) based solution volume to cryo-EM maps (mrc, situs types)^4,49. Additionally, evolutionarily conserved sequence information can guide distance pair constraints⁵⁰. These sequence-based distance constraints are similar to in distance constraints from NMR NOESY. They effectively reduce needed conformational sampling space allowing faster and more accurate simulation. Practically speaking, constraint information can be downloaded from the Gremlin webserver⁵¹ and can be incorporated easily into most Rosetta applications²⁷. For co-evolution information to be helpful, at least 3~100 times of more number of sequences in the family than the length of the query protein are required⁵². The new Robetta webserver²¹ conveniently answers whether there are enough numbers of homologous sequences to the query sequence.

Usually, selecting “better” (negatively lower) Rosetta total energy/score alone designs intended protein models successfully⁵³. However, to choose a more native-like model after simulation, one needs to pay attention to other factors as well including a protein function, stability, solubility and random mutations⁵⁴. Therefore, one needs to select a frequently sampled⁵⁵ and preferably sampled (via biased forward folding)^44,38 decoy. To select a tightly packed design as in a native protein⁵⁶, relaxing in dualspace (combined internal coordinate and cartesian relaxation)⁵⁷ is recommended. Especially, one needs to make sure that the model does not have any unsatisfied polar bonds (making structures unstable)⁵⁸ which is not easily captured by the Rosetta score function alone. Related RosettaScripts⁵⁹ conveniently filter most of these traits (packing, unsatisfied polar bond) from many other decoys/candidate models. Upon final visual inspection, Foldit standalone⁶⁰ conveniently highlights any remaining unsatisfied polar bonds or near-unsatisfied polar bonds. To analyze any clashes, PyMol⁶¹ can visualize Van der Waals (VDW) radii of each atom as spheres. These visual inspections can help to ensure correct chain connectivity/contact order of the models as well. For example, although Rosetta is very good at linking missing protein loop regions, sometimes it results in a structure whose continuous backbone protein chain is intertwined (not observed in native/foldable proteins). We needed to remove models with this intertwined continuous chain when modeling RNA structure as well⁶².

Flexible fitting of atomistic models

Once the initial atomistic model is obtained, one needs to fit the model into cryo-EM map. Selection of fitting methods depend on starting similarity between the initial atomistic model and cryo-EM map (Figure 2). For example, if the initial atomistic model is quite different structurally from the map (requiring larger radius of convergence), then rigid body fitting is better to be applied first. Once overall/global alignment by rigid body fitting between the model and map is high, then flexible fitting can be applied. Next, the local structure can be refined. However, if the initial atomistic model is close to the map, then flexible fitting, followed by refinement is often sufficient. Of course, if the backbone of initial atomistic model is already quite similar to the map, local structure refinement alone suffices. In our experience, rigid body fitting alone often cannot fit local geometrical features. Flexible fitting directly without rigid-body global fitting often fails due to too steep an energy gradient when the initial atomistic model is quite different structurally from the map⁴ (note that a newer version of phenix.cryo_fit⁶³ helps to overcome this limitation). A similar analogy can be made for refinement only. For example, using refinement alone⁶⁴ with cases where the model is significantly different from the map, the structure refinement can be trapped in a local energy minimum and not fit properly⁴. In our experience, successive running of rigid body fitting, flexible fitting and refinement is an optimal strategy. If the initial correlation between model and map is already high (the RMSD between initial atomistic model and map is less than a few Angstroms), refinement only⁶⁴ achieves faster simulation speed, while producing fits similar to flexible fitting followed by refinement.

Figure 2.

Atomistic structure fitting flowchart

Rigid body fitting.

As we implemented rigid body fitting, both ‘Fit in map’ module in UCSF Chimera⁶⁵ and phenix.dock_in_map⁶⁶ resulted in similar results (alignments) in almost all cases. The ‘Fit in map’ module in UCSF Chimera was sometimes faster and does not require map resolution information. However, when the initial model and map are far apart, phenix.dock_in_map automatically fits well, bypassing the requirement of manual initial placement. Therefore, when one needs to perform rigid body fitting for many structure-map pairs (hundreds to thousands), phenix.dock_in_map⁶⁶, which enables commandline execution, is an excellent choice.

Flexible fitting.

In flexible fitting, a common approach is a cross-correlation coefficient-based method, where the MD potentials are biased by correlation between the experimentally derived map and a simulated map, achieving atomistic models consistent with the cryo-EM map. In terms of parallelization, atomic decomposition MD is suitable for small proteins, while the domain decomposition MD is better for larger systems such as ribosome⁶⁷. Automatic running of this flexible fitting is provided by phenix.cryo_fit which can produce consistent results without the need of learning MD simulation⁴ (Figures 3–4 and Supporting Movie 1). This method has successfully fitted many DNA, RNA, and protein structures, including the ribosome and nucleosome, while preserving geometrical stereochemistry. Phenix.cryo_fit2⁶³, which runs entirely within the PHENIX software suite⁶⁸ is currently under development. This package synergistically works with other PHENIX applications. For example, it can deal with most chemical entities easily using phenix.eLBOW⁶⁹ and incorporates user designated distance constraints that come from other sources such as NMR, SHAPE⁷⁰, FRET⁷¹ and SAXS⁴⁹ to further improve model correctness while minimizing the sampling requirement. The method maintains secondary structure by using automatically generated secondary structure restraints⁷² and model idealization⁷³. All these enhanced features of phenix.cryo_fit2 allows it to outperform phenix.cryo_fit in 7 cases out of 10 cases (for cryo-EM maps with 3~24 Å resolutions). In figures 3–4, we display early results for fits with quite large radii of convergence (i.e., the initial structure is quite far from the cryo-EM map. Despite these large radii of convergence, both the tRNA and full ribosome complexes undergo large conformational changes, enabling a close fit with the high resolution cryo-EM map⁷⁴.

Figure 3.

Flexible fitting of tRNA atomistic model into cryo-EM map a) before phenix.cryo_fit b) after phenix.cryo_fit (see supporting movie S1 for movie illustration). Flexible fitting produces a close fit despite a large radius of convergence.

Figure 4.

Flexible fitting of ribosome complex atomistic model into cryo-EM map a) before phenix.cryo_fit b) after phenix.cryo_fit. Flexible fitting of the ribosome complex using phenix.cryo_fit produces a close fit despite a large radius of convergence. A large conformational change of the L1 stalk of the ribosome is required to achieve the fit.

Refinement.

Refinement aims to obtain a model with as high quality a fit as possible while possessing an expected geometry (no outliers). To achieve this, phenix.real_space_refine⁶⁴ performs 5 macro-cycles of global real-space refinement with rotamer, Ramachandran plots and C-beta deviation restraints. On top of rotamer searching and structure minimization, Rosetta based refinement employs “fragment insertion” (protein backbone torsion angle replacement from probable/predicted protein fragments) for effective sampling of backbone dihedral bonds⁷⁵. For refinement that needs a larger radius of convergence, MD simulation based methods, including those described above^4,67 can be used as complementary techniques^76,77.

Conclusion

Here, we briefly reviewed practical considerations of atomistic structure building and fitting with respect to single particle analysis based cryo-EM reconstruction maps. We note that although we focus on modeling with single particle analysis based maps, similar approaches are being developed for tomography based cryo-EM maps³², and SAXS based solution volumes⁴⁹. Additionally, while we focus on atomistic structure modeling in this article, other primary tools for cryo-EM (e.g. from map-quality assessment to validation) are well summarized by Liebschner et al⁶⁸. With more and more synergistic methods in development, fully automated atomistic structure modeling of cryo-EM maps may be possible in the near-future.